Method for drone-assisted thermal-aware mec network task scheduling and resource allocation

Abstract

The application discloses a method for unmanned aerial vehicle assisted thermal sensing MEC network task scheduling and resource allocation. In the case of considering CPU thermal constraints, the cruising time of the unmanned aerial vehicle is minimized by jointly optimizing user admission policy, task scheduling, unmanned aerial vehicle hovering trajectory, flight time, and computing and communication resource allocation. The optimization problem is further decomposed into three sub-problems and iteratively processed in order. Simulation results show that the method has good performance improvement effect compared with other algorithms.

Description

Technical Field

[0001] This application relates to the field of communication technology, and in particular to a method for task scheduling and resource allocation in a drone-assisted thermal sensing MEC network. Background Technology

[0002] Due to the unique mobility and cost advantages of unmanned aerial vehicles (UAVs), existing technologies utilize UAVs as aerial MEC (Mobile Edge Computing) servers, effectively expanding their service coverage in resource-scarce areas. UAV networks can provide various services to IoT devices by adjusting the UAV's trajectory, such as computational offloading, data collection, and content caching.

[0003] However, compared to MEC servers integrated into base stations, airborne MEC servers are relatively smaller, lighter, and have less computing power, size, and weight due to factors such as hardware cost, user experience design, and deployment environment. To provide better computing services for drones and ensure greater maneuverability, the computing power, size, and weight of airborne MEC servers need to be comprehensively considered. For example, the DJI Manifold 2 platform uses an Intel Core i7-8550U processor with a CPU clock speed of 1.8GHz. Meanwhile, the DJI Manifold 2 is only 11 cm long (width) and 2.6 cm thick, and weighs less than 200 grams. On the other hand, with the development of semiconductor technology used in chips and its exponential performance improvements, more and more embedded real-time systems are expected to be implemented on these power density computing platforms, which further brings new challenges to chip heat dissipation and temperature control.

[0004] Existing literature on UAV-supported MEC networks primarily focuses on the energy consumption of UAVs during computing, hovering, or flight, without disclosing how to address the issue of excessive CPU temperature in these situations. Although DVFS scheduling can effectively alleviate or resolve the hardware limitations (i.e., CPU temperature) of airborne MEC servers and improve the efficiency of computing resource allocation, bottlenecks in computing resources and energy supply still hinder the promotion and application of UAV-supported MEC networks.

[0005] To further improve the utilization efficiency of communication and computing resources in the MEC network supported by drones, the drones adopt a "hover-fly-hover" mode to access all hovering points. While this mode can help minimize total flight time and total energy consumption, or maximize the minimum average throughput for all users, it wastes the drone's computing resources during flight by ignoring them.

[0006] Mobile edge computing (MEC) powered by unmanned aerial vehicles (UAVs) can provide services to resource-scarce areas, such as rural areas and disaster relief. However, the computing power and energy supply of UAVs mounted on MEC servers are limited. When used in computationally intensive applications, hardware limitations (such as size and weight restrictions) can lead to severe heat dissipation, resulting in excessively high CPU temperatures.

[0007] Existing methods do not consider the impact of CPU temperature on the calculation results and hovering time of the "hover-flight-hover" mode, and cannot achieve the minimum hovering time of the drone while taking CPU temperature into account. Summary of the Invention

[0008] This application provides a method for task scheduling and resource allocation in a drone-assisted thermal sensing MEC network to address the technical problems in the prior art where the computing power and energy supply of drone-supported mobile edge computing servers are limited, and when used for a long time in computationally intensive applications, the CPU temperature will be too high due to hardware limitations, and the drone's cruise time in the existing "hover-flight-hover" mode cannot reach the shortest possible time.

[0009] This application provides a method for task scheduling and resource allocation in a drone-assisted thermal sensing MEC network, including the following steps:

[0010] Step S1: For Given a hovering trajectory Q, allocate computational and communication resources. Flight time The optimal user admission strategy is obtained by using the K-means clustering algorithm. The optimal task scheduling strategy is obtained by combining the greedy and B&B algorithms. The forwarding decision is obtained using a greedy algorithm. The B&B algorithm is used for calculation and caching decisions. ,get In obtaining Later update And based on the newly obtained Design scheduling strategies for unforwarded tasks to enable immediate computation or temporary caching.

[0011] Step S2: For Based on the optimal user admission policy Optimal task scheduling strategy Substitute The optimal flight time was calculated using the CVX solver. ;

[0012] Step S3: For : Transform non-convex constraints into convex constraints using continuous convex approximation, within a given global lower bound. Optimize the temporary hover trajectory during the i-th iteration in the inner loop. This continues until convergence to a tolerable accuracy, at which point the optimal result is obtained. And update again For subsequent optimization;

[0013] Step S4: For : Using given conditions The next first-order Taylor expansion will The problem is transformed into a convex problem, and then the optimal allocation strategy for computational and communication resources is obtained iteratively using CVX based on auxiliary variables. ;

[0014] Step S5: Sequential Iterative Optimization , , , And update the relevant variables to obtain the optimal cruise conditions for the UAV-based MEC network;

[0015] Among them, 1) User access control policy and task scheduling policy; 2) Flight schedule; 3) 4) Hover trajectory optimization; , and strategies for allocating computing and communication resources.

[0016] Preferably, step S1 includes the following steps:

[0017] Step S11: Use the K-means clustering algorithm to obtain the optimal user admission strategy for each hovering point. ;

[0018] Step S12: From the unloaded task Select the task to be processed and send it to the base station for processing using a greedy algorithm;

[0019] Step S13: Use the B&B algorithm to obtain the optimal scheduling decision for computation or caching. .

[0020] Preferably, step S11: The N observations / users are divided into K clusters / groups using the K-means clustering method, where K≤N, each observation belongs to the cluster center closest to the mean, and the distance between different cluster centers is the greatest.

[0021] Given a set of user locations and user group collection K-means clustering minimizes the sum of squares within a cluster, with the goal of finding:

[0022] (9)

[0023] in, It is clustering The average location of users in the middle;

[0024] A two-stage heuristic is used to find the clustering solution to initialize the positions of the K centroids:

[0025] Step S111: In the t-th iteration, assign the n-th user to the nearest center, based on:

[0026]

[0027] Among them, the nth user belongs to the kth cluster;

[0028] Step S112: Update the cluster using the average value of user coordinates in the group according to the following formula. The location of the center point:

[0029]

[0030] in, It is the cluster at the t-th iteration. The number of users in the middle;

[0031] Step S113: Repeat steps S111 to S112 until the clustering is stable.

[0032] Preferably, for time-sensitive computational tasks, step S2 is obtained through calculation by the ground base station integrated server;

[0033] Step S1 includes the following steps: making forwarding decisions using a greedy algorithm, then sending the most recently unloaded task to the base station, and reducing the hover time based on its latency requirements;

[0034] Step S114: Based on the optimal user admission policy in step S1 The clustering user set of the j-th hovering point express, and These are the size and index of the user in the cluster, respectively;

[0035] Step S115: and Representing the user's forwarding and local computation sets respectively, for all ,from Select the first The number of users, and the number of users The task of unloading data from a user is forwarded to the base station, and it is determined whether the operation simultaneously satisfies constraints (7b)-(7r):

[0036] (7b)

[0037] (7c)

[0038] (7d)

[0039] (7e)

[0040] (7f)

[0041] (7g)

[0042] (7h)

[0043] (7i)

[0044] (7j)

[0045] (7k)

[0046] (7l)

[0047] (7m)

[0048] (7n)

[0049] (7o)

[0050] (7p)

[0051] (7q)

[0052] (7r)

[0053] If the judgment result is yes, then it makes Then let ;

[0054] Step S116: Repeat steps S114-115 until there are no more forwardable offload tasks, and reduce the optimal flight time obtained in step S2. Derive the j-th point ;

[0055] Step S117: Repeat steps S114-117 for all hover points to make a forwarding decision.

[0056] Preferably, step S4 includes the following steps:

[0057] Based on income Under the constraints of CPU temperature, computation, and cache, the following equation is solved using the CVX solver:

[0058] (22a)

[0059] (22b)

[0060] (22c)

[0061] (22d) (22e)

[0062] (22f)

[0063]

[0064] To obtain the optimal communication and computing resource allocation strategy to support data transmission for time-sensitive computing tasks.

[0065] Preferably, it includes the following steps:

[0066] Step S6: Set the initial temperature , and This represents the number of iterations in the inner and outer loops at the current temperature. As a cooling factor, based on the cluster user's initialized hover trajectory And ensure that each user can access the drone from any hovering point;

[0067] Step S7: Based on the current trajectory satisfying constraints (7i), (7k), (7l), (7q), (7r), and (12b) Generate hover trajectory in the i-th iteration ,calculate , It is given The cruise duration of the drone;

[0068] Step S8: If ,use Update the initial hover trajectory. If Probability can be used Will Replace with Specifically, a random value It is generated by the standard normal distribution, when Time settings Otherwise discard .

[0069] The beneficial effects that this application can produce include:

[0070] 1) The UAV-assisted thermal sensing MEC network task scheduling and resource allocation method provided in this application minimizes the UAV's cruise time by jointly optimizing user admission policies, task scheduling, UAV hovering trajectory, flight time, and computational and communication resource allocation, while considering CPU thermal constraints. Simulation results using the method provided in this application show that, compared with other algorithms, the proposed method can guarantee the UAV's cruise time while ensuring CPU temperature, demonstrating a significant performance improvement.

[0071] 2) This application provides a method for task scheduling and resource allocation in a UAV-assisted thermal sensing MEC network. This method offers a framework for a UAV-assisted cloud edge MEC network to minimize cruise duration under CPU temperature constraints, where user admission policy, task scheduling, hovering trajectory, and computational and communication resource allocation strategies are jointly optimized. This application is the first to attempt to study the impact of the CPU temperature of an airborne MEC server on computational resource allocation strategies. Simultaneously, the MEC server is equipped with a UAV that provides services to the device in a "hover-hover" mode and assists in calculating tasks unloaded during flight.

[0072] 3) The method for thermal sensing task scheduling and resource allocation based on UAV-based MEC networks provided in this application utilizes Continuous Convex Approximation (SCA) and Block Coordinate Descent (BCD) techniques to derive a dual-loop iterative algorithm to solve the defined MINLP problem. Furthermore, due to the branch and bound (B&B) algorithm... The computational complexity is high, especially for a large number of users and hover points. Therefore, a low-complexity algorithm based on K-means clustering and a greedy algorithm was designed to reduce the computational complexity. The complexity.

[0073] 3) The UAV-assisted thermal sensing MEC network task scheduling and resource allocation method provided in this application demonstrates that WSABC outperforms other schemes, further proving the significance and benefits of computation and cache scheduling, in-flight computation, and BS collaboration. Compared with WCWF and WBC, WSABC can significantly shorten cruise time by 14.9%-18.8% and 24.2%-29.8%, respectively. Although this method sacrifices at most 14.4% of cruise time compared to WTC, considering the optimization effect on CPU temperature, it will benefit the performance and reliability of the computing platform. Attached Figure Description

[0074] Figure 1 This is a schematic diagram of the MEC system structure for the hybrid activation of UAV-base station in a specific embodiment of this application;

[0075] Figure 2 This is a line graph comparing cruise duration with different CPU frequencies under different schemes in the embodiments of this application;

[0076] Figure 3 Line graphs showing the comparison of CPU temperature (and cruise time) with different CPU frequencies under different schemes in the embodiments of this application;

[0077] Figure 4 This is a bar chart comparing the speeds of UAVs under the WSABC and WCWF schemes in the embodiments of this application;

[0078] Figure 5 This is a line graph comparing the cruise times of different schemes under different bandwidths in the embodiments of this application;

[0079] Figure 6 Line graphs showing the comparison of cruising time and different cache capacities under different schemes in the embodiments of this application;

[0080] Figure 7 This is a diagram showing the convergence results of the WSABC scheme in the embodiments of this application;

[0081] Figure 8 This is a trajectory diagram of a WSABC-based UAV relative to a base station at different positions in an embodiment of this application;

[0082] Figure 9 This is a bar chart showing the data volume of different operations at different locations of the base station based on WSABC in the embodiments of this application. Detailed Implementation

[0083] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0084] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0085] Technical features not intended to solve the technical problems of this application are all set or installed according to common methods in the prior art, and will not be described in detail here.

[0086] This application provides a method for task scheduling and resource allocation in UAV-assisted thermal sensing MEC networks. This method proposes a thermal sensing task scheduling and resource allocation strategy to minimize UAV cruise time. Since the problem is a complex mixed-integer nonlinear programming (MINLP) problem with strongly coupled variables, it is transformed into three more manageable subproblems to further decouple these variables:

[0087] 1) User access control policies and task scheduling policies;

[0088] 2) Flight schedule;

[0089] 3) Hover trajectory optimization;

[0090] 4) The calculation of communication resource allocation strategy is performed and processed iteratively in sequence.

[0091] Driven by the aforementioned pioneering work, the MEC server located at the BS will be regarded as an additional auxiliary to the airborne server in handling high-energy-consuming and time-sensitive tasks, with the device offloading its tasks to the drone for further processing (either computed or cached on the drone, or forwarded to the base station).

[0092] I. System Model and Problem Formulation

[0093] See Figure 1 The UAV-base station hybrid MEC system used in this application consists of a rotary-wing UAV, a base station (BS), and a set of [unclear - possibly related to a device or system]. The components of the Internet of Things (IoT) are represented.

[0094] Specifically, the drone has an onboard computing processor with limited computing power. The base station, on the other hand, is equipped with a high-performance processing server to provide high-speed computing services. Furthermore, assume that the nth device has a time-sensitive computing task. , represented as { , , }, ∀n, where Indicates the amount of input data to be processed (unit: bits). It is the CPU cycles required to compute one bit of input data, and This is the maximum permitted delay to complete the mission. The drone is for each... IoT devices at marked hover points provide computing and caching services. From the initial hover point To the final hovering point The drone flies at a fixed altitude H during its cruise. Without sacrificing versatility, a three-dimensional Cartesian coordinate system is used to describe the positions of the drone and nodes (base stations and IoT devices). Specifically, = , = and Let $\mathbf{n}$ represent the location of the base station, the location of the $n$-th device, and the location of the drone at the $j$-th hovering point, respectively. For simplicity, the base station is represented as the 0th node; therefore, the distance between the $n$-th node and the drone at the $j$-th hovering point is $\mathbf{n}$. .

[0095] remember and for The decision to unload and forward the drone when it hovers at point j. Specifically, =1 indicates that the device will Unload it onto the drone. Otherwise, =0. Meanwhile. =1 indicates that the received task will be forwarded from the drone to the base station; otherwise... =0. Meanwhile. =1 means the received task is forwarded from the drone to the base station. Otherwise, =0. Furthermore, for cases where there is no forwarding, [the following will be displayed:] Represented as a computation or caching task Based on the scheduling decision, the task will not be forwarded to the base station. Specifically, =1 indicates a task that computes the nth device, which was recently uninstalled or previously cached. Otherwise, when The drone then chose to cache the task for later processing. This was due to the unloading of the task. It will be further scheduled for forwarding, computation, or caching, and must be computed on the drone or base station. It should be uploaded to the drone at a certain hovering point and meet the following requirements. .

[0096] also, The forwarding decision can only be made after the unloading process is completed at the j-th hovering point, satisfying the condition that... In other words, it is not possible to make a decision after the j-th hovering point. The forwarding decision. Due to It is indivisible and must be calculated. It can be processed on the drone or forwarded to the base station via the backhaul link for further calculation to meet the requirements. Due to the causal relationship of information, Calculations, caching, or forwarding can only be performed after the upload and transmission are complete, thus satisfying the requirement. .

[0097] Due to the strict latency requirements of time-sensitive computational tasks, forwarded tasks must be processed immediately, but cached tasks can be scheduled for subsequent processing on the drone. Furthermore, the buffer space occupied cannot exceed the maximum cache capacity of the drone, denoted by C, and is subject to the following limitations: ,in This represents the space occupied at point (j-1), and the last three terms represent the tasks currently unloaded by the user, the tasks forwarded to the base station, and the computational tasks at point j, respectively.

[0098] Communication mode

[0099] Assuming the wireless channel between the drone and the device is dominated by a line-of-sight loss link, the offloading channel gain between the nth device and the drone at the j-th point is... ,in, For reference distance The channel gain at point j is... Similarly, the forwarding channel gain between the UAV and the base station at point j is... .

[0100] Since the considered drone-base station hybrid-enabled MEC system is based on FDMA (Frequency Division Multiple Access), the drone and equipment will share the common bandwidth W during offloading and forwarding. ∈[0,1] and ∈[0,1] represents the bandwidth portion of the backhaul link and the upload link from the nth user to the jth point drone, satisfying... Therefore, the achievable rates for offloading and forwarding from the nth user to the drone and from the drone to the base station can be expressed as follows: ,in , Indicates the transmission power of the user and the drone. This represents the noise power spectral density of drones and base stations.

[0101] To maintain good end-to-end performance, the total data rate of all access links from the drone to the base station should not exceed the achievable rate of the backhaul link at each point, given... .

[0102] B. Delay

[0103] Due to limited computing power, it is assumed that the device will offload its tasks for further processing, rather than computing them locally. Based on this, the following is given: Unloading time and its corresponding calculation time:

[0104] (1)

[0105] (2)

[0106] in, and Let represent the computing resources allocated to the nth user by the drone and the base station at the j-th point, respectively. Since the cache write and read operations are negligible, therefore... The execution delay at point j is (3)

[0107] according to ,here and Both cannot be 1 simultaneously. Furthermore, the computing resources allocated to the device cannot exceed the computing capabilities of the drone and the base station. and This indicates that they respectively satisfy and .

[0108] Because stable offloading and forwarding channels facilitate transmission during drone hovering, both the device and the drone can upload and forward their tasks. In other words, the drone can only fly to the next hovering point after the offloading and forwarding process is complete. Therefore, the drone's cruise time between point j and point (j+1) can be written as... = + It includes the following two parts:

[0109] 1) Hovering time: Its lower bound can be represented as This ensures stability during uploads. It's important to note that the forwarding and unloading phases occur simultaneously to transmit data that needs to be forwarded at point j, therefore the time is equivalent.

[0110] 2) Flight time: It is given by the following formula ,in, Let be the average velocity from point j to point j+1. This indicates the maximum flight speed of the drone.

[0111] To further improve the efficiency of computing resource utilization, unloading tasks can also be handled while the drone is flying between hovering points. Furthermore, in addition to cached tasks, computational and forwarding tasks must be fully processed by the drone and base station before reaching the next hovering point, meeting [the required parameters]. .

[0112] To better understand, The points that are unloaded and calculated are represented as and , respectively by and Given. Note that the offloading and computation points are not necessarily the same, but the offloading and forwarding points must be the same because the forwarded task has strict latency requirements for processing. Therefore, The entire execution delay consists of the following two parts:

[0113] l Calculated waiting time for

[0114]

[0115] The first two conditions indicate that the unloading and calculation processes are executed at the same point.

[0116] The calculation time is .

[0117] The entire execution delay can be further expressed as: (4)

[0118] C. Energy consumption of drones

[0119] Because limited onboard energy restricts the drone's cruise time, the energy consumed by computation, buffering, relaying, hovering, and flight must be carefully managed.

[0120] Calculation: Using Dynamic Voltage Frequency Scaling (DVFS) technology, the energy consumption of user tasks is calculated as follows: , The computational energy efficiency coefficient is related to the drone processor chip.

[0121] Forwarding: The energy consumption for forwarding user tasks to the base station is... ,in, It is the transmission power of the forwarding task.

[0122] Caching: If the offloading task is not calculated or forwarded to the base station on the drone, it must be temporarily cached on the drone. =0, the corresponding energy consumption of the cache is , Cache energy factor related to the storage hard drive of drones

[0123] Hovering: The total energy consumed by hovering is Among them, the hovering function It is a constant value.

[0124] Flight: Based on the results obtained from the hovering step, the energy consumption of the rotary-wing UAV during flight is... Here M is the mass of the drone.

[0125] Therefore, the total energy consumption of a drone can be written as This formula should satisfy , It's the energy budget for drones.

[0126] D. Power and thermal energy model of UAVs

[0127] Since the CPU processor is usually packaged inside the drone's fuselage, and given the drone's small size and poor heat dissipation, overheating can cause permanent damage to the CPU. This article considers CPU temperature to protect the processor.

[0128] Power Model: Since a processor is either idle or active, dynamic power management needs to be considered. Specifically, if the processor is currently executing a task, it is considered active; otherwise, it is idle. Its power consumption is expressed as... ,in, This is the power dissipation of the processor in preparation for computational tasks, and can be considered a constant value. Given a set of tasks... , The additional dynamic capabilities for computational tasks are represented as .

[0129] Thermal Model: Driven by pioneering work, a well-known thermal circuit model was used to convert the processor's power consumption into its temperature. and Let S(t) represent the average processor power and ambient temperature over a time period t. Therefore, the processor temperature S(t) at the end of this period is:

[0130] (5)

[0131] in, and These are thermal resistance and capacitance, respectively. This is the processor's initial temperature. It can be seen that the temperature will rise / fall, eventually reaching... The stable temperature S(∞) of the CPU processor is defined as... (6)

[0132] To meet thermal requirements, the stable temperature should meet the following conditions. ,in This is the highest temperature of the CPU processor.

[0133] Problem Description

[0134] Will This is represented as an uninstallation decision (also known as user permission), which will This is indicated as a forwarding decision. Represented as a scheduling decision for computation or caching, Represented as the hovering trajectory of the drone, This represents the CPU frequency allocated to the user by the drone and base station. This is represented as the bandwidth allocated to users and drones. Indicated as hover time, This refers to the flight time of the drone as it flies between different points. It's important to note that in the following sections... These can be considered auxiliary variables. In this paper, we jointly optimize the above variables to minimize the drone's cruise duration, as shown in the formula:

[0135] (7a)

[0136] (7b)

[0137] (7c)

[0138] (7d)

[0139] (7e)

[0140] (7f)

[0141] (7g)

[0142] (7h)

[0143] (7i)

[0144] (7j)

[0145] (7k)

[0146] (7l)

[0147] (7m)

[0148] (7n)

[0149] (7o)

[0150] (7p)

[0151] (7q)

[0152] (7r)

[0153] (7s)

[0154] (7t)

[0155] Due to hardware limitations, the UAV's storage space (7j), energy consumption (7k), computing resources (7n), and CPU stable temperature (7s) are all restricted. Furthermore, constraint (7l) stipulates that each user's task must be fully computed within a given latency requirement, and service can only be provided to the nth user if the user is within the coverage area of ​​the j-th point (7p) of the UAV. The service radius of the drone.

[0156] II. Algorithm Proposal

[0157] According to Observations show that due to the strong coupling variables in the objective function and constraints... This demonstrates that this is a very challenging MINLP problem. Furthermore, these integer variables... This further increases the difficulty of solving the problem. Therefore, it is impossible to directly obtain the current form. The solution. Inspired by iterative design, the BCD method is applied to decouple these variables, and... It is divided into the following four sub-problems:

[0158] 1) User access control policies and task scheduling policies;

[0159] 2) Flight schedule;

[0160] 3) Hover trajectory optimization;

[0161] 4) , and strategies for allocating computing and communication resources.

[0162] First, given the hovering trajectory Q, allocate computational and communication resources. Flight time The optimal user access method is obtained through K-means clustering algorithm. Furthermore, by combining the applications of the Greedy and B&B algorithms, an optimal task scheduling strategy is proposed. .

[0163] Specifically, forwarding decision The result is obtained from a greedy algorithm, and the calculation and caching decisions are made accordingly. It is then obtained through the B&B algorithm.

[0164] It is worth noting that, You can get To be updated later.

[0165] Second, under the newly acquired user access policy and task scheduling policy, the optimal flight time It can be obtained using the CVX solver because the corresponding subproblem is convex.

[0166] Third, the non-convex constraint is transformed into a convex constraint using the Continuous Convex Approximation (SCA) by introducing a given global lower bound. The temporary hovering trajectory during the i-th iteration is optimized within its inner loop. Optimal results can be obtained before reaching the allowable accuracy. .

[0167] also, It can be updated again and used for subsequent optimizations.

[0168] fourth, You can first use the given conditions The first-order Taylor expansion is transformed into a convex problem. Then, with the help of auxiliary variables, the optimal allocation strategy for computational and communication resources can be obtained iteratively based on CVX. .

[0169] Finally, the four sub-problems will be optimized sequentially, and the relevant variables will be updated. This process is called the BCD method.

[0170] We will use an iterative method based on the BCD approach to jointly optimize user access, task scheduling, hovering trajectory, computation, and communication resource allocation, and give the computational complexity of the algorithm.

[0171] A. User access control and task scheduling

[0172] As mentioned above, the uninstallation task The drone will be dispatched based on the following two scenarios:

[0173] 1) Forwarding, transmitting to the base station for calculation;

[0174] 2) Non-forwarding: immediately computed or cached in the drone, and computed in the future.

[0175] Because the unloading decision, computation decision, and caching decision are strongly coupled and have a significant impact on each other, it is necessary to solve for the given conditions. Below By jointly optimizing user access control and task scheduling strategies, the hovering time of drones can be minimized.

[0176] (8a)

[0177]

[0178] Due to the coupling of the two variables, it is strongly nonconvex. To handle this integer-related subproblem, the B&B method is used. By constructing a search tree and calculating boundary values, the optimal solution can be obtained. .

[0179] Although the B&B algorithm can find the global optimum by recursively traversing the solution space, Increasing the complexity of the algorithm will make it less suitable for practical applications.

[0180] In order to To make it easier to handle, a combined approach is proposed to optimize and update the three variables in sequence.

[0181] First, for each hovering point, the optimal user admission strategy is obtained using the K-means clustering algorithm. .

[0182] Second, optimal forwarding decisions can be designed. The task is selected from the unloaded tasks and then sent to the base station for further processing according to a greedy algorithm.

[0183] Third, the B&B algorithm is further applied to achieve optimal scheduling decisions for computation or caching. This combined method does not directly use the B&B algorithm for solution. Its computational complexity is low, which helps to simplify the subsequent processing.

[0184] 1) User Admission Strategy Based on K-means Clustering Algorithm: As the number of users grows, the difficulty of designing a user admission strategy using the B&B algorithm gradually increases. Inspired by prototype-based clustering algorithms, the K-means clustering method is used to divide N observations / users into K clusters / groups, where K≤N. Moreover, each observation belongs to the cluster (cluster center) closest to the average value, and the distance between different cluster centers is the greatest.

[0185] Given a set of user locations and user group collection K-means clustering minimizes the sum of squares (squared Euclidean distance) within clusters. In general, the goal is to find...

[0186] (9)

[0187] It is clustering The mean of the user locations is used. However, due to NP-hardness, it is difficult to obtain the minimum within-class variance of the K clusters. To address this computationally expensive problem, a two-stage heuristic is used to find the clustering solution to initialize the locations of the K centroids:

[0188] Step 1: In the t-th iteration, assign the n-th user to the nearest center, based on:

[0189]

[0190] The nth user belongs to the kth cluster.

[0191] Step 2: Update the cluster using the average coordinates of users in the group, according to the following formula. The location of the center point.

[0192]

[0193] It is the cluster at the t-th iteration. The number of users in the country.

[0194] Step 3: Repeat the above steps until the clusters stabilize. In other words, keep repeating until the cluster centers and their respective users no longer change.

[0195] Therefore, users will be able to determine their geographic coordinates. Users are assigned to appropriate clusters, each containing different hover points. Within each cluster, users can assign tasks to drones, thus deriving the optimal user admission strategy. In this way, the computational complexity of finding the optimal user admission strategy will be greatly reduced.

[0196] 2) Forwarding decision based on greedy algorithm: Due to the limited computing power of the airborne MEC server, the UAV cannot meet all the requirements of time-sensitive computing tasks. Therefore, these computing burdens need to be forwarded to the base station for further processing.

[0197] As a complement to computation, base station integrated servers are more powerful and better suited for handling low-latency tasks compared to airborne MEC servers, thus saving time. To fully utilize communication and computing resources, it is beneficial to forward as many tasks as possible within a given backhaul link bandwidth. Therefore, a greedy algorithm-based method is proposed to make forwarding decisions, then send the most recently offloaded tasks to the base station, reducing hover time based on its latency requirements. Furthermore, until there is no way to reduce hover time, the forwarding task continues to be selected at the current moment.

[0198] For greater clarity, the details of the greedy algorithm-based method are as follows:

[0199] Step 1: Based on the user admission policy, cluster the users at the j-th hover point using the set. express, and These are the size and index of the user in the cluster, respectively;

[0200] Step 2: For simplicity, use and These represent the user's forwarding and local computation sets, respectively. For all ,from Select the first Each user, and forwards its offloaded task to the base station, provided that the decision not only satisfies constraints (30b)-(30r), but also helps reduce hover time, so that... Then let ;

[0201] Step 3: Repeat the above steps until there are no more forwardable unloading tasks, and reduce the hover time in step 2 to export point j. ;

[0202] Step 4: In a similar manner, forwarding decisions can be made for all hover points.

[0203] 3) Computation or cache scheduling based on the B&B algorithm: Next, we will utilize the newly obtained... Design scheduling strategies for unforwarded tasks, allowing them to be computed immediately or temporarily cached for further processing on the drone. Therefore, the scheduling-related problem can be represented as:

[0204] (10a)

[0205]

[0206] This is an integer programming problem that can be solved using the B&B algorithm.

[0207] In summary, user access policies and task scheduling strategy Alternating optimization can be performed based on the above process. This method no longer directly applies the B&B method for derivation. This is because the process becomes difficult and time-consuming as the number of users and hovering points increases, and the limited embedded battery capacity further hinders the implementation of this method on drones equipped with MEC servers. Additionally, auxiliary variables... The newly obtained variables can be updated accordingly.

[0208] B. Flight time and hovering trajectory

[0209] Newly Obtained and given Under these conditions, designing flight and hovering trajectory planning to authorize user access and satisfy different constraints (temperature limits, computing power, and storage capacity) at corresponding points is crucial. Next, we will attempt to find the optimal flight time arrangement for a given initial hovering trajectory by solving the following problem. :

[0210] (11a)

[0211]

[0212] Because it is for It's convex, so it can be solved directly using CVX. Optimal flight time. This not only meets the flight time requirements but also the speed constraints of the drone, making hovering trajectory planning more intuitive. Therefore, it allows for the determination of the optimal hovering trajectory. Optimize, auxiliary variables Updated based on the following issues:

[0213] (12a)

[0214] (12b)

[0215] Since the relevant constraints (7i), (7k), (7l), (7q), and (7r) are all non-convex with respect to Q, they cannot be solved using existing formulas. To address the non-convexity of the constraints, an algorithm based on SCA is proposed, which uses appropriate convex approximations to iteratively solve the non-convex constraints. Value. Specifically, a relaxation auxiliary variable is introduced. To handle constraints and The relevant non-convex part, in which It is given by the following formula:

[0216] (13)

[0217] .Although about It is neither convex nor concave, but for the above constraints, it is... It is convex. Using the framework of convex approximation, the first-order Taylor approximation of a convex function is the global lower bound of the function. Therefore, for a given local point at the i-th iteration... , It can have a lower lower bound.

[0218] (14)

[0219] and These are the constant rate and first derivative at the i-th iteration. .final, This can be further expressed as the following problem:

[0220] (15a)

[0221] (15b)

[0222] (15c)

[0223] (15d)

[0224] (15e)

[0225] (15f)

[0226] (15g)

[0227] Due to (15b) This ensures that it remains non-convex. To address this issue, an auxiliary variable is introduced. To be further processed ,in It is for It is convex. Using the first-order Taylor approximation, The lower limit is:

[0228] (16)

[0229] Unlike the above operations for processing non-convex parts Its upper bound is introduced to replace the left-hand side of (38b) to achieve scale consistency, denoted as .because for It is monotonically decreasing, therefore its upper bound is... It can be represented as:

[0230] (17)

[0231] therefore, This can be rephrased as the following question:

[0232] (18a)

[0233] (18b)

[0234] (18c)

[0235] It is convex and can be solved directly using a solver, such as CVX.

[0236] C. Allocation of communication and computing resources

[0237] Based on the newly derived The next step is to design optimal communication and computing resource allocation strategies to support data transmission and meet the requirements of time-sensitive computing tasks under CPU temperature, computation, and cache constraints. Following this, we can obtain... Update auxiliary variables To further minimize the drone's cruise time,

[0238] (19a)

[0239] ;

[0240] Due to the backhaul link (7i), energy supply (7k), and delay requirements (7l), The constraints are non-convex and nonlinear, therefore an auxiliary variable is introduced. As The lower bound is given as follows:

[0241] (20)

[0242] The left side of the inequality relative to It is concave because of its second derivative. It is not true, .

[0243] In other words, due to The feasible area is With the help of [the constraint], the objective function value is reduced, and the obtained objective function value is the upper bound of the original objective function value. Therefore, if the equality in equation 20 holds, the objective function values ​​are equal. Along this path, the left side of constraint inequality (7i) [is constrained]. A larger value should be substituted. Because it is not positive. Using SCA technology, auxiliary variables Considered as At any given local point The upper bound below is given as:

[0244] (twenty one)

[0245] Add the two constraint expressions (20) and (21) to In this context, it can be rephrased as the following question:

[0246] (22a)

[0247] (22b)

[0248] (22c)

[0249] (22d) (22e)

[0250] (22f)

[0251]

[0252] The optimal solution can be obtained effectively through CVX.

[0253] D. Overall Algorithm and Complexity Analysis

[0254] Through the , , ,and Iterative optimization, until a tolerable accuracy is achieved in the outer loop, yields a suboptimal solution to the original algorithm.

[0255] Algorithm 1 summarizes the details of the proposed algorithm, and its complexity analysis is as follows. , and They represent and The number of iterations for the outer and inner rings. First, The computational complexity consists of three parts: 1) The first part is obtaining the user admission strategy through the K-means clustering algorithm, the complexity of which is... , , 1) The number of users in the k-th cluster; 2) The second part derives the forwarding strategy using a greedy algorithm, with a complexity of O(n log n). 3) The third part is to formulate a scheduling strategy for computation or caching using the B&B algorithm. The algorithm complexity is... .because Greater than ,therefore The algorithm complexity is Second, because CVX is used for processing... and Therefore, the corresponding computational complexity can be expressed as and .third, The computational complexity is It can be simplified to Finally, the overall algorithm complexity is O(n log n). It can be abbreviated as .

[0256] This demonstrates that the complexity of the algorithm proposed in this application is within an acceptable range, and its complexity is lower than that of directly using the bnb solution method.

[0257] The method provided in this application includes the following steps:

[0258] Input: Given initial variables Given z=0.

[0259] Output: Optimization variables .

[0260] 1: Repeat

[0261] 2: Yes Given The optimal user admission strategy is obtained by using the K-means clustering algorithm. The optimal forwarding strategy is obtained using a greedy algorithm. The optimal scheduling method for computation or caching is derived using the B&B algorithm. Finally, by utilizing the newly acquired user access control and task scheduling policies, the hover time can be updated. .

[0262] 3: Yes : The newly obtained Substitute and give Optimal flight schedule It can be obtained from the CVX solver.

[0263] 4: To By updating This allows us to obtain the optimal hovering trajectory of the drone. (Steps 5-8) Initialize i=0 in the inner loop.

[0264] 5: Repeat

[0265] 6: In Introduced in China and As the upper and lower bounds of the corresponding terms in the constraint, we attempt to find the appropriate local point. Under the condition of obtaining the optimal This addresses the convexity problem of restated representation.

[0266] 7: Update and .

[0267] 8: Until convergence reaches a tolerable level of accuracy.

[0268] 9: Regarding Given the newly obtained Optimal allocation of communication bandwidth and computing resources In the following steps (10-13), we can initialize i = 0 and set it in the inner loop. .

[0269] 10: Repetition

[0270] 11: Similarly, will and Introduction As the lower and upper bounds of the corresponding terms in the constraints, we attempt to find the appropriate local point. Under the condition of obtaining the optimal This addresses the convexity problem of restated representation.

[0271] 12: Update and .

[0272] 13: Until convergence to a tolerable level of accuracy.

[0273] 14: Update variables: and obtain .

[0274] 15: Until convergence to tolerable accuracy. Example

[0275] The simulation in this embodiment, performed according to the method provided in this application, yields the following results:

[0276] Numerical results validate the effectiveness and performance of the proposed WSABC (Scheduling and Base Station Cooperation) scheme.

[0277] Parameter settings:

[0278] Multiple devices are randomly distributed within a 1000m × 1000m area. Each device has a task awaiting computation, N=45, J=4. In the scenario under consideration, the drone will... From start to finish Meanwhile, the base station, acting as a cloud server, is located... This helps users calculate forwarding tasks. Unless otherwise stated, the simulation parameters are summarized in Table 1.

[0279] Table 1: Simulation Parameters

[0280]

[0281] 1. Hover trajectory initialization

[0282] While K-means clustering can determine the number of hovering points based on user groups, the service (flight) sequence of drones significantly impacts cruise duration performance, making planning challenging. Therefore, this study aims to initialize a suitable flight sequence considering constraints to meet requirements for latency, energy consumption, computational power, communication bandwidth, and buffer capacity. In practical applications, obtaining the initial service sequence is a type of Traveling Salesman Problem, which can be solved using the Simulated Annealing (SA) algorithm, as shown below:

[0283] Step 1: Initial temperature express, and This represents the number of iterations in the inner and outer loops at the current temperature. As a cooling factor, initialize a hover trajectory based on cluster users. And ensure that each user can access the drone from any hovering point.

[0284] Step 2: Based on the current trajectory satisfying constraints (7i), (7k), (7l), (7q), (7r), and (12b) Generate hover trajectory in the i-th iteration ,calculate , It is given The cruise duration of the drone.

[0285] Step 3: If ,use Update the initial hover trajectory. If Probability can be used Will Replace with Specifically, a random value It is generated by the standard normal distribution, when Time settings Otherwise discard .

[0286] Simulation results and analysis:

[0287] To evaluate the performance of the proposed overall algorithm based on scheduling and base station cooperation WSABC, the following reference scheme is considered as a benchmark, named

[0288] 1) Drones as relays (UAR): Drones act as relays, forwarding user tasks to base stations.

[0289] No base station cooperation (WS) is available in this scenario, where no base station is deployed.

[0290] Without scheduling (WS), the offloaded task must be calculated immediately at the drone or the base station at the current hovering point.

[0291] There is no thermal constraint (WTC), and the CPU temperature limit of the onboard MEC server is not taken into account.

[0292] The cruise duration results for each scheme are listed below. Figure 2 In our study, we evaluated the impact of the maximum CPU frequency of the airborne MEC server on cruise time performance. The optimization methodology used in different benchmark scenarios was the same.

[0293] The results show that the cruise time decreases with increasing onboard MEC server CPU frequency. This is because when the computing power of the MEC increases, the UAV can allocate more computing resources to handle offloaded tasks, thus saving computing time and cruise time, as shown in Equations (1) and (7r). Compared with WBC (or WS, or WCWF), it can be seen that as the CPU frequency increases, the percentage of cruise time saved increases from 24.5% to 31.8% (or from 6.4% to 16.6%, or from 15.5% to 17.4%). It is worth noting that if the maximum CPU frequency of the onboard MEC server is less than 4GHz, the WBC scheme cannot meet the user's latency requirements. In addition, when the CPU frequency of the onboard MEC server increases, the rate of reduction in cruise time slows down or saturates. This is because the UAV prevents its CPU from being permanently damaged due to overheating. To understand the impact of CPU temperature on UAV cruise time more deeply, we emphasize Figure 3The temperature-time trade-off. Compared to the significant impact of WTC on cruise time, increasing the CPU frequency of the onboard MEC server does not significantly help reduce the cruise time of WSABC. Considering the thermal constraints of the CPU, WSABC will control the CPU temperature, keeping it near its peak (60°C) to protect the performance and reliability of the computing platform, although it will sacrifice up to 14.4% of cruise time compared to WTC.

[0294] Figure 4 This displays the drone's speed under the WSABC and WCWF schemes. In WCWF, to conserve cruising time, the drone will fly at maximum speed to the next hovering point. A speed of 0 indicates that the drone is circling to provide computing, relay, or caching services to the user. From Figure 4 It can be seen that, compared to WSABC, drones in WCWF mode spend more time cruising and hovering because the drone must handle the same number of unloading tasks even if it cannot handle tasks in flight mode. Furthermore, Figure 5 This demonstrates the impact of total bandwidth on cruise time performance. As communication resources become more abundant, offloading and forwarding times gradually decrease, as shown in Equation 1, and the drone's cruise time also gradually decreases. Clearly, WSABC outperforms other solutions in cruise time performance because it achieves the optimal user admission and task scheduling strategies mentioned above.

[0295] also, Figure 6 In this study, the drone's cruise duration decreases as the drone's cache capacity increases. This is because the drone has more buffer space to temporarily store additional tasks uploaded by the user, rather than hovering and waiting to process them. It's important to note that as the drone's cache capacity approaches a certain threshold, the reduction in cruise duration relative to cache capacity gradually decreases; that is, the impact of continuously increasing cache capacity on cruise duration performance becomes less significant. Furthermore, WSABC outperforms other schemes, demonstrating the importance and benefits of computation and cache scheduling, in-flight computation, and base station cooperation. Specifically, compared to WCWF and WBC, considering in-flight computation or leveraging base stations integrated with MEC servers, WSABC can significantly reduce cruise duration by approximately 14.9%–18.8% and 24.2%–29.8%, respectively.

[0296] Figure 7 The convergence performance of WSABC is described; it converges quickly to the iteration time for different numbers of users. Note that the cruise time increases with the number of users. This is because more and more users will share the same amount of computing and communication resources, leading to insufficient resources for performing the offloading task, thus prolonging the UAV's cruise time.

[0297] Figure 8and Figure 9 This explains the impact of base station location, namely and Impact on UAV flight trajectory and task scheduling strategy. For convenience, CP, FP, and OP represent the computation portion, forwarding portion, and total task offloading amount at the current hovering point, respectively. It can be observed that, due to task scheduling, at the first (or third) point, the offloading amount of tasks is greater than (or less than) the computation and forwarding amount. That is, some tasks are cached at the first point and computed at subsequent points. When the UAV moves away from the base station, i.e., at the first and second points, only a small portion of tasks are forwarded to the base station for processing. Simultaneously, the UAV tends to forward all offloading tasks to the base station for further computation at the third and fourth points, because the offloading time is relatively small compared to the first and second points, or because the UAV aims to prevent the CPU from overheating due to high computational load. Figure 8 The "user" in the middle refers to the user; "BS" represents the base station. The first red marker indicates the position of the BS in the upper right corner, the first blue marker indicates the position of the BS in the upper left corner, the second red marker indicates the trajectory of the drone when it is in the upper right corner of the BS, and the second blue marker indicates the trajectory of the drone when it is in the upper right corner of the BS.

[0298] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for task scheduling and resource allocation in a drone-assisted thermal sensing MEC network, characterized in that, Includes the following steps: Step S1: For Given a hovering trajectory Q, allocate computational and communication resources. Flight time The optimal user admission strategy is obtained by using the K-means clustering algorithm. The optimal task scheduling strategy is obtained by combining the greedy algorithm and the B&B algorithm. The forwarding decision is obtained using a greedy algorithm. The B&B algorithm is used for calculation and caching decisions. ,get In obtaining Later update And based on the newly obtained Design scheduling strategies for unforwarded tasks to enable immediate computation or temporary caching. Step S2: For Based on the optimal user admission policy Optimal task scheduling strategy Substitute The optimal flight time was calculated using the CVX solver. ; Step S3: For : Transform non-convex constraints into convex constraints using continuous convex approximation, within a given global lower bound. Optimize the temporary hover trajectory during the i-th iteration in the inner loop. This continues until convergence to a tolerable accuracy, at which point the optimal result is obtained. And update again For subsequent optimization; Step S4: For : Using given conditions The next first-order Taylor expansion will The problem is transformed into a convex problem, and then the optimal allocation strategy for computational and communication resources is obtained iteratively using CVX based on auxiliary variables. ; Step S5: Sequential Iterative Optimization , , , And update the relevant variables to obtain the optimal cruise conditions for the UAV-based MEC network; Among them, 1) User access control policy and task scheduling policy; 2) Flight schedule; 3) 4) Hover trajectory optimization; Computing and communication resource allocation strategies; Step S4 includes the following steps: Based on income Under the constraints of CPU temperature, computation, and cache, the following equation is solved using the CVX solver: （22a） （22b） （22c） in, Task size （22d） （22e） （22f） （7m） （7n） （7o） （7p） （7q） （7r） （7s） （7t） (20) (21) To obtain the optimal communication and computing resource allocation strategy to support data transmission for time-sensitive computing tasks.

2. The method according to claim 1, characterized in that, Step S1 includes the following steps: Step S11: Use the K-means clustering algorithm to obtain the optimal user admission strategy for each hovering point. ; Step S12: From the unloaded task Select the task to be processed and send it to the base station for processing using a greedy algorithm; Step S13: Use the B&B algorithm to obtain the optimal scheduling decision for computation or caching. .

3. The method according to claim 2, characterized in that, Step S11: Use the K-means clustering method to divide the N observations / users into K clusters / groups, where K≤N, each observation belongs to the cluster center closest to the mean, and the distance between different cluster centers is the greatest. Given a set of user locations and user group collection K-means clustering minimizes the sum of squares within a cluster, with the goal of finding: （9） in, It is clustering The average location of users in the middle; A two-stage heuristic is used to find the clustering solution to initialize the positions of the K centroids: Step S111: In the t-th iteration, assign the n-th user to the nearest center, based on: Among them, the nth user belongs to the kth cluster; Step S112: Update the cluster using the average value of user coordinates in the group according to the following formula. The location of the center point: in, It is the cluster at the t-th iteration. The number of users in the middle; Step S113: Repeat steps S111 to S112 until the clustering is stable.

4. The method according to claim 1, characterized in that, For time-sensitive computational tasks, step S2 is obtained through calculation by the ground base station integrated server; Step S1 includes the following steps: making forwarding decisions using a greedy algorithm, then sending the most recently unloaded task to the base station, and reducing the hover time based on its latency requirements; Step S114: Based on the optimal user admission policy in step S1 The clustering user set of the j-th hovering point express, and These are the size and index of the user in the cluster, respectively; Step S115: and Representing the user's forwarding and local computation sets respectively, for all ,from Select the first The number of users, and the number of users The task of unloading data from a user is forwarded to the base station, and it is determined whether the operation simultaneously satisfies constraints (7b) to (7r): （7b） （7c） （7d） （7e） (7f) (7g) （7h） （7i） （7j） （7k） （7l） （7m） （7n） （7o） （7p） （7q） （7r） If the judgment result is yes, then: Then let ; Step S116: Repeat steps S114-115 until there are no more forwardable offload tasks, and reduce the optimal flight time obtained in step S2. Derive the j-th point ; Step S117: Repeat steps S114-117 for all hover points to make a forwarding decision.

5. The method according to claim 4, characterized in that, Includes the following steps: Step S6: Set the initial temperature , and This represents the number of iterations in the inner and outer loops at the current temperature. As a cooling factor, based on the cluster user's initial hover trajectory And ensure that each user can access the drone from any hovering point; Step S7: Based on the current trajectory satisfying constraints (7i), (7k), (7l), (7q), (7r), and (12b) Generate hover trajectory in the i-th iteration ,calculate , It is given The cruise duration of the drone; （7i） （7k） （7l） （7q） in, （7r） ； Where, q j+1 : The coordinates of the UAV at the (j+1)th hovering point; q j : The coordinates of the UAV at the j-th hovering point; v j : The flight speed of the UAV between the j-th hovering point and the (j+1)-th hovering point; v max The maximum flight speed of a UAV; Step S8: If ,use Update the initial hover trajectory; if Probability can be used Will Replace with Specifically, a random value It is generated by the standard normal distribution, when Time settings Otherwise discard .

Images