Mobile QR Code QR CODE

  1. (China Mobile Group Device Co., Ltd., Beijing, 100053, China )



Terminal equipment, Site level AI, End side computing network, Calculate energy consumption

1. Introduction

A computing network is a new architecture to adapt to the development trend of computing network convergence. The network connects computing resources scattered in different places. In addition, it enables different applications to schedule computing resources in different places in real-time according to the needs through unified coordination scheduling to optimize network and computing power and improve user experience [1]. Njitucke Z T et al. analyzed a new three-neuron small network to solve the problem that analog integrated circuits can simulate brain functions. A heterogeneous coupling computational force network was proposed [2]. Chen L et al. used reinforcement learning and migration learning to optimize the performance of the network system and reduce energy loss [3], aiming at the computational energy efficiency of the mobile edge computing network system of the future UAV. Chen Y et al. conducted a detailed analysis of mobile edge computing networks and proposed a hybrid energy supply model by integrating energy into the Internet of Things terminal devices [4]. Accordingly, research has focused on analyzing the field-level AI reasoning technology of terminal equipment of end-side computing networks. In particular, the present study assesses the cooperation of multiple UAVs, aiming to reduce the energy loss of the end-side computing network through this method and improve real-time computing migration and resource allocation.

2. Related Work

The end-to-end computing network focuses on the existing cellular network, wireless network, and other network architectures to solve the problems of decentralized, inconsistent distributed computing and private storage, which provides customers with more intelligent and personalized application services [5]. Therefore, several studies have conducted in-depth research on developing mobile communication networks. Mustafa E et al. investigated the high latency problem of mobile edge computing in the end-side computing network using the integrated services of mobile edge computing to transmit the wireless power to the terminal equipment. Low latency to reduce the delay limit was also examined [6]. Do Duy T et al. used digital twin technology in industrial automation to assist the mobile edge computing model, minimizing the end-to-end delay, reducing the end-to-end delay in end-to-end computing networks [7]. Jiang M et al. proposed a multi-intersection vehicle cooperation control method based on the end-side computing network to improve road safety and strengthen traffic management, optimize vehicle cooperation, improving the driving efficiency of multi-intersection [8]. Khan LU et al. proposed a new framework using the end-side computing network to improve the reliability of the Internet of Things application of the sixth-generation wireless system, providing help for implementing the sixth-generation wireless system supporting digital twins [9]. Hasan M K et al. proposed a new method of blockchain based on the end-side computing network to address the security threat of the Internet of Things in the fifth-generation wireless system, strengthening the security of the medical Internet of Things and improving the quality of service [10].

Wang L et al. proposed integrating supply chain financing using edge computing in the end-to-end computing network to promote the rapid development of Internet finance. The data sharing and data processing capabilities between commercial banks and enterprises were enhanced [11]. Alqahtani F et al. introduced machine learning based on mobile edge computing to reduce the response time of the distribution network. The service rate used was improved and the unloading technology was strengthened [12]. Rajavel R et al. proposed an object tracking and behavior recognition system based on the end-to-end computing network to solve the problem of managing distributed intelligent monitoring. The system reduced the network bandwidth and response time and promoted the development of intelligent medicine [13]. Talusan J et al. analyzed the edge computing of roadside units in detail based on the end-side computing network for the problem of urban data-intensive traffic networks to improve the response speed and accuracy of the network [14]. Morimoto M et al. applied edge computing networks to conduct a detailed analysis of the inner part of the neural network and solve the problems related to machine learning in fluid dynamics, providing help for fully solving fluid problems [15].

The current end-to-end computing network is a relatively cutting-edge technology, but there has been less research on the key technologies. Therefore, the research on the field-level AI technology of its terminal equipment has certain innovations. This study examines the computational energy efficiency of multi-UAV-assisted edge computing networks, making this technology can help UAV flight obstacle avoidance.

3. Analysis of Key Technologies of End to side Computing Network based on Field Level AI Reasoning of Terminal Equipment

3.1 UAV-assisted MEC System Model and Problem Analysis

A technical algorithm of computing migration and resource allocation for the end-to-side computing network is proposed, and a multi-UAV cooperative edge computing (MEC) system model is constructed to reduce the energy consumption of the end-to-side computing network of UAV (Unmanned Aerial Vehicle) terminal equipment under the field level Artificial Intelligence (AI) reasoning. As a communication terminal, the high-speed mobility of UAVs makes the wireless communication between UAVs and ground terminal equipment unpredictable. In the process of UAV route planning, UAVs can assist in effective energy distribution in MEC networks owing to the rapid development of field-level AI [16]. In actual work, the MEC network is usually composed of multiple UAVs and users. Fig. 1 presents a schematic diagram.

From Fig. 1, the MEC network assisted by multiple UAVs includes multiple UAV communication terminals and ground terminal equipment. Among them, the UAV communication terminal has many edge ECS and computing resources. It has a flight mission from the initial point to the target point, while the ground terminal equipment monitors the surrounding environment. Based on limited computing and storage resources, the sensor data collected by the terminal can be transmitted to the UAV through the wireless uplink. Therefore, according to the European theorem, the distance between the UAV communication terminal and the ground terminal equipment in the time slot is expressed in Eq. (1).

(1)
$ D_{i,j}\left(n\right)=\sqrt{h^{2}+\left\| \omega _{j}\left(n\right)-U_{i}\right\| ^{2}} $

In Eq. (1), $D$ represents the distance between the UAV communication terminal and ground terminal equipment; $n$ represents time slot; $h$ represents UAV flight altitude; $\omega $ represents the coordinates projected on the horizontal plane during UAV flight; $U$ represents a horizontal coordinate of ground terminal equipment; $i$ represents ground terminal equipment; $j$ indicates UAV communication terminal. Generally, the wireless transmission line between the UAV communication terminal and ground terminal equipment can be considered a typical line of sight (LoS) transmission [17]. Therefore, the gain expression of the wireless signal channel is expressed as Eq. (2).

(2)
$ h_{i,j}\left(n\right)=\beta _{0}\left(D_{i}\left(n\right)\right)^{-2} $

In Eq. (2), $\beta _{0}$ represents the power gain of the wireless signal channel at a distance of one meter. For conventional ground terminal equipment, the calculation task when using field-level AI estimation can be performed at the local and UAV communication ends. When computing at the local end, the number of computing tasks that can be performed by the terminal device and the energy consumption generated by executing tasks can be calculated using Eq. (3).

(3)
$ \left\{\begin{array}{l} R_{i}^{l}\left(n\right)=\frac{\zeta _{i}\left(n\right)\tau }{\rho }\\ E_{i}^{l}\left(n\right)=\kappa \zeta _{i}^{3}\left(n\right)\tau \end{array}\right. $

In Eq. (3), $R_{i}^{l}\left(n\right)$ represents the number of executable computing tasks of the terminal device; $\zeta _{i}\left(n\right)$ is the calculation rate; $\rho $ is the number of CPU calculation cycles; $E_{i}^{l}\left(n\right)$ is the energy loss; $\kappa $ is the energy coefficient; $\tau $ indicates the length of the sub timeslot. When the terminal device migrates the computing task to the edge cloud for execution, it needs to transfer the input data of the computing task to the UAV. The computing task amount and energy consumption are expressed as Eq. (4).

(4)
$ \left\{\begin{array}{l} R_{i,j}^{c}\left(n\right)=\gamma _{i,j}\left(n\right)W\tau \log _{2}\left(1+\frac{\lambda _{i,j}\left(n\right)h_{i,j}\left(n\right)}{\gamma _{i,j}\left(n\right)WN_{0}}\right)\\ E_{i,j}^{c}\left(n\right)=\lambda _{i,j}\left(n\right)\tau \end{array}\right. $

In Eq. (4), $R_{i,j}^{c}\left(n\right)$ is the task amount calculated at the UAV end; $\gamma $ is the proportion of spectrum resources; $W$ is the total spectrum bandwidth; $\lambda $ is transmission power; $N_{0}$ is the density of background noise power spectrum; $E_{i,j}^{c}\left(n\right)$ indicates the energy consumption of the UAV terminal for task execution. Therefore, the expression of total computing tasks and total energy consumption of the MEC network assisted by multiple UAVs is expressed as Eq. (5).

(5)
$ \left\{\begin{array}{l} R\left(s\left(n\right),\zeta \left(n\right),\lambda \left(n\right),\gamma \left(n\right),\omega \left(n\right)\right)=\sum _{n=1}^{N}\sum _{i=1}^{M}\left\{\left.R_{i}^{l}\left(n\right)+\sum _{j=1}^{F}s_{i,j}\left(n\right)R_{i,j}^{c}\left(n\right)\right\}\right.\\ E\left(s\left(n\right),\zeta \left(n\right),\lambda \left(n\right)\right)=\sum _{n=1}^{N}\sum _{i=1}^{M}\left\{\left.\kappa f_{i}^{3}\left(n\right)\tau +\sum _{j=1}^{F}s_{i,j}\left(n\right)\lambda _{i,j}\left(n\right)\tau \right\}+P_{c}MT\right. \end{array}\right. $

In Eq. (5), $N$ is the number of sub timeslots; $M$ is the number of ground terminal equipment; $F$ is the number of UAVs; $P_{c}$ is the static power consumption constant; $T$ is the total time for UAV to execute tasks. Therefore, the energy efficiency expression of the system is expressed as Eq. (6).

(6)
$ \eta _{CE}=\frac{R\left(s\left(n\right),f\left(n\right),\lambda \left(n\right),\gamma \left(n\right),\omega \left(n\right)\right)}{E\left(s\left(n\right),f\left(n\right),\lambda \left(n\right)\right)} $

In Eq. (6),$\eta _{CE}$ represents the calculated energy efficiency of the system. Aiming to maximize system energy efficiency, this study examines the model construction based on the partial task migration model, combined with the rate-constrained transmission power value calculated locally and the planning of spectrum resource allocation UAV aircraft. Fig. 2 presents the modeling of the computing energy efficiency maximization in the MEC network assisted by multiple UAVs.

This study evaluates question $Q$ when building the problem model (Fig. 2), wherein $\varphi _{i,j}\left(n\right)$ represents an integer variable; $l$ is a UAV communication terminal that is not equal to $j$; $C1$ is the actual speed efficiency of local calculation of terminal equipment; $C2$ is the upper limit of the maximum transmission power of the terminal equipment; $C3$ is that each terminal device can connect at most one UAV in the same time slot for task migration, $C4$ is the spectrum resource constraints for all terminal devices associated with the same UAV; $C5$ is the total amount of computing tasks completed by each terminal device in the total time of UAV task execution cannot be less than the minimum amount of computing tasks; $C6$ means that the flight speed of the UAV cannot exceed the maximum speed upper limit; $C7$ is the shortest safe distance between multiple UAVs; $C8$ is the shortest safe distance between UAV and obstacle; $\omega _{I,j}$ in $C9$ represents the initial point of a UAV; $\omega _{F,j}$ represents the target point of UAV. In addition, the problem becomes a mixed integer programming problem that is difficult to solve because of the integer of $\varphi _{i,j}\left(n\right)$.

Fig. 1. MEC network assisted by multiple UAVs.
../../Resources/ieie/IEIESPC.2023.12.6.472/fig1.png
Fig. 2. Modeling of the computational energy efficiency maximization in MEC networks assisted by multiple unmanned aerial vehicles.
../../Resources/ieie/IEIESPC.2023.12.6.472/fig2.png

3.2 Research on Computational Energy Efficiency Maximization Algorithm based on Two-level Iteration

The research adopts the Dinkelbach method to eliminate fractional structure, which makes the problem easy to solve because the problem is nonlinear as a mixed integer programming problem [18]. The research decomposes the problem into multiple sub-problems considering that the actual UAV communication terminal and ground terminal equipment makes the problem associate with the users, resource allocation of communication computing, and UAV route planning, under the reasoning of field-level AI. Based on the two-level iteration, a computational energy efficiency maximization algorithm is proposed to solve each independent sub-problem. Dinkelbach is used to transform the problem into a parameter-planning problem. At this time, the reconstruction expression of the system computing energy efficiency can be expressed as Eq. (7).

(7)
$ \eta =\frac{R\left(X\left(n\right)\right)}{E\left(X\left(n\right)\right)} $

In Eq. (7), $\eta $ is the calculated energy efficiency after reconstruction; $X\left(n\right)$ is the set of variables such as local calculation rate and transmission power. Fig. 3 shows the flow of the computational energy efficiency maximization (ICEM) algorithm using the Dinkelbach method.

In Fig. 3, $k$ represents the number of iterations, and $\sigma $ is the threshold parameter. The Dinkelbach method establishes a set of increasing energy efficiency coefficients $\eta $ and tries to close the value of the $\zeta \left(\eta \right)$ function to 0 to obtain the best energy efficiency $\eta ^{*}$. Under the given conditions of $\eta $, the optimal solution of a group of $X\left(n\right)$ can be obtained using question $Q1$. According to Eq. (7), the value of $\eta $ can be updated according to the available solution $X\left(n\right)$. The best energy efficiency value $\eta ^{*}$ can be obtained after one iteration, and $\zeta \left(\eta ^{*}\right)=0$. Although the converted problem $Q1$ makes the objective function solution easier, the coupling relationship between UAV and other variables in the actual route planning makes the problem non-convex. Therefore, the solution to a non-convex problem $Q1$ is divided into user association, resource allocation, UAV route planning, and other sub-problems. In the user association optimization, $Q1$ can be converted to a user association problem if the local calculation rate, the transmission power of the terminal device, and other conditions are known. The expression is written as Eq. (8).

In Eq. (8), in question $SQ1$, the continuous variable is completely separated from the original question $Q1$, leaving only the integer variable of 0-1. At this point, the problem can be identified as a standardized linear programming problem about integers, and current optimization algorithms can solve problem $Q1$. In the process of a reasonable allocation of resources, any designated user is associated with the route planning of UAVs. The resource optimization configuration problem in question $Q1$ is expressed as Eq. (9).

(8)
$ \begin{equation*} \begin{array}{l} SQ1\colon \max _{s\left(n\right)}\sum _{n=1}^{N}\sum _{i=1}^{M}\sum _{j=1}^{F}\varphi _{i,j}\left(n\right)\gamma _{i,j}\left(n\right)W\tau \log _{2}\left(1+\frac{\lambda _{i,j}\left(n\right)h_{i,j}\left(n\right)}{\gamma _{i,j}\left(n\right)WN_{0}}\right)\\ \begin{array}{ll} & -\eta \sum _{n=1}^{N}\sum _{i=1}^{M}\sum _{j=1}^{F}\varphi _{i,j}\left(n\right)\lambda _{i,j}\left(n\right)\tau \end{array} \end{array} \end{equation*} $
(9)
$ \begin{equation*} \begin{array}{l} SQ2\colon \max _{X_{1}\left(n\right)}\sum _{n=1}^{N}\sum _{i=1}^{M}\left\{\left.\frac{\zeta _{i}\left(n\right)\tau }{\rho }+\sum _{j=1}^{F}\delta _{i,j}\left(n\right)\alpha _{i,j}\left(n\right)W\tau \log _{2}\left(1+\frac{\lambda _{i,j}\left(n\right)h_{i,j}\left(n\right)}{\gamma _{i,j}\left(n\right)WN_{0}}\right)\right\}\right.\\ \begin{array}{ll} & -\eta \sum _{n=1}^{N}\sum _{i=1}^{M}\left\{\left.\kappa {\zeta _{i}}^{3}\left(n\right)\tau +\sum _{j=1}^{F}\varphi _{i,j}\left(n\right)\lambda _{i,j}\left(n\right)\tau \right\}\right. \end{array} \end{array} \end{equation*} $

In Eq. (9), the objective function in question $SQ2$ is convex, and the conditions of each constraint are convex sets. Therefore, the standard convex optimization algorithm can be selected to solve problem $SQ2$. Finally, in the UAV route planning problem, the optimized problem expression is expressed as Eq. (10).

(10)
$ SQ3\colon \max _{q\left(n\right)}\sum _{n=1}^{N}\sum _{i=1}^{M}\sum _{j=1}^{F}\varphi _{i,j}\left(n\right)\gamma _{i,j}\left(n\right)W\tau \log _{2}\left(1+\frac{\lambda _{i,j}\left(n\right)\beta _{0}}{\gamma _{i,j}\left(n\right)WN_{0}\left(h^{2}+\left\| \omega _{j}\left(n\right)-U_{i}\right\| ^{2}\right)}\right) $

In Eq. (10), problem $SQ3$ is still a non-convex problem because of the association of users, reasonable allocation of resources, and decoupling of route delineation. For the UAV flight path, the left equation of the cost function and constraint condition $C5$ is not convex or concave. Therefore, $SQ3$ is a non-convex problem, which is difficult to solve. The problem of $SQ3$ is solved by selecting the continuous convex approximation method to transform its non-convex functions. The uplink achievable spectrum efficiency of the ground terminal equipment is expressed as Eq. (11).

(11)
$ R_{i,j}\left(n\right)=\log _{2}\left(1+\frac{\lambda _{i,j}\left(n\right)g_{0}}{\left(h^{2}+\left\| \omega _{j}\left(n\right)-U_{i}\right\| ^{2}\right)}\right) $

In Eq. (11), $g_{0}$ represents the signal-to-noise ratio. $R_{i,j}\left(n\right)$ is a convex function associated with $\left\| \omega _{j}\left(n\right)-U_{i}\right\| ^{2}.$ Therefore, any point of $R_{i,j}\left(n\right)$ on $\left\| q_{j}\left(n\right)-u_{i}\right\| ^{2}$can be expanded by the corresponding first-order Taylor expansion to obtain the global lower bound. The low bound expression is expressed as Eq. (12).

(12)
$ \begin{equation*} \hat{R}_{i,j}\left(n\right)=R_{i,j}^{k}\left(n\right)+\nabla {R^{k}}_{i,j}\left(n\right)\left(\left\| \omega _{j}\left(n\right)-u_{i}\right\| -\left\| \omega _{j}^{k}\left(n\right)-U_{i}\right\| \right) \end{equation*} $

In Eq. (12), $\hat{R}_{i,j}\left(n\right)$ is the global lower bound of $R_{i,j}\left(n\right)$; $\omega _{j}^{k}\left(n\right)$ is the UAV flight path obtained after the $k$ iteration; $R_{i,j}^{k}\left(n\right)$ is the spectral efficiency obtained after the $k$ iteration; $\nabla {R^{k}}_{i,j}\left(n\right)$ is the first order partial derivative associated with $\omega _{j}^{k}\left(n\right)-u_{i}\,.$ $R_{i,j}^{k}\left(n\right)$ and $\nabla {R^{k}}_{i,j}\left(n\right)$ are expressed as Eq. (13).

(13)
$ \left\{\begin{array}{l} R_{i,j}^{k}\left(n\right)=\log _{2}\left(1+\frac{\lambda _{i,j}\left(n\right)g_{0}}{\left(h^{2}+\left\| \omega _{j}^{k}\left(n\right)-U_{i}\right\| ^{2}\right)}\right)\\ \nabla {R^{k}}_{i,j}\left(n\right)=-1+\frac{\lambda _{i,j}\left(n\right)g_{0}\log _{2}e}{\left(h^{2}+\left\| \omega _{j}^{k}\left(n\right)-U_{i}\right\| ^{2}\right)\left(h^{2}+\left\| \omega _{j}^{k}\left(n\right)-U_{i}\right\| ^{2}+\lambda _{i,j}\left(n\right)g_{0}\right)} \end{array}\right. $

In Eq. (13), the constraint conditions $C7$ and $C8$, the continuous convex approximation method is selected to relax it. The inequality obtained after the first order Taylor expansion of $\omega _{j}^{k}\left(n\right)$ and $\omega _{l}^{k}\left(n\right)$ can be written as Eq. (14).

(14)
$ \begin{equation*} \left\{\begin{array}{l} \left\| \omega _{j}^{k}\left(n\right)-\omega _{l}^{k}\left(n\right)\right\| ^{2}\geq -\left\| \omega _{j}^{k}\left(n\right)-\omega _{l}^{k}\left(n\right)\right\| ^{2}+2\left(\omega _{j}^{k}\left(n\right)-\omega _{l}^{k}\left(n\right)\right)^{T}\left(\omega _{j}\left(n\right)-\omega _{l}\left(n\right)\right)\\ \left\| \omega _{j}\left(n\right)-b_{m}\right\| ^{2}\geq \left\| \omega _{j}^{k}\left(n\right)-b_{m}\right\| ^{2}+2\left(\omega _{j}^{k}\left(n\right)-b_{m}\right)^{T}\left(\omega _{j}\left(n\right)-\omega _{j}^{k}\left(n\right)\right) \end{array}\right. \end{equation*} $

In Eq. (14), $b_{m}$ represents the offset, and $T$ is the transposition. According to Eqs. (13) and (14), problem $SQ3$ can be transformed into an approximate convex problem, which is expressed as Eq. (15).

(15)
$ SQ3^{'}\colon \max _{q\left(n\right)}\sum _{n=1}^{N}\sum _{i=1}^{M}\sum _{j=1}^{F}\varphi _{i,j}\left(n\right)\alpha _{i,j}\left(n\right)W\tau \hat{R}_{i,j}\left(n\right) $

In Eq. (15), it is an approximate convex problem because the cost function and restraint condition are both approximate convex functions. Hence, a standard convex optimization tool can be selected to solve it. Combining Dinkelbach with the optimization of the three subproblems, the two-layer iterative algorithm is used to maximize the computational efficiency, which is also called the iterative optimization algorithm of the two-layer loop structure. Among them, its external circular structure uses the relevant Dinkelbach methods to plan the parameters for calculating the efficiency, and the inner loop structure is mainly a joint optimization problem of the three subproblems. Theoretically, the internal loop structure of the algorithm achieves local optimization. The actual requirements for the convergence speed of the algorithm is greatly reduced under the actual terminal equipment field-level AI reasoning.

In terms of computational complexity, when the actual end-to-side computing force network is expanded, the computational complexity of the ICEM algorithm increases significantly, which poses a huge challenge to the computing power of field-level AI reasoning of terminal equipment. Hence, the threshold parameters can be adjusted to alleviate the problem in practical situations, i.e., increasing the threshold of the algorithm proposed in the study can reduce the number of iterations and computational energy efficiency. Therefore, under the field-level AI reasoning oriented to UAV communication terminals and ground terminal equipment, an average point can be selected in terms of the algorithm complexity and accuracy.

Fig. 3. Flow chart of the energy efficiency maximization algorithm using the Dinkelbach method.
../../Resources/ieie/IEIESPC.2023.12.6.472/fig3.png

4. Simulation Performance Analysis on Key Technologies of End-side Computing Network of UAV Communication Terminal

The simulation analysis is conducted to verify the effectiveness of the ICEM algorithm in optimizing the end-side computing network of UAV communication terminals and ground terminal equipment. Each UAV has its flight mission in the MEC network assisted by multiple UAVs. Table 1 lists the simulation parameters of the UAV communication terminal and ground terminal equipment.

From Table 1, five ground terminal equipment are set up and distributed randomly in a 1km ${\times}$ 1km two-dimensional area. The maximum transmit power is 200 mW. The power gain of the information channel is -50 dB, and the density of calculation processing is 1000 cycles/bit. The UAV flight height is fixed at 100m from the ground, and the speed is limited to 50 m/s, while the shortest safety distance between multiple UAVs is 60 m. Therefore, the convergence of the algorithm is analyzed. Among them, four benchmark algorithms are introduced to compare with the ICEM algorithm, including the Fixed Track flight (FixT), Allof, SNR_based that are based on the signal-to-noise ratio (SNR), and Computation Bits Maximization (CBM). In addition, the CBM scheme does not involve energy consumption but only considers the maximum computing task, i.e., when the energy efficiency of the proposed system model is zero. Fig. 4 presents the results.

In Fig. 4, the number of iterations on the horizontal axis only refers to the number of iterations after the energy efficiency solution is updated. The main reason is to consider the actual situation that the calculated energy efficiency is updated in the external loop structure of the ICEM algorithm, whereas the internal loop structure does not change. Fig. 4(b) introduces the cumulative distribution function (CDF) of the iteration number. In the first 15 iterations, the calculation energy efficiency of the other three algorithms, except the FixT scheme, increases sharply as the number of iterations is increased and tends to be stable when it increases from 12 (${\times}$ 10$^{8}$) to 15 (${\times}$ 10$^{8}$) bits/Joule. FixT tends to be stable after the 10th iteration, but its computational efficiency is also the least, with a maximum of 8 (${\times}$ 10$^{8}$) bits/Joule. The system computing efficiency obtained by the ICEM algorithm is approximately 15 (${\times}$ 10$^{8}$) bits/Joule higher than that of the other three schemes.

When the CDF of the number of iterations is 0.9, the threshold can be increased 10-fold to reduce 50 iterations, while the number of UAVs and mobile terminal devices can be doubled, and the number of iterations exceed 20. In summary, the ICEM algorithm can obtain more computational efficiency when the convergence speed is not low. Moreover, for different actual situations, the end-side computing force network system under the field-level AI reasoning of the terminal equipment can choose to adjust the algorithm to terminate the threshold to achieve the balance between the convergence speed of the algorithm and the accuracy of the objective function. Therefore, the impact of the total time of different tasks on UAV route planning is analyzed, as shown in Fig. 5.

From Fig. 5, the UAV flight path almost overlaps with the total time of different missions, but the UAV is not always flying. Therefore, they will remain static with the service terminal equipment for a certain time under the reasoning of the field-level AI when the total time is 60s and 90s to complete the calculation task of the actual end-side computing network as much as possible. In addition, when the total task execution time is 30 s, the UAV will fly at the maximum speed in most time slots. On the other hand, the UAV has only a 50% probability of flying at the maximum speed when the total time is 60s and 90s. Meanwhile, when the total time is 60s and 90s, the probability that the speed is zero is far greater than the probability at 30s. Therefore, after using the ICEM algorithm, the terminal device side computing network resources under the field level AI reasoning are effectively scheduled, greatly reducing the energy consumption to complete the task. In addition, this study also analyzes the impact of the minimum amount of computing tasks completed by the terminal equipment on the computing energy efficiency of the end-to-end computing network system, as shown in Fig. 6.

The increase in the minimum computing tasks promotes the ICEM algorithm, FixT, and SNR (Fig. 6). According to the based scheme, the computing energy efficiency drops sharply when the minimum computing workload is less than 20 Mbits, and then slowly declines. If all computing jobs are placed on the terminal for local processing, the maximum computing workload of the terminal device can reach 15 Mbits when the total execution time is 30 seconds. That is, when the minimum task quantity is greater than 15 megabits, it is necessary to transfer some computing tasks to the UAV owing to the limited computing power of the terminal equipment. On the other hand, with increasing computing migration, the limited spectrum resources will be incapable of meeting the large-scale computing task transfer, reducing the computing efficiency. In summary, the ICEM algorithm has the highest computational efficiency among the four schemes. Hence, the research deeply observes the changes in the computing tasks and energy consumption of the end-to-end computing network system regarding the minimum tasks under different schemes, as shown in Fig. 7.

The number of computing tasks obtained by the based scheme is the same and close to the minimum amount to obtain the optimal computing energy efficiency (Fig. 7). In addition, the energy consumption of the ICEM algorithm is slightly lower than SNR_ Based scheme when the minimum task size is 5Mbits and 10Mbits. The calculation workload and energy consumption obtained by all in the scheme shows approximate linear growth as the minimum task amount increases. The total calculation and energy loss of the FixT scheme are higher than those of the other three algorithms. The CBM archives can obtain the maximum number of computing tasks and energy consumption. Therefore, the ICEM algorithm has the minimum amount of computing tasks and energy consumption. Fig. 8 shows the change in the terminal equipment and UAV number in calculating the energy efficiency under different schemes.

The calculated energy efficiency is greater than that when the number of UAVs is 2 when the number of UAVs is 4, as shown in Fig. 8. The calculated amount of the system also increase as the number of UAVs increases so that the terminal equipment can complete more calculations better. SNR at F=2_ Based on Allof and the proposed algorithm, the energy efficiency decreases as the number of terminals increases. When F=4, the energy efficiency of the three methods increases as the number of terminals increases. When M>8, the energy efficiency decreases rapidly. Overall, ICEM computing is always the most energy efficient. Finally, the impact of obstacles on the UAV route trajectory is studied; Fig. 9 presents the results.

The UAV will fly as close to the terminal equipment as possible and avoid all obstacles in the area to effectively assist the ground terminal to complete the calculation migration and task execution (Fig. 9). In addition, in the case of obstacles, UAVs are more likely to fly at maximum speed than those without obstacles. In the first and final 10 s of the mission, the flight speeds of the two UAVs are the same, both dropping from 50 m/s to 0 m/s. The main reason is that the UAV will save more time slots to complete the flight mission under the reasoning of the field-level AI. Overall, the ICEM algorithm enables UAVs to avoid obstacles perfectly in the scene with obstacles without exceeding the maximum speed.

Table 1. Corresponding simulation parameters of the UAV communication terminal and ground terminal equipment.

Number of terminal devices

2D area

Total spectrum resources (MHz)

Communication bandwidth (MHz)

5

1km×1km

20

20

Maximum transmitting power of terminal equipment (mW)

Channel power gain (dB)

Power spectral density of background noise (dBm/Hz)

Calculate processing density (cycles/bit)

200

-50

-150

1000

Energy coefficient

Maximum local computing rate (GHz)

Maximum UAV flight speed (m/s)

UAV flight altitude

10-28

0.5

50

100m

Minimum safe distance between unmanned aerial vehicles (m)

-

60

-

Fig. 4. Comparative analysis of algorithm convergence.
../../Resources/ieie/IEIESPC.2023.12.6.472/fig4.png
Fig. 5. Analysis of UAV path planning under different mission times.
../../Resources/ieie/IEIESPC.2023.12.6.472/fig5.png
Fig. 6. Variation diagram of the different minimum computing tasks.
../../Resources/ieie/IEIESPC.2023.12.6.472/fig6.png
Fig. 7. Statistical graph of system energy consumption with respect to the minimum calculation task under different schemes.
../../Resources/ieie/IEIESPC.2023.12.6.472/fig7.png
Fig. 8. Change in the number of terminal equipment and UAVs for calculating energy efficiency under different schemes.
../../Resources/ieie/IEIESPC.2023.12.6.472/fig8.png
Fig. 9. Analysis of the Influence of Obstacles on UAV Flight Path.
../../Resources/ieie/IEIESPC.2023.12.6.472/fig9.png

5. Conclusion

The system computing energy efficiency optimization effect under the cooperation of multiple UAVs is studied and analyzed, aiming at the field level AI reasoning oriented to terminal devices in the end-to-side computing force network. A two-level iterative computing maximization algorithm is introduced and verified by simulations. The system computing efficiency obtained by the ICEM algorithm is higher than the other three schemes, and the maximum is approximately 15 (${\times}$ 10$^{8}$) bits/Joule. When the CDF value of the iterations is 0.9, a 10-fold increase in the threshold value can decrease the number of iterations by more than 50 while doubling the number of UAVs and mobile terminal devices will increase the number of iterations by more than 20. When the total task execution time is 30s, the UAV can fly at the maximum speed in most time slots, and the maximum computing workload of the terminal device can reach 15 Mbits. When the total time is 60 s and 90 s, the UAV has only a 50% probability of flying at the maximum speed. The number of UAVs is set to 2 and 4, and the ICEM calculation energy efficiency is always the highest, at 7-12 bits/Joule and 12-14 bits/Joule, respectively. In addition, the flight speeds of the two UAVs are the same, within 10s at the beginning and 10s at the end of the mission. In summary, the ICEM algorithm can obtain the optimal value with less convergence time and performs better than other benchmark algorithms when meeting the user's quality of service requirements. On the other hand, the research only considers the high-speed mobility of UAVs to provide real-time mobile computing services for terminal devices but does not consider the mobility of terminal devices, which requires follow-up research.

REFERENCES

1 
M. Żarski, B. Wójcik, K. Książek, and J. A. Miszczak, “Finicky transfer learning—A method of pruning convolutional neural networks for cracks classification on edge devices,” Computer‐Aided Civil and Infrastructure Engineering, Vol. 37, No. 4, pp. 500-515. 2022.DOI
2 
Z.T. Njitacke, et al., “Hamiltonian energy computation and complex behavior of a small heterogeneous network of three neurons: circuit implementation,” Nonlinear Dynamics, Vol. 107, No. 3, pp. 2867-2886. 2022.DOI
3 
L. Chen, et al., “Intelligent ubiquitous computing for future UAV-enabled MEC network systems,” Cluster Computing, Vol. 25, No. 4, pp. 2417-2427. 2022.DOI
4 
Y. Chen, et al., “Dynamic task offloading for mobile edge computing with hybrid energy supply,” Tsinghua Science and Technology, Vol. 28, No. 3, pp. 421-432. 2022.DOI
5 
M, Kamel, et al., “Distributed Address Table (DAT): A decentralized model for end-to-end communication in IoT,” Peer-to-Peer Networking and Applications, Vol. 15, No. 1, pp. 178-193. 2022.DOI
6 
E. Mustafa, et al., “Joint wireless power transfer and task offloading in mobile edge computing: a survey,” Cluster Computing, Vol. 25, No. 4, pp. 2429-2448. 2022.DOI
7 
T. Do-Duy, et al., “Digital Twin-Aided Intelligent Offloading with Edge Selection in Mobile Edge Computing,” IEEE Wireless Communications Letters, Vol. 11, No. 4, pp. 806-810. 2022.DOI
8 
M. Jiang, et al., “A multi-intersection vehicular cooperative control based on end-edge-cloud computing,” IEEE Transactions on Vehicular Technology, Vol. 71, No. 3, pp. 2459-2471. 2022.DOI
9 
L. U. Khan, et al., “Digital-twin-enabled 6G: Vision, architectural trends, and future directions. IEEE Communications Magazine, Vol. 60, No. 1, pp. 74-80. 2022.DOI
10 
M. K. Hasan, et al., “A review on security threats, vulnerabilities, and counter measures of 5G enabled Internet-of-Medical-Things,” IET Communications, Vol. 16, No. 5, pp. 421-432. 2022.DOI
11 
L. Wang, Y. Wang, “Supply chain financial service management system based on block chain IoT data sharing and edge computing,” Alexandria Engineering Journal, Vol. 61, No. 1, pp. 147-158. 2022.DOI
12 
F. Alqahtani, M. Al-Maitah, and O. Elshakankiry, “A proactive caching and offloading technique using machine learning for mobile edge computing users,” Computer Communications, Vol. 181, pp. 224-235. 2022.DOI
13 
R. Rajavel, et al., “IoT-based smart healthcare video surveillance system using edge computing,” Journal of Ambient Intelligence and Humanized Computing, Vol. 13, No. 6, pp. 3195-3207. 2022.DOI
14 
J. Talusan, et al., “Route Planning Through Distributed Computing by Road Side Units,” IEEE Access, Vol. 8, pp. 176134-176148. 2020.URL
15 
M. Morimoto, et al., Generalization techniques of neural networks for fluid flow estimation,” Neural Computing and Applications, Vol. 34, No. 5, pp. 3647-3669. 2022.DOI
16 
Z. Liu, et al., “Robust Edge Computing in UAV Systems via Scalable Computing and Cooperative Computing,” IEEE Wireless Communications, Vol. 28, No. 5, pp. 36-42. 2021.DOI
17 
M.A. Kumar, and S. Kanthalakshmi, “H∞Control law for line of sight stabilization in two-axis gimbal system,” Journal of Vibration and Control, Vol. 28, No. 1-2, pp. 182-191. 2022.DOI
18 
S. Rezvani, et al., “Optimal Power Allocation in Downlink Multicarrier NOMA Systems: Theory and Fast Algorithms,” IEEE Journal on Selected Areas in Communications, Vol. 40, No. 4, pp. 1162-1189. 2022.DOI

Author

Mao Ni
../../Resources/ieie/IEIESPC.2023.12.6.472/au1.png

Mao Ni obtained a Master's degree in Computer Technology from Beijing University of Posts and Tele-communications in 2012. He is currently responsible for product development at China Mobile Communications Group Terminal Co., Ltd. He has led multiple major research and development projects for China Mobile and participated in national major projects related to Computing Force Network. Wang Lei has authored more than 10 patents, and his research interests include Computing Force Network, 5G terminal technology, operating systems, and artificial intelligence.

Ting Zhou
../../Resources/ieie/IEIESPC.2023.12.6.472/au2.png

Ting Zhou obtained a Master's degree in Information and Communication Engineering from Beijing University of Posts and Telecommunications in 2014. She currently works at China Mobile Communications Group Terminal Co., Ltd. Engaged in long-term research on mobile terminal product planning, Terminal Communication Technology, and Computing Force Network technology. She led multiple major projects of China Mobile in Computing Force Network, participated in national major projects related to Terminal-side Computing Force Network. She has published 5 papers, and wrote over 10 patents. Her areas of interest include Computing Force Network, 5G terminal technology, etc.

Hengjiang Wang
../../Resources/ieie/IEIESPC.2023.12.6.472/au3.png

Hengjiang Wang obtained his Master's degree in Computational Engineering from Tokyo Institute of Technology in 2005. He is currently the Deputy General Manager and Senior Engineer at China Mobile Device Company, responsible for technological innovation, technical research, terminal software/hardware R&D in the terminal field. Meanwhile, he serves as a member of MIIT Economic Expert Committee, a member of CAICT Technical Committee, and Vice Chairman of TAF. He has presided over terminal-related tasks in many major projects under the Ministry of Science and Technology, MIIT, and NDRC. He has published 5 papers, authorized over 30 patents, led the development of 12 industry standards and 35 enterprise standards, as well as 44 software copyrights. His areas of interest include end-to-end communication protocols for network terminals, multimedia processing, terminal software/hardware development, distributed operating systems, cloud terminals, Computing Force Network, artificial intelligence, etc

Fang Cui
../../Resources/ieie/IEIESPC.2023.12.6.472/au4.png

Fang Cui obtained her Master's degree in Communication and Information System from the Institute of Telecommunication Science and Technology, MIIT in 2004. She currently serves as the General Manager and Senior Engineer in the Technology Department of China Mobile Device Company. She has been engaged in the work of mobile terminal chips, terminal technology standards, and terminal industry promotion. She has presided over and participated in many major projects related to the terminal field under MIIT and NDRC. She has published 5 papers, authorized over 10 patents, and led teams to develop the XR device industry standard. Her areas of interest include 5G terminals, eSIM, multimedia processing, cloud terminals, computing power networks, quantum computing, XR, etc.