Edge calculation unloading method based on atomic orbit search optimization strategy

Abstract

The invention discloses an edge computing unloading method based on an atomic orbit search optimization strategy, which researches that under the scene of a single unmanned aerial vehicle and multiple terminal devices, limitation of communication, computing and cache resources is considered, an optimization model is established, time delay energy consumption of various unloading schemes is analyzed, and the target is to minimize system cost. Namely, the terminal equipment calculates the weighted sum of the time delay and the energy consumption of the task. Meanwhile, an improved atomic orbital search algorithm is provided, the global search capability is improved through adaptive electron distribution, the local search precision is improved in combination with crossover mutation operation, and the defect that a traditional atomic orbital search algorithm is prone to premature and is caught in a locally optimal solution too early is overcome. According to the method, the unloading decision speed and accuracy of the algorithm and other algorithms under different weighting coefficients, different terminal device numbers and different iterations are compared, it is proved that the unloading scheme can remarkably reduce the system cost, and it is proved that the proposed scheme has the advantages in the aspects of reducing energy consumption and time delay.

Description

Technical Field

[0001] The present invention belongs to the field of Internet of Things, and in particular relates to an edge computing offloading method based on atomic orbital search optimization strategy. Background Art

[0002] With the popularization of 5G and IoT devices, traditional cloud computing is unable to cope with low-latency, high-bandwidth applications. For real-time scenarios such as smart transportation and driverless driving, centralized cloud computing faces communication latency and bandwidth bottlenecks, making it difficult to meet the needs of fast processing and instant feedback, affecting system performance and experience. Mobile Edge Computing (MEC) has emerged, deploying data processing and storage capabilities to the edge of the network, close to the user end. In this way, the data transmission distance and time can be greatly shortened, latency can be reduced, and bandwidth can be saved, which not only improves the system response speed, but also extends the device life and reduces energy consumption, laying a solid foundation for the long-term and efficient operation of drones. With its advantages in reducing latency and energy consumption, MEC has great application value and prospects in the field of smart devices such as drones. Summary of the invention

[0003] The purpose of the present invention is to solve the task offloading scenario of single drone and multi-terminal devices, and to build a collaborative architecture of integrated communication, computing and cache resources by utilizing the flexible deployment and caching mechanism of drones, so as to realize local computing of tasks and drone-assisted computing, thereby improving the utilization rate and responsiveness of system resources, solving the non-convex optimization problem with the goal of minimizing system cost, and providing an edge computing offloading method based on atomic orbital search optimization strategy. By setting up control experiments, multiple batches of comparisons were carried out with GWO, AOS, PSO, SSA, and BA algorithms, and relevant tests and verifications were carried out in real scenarios, proving that the OAOS algorithm can obtain the best fitness in a smaller number of iterations, and has good stability.

[0004] The edge computing offloading method based on the atomic orbital search optimization strategy of the present invention mainly includes the following key steps:

[0005] First, the construction of the UAV system model structure:

[0006] Section 1.1. Establish UAV communication model;

[0007] 1.2. Establish the UAV computing model;

[0008] Section 1.3: Establishing a drone cache model;

[0009] Second, the construction of mathematical model structure of optimization problem:

[0010] Section 2.1, optimization function formation;

[0011] Third, the design of edge computing offloading method based on atomic orbital search optimization strategy:

[0012] Section 3.1, Atomic Orbital Search Algorithm;

[0013] Section 3.2: Improvement of atomic orbital search algorithm;

[0014] Section 3.3. UAV-assisted edge computing offloading method based on atomic orbital search optimization strategy.

[0015] Furthermore, in step 1.1, a UAV communication model is established, that is, a probabilistic line-of-sight channel is used to establish a model, and the line-of-sight communication probability between terminal device i and the UAV is and non-line-of-sight communication probability They are:

[0016]

[0017] where ρ and is the environmental parameter, θ i represents the elevation angle between device i and the drone, and the average path loss from terminal device i to the drone is expressed as:

[0018]

[0019] The average channel gain h between terminal device i and drone i =1 / L i , where α is the path loss exponent, P i is the transmission power of terminal device i, f c is the signal frequency, c is the speed of light, d i is the distance between the ground IoT device i and the drone;

[0020] When the terminal device channels are independent of each other and the device signals do not interfere with each other, the transmission rate from device i to the drone is:

[0021]

[0022] The transmission power of device i is P i , the Gaussian white noise power is N 0 , B i is the channel band allocated to terminal device i, using the partial offloading decision, x i is the task offloading rate of terminal device i, and the amount of data offloaded is x i I i , then the uplink transmission time of device i offloading the computing task to the drone is:

[0023]

[0024] Among them Ii is the task data volume of terminal device i, and the sum of bandwidth allocated by the drone to all terminal devices cannot exceed the maximum uplink bandwidth B, so there are the following constraints:

[0025]

[0026] Then the uplink transmission energy consumption of device i offloading the computing task to the drone is:

[0027]

[0028] Furthermore, the method for establishing the UAV computing model in step 1.2 is as follows: the total number of CPU instruction cycles required to complete the computing task of device i is F i =I i C i , where C i For CPU instruction cycles, computing tasks are divided into the following two ways:

[0029] (1) Local computing mode: In the local computing mode, the total computing resources of the terminal device are used with clock frequency f. At this time, the computing time required for the task of the i-th device to be calculated locally is It is expressed as:

[0030]

[0031] The energy consumption of local execution of the task of the i-th device is expressed as:

[0032]

[0033] where k i is a constant related to the terminal device structure;

[0034] (2) UAV computing mode: In the scenario of a single UAV with multiple terminal devices, f i Indicates the computing resources allocated by the drone to the i-th device. At this time, the time consumed by the drone MEC to execute the i-th device task It is expressed as:

[0035]

[0036] When considering the energy consumption of offloading tasks to drones, the energy consumption of uploading data is mainly considered, and the energy consumption of the terminal equipment standby is negligible. The energy consumption is expressed as:

[0037]

[0038] Among them, k u It is a constant related to the MEC server structure, usually k i =ku =k=10 -27 , k is a constant related to the CPU hardware structure;

[0039] At a specific time, the drone only serves one terminal device to ensure the sequential execution of tasks, and its computing resources are limited and need to be constrained:

[0040]

[0041] Among them, F max It represents the maximum computing resources that can be allocated to the UAV, and K is the set of terminal devices.

[0042] Furthermore, the method of establishing the drone cache model in step 1.3 is as follows: max Indicates the maximum available cache space of the drone. Whether the drone caches the content of a certain device is determined by a binary variable c i Indicates that c i ={0,1} (1 means cache, 0 means no cache), then the cache resource constraint of the drone is:

[0043]

[0044] When device i needs to calculate and cache the process, the drone can take over the task. After taking on the task, it will complete all the calculation work and send the results to device i. At this time, the calculation task Task i The cost only includes the time cost of the drone completing the computing task and energy costs They are:

[0045]

[0046] If the content is not cached, the terminal device needs to undertake all the work by itself when performing the computing task, which involves the time and energy consumption of the computing process. i Time cost and energy costs They are:

[0047]

[0048]

[0049] Generally, cache is a long statistical result. In this paper, we consider that all terminal devices have the same task cache strategy. According to equations (15)-(16), terminal device i calculates task Task i The time required i and energy consumption E i They are:

[0050]

[0051]

[0052] Furthermore, the optimization function in step 2.1 is proposed to reduce the system cost to improve the service quality, where the system cost Time Cost Let ξ represent the weighted coefficient of latency and energy consumption, then device i calculates task Task i Required cost energy consumption cost i It is expressed as:

[0053]

[0054] Then the system cost is expressed as:

[0055]

[0056] According to the above description, the formula of the optimization problem is constructed as:

[0057]

[0058]

[0059]

[0060]

[0061]

[0062]

[0063]

[0064]

[0065] Among them, 0≤ξ≤1, constraint (22) is that the channel bandwidth allocated to all terminal devices does not exceed the total bandwidth of the system channel, constraint (23) is that the total computing resources allocated to each terminal device does not exceed the maximum computing resources of the MEC server, and constraint (25) is that each device must be within the maximum tolerable delay To complete the task within 100m, constraint (28) is the UAV cache resource constraint.

[0066] Further, the atomic orbital search algorithm described in step 3.1, the steps for finding the optimal solution are as follows:

[0067] Step 1: Initialization, randomly generate candidate solution sets (electrons) and calculate energy levels Sorting;

[0068] Step 2: Use normal distribution to determine candidate solution orbitals, where the central orbital has a low energy level;

[0069] Step 3: Calculate the combined state BS of each layer of orbital candidate solutions k (position mean), binding energy BE k (Fitness mean) and the lowest energy candidate solution LE k ;

[0070] Step 4: Transition decision, based on the preset probability, decide whether each candidate solution should transition or generate a new individual;

[0071] Step 5: Determine the transition direction, compare the energy level with the binding energy, consider photon emission when the energy level is higher than the binding energy, and use the lowest energy level candidate solution LE and the position average BS to update the current candidate solution position according to the formula. The execution formula is:

[0072]

[0073] in and It represents the current position and next active position of the i-th candidate solution in the k-th layer, and the random number vector α i , β i , γ i Uniformly distributed in the interval (0, 1) and used to calculate the released energy. If the energy of a candidate solution at a certain layer Less than the binding energy BE of the layer k , then considering the absorption of photons, the execution formula is:

[0074]

[0075] Furthermore, the method of using the improved atomic orbital search algorithm described in step 3.2 is as follows. In order to solve the problem that the AOS algorithm falls into the local optimum, an adaptive electron distribution strategy is proposed (uniformly distributing electrons in the initial stage to achieve global search, and concentrating electrons to the inner orbits in the later iteration to improve the solution accuracy) and the crossover and mutation operators are introduced to improve the search performance. The two improvement methods are as follows:

[0076] (1) Adaptive electron distribution

[0077] Based on the normal distribution principle of electrons in the AOS algorithm, a strategy is proposed, that is, electrons are evenly distributed in each orbital layer in the early stage, and concentrated in the inner orbital layer in the later stage. n = MaxLayer represents the maximum number of atomic orbital layers, t = Iter represents the current number of iterations, and s = s 0 ×(1-t / MaxIter) represents the adjustment factor of the adaptive electron distribution strategy, s 0=MaxIter / 2 (where MaxIter represents the maximum number of iterations), and the proportion of electrons in the current orbit is expressed as:

[0078] arr=[s+(n-1)×t,s+(n-2)×t,…,s+(ni)×t,…,s] (31)

[0079]

[0080] prob is a fractional array, whose dimension value corresponds to the ratio of the number of orbital electrons to the total number of electrons. When the maximum number of orbital layers MaxLayer is set to 6, the maximum number of iterations MaxIter is 400, and the number of electrons is 120, as the number of iterations increases, the number of electrons in the inner orbit is significantly greater than that in the outer orbit;

[0081] (2) Introduction of crossover and mutation operators

[0082] The crossover operator is introduced in the position update link after the photon absorption stage of the original algorithm. The position codes of the current solution and the global optimal solution are converted into binary. The random generated decimal is compared with the crossover probability to decide whether to exchange the corresponding bits. Figure 3 It is the exchange graph when the crossover probability is 0.5. If the random value is ≥ 0.5, it is exchanged;

[0083] If no photon hits, a mutation operator is introduced, which also encodes the position of the current solution into binary form, generates a random value for each solution and compares it with the mutation probability, and performs an XOR operation when it is greater than the mutation probability;

[0084] In order to find the optimal unloading decision in the UAV scenario, an optimized atomic orbital search algorithm is proposed based on the above two improvements. The specific steps of the algorithm are as follows:

[0085] Step 1: Initialize the population position, pass in parameters according to the optimization problem, use the Cost function to calculate the individual fitness and sort the candidate solutions by fitness, and set MaxIter, MaxLayer and photon action probability PR;

[0086] Step 2: For each iteration, randomly generate the current track layer number L from the range 1 to MaxLayer;

[0087] Step 3: For each orbital, calculate the bound state BS k and binding energy BE k ;

[0088] Step 4: Generate a random number rand_val for each electron and compare it with PR. If rand_val is less than PR and the current electron energy is greater than BE kThen emit a photon and use formula (31) to update the position. If the current electron energy is less than BE k Then absorb the photon and use formula (30) to update the position and perform crossover operation. If rand_val is greater than PR, perform mutation operation;

[0089] Step 5: Update the positions of all electrons, recalculate the fitness, and sort them according to the fitness;

[0090] Step 6: Determine whether Maxiter is reached, if so, stop and return the optimal solution, otherwise return to step 2;

[0091] The initial population size is N. Because we need to calculate the fitness of each individual (time complexity is O(f)), the initial time complexity is O(N×f), the sorting time complexity is O(N×log N), and the time complexity of calculating the binding state and binding energy is negligible, so the total time complexity is O(MaxIter×N×(log N+f)), and the space complexity is O(N×d);

[0092] Furthermore, the algorithm of the drone-assisted edge computing offloading method based on the atomic orbital search optimization strategy described in step 3.3 is as follows:

[0093] Step 1: Initialize the available computing resources of the drone, the offloaded tasks of each device, and the parameters required by the OAOS algorithm;

[0094] Step 2: Generate an initial offloading plan for each device, evaluate each offloading plan based on expected latency and energy consumption, and add the evaluated offloading plan to the initial population, which will serve as the starting point for the OAOS algorithm optimization process;

[0095] Step 3: Call Algorithm 1 and use the OAOS algorithm to optimize the initial offloading plan;

[0096] Step 4: Use the best offloading solution for each device, calculate the weighted sum of actual delay and energy consumption, and record the results;

[0097] Step 5: After applying the optimal solution, check whether the constraints are met. If so, return it as the global optimal strategy.

[0098] The time complexity is dominated by the optimization step of the OAOS algorithm, and its time complexity is approximately O(MaxIter×N×(MaxLayer+log n)); the space complexity depends on the population storage and fitness value storage, so the space complexity is O(N×K), where K is the number of terminal devices.

[0099] The advantages and positive effects of the present invention are:

[0100] See attached Fig.10, the present invention designs an edge computing offloading method based on atomic orbital search optimization. For single UAV and multi-terminal device task offloading scenarios, it uses the characteristics of UAVs to build a collaborative architecture, realizes multivariate computing, solves the problem of minimizing system costs, and improves resource utilization and responsiveness. Although the atomic orbital search algorithm has high accuracy and fast convergence, and can find the optimal parameters based on quantum chemistry theory, it is easy to fall into local optimality in single UAV scenarios. Therefore, this method introduces adaptive strategies and crossover mutation operators to adjust individual distributions, break through limitations, and strengthen global optimization, with significant practical value. BRIEF DESCRIPTION OF THE DRAWINGS

[0101] Figure 1 It is a scenario where a single drone assists the MEC system to serve terminal devices;

[0102] Figure 2 is the adaptive electron distribution map;

[0103] Figure 3 It is a cross-process diagram;

[0104] Figure 4 is the delay and energy consumption variation curve under different weighting coefficients ξ;

[0105] Figure 5 is the relationship between the algorithm time and the number of terminal devices (ξ=0.3);

[0106] Figure 6 is the relationship between the number of iterations and fitness (K = 30);

[0107] Figure 7 is the relationship between the number of devices and fitness (ξ=0.3);

[0108] Figure 8 is the relationship between system cost and the cache capacity of the drone;

[0109] Fig. 9 is the relationship between the number of algorithm iterations and fitness (ξ=0.3);

[0110] Fig.10 It is a flow chart of the edge computing offloading method based on the atomic orbital search optimization strategy of the present invention. DETAILED DESCRIPTION

[0111] Example 1

[0112] This embodiment is based on the MATLAB R2024a simulation platform and is studied under the hardware conditions of 12th Gen Intel(R)Core(TM)i5-12490F 3.00GHz and 16GB memory. The OAOS algorithm is compared with AOS, PSO, GWO, SSA, and BA to verify the feasibility of the improved AOS algorithm. The OAOS performance is analyzed from multiple dimensions: such as the impact of different weighting coefficients on the offloading scheme, the change in algorithm time caused by different terminal device quantities, the correlation between the number of iterations and the number of devices on fitness, and the relationship between latency and cache capacity and system cost.

[0113] The edge computing offloading method based on the atomic orbital search optimization strategy of the present invention mainly includes the following key steps:

[0114] First, the construction of the UAV system model structure:

[0115] Section 1.1. Establish UAV communication model;

[0116] 1.2. Establish the UAV computing model;

[0117] Section 1.3: Establishing a drone cache model;

[0118] Second, the construction of mathematical model structure of optimization problem:

[0119] Section 2.1, optimization function formation;

[0120] Third, the design of edge computing offloading method based on atomic orbital search optimization strategy:

[0121] Section 3.1, Atomic Orbital Search Algorithm;

[0122] Section 3.2: Improvement of atomic orbital search algorithm;

[0123] Section 3.3. UAV-assisted edge computing offloading method based on atomic orbital search optimization strategy.

[0124] Furthermore, in step 1.1, a UAV communication model is established, that is, a probabilistic line-of-sight channel is used to establish a model, and the line-of-sight communication probability between terminal device i and the UAV is and non-line-of-sight communication probability They are:

[0125]

[0126] where ρ and is the environmental parameter, θ i represents the elevation angle between device i and the drone. The average path loss from terminal device i to the drone is expressed as:

[0127]

[0128] The average channel gain h between terminal device i and drone i =1 / L i . Where α is the path loss exponent, P i is the transmission power of terminal device i, f c is the signal frequency, c is the speed of light, d i is the distance between the ground IoT device i and the drone.

[0129] When the terminal device channels are independent of each other and the device signals do not interfere with each other, the transmission rate from device i to the drone is:

[0130]

[0131] The transmission power of device i is P i , the Gaussian white noise power is N 0 , B i is the channel band allocated to terminal device i. Using the partial offloading decision, x i is the task offloading rate of terminal device i, and the amount of data offloaded is x i I i , then the uplink transmission time of device i offloading the computing task to the drone is:

[0132]

[0133] Among them I i is the task data volume of terminal device i, and the sum of bandwidth allocated by the drone to all terminal devices cannot exceed the maximum uplink bandwidth B, so there are the following constraints:

[0134]

[0135] Then the uplink transmission energy consumption of device i offloading the computing task to the drone is:

[0136]

[0137] The method of establishing the UAV computing model in step 1.2 is as follows. The total number of CPU instruction cycles required to complete the computing task of device i is F i =I i C i , where C i is the CPU instruction cycle. The computing tasks are divided into the following two ways:

[0138] (1) Local computing mode: In the local computing mode, the total computing resources of the terminal device are used with clock frequency f. At this time, the computing time required for the task of the i-th device to be calculated locally is It is expressed as:

[0139]

[0140] The energy consumption of local execution of the task of the i-th device is expressed as:

[0141]

[0142] where k i is a constant related to the terminal device structure.

[0143] (2) UAV computing mode: Figure 1 In the scenario of a single drone with multiple terminals, f i Indicates the computing resources allocated by the drone to the i-th device. At this time, the time consumed by the drone MEC to execute the i-th device task It is expressed as:

[0144]

[0145] When considering the energy consumption of offloading tasks to drones, the energy consumption of uploading data is mainly considered, and the energy consumption of the terminal equipment standby is negligible. The energy consumption is expressed as:

[0146]

[0147] Among them, k u It is a constant related to the MEC server structure, usually k i =k u =k=10 -27 , k is a constant related to the CPU hardware structure.

[0148] At a specific time, the drone only serves one terminal device to ensure the sequential execution of tasks, and its computing resources are limited and need to be constrained:

[0149]

[0150] Among them, F max It represents the maximum computing resources that can be allocated to the UAV, and K is the set of terminal devices.

[0151] The method of establishing the drone cache model in step 1.3 is as follows: max Indicates the maximum available cache space of the drone. Whether the drone caches the content of a certain device is determined by a binary variable c i Indicates that c i ={0,1}, where 1 means caching and 0 means no caching. The cache resource constraint of the drone is:

[0152]

[0153] When device i needs to calculate and cache the process, the drone can take over the task. After taking on the task, it will complete all the calculation work and send the results to device i. At this time, the calculation task Task i The cost only includes the time cost of the drone completing the computing task and energy costs They are:

[0154]

[0155]

[0156] If the content is not cached, the terminal device needs to undertake all the work by itself when performing the computing task, which involves the time and energy consumption of the computing process. i Time cost and energy costs They are:

[0157]

[0158]

[0159] Generally, cache is a long statistical result. In this paper, we consider that all terminal devices have the same task cache strategy. According to equations (15)-(16), terminal device i calculates task Task i The time required i and energy consumption E i They are:

[0160]

[0161]

[0162] Furthermore, the optimization function in step 2.1 is proposed to reduce the system cost to improve the service quality, where the system cost Time Cost Let ξ(0≤ξ≤1) represent the weighted coefficient of latency and energy consumption, then device i calculates task Task i Required cost energy consumption cost i It is expressed as:

[0163]

[0164] Then the system cost is expressed as:

[0165]

[0166] According to the above description, the present invention constructs the formula of the optimization problem as follows:

[0167]

[0168]

[0169]

[0170]

[0171]

[0172]

[0173]

[0174]

[0175] Among them, constraint (22) is that the channel bandwidth allocated to all terminal devices does not exceed the total bandwidth of the system channel, constraint (23) is that the total computing resources allocated to each terminal device does not exceed the maximum computing resources of the MEC server, and constraint (25) is that each device must be within the maximum tolerable delay. To complete the task within 100m, constraint (28) is the UAV cache resource constraint.

[0176] Further, the atomic orbital search algorithm described in step 3.1, the steps for finding the optimal solution are as follows:

[0177] Step 1: Initialization, randomly generate candidate solution sets (electrons) and calculate energy levels Sorting;

[0178] Step 2: Use normal distribution to determine candidate solution orbitals, where the central orbital has a low energy level;

[0179] Step 3: Calculate the combined state BS of each layer of orbital candidate solutions k (position mean), binding energy BE k (Fitness mean) and the lowest energy candidate solution LE k ;

[0180] Step 4: Transition decision, based on the preset probability, decide whether each candidate solution should transition or generate a new individual;

[0181] Step 5: Determine the transition direction, compare the energy level with the binding energy, consider photon emission when the energy level is higher than the binding energy, and use the lowest energy level candidate solution LE and the position average BS to update the current candidate solution position according to the formula. The execution formula is:

[0182]

[0183] in and It represents the current position and next active position of the i-th candidate solution in the k-th layer, and the random number vector α i , β i , γ i is evenly distributed in the interval (0, 1) and is used to calculate the released energy. Less than the binding energy BE of the layer k , then considering the absorption of photons, the execution formula is:

[0184]

[0185] The method of using the improved atomic orbital search algorithm described in step 3.2 is as follows. In order to solve the problem of the AOS algorithm falling into the local optimum, an adaptive electron distribution strategy is proposed (uniformly distribute electrons in the initial stage to achieve global search, and concentrate electrons to the inner orbits in the later iteration to improve the solution accuracy) and the crossover and mutation operators are introduced to improve the search performance. The two improvement methods are as follows:

[0186] (1) Adaptive electron distribution

[0187] Based on the normal distribution principle of electrons in the AOS algorithm, a strategy is proposed, that is, electrons are evenly distributed in each orbital layer in the early stage, and concentrated in the inner orbital layer in the later stage. n = MaxLayer represents the maximum number of atomic orbital layers, t = Iter represents the current number of iterations, and s = s 0 ×(1-t / MaxIter) represents the adjustment factor of the adaptive electron distribution strategy, s 0 =MaxIter / 2 (where MaxIter represents the maximum number of iterations), and the proportion of electrons in the current orbit is expressed as:

[0188] arr=[s+(n-1)×t,s+(n-2)×t,…,s+(ni)×t,…,s] (31)

[0189]

[0190] prob is a fractional array, whose dimension value corresponds to the ratio of the number of orbital electrons to the total number of electrons, as shown in the attached Figure 2 As shown in the figure, when the maximum number of orbital layers MaxLayer is set to 6, the maximum number of iterations MaxIter is 400 and the number of electrons is 120, as the number of iterations increases, the number of electrons in the inner orbit is significantly greater than that in the outer orbit.

[0191] (2) Introduction of crossover and mutation operators

[0192] The crossover operator is introduced in the position update link after the photon absorption stage of the original algorithm. The position codes of the current solution and the global optimal solution are converted into binary. The random generated decimal is compared with the crossover probability (0.5) to decide whether to exchange the corresponding bits. Figure 3 is the exchange diagram when the crossover probability is 0.5. If the random value is ≥ 0.5, the exchange is performed. The coding position 703 of individual 1 is randomly generated. Individual 2 is the global optimal individual. The crossover process is shown in the attached Figure 3 .

[0193] If no photon hits the solution, a mutation operator is introduced. The position of the current solution is encoded into binary form. A random value is generated for each solution and compared with the mutation probability. If the value is greater than the mutation probability, an XOR operation is performed.

[0194] In order to find the optimal unloading decision in the UAV scenario, the present invention proposes an optimized atomic orbital search algorithm based on the above two improvements, and the specific steps of the algorithm are as follows:

[0195] Step 1: Initialize the population position, pass in parameters according to the optimization problem, use the Cost function to calculate the individual fitness and sort the candidate solutions by fitness, and set MaxIter, MaxLayer and photon action probability PR;

[0196] Step 2: For each iteration, randomly generate the current track layer number L from the range 1 to MaxLayer;

[0197] Step 3: For each orbital, calculate the bound state BS k and binding energy BE k ;

[0198] Step 4: Generate a random number rand_val for each electron and compare it with PR. If rand_val is less than PR and the current electron energy is greater than BE k Then emit a photon and use formula (31) to update the position. If the current electron energy is less than BE k Then absorb the photon and use formula (30) to update the position and perform crossover operation. If rand_val is greater than PR, perform mutation operation;

[0199] Step 5: Update the positions of all electrons, recalculate the fitness, and sort them according to the fitness;

[0200] Step 6: Determine whether Maxiter is reached. If so, stop and return the optimal solution. Otherwise, return to step 2.

[0201] The initial population size is N. Since the fitness of each individual needs to be calculated (the time complexity is O(f)), the initial time complexity is O(N×f), the sorting time complexity is O(N×log N), and the time complexity of calculating the binding state and binding energy is negligible. Therefore, the total time complexity is O(MaxIter×N×(log N+f)) and the space complexity is O(N×d).

[0202] Step 3.3 describes the drone-assisted edge computing offloading method based on atomic orbital search optimization strategy, and the specific algorithm steps are as follows:

[0203] Step 1: Initialize the available computing resources of the drone, the offloaded tasks of each device, and the parameters required by the OAOS algorithm;

[0204] Step 2: Generate an initial offloading plan for each device, evaluate each offloading plan based on expected latency and energy consumption, and add the evaluated offloading plan to the initial population, which will serve as the starting point for the OAOS algorithm optimization process;

[0205] Step 3: Call Algorithm 1 and use the OAOS algorithm to optimize the initial offloading plan;

[0206] Step 4: Use the best offloading solution for each device, calculate the weighted sum of actual delay and energy consumption, and record the results;

[0207] Step 5: After applying the optimal solution, check whether the constraints are met. If so, return it as the global optimal strategy.

[0208] The time complexity is dominated by the optimization step of the OAOS algorithm, and its time complexity is approximately O(MaxIter×N×(MaxLayer+log n)); the space complexity depends on the population storage and fitness value storage, so the space complexity is O(N×K), where K is the number of terminal devices.

[0209] In this example, a simulation scenario is constructed. In a square area of ​​500m×500m, a drone and 20 terminal devices are deployed. Both the drone and the terminal devices use communication devices that support the IEEE 802.11ax protocol. The drone first finds a suitable hovering position and then allocates a communication channel for the terminal device. The system cost is calculated under different latency and energy consumption ratios, and the performance of different algorithms is compared. Two simulation experiments will be conducted.

[0210] 1) Experiment A: OAOS is compared with AOS, PSO, GWO, SSA, and BA to verify its feasibility. The performance of OAOS under different conditions is analyzed from the aspects of weight coefficient ξ, number of terminal devices, number of iterations, and tolerable delay.

[0211] 2) Experiment B: A simulation test in a real-world scenario. The system cost is calculated with a focus on latency and energy consumption when the weighting coefficients ξ are 0.3, 0.5, and 0.7, and the performance of different algorithms is compared.

[0212] The performance of the proposed algorithm is evaluated through simulation experiments. The basic parameters used in the experiment are shown in Table 1.

[0213] Table 1 Simulation parameters

[0214] parameter Value K [10,90] <![CDATA[F max ]]> 40GHz f 1GHz <![CDATA[σ 2 ]]> -174dBm B 20MHz <![CDATA[I i ]]> [0.5,1.5]Mbits α 2.7 <![CDATA[P i ]]> 0.2W <![CDATA[k i ,k u ]]> <![CDATA[10 -27 ]]> <![CDATA[C i ]]> <![CDATA[[0.8,1.2]×10 3 cycles / bit]]> ξ (0,1) <![CDATA[D max ]]> [0.5,4.5]Mbits

[0215] This simulation experiment will consider four performance indicators, which are:

[0216] 1. Latency and energy consumption. Latency refers to the time delay a signal experiences when it is transmitted from the sender to the receiver; energy consumption refers to the total amount of energy consumed by a device within a certain period of time.

[0217] 2. Iteration speed. Iteration speed is an indicator to measure the speed of status update during the iteration process, which can be reflected in the number of iterations per unit time or the iteration time required to achieve a specific goal.

[0218] 3. System cost. System cost is defined as the total cost consumption of all terminal device computing tasks.

[0219] 4. Algorithm fitness: Algorithm fitness is used to measure the algorithm's ability to solve and adapt to problems in specific scenarios.

[0220] The simulation results of this example are as follows:

[0221] 1. Performance of OAOS algorithm in simulation platform

[0222] 1) Variation of delay and energy consumption under different weighting coefficients ξ

[0223] Attached Figure 4 The figure shows the impact of different weighted coefficients ξ on latency and energy consumption. It can be seen that the larger the ξ, the higher the energy consumption, but the latency will be reduced. Therefore, when designing an offload decision plan, the weighted coefficient can be set according to the user's requirements for latency and energy consumption to meet the user's personalized requirements.

[0224] 2) Relationship between algorithm time and number of terminal devices

[0225] The parameter ξ was set to 0.3, 0.5, and 0.7, and the relationship between the algorithm calculation time and the number of terminal devices was observed from different delay and energy consumption ratios. The results are shown in the attached figure. Figure 5 It can be seen that under different emphasis ratios, the OAOS algorithm performs steadily, and most of the time, its time consumption is less than that of other algorithms.

[0226] 3) Relationship between number of iterations and fitness

[0227] Under the same setting of latency and energy consumption, the number of terminal devices is set to 30, 60, and 90 to observe the change of fitness of each algorithm with iteration. The experimental results are attached. Figure 6It can be seen that when the number of devices increases, the iteration of the OAOS algorithm is fast and stable, and the fitness is continuously low. In contrast, although AOS has strong computational power, its search is highly random and prone to local optimality; PSO has fast iterations but poor fitness; BA has poor stability; GWO has long iterations; and SSA has higher fitness than OAOS. Therefore, OAOS can effectively reduce costs.

[0228] 4) Relationship between the number of devices and fitness

[0229] When analyzing the relationship between the number of terminal devices and fitness, we still experimented with the parameter ξ at 0.3, 0.5, and 0.7 to observe the relationship between the number of devices and fitness in different scenarios. The results are shown in the attached figure. Figure 7 As shown in the figure, it can be seen that under different parameters ξ, as the number of devices increases, the OAOS algorithm always maintains the minimum fitness value. At the same time, when ξ is set to 0.3, the AOS algorithm falls into a local optimum.

[0230] 5) Relationship between system cost and drone cache capacity

[0231] When other conditions remain constant, the cache capacity of the drone is changed, and the weighted sum of the delay and energy consumption of the OAOS algorithm is observed. Figure 8 ), D max = 0 (drone has no cache). As can be seen from the figure, as the maximum tolerable delay or cache space increases, the system cost is significantly reduced, because low delay requires fast processing, which consumes more resources; the cost changes are consistent under different delays, which is due to the fixed data volume.

[0232] 2. Performance of OAOS algorithm in real scenarios

[0233] 1) Relationship between algorithm iterations and fitness

[0234] By the attached Fig. 9 As we know, due to environmental interference in real scenes, the iteration time of each algorithm is long. GWO and SSA have slow iterations and are difficult to meet real-time requirements; BA is unstable and does not meet the stability requirements; the fitness calculated by AOS and PSO is significantly higher than that of OAOS algorithm.

[0235] The simulation results show that the performance of the OAOS algorithm is verified under ideal conditions and in real scenarios, and the relationship between the OAOS algorithm and the GWO, AOS, PSO, SSA, and BA algorithms in terms of the number of terminal devices, number of iterations, and weighted coefficients and fitness is compared, proving the advantages of the OAOS algorithm.

Claims

1. An edge computing offloading method based on atomic orbital search optimization strategy, characterized in that The method mainly includes the following steps: First, the construction of the UAV system model structure: Section 1.

1. Establish UAV communication model; 1.

2. Establish the UAV computing model; Section 1.3: Establishing a drone cache model; Second, the construction of mathematical model structure of optimization problem: Section 2.1, optimization function formation; Third, the design of edge computing offloading method based on atomic orbital search optimization strategy: Section 3.1, Atomic Orbital Search Algorithm; Section 3.2: Improvement of atomic orbital search algorithm; Section 3.

3. UAV-assisted edge computing offloading method based on atomic orbital search optimization strategy.

2. The edge computing offloading method based on atomic orbital search optimization strategy as claimed in claim 1, characterized in that: In step 1.1, the UAV communication model is established. The method of establishing the model using the probabilistic line-of-sight channel is as follows: the line-of-sight communication probability between the terminal device i and the UAV is defined as and non-line-of-sight communication probability They are: where ρ and is the environmental parameter, θ i represents the elevation angle between device i and the drone, and the average path loss from terminal device i to the drone is expressed as: The average channel gain h between terminal device i and drone i =1 / L i , where α is the path loss exponent, P i is the transmission power of terminal device i, f c is the signal frequency, c is the speed of light, d i is the distance between the ground IoT device i and the drone; When the terminal device channels are independent of each other and the device signals do not interfere with each other, the transmission rate from device i to the drone is: The transmission power of device i is P i , Gaussian white noise power is N0, B i is the channel band allocated to terminal device i, using the partial offloading decision, x i is the task offloading rate of terminal device i, and the amount of data offloaded is x i I i , then the uplink transmission time of device i offloading the computing task to the drone is: Among them I i is the task data volume of terminal device i, and the sum of bandwidth allocated by the drone to all terminal devices cannot exceed the maximum uplink bandwidth B, so there are the following constraints: Then the uplink transmission energy consumption of device i offloading the computing task to the drone is:

3. The edge computing offloading method based on atomic orbital search optimization strategy as claimed in claim 1, characterized in that: The method of establishing the UAV computing model in step 1.2 is as follows. The total number of CPU instruction cycles required to complete the computing task of device i is F i =I i C i , where C i For CPU instruction cycles, computing tasks are divided into the following two ways: (1) Local computing mode: In the local computing mode, the total computing resources of the terminal device are used with clock frequency f. At this time, the computing time required for the task of the i-th device to be calculated locally is It is expressed as: The energy consumption of local execution of the task of the i-th device is expressed as: where k i is a constant related to the terminal device structure; (2) UAV computing mode: In the scenario of a single UAV with multiple terminal devices, f i Indicates the computing resources allocated by the drone to the i-th device. At this time, the time consumed by the drone MEC to execute the i-th device task It is expressed as: When considering the energy consumption of offloading tasks to drones, the energy consumption of uploading data is mainly considered, and the energy consumption of terminal equipment standby is negligible. The energy consumption is expressed as: Among them, k u It is a constant related to the MEC server structure, usually k i =k u =k=10 -27 , k is a constant related to the CPU hardware structure; At a specific time, the drone only serves one terminal device to ensure the sequential execution of tasks, and its computing resources are limited and need to be constrained: where F max It represents the maximum computing resources that can be allocated to the UAV, and K is the set of terminal devices.

4. The edge computing offloading method based on atomic orbital search optimization strategy as claimed in claim 1, characterized in that: The method for establishing the drone cache model in step 1.3 is as follows: max Indicates the maximum available cache space of the drone. Whether the drone caches the content of a certain device is determined by a binary variable c i Indicates that c i ={0,1}, where 1 means caching and 0 means no caching. The cache resource constraint of the drone is: When device i needs to calculate and cache the process, the drone can take over the task. After taking on the task, it will complete all the calculation work and send the results to device i. At this time, the calculation task Task i The cost only includes the time cost of the drone completing the computing task and energy costs They are: If the content is not cached, the terminal device needs to undertake all the work by itself when performing the computing task, which involves the time and energy consumption of the computing process. i Time cost and energy costs They are: According to equations (15)-(16), terminal device i calculates task Task i The time required i and energy consumption E i They are:

5. The edge computing offloading method based on atomic orbital search optimization strategy as claimed in claim 1, characterized in that: The optimization function in step 2.1 is proposed to reduce the system cost to improve the service quality, where the system cost Time Cost Let ξ represent the weighted coefficient of latency and energy consumption, then device i calculates task Task i Required cost energy consumption cost i It is expressed as: Then the system cost is expressed as: According to the above description, the formula of the optimization problem is constructed as: Among them, 0≤ξ≤1, constraint (22) is that the channel bandwidth allocated to all terminal devices does not exceed the total bandwidth of the system channel, constraint (23) is that the total computing resources allocated to each terminal device does not exceed the maximum computing resources of the MEC server, and constraint (25) is that each device must be within the maximum tolerable delay To complete the task within 100m, constraint (28) is the UAV cache resource constraint.

6. The edge computing offloading method based on atomic orbital search optimization strategy according to claim 1, characterized in that: The steps for finding the optimal solution using the atomic orbital search algorithm described in step 3.1 are as follows: Step 1: Initialization, randomly generate candidate solution sets (electrons) and calculate energy levels Sorting; Step 2: Use normal distribution to determine candidate solution orbitals, where the central orbital has a low energy level; Step 3: Calculate the combined state BS of each layer of orbital candidate solutions k (position mean), binding energy BE k (Fitness mean) and the lowest energy candidate solution LE k ; Step 4: Transition decision, based on the preset probability, decide whether each candidate solution should transition or generate a new individual; Step 5: Determine the transition direction, compare the energy level with the binding energy, consider photon emission when the energy level is higher than the binding energy, and use the lowest energy level candidate solution LE and the position average BS to update the current candidate solution position according to the formula. The execution formula is: in and It represents the current position and next active position of the i-th candidate solution in the k-th layer, and the random number vector α i , β i , γ i Uniformly distributed in the interval (0, 1) and used to calculate the released energy. If the energy of a candidate solution at a certain layer Less than the binding energy BE of the layer k , then considering the absorption of photons, the execution formula is:

7. The edge computing offloading method based on atomic orbital search optimization strategy as claimed in claim 1, characterized in that: The method of using the improved atomic orbital search algorithm described in step 3.2 is as follows. In order to solve the problem that the AOS algorithm falls into the local optimum, an adaptive electron distribution strategy is proposed and crossover and mutation operators are introduced to improve the search performance. The two improved methods are as follows: (1) Adaptive electron distribution Based on the principle of normal distribution of electrons in the AOS algorithm, a strategy is proposed, that is, electrons are evenly distributed in each orbital layer in the early stage, and concentrated in the inner orbital layer in the later stage. n = MaxLayer represents the maximum number of atomic orbital layers, t = Iter represents the current number of iterations, s = s0 × (1-t / MaxIter) represents the adjustment factor of the adaptive electron distribution strategy, s0 = MaxIter / 2, where MaxIter represents the maximum number of iterations, and the proportion of electrons in the current orbital layer is expressed as: arr=[s+(n-1)×t,s+(n-2)×t,…,s+(ni)×t,…,s] (31) prob is a fractional array, whose dimension value corresponds to the ratio of the number of orbital electrons to the total number of electrons. When the maximum number of orbital layers MaxLayer is set to 6, the maximum number of iterations MaxIter is 400, and the number of electrons is 120, as the number of iterations increases, the number of electrons in the inner orbit is significantly greater than that in the outer orbit; (2) Introduction of crossover and mutation operators A crossover operator is introduced in the position update link after the photon absorption stage of the original algorithm. The position codes of the current solution and the global optimal solution are converted into binary. The random generated decimal is compared with the crossover probability to decide whether to exchange the corresponding bits. Figure 3 is an exchange diagram when the crossover probability is 0.

5. If the random value is ≥ 0.5, the bits are exchanged. If no photon hits, a mutation operator is introduced, which also encodes the position of the current solution into binary form, generates a random value for each solution and compares it with the mutation probability, and performs an XOR operation when it is greater than the mutation probability; Based on the above two improvements, an optimized atomic orbital search algorithm is proposed. The specific steps of the algorithm are as follows: Step 1: Initialize the population position, pass in parameters according to the optimization problem, use the Cost function to calculate the individual fitness and sort the candidate solutions by fitness, and set MaxIter, MaxLayer and photon action probability PR; Step 2: For each iteration, randomly generate the current track layer number L from the range 1 to MaxLayer; Step 3: For each orbital, calculate the bound state BS k and binding energy BE k ; Step 4: Generate a random number rand_val for each electron and compare it with PR. If rand_val is less than PR and the current electron energy is greater than BE k Then emit a photon and use formula (31) to update the position. If the current electron energy is less than BE k Then absorb the photon and use formula (30) to update the position and perform crossover operation. If rand_val is greater than PR, perform mutation operation; Step 5: Update the positions of all electrons, recalculate the fitness, and sort them according to the fitness; Step 6: Determine whether MaxIter is reached, if so, stop and return the optimal solution, otherwise return to step 2; The initial population size is N. Since the fitness of each individual needs to be calculated, the time complexity is O(f), the initial time complexity is O(N×f), the sorting time complexity is O(N×logN), and the time complexity of calculating the binding state and binding energy is negligible. Therefore, the total time complexity is O(MaxIter×N×(logN+f)), and the space complexity is O(N×d).

8. The edge computing offloading method based on atomic orbital search optimization strategy as claimed in claim 1, characterized in that: Step 3.3 describes the drone-assisted edge computing offloading method based on atomic orbital search optimization strategy, and the specific algorithm steps are as follows: Step 1: Initialize the available computing resources of the drone, the offloaded tasks of each device, and the parameters required by the OAOS algorithm; Step 2: Generate an initial offloading plan for each device, evaluate each offloading plan based on expected latency and energy consumption, and add the evaluated offloading plan to the initial population, which will serve as the starting point for the OAOS algorithm optimization process; Step 3: Call Algorithm 1 and use the OAOS algorithm to optimize the initial offloading plan; Step 4: Use the best offloading solution for each device, calculate the weighted sum of actual delay and energy consumption, and record the results; Step 5: After applying the optimal solution, check whether the constraints are met. If they are met, return it as the global optimal strategy; The time complexity is dominated by the optimization step of the OAOS algorithm, and its time complexity is approximately O(MaxIter×N×(MaxLayeer+log n)); the space complexity depends on the population storage and fitness value storage, so the space complexity is O(N×K), where K is the number of terminal devices.

Images