Joint positioning and computing support method for internet of things devices

Abstract

The application discloses a joint positioning and computing support method of an Internet of Things (IoT) device, which comprises the following steps: collecting distance information and residual power information of the IoT device and anchor nodes, and obtaining a set of available anchor nodes of the IoT device; when the number of the available anchor nodes is less than 3, collecting distance information of the IoT device and other IoT devices; constructing a positioning model by a UAV according to the collected information; converting a positioning problem into a semi-definite programming problem by weighted least squares and semi-definite relaxation; and solving the coordinates of the IoT device; taking minimization of total system energy consumption and delay as an optimization target, and obtaining joint decisions of computing task offloading, UAV trajectory planning and UAV computing resource allocation by using a deep reinforcement learning method; obtaining an optimal downlink power allocation decision by using a successive convex approximation method, and the target is to maximize the minimum downlink throughput; and the UAV executes the optimal joint decision to support the offloading request of the IoT device, and executes the optimal downlink power allocation decision to improve the downlink throughput of the system.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of Internet of Things, and relates to a joint positioning and computing support method for Internet of Things devices. BACKGROUND

[0002] The rapid development of Internet of Things (IoT) technology has driven the exponential growth of the number of IoT devices, giving rise to a large number of computationally intensive and delay-sensitive applications. In some scenarios, IoT devices are widely deployed in challenging or sparsely ground-based station areas, such as forests, mountains, deserts, and underwater locations, and need to constantly move and perform some computationally intensive work, including disaster warning, long-pipeline infrastructure detection, underwater infrastructure detection, and military operations, etc. Most of these works need to track the monitoring target, which involves positioning functions. However, due to high costs, IoT devices cannot continuously and stably obtain real-time positions by carrying satellite positioning modules. The mobility, privacy, limited power of IoT devices, and the complexity of the environment (e.g., out of line-of-sight range) make the positioning of IoT devices extremely challenging. In order to reduce the cost of IoT devices, protect the privacy of IoT devices, and at the same time provide stable real-time computing support for IoT devices, a multi-device collaborative Time of Arrival (TOA) positioning scheme based on SemiDefinite Relaxation (SDR) needs to be constructed.

[0003] On the other hand, the battery power, computing resources, and storage resources of IoT devices are very limited, and cannot handle a large amount of computationally intensive and delay-sensitive tasks. Due to the limitations of hardware conditions and the environment, the battery life of IoT devices is limited and it is difficult to obtain real-time power supply. In addition, the computing and storage resources owned by IoT devices may not be sufficient to handle some large tasks, and devices with strong processing capabilities are needed to provide computing support. To solve the above problems, researchers have proposed a new computing paradigm, Mobile Edge Computing (MEC), to expand the computing capacity of IoT devices. In the considered challenging environment, the ground infrastructure is sparsely distributed and cannot provide stable and reliable computing support for IoT devices, and it is necessary to study the computing offloading based on flexibly deployed Unmanned Aerial Vehicle (UAV).

[0004] The Chinese patent "CN114745389A mobile edge computing system computing offloading method" designs a computing offloading scheme for a mobile edge computing system. In order to reduce the average information age in the Internet of Things system, first, the state update task at the Internet of Things device end is analyzed, the computing task's computing resource demand and space demand for the edge device are obtained, the time delay at the local and edge server end is calculated, and then the information age of each device and the information age of the entire system are obtained; according to the task demand and environmental conditions of each device itself, the optimal information age computing offloading strategy is selected; a computing offloading scheme based on game theory optimal information age is proposed, all devices compete for the opportunity to update the offloading strategy in each iteration, until the offloading strategy of all devices converges, the final computing offloading method is obtained, effectively reducing the average information age of the system, to meet the information freshness demand of different types of Internet of Things devices. However, in areas with challenging or sparse ground base stations, such as forests, mountains, and underwater locations, when facing a large number of computing-intensive and delay-sensitive tasks, this scheme tends to offload tasks to local processing, and cannot provide sufficient computing support for related IoT devices through sparse deployment of edge servers, which will increase the energy consumption of IoT devices with limited power; in addition, considering the mobility of IoT devices, this scheme cannot provide stable real-time computing support.

[0005] The Chinese patent "CN114124955A a computing offloading method based on multi-agent game" designs a computing offloading method based on two-stage multi-agent game. This patent establishes a network channel model for 5G hybrid dual-network mode, calculates the channel transmission rate of users according to Shannon's theorem, and calculates the time delay and energy consumption of users performing task offloading according to the channel transmission rate; according to Stackelberg game, an edge offloading model is established, the private network users are regarded as leaders, and the public network users are regarded as followers, the offloading amount is taken as the strategy, and the utility function of the private network users and the utility function of the public network users are set; in the case of complete information game, the existence of Stackelberg equilibrium is verified, and the first optimal offloading strategy of the private network users and the public network users is solved; in the case of incomplete information game, the TSDRL algorithm is used to obtain the second optimal offloading strategy, and the convergence of the second optimal offloading strategy is evaluated using the first optimal offloading strategy. This technical solution optimizes the task offloading amount in the scenario where the ground base stations are sparsely distributed, but ignores the overall system delay and energy consumption, so this scheme cannot provide high quality of experience for users in some challenging scenarios. SUMMARY

[0006] To address the aforementioned technical problems, the present invention aims to provide a joint positioning and computing support method for IoT devices, which reduces the positioning energy consumption of IoT devices, protects the privacy of IoT devices, and provides stable and reliable computing support for IoT devices in challenging environments with sparsely distributed terrestrial base stations.

[0007] This invention provides a joint positioning and computing support method for Internet of Things (IoT) devices, comprising the following steps:

[0008] Step 1: Collect distance information between IoT devices and anchor nodes within communication range, as well as the remaining battery power information of the anchor nodes, to obtain the set of available anchor nodes for each IoT device;

[0009] Step 2: For cases where the number of available anchor nodes is less than 3, construct a set of auxiliary IoT devices based on the communication range of the target IoT device, and calculate the distance information between the target IoT device and all auxiliary IoT devices;

[0010] Step 3: The IoT device sends the collected information to the UAV. The UAV builds a positioning model based on the location and distance information. The positioning problem is transformed into a semidefinite programming problem through weighted least squares and semidefinite relaxation. Finally, a convex optimization tool is used to solve for the position coordinates of the target IoT device.

[0011] Step 4: The UAV edge server dynamically plans the flight trajectory and allocates computing resources to handle offloaded task requests. The goal is to minimize the total system energy consumption and latency. A deep reinforcement learning method is used to obtain joint decisions on computing task offloading, UAV trajectory planning, and UAV computing resource allocation.

[0012] Step 5: The UAV edge server obtains the optimal downlink power allocation decision through a successive convex approximation method, with the goal of maximizing the minimum downlink throughput;

[0013] Step 6: The UAV performs optimal joint decision-making to support the offloading requests of IoT devices, while performing optimal downlink power allocation decision-making to improve the downlink throughput of the system.

[0014] A joint positioning and computing support method for Internet of Things (IoT) devices has at least the following beneficial effects:

[0015] 1. This invention utilizes mobile edge computing technology to deploy edge servers on UAVs, constructing mobile edge servers that can proactively plan movement trajectories to better support the unloading tasks of IoT devices, increase the flexibility of the computing support system, and improve the utilization rate of UAV computing resources. It solves the problems that IoT devices cannot be equipped with high-cost GPS systems due to cost constraints, and that stable and reliable communication and computing support cannot be obtained due to the sparse distribution of ground base stations and challenging environments.

[0016] 2、The joint positioning method in the joint positioning and computing support technical solution of the Internet of Things device can obtain more accurate real-time positioning information, the computing support method accelerates the network training speed, improves the task completion rate, and improves the resource utilization of the system. In the experiment, the algorithm convergence performance, system total delay and energy consumption, task uncompleted proportion, average unloaded task data size, UAV computing resource utilization and other performance indicators are evaluated, and good results are achieved, which can greatly improve the practicability of the system. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 is a system architecture diagram;

[0018] Figure 2 is a joint positioning flowchart of Internet of Things devices;

[0019] Figure 3 is a joint positioning and computing support flowchart of Internet of Things devices;

[0020] Figure 4 is a graph of the change of the mean square positioning error with the average number of anchor nodes;

[0021] Figure 5 is a graph of the change of the positioning energy consumption with the average number of anchor nodes;

[0022] Figure 6 is a convergence performance comparison graph of different algorithms;

[0023] Figure 7 is a graph of the change of the total delay and energy consumption of different algorithms with the network training process;

[0024] Figure 8 is a graph of the change of the uncompleted task proportion of different algorithms with the network training process;

[0025] Figure 9a is a graph of the change of the average unloaded task data size with the task data size when the training round is 50;

[0026] Figure 9b is a graph of the change of the average unloaded task data size with the task data size when the training round is 6000;

[0027] Figure 10 is a graph of the change of the task unloading proportion after state coordination with the network training process;

[0028] Figure 11 is a graph of the change of the computing resource utilization of the UAV with the number of IoT devices. DETAILED DESCRIPTION

[0029] The present application provides a joint positioning and computing support method for Internet of Things devices.

[0030] First, according to the distance information of the IoT device and the anchor nodes within the communication range and the residual capacity information of the anchor nodes, a set of available anchor nodes is obtained, and for the case that the number of available anchor nodes is small, a set of auxiliary IoT devices is obtained according to the distance information of the IoT device and other IoT devices within the communication range. In addition, the IoT device obtains its position coordinates by a geometric method according to the position and distance information in the available anchor nodes and auxiliary IoT devices.

[0031] Then, the mobile edge computing technology and the UAV equipped with an edge server are used to perform flight trajectory planning, task offloading and computing resource allocation, to provide computing support for the IoT device and minimize the weighted sum of the total energy consumption and delay of the system. The UAV agent is trained in a distributed manner, and the state conflict problem is solved by a coordination normalization method to speed up the training process.

[0032] Finally, the joint trajectory planning, task offloading and computing resource allocation actions obtained by training the UAV agent are executed, and the optimal downlink power allocation decision is obtained by using the successive convex approximation technique and executed.

[0033] The joint positioning and computing support method of the Internet of Things device can provide strong computing resource expansion for the Internet of Things device, increase the flexibility of the computing support system, and improve the resource utilization efficiency of the system.

[0034] As shown in Figure 1 The network model of the application is divided into two layers: a device layer and a server layer. The device layer includes a large number of IoT devices with computationally intensive and delay-sensitive tasks, including disaster warning, long-pipeline infrastructure detection, underwater infrastructure detection and military operations. The IoT devices move according to the Gaussian Markov random motion model and randomly generate task requests with a certain probability. The computing resources of the IoT devices are very limited, and complex tasks can be offloaded to edge servers for processing as much as possible. In addition, a certain number of anchor nodes are fixedly deployed near the IoT devices for positioning. The server layer includes a plurality of UAVs equipped with edge servers, and each UAV receives offloading requests from IoT devices while planning flight trajectories to better provide computing support for IoT devices.

[0035] The application scheme mainly includes two parts: joint positioning and computing support of Internet of Things devices. As shown in Figure 2 The joint positioning of Internet of Things devices mainly includes collecting anchor node energy consumption information, obtaining distance information between IoT devices and anchor nodes, forming a set of available anchor nodes and a set of auxiliary IoT devices, and obtaining positioning coordinates by a geometric method. As shown in Figure 3The computing support includes determining joint task offloading, UAV trajectory planning and computing resource allocation decisions, and determining optimal downlink power allocation decisions, as shown. Specifically, it includes:

[0036] Step 1: Collect the distance information of IoT devices from anchor nodes within the communication range and the residual energy information of anchor nodes, obtain the available anchor node set of each IoT device, and step 1 is specifically:

[0037] Step 1.1: For IoT device u i Construct a candidate anchor node set for positioning Including u i All anchor nodes within the communication range;

[0038] Step 1.2: IoT device u i Send a positioning request signal q to the candidate anchor node i Containing the unique identifier of the IoT device and the positioning request information;

[0039] Step 1.3: After the positioning request signal reaches the candidate anchor node, the anchor node n k According to the sending timestamp and the arrival timestamp of the request, calculate the Euclidean distance d i Between the IoT device u k And anchor node n ik And return its own identifier information, request arrival timestamp information, distance information from the IoT device, its own location information and residual energy information to the IoT device;

[0040] Step 1.4: For the candidate anchor nodes in the set IoT device u i According to the collected anchor node residual energy information, filter the anchor nodes with residual energy greater than the threshold e, form the available anchor node set And calculate the number c i Of anchor nodes in the available anchor node set.

[0041] Step 2: For the case where the number of available anchor nodes is less than 3, construct the auxiliary IoT device set of the target IoT device according to its communication range, and calculate the distance information between the target IoT device and all auxiliary IoT devices, step 2 is specifically:

[0042] Step 2.1: For IoT device u i Construct auxiliary IoT device set Including u i All IoT devices within the communication range;

[0043] Step 2.2: IoT device u iSending a positioning request signal q to an auxiliary IoT device i , wherein the unique identifier of the IoT device u i and the positioning request information are contained;

[0044] Step 2.3: After the positioning request signal reaches the auxiliary IoT device u j , the auxiliary IoT device u i calculates the Euclidean distance d ij between itself and the corresponding IoT device u i according to the requested sending timestamp and the arrival timestamp, and returns the information such as its own identifier information, the timestamp information of the request arrival, and the distance information with the IoT device u i to the target IoT device u ik .

[0045] Step 3: The IoT device sends the collected information to the UAV, and the UAV constructs a positioning model according to the position and distance information, converts the positioning problem into a semi-definite programming (SDP) problem through weighted least squares and semi-definite relaxation (SDR), and finally solves the position coordinates of the target IoT device by using a convex optimization tool. Step 3 is specifically as follows:

[0046] Step 3.1: For each IoT device, the positioning problem is defined as estimating the position of the target IoT device according to the measured distance containing measurement noise and NLOS scene noise, given the available anchor nodes and auxiliary IoT device positions, and the positioning model of the IoT device is constructed as follows:

[0047]

[0048] where d ik represents the observed distance of the IoT device u i to the anchor node n k or the auxiliary IoT device u k , represents the true distance of the IoT device u i to the anchor node n k or the auxiliary IoT device u k , n ik is the measurement noise following a Gaussian distribution with a mean of 0 and a variance of σ 2 , m ik is the NLOS scene noise, and m ik >> n ik ; is the set of available anchor nodes, is the set of auxiliary IoT devices; the set stores the distance information between IoT devices, and the set stores the distance information between IoT devices and anchor nodes.

[0049] Step 3.2: square both sides of the positioning model, omit the much smaller remaining term Meanwhile let The above positioning model can be transformed into:

[0050]

[0051] Step 3.3: Let the IoT device coordinates be The anchor node coordinates are where is the set of IoT devices, is the set of anchor nodes, according to the positioning model obtained in step 3.2, the unknown parameter S i and p ik Use weighted least squares to estimate, which is converted into the following nonlinear weighted least squares problem:

[0052]

[0053]

[0054] where the weight parameter w ik = 1 / (d ik · σ ik ) 2 , σ ik denotes the variance corresponding to the measurement noise, when from the set , when from the set ,

[0055] Step 3.4: Introduce auxiliary variables i.e. and The positioning problem can be converted as follows:

[0056]

[0057]

[0058]

[0059]

[0060]

[0061] Step 3.5: Relax the constraints using SDR, combine the Schur complement theorem, convert the non-convex problem in step 3.4 into an SDP problem, and solve it through convex optimization tools such as CVX, and through continuous iteration to obtain relatively stable IoT device positioning coordinates Si .

[0062] So far, the positioning process of the IoT device is completed. Next, at each time slot t, the IoT device sends its location to the UAV while offloading tasks, so that the UAV can better provide computing support for it.

[0063] Step 4: The UAV edge server dynamically plans a flight trajectory and allocates computing resources to process the offloaded task request, with the goal of minimizing the total energy consumption and delay of the system, and adopts a deep reinforcement learning method to obtain the joint decision of computing task offloading, UAV trajectory planning and UAV computing resource allocation. Step 4 is specifically:

[0064] Step 4.1: Define the total delay and energy consumption weighted sum of the UAV computing support system, and then jointly optimize the offloading decision, UAV path planning decision and UAV computing resource allocation decision with the goal of minimizing the total energy consumption and delay of the system;

[0065] The present application proposes a joint optimization scheme based on deep reinforcement learning, and the optimization goal is the weighted sum of the total energy consumption and delay of the system, including the transmission and computing energy consumption, transmission and computing delay generated by the IoT device, and the flight and computing energy consumption, computing delay generated by the UAV edge server:

[0066]

[0067] Wherein, t is used to indicate the time slot, m is used to indicate the UAV, n is the total time slot length, M is the number of UAVs, alpha is the computing offloading proportion variable, U is the UAV position coordinate, f is the computing resource allocation variable, In order to keep the delay t and the energy consumption E in the same order of magnitude; T t m The computing delay of the UAV v m The total task delay T The total transmission and computing energy consumption of the UAV flight energy consumption and computing energy consumption and the corresponding transmission and computing energy consumption of the task request processed by the UAV.

[0068] The computing resource f t i，m The computing resource f

[0069]

[0070] The total task delay T t m The total task delay T

[0071]

[0072] wherein the UAV has the following restrictions on the flight speed:

[0073]

[0074] for the UAV flight speed, v min and v max are the lower and upper bounds of the UAV flight speed.

[0075] Step 4.2: Each UAV edge server deploys a deep reinforcement learning module as an agent, defines a Markov decision process corresponding to the joint optimization problem, including the agent Agent, the environment state State, the action Action, and the reward Reward, to provide a basis for neural network learning and training on the UAV edge server side;

[0076] Agent: Each UAV edge server is regarded as an Agent, and the environment is considered completely observable to the Agent, and the observation value is equivalent to the state. Each Agent contains Actor and Critic networks, which act as action policy makers and policy critics, respectively. The parameters of the Critic value network are μ, and the corresponding parameters of its target network are μ - ; The parameters of the Actor policy network are θ, and the corresponding parameters of its target network are θ - .

[0077] State: Includes IoT device information and UAV edge server information, and the UAV edge server v m observes the state at time slot t as where f t m is the idle computing resource of the UAV at time slot t, is the position coordinate of the UAV at time slot t, R t,u is the uplink transmission rate vector of the IoT device, C t is the task information set of the IoT device, Req t is the set of IoT device offloading requests, and the elements in the set satisfy where 0 represents no offloading request, 1 represents a request, and -1 represents a request after state coordination.

[0078] Action: Includes offloading ratio decision, computing resource allocation decision, and next time slot position coordinate of the UAV, and the UAV edge server v m performs the action at time slot t as where By the speed angle change amount during the UAV flight and the speed magnitude , a t is redefined as

[0079] Reward: When an agent performs an action, the system will move from one environment state to another and obtain a reward, which will guide each agent to reach its optimal policy. The definition of the reward function is usually related to the optimization goal of the system, so the reward function is defined as the negative of the weighted sum of energy consumption and latency:

[0080]

[0081] where the negative is taken to convert the cost into a reward, and the log(·) function is used to smooth the reward, is a penalty when multiple agents produce different offloading actions for the same IoT device, resulting in state conflicts.

[0082] Step 4.3: Each UAV edge server agent selects an action a t through the actor policy network.

[0083] Step 4.4: The agent performs the action a t obtained in step 4.3, and observes the reward r t and the next state s t+1 .

[0084] Step 4.5: The agent stores the experience tuple R(s t , a t , r t , s t+1 ) in the experience buffer, and samples a small batch of experiences from the experience buffer to update the neural network to speed up the training process, updates the main actor network by minimizing the policy objective function, and updates the main critic network by minimizing the loss function;

[0085] Step 4.6: The agent updates the target actor network and the target critic network through soft policy updates at each time slot;

[0086] Step 4.7: For the case where the IoT device is in multiple UAV coverage areas, the action conflict problem caused by the state conflict of related agents and the amplitude difference problem between input state elements, each agent communicates after generating an action, and for the case where conflicting actions are generated, the corresponding penalty is added to its reward to achieve state coordination and normalization;

[0087] Step 4.8: The agent repeatedly performs steps 4.3-4.6 to try and learn, and finally obtains the optimal joint decision of task offloading, UAV trajectory planning, and computing resource allocation {a, U, f}.

[0088] Step 5: The UAV edge server obtains the optimal downlink power allocation decision by a method of successive convex approximation, with the goal of maximizing the minimum downlink throughput, and step 5 is specifically:

[0089] Step 5.1: Define the minimum average downlink transmission rate of the UAV computing support system to optimize the downlink power allocation and thus improve the downlink throughput;

[0090]

[0091] where P d is the downlink transmission power allocation variable, represents the proportion of the IoT device task that is offloaded to the UAV edge server for processing, is the downlink transmission rate between the IoT device u k and the UAV edge server v m , and is defined as follows:

[0092]

[0093] where W represents the bandwidth, represents the noise power, represents the channel gain, represents the transmission power of the UAV v m , and represents the co-channel interference caused by other UAVs at time slot t.

[0094] By introducing an auxiliary variable R dmin , the problem is redefined as a problem of maximizing the minimum average downlink transmission rate:

[0095]

[0096]

[0097] where the downlink transmission power of the UAV edge server cannot exceed its specified maximum downlink transmission power:

[0098]

[0099] Step 5.2: The maximum minimum average downlink transmission rate problem is converted into a convex problem by using the properties of the logarithmic function in the downlink transmission power and the first-order Taylor expansion, and the optimal power value is obtained in each iteration by using the successive convex approximation algorithm SCA until the difference between the optimal values of two consecutive iterations is less than a certain threshold, and the optimization of the maximum downlink throughput is realized, and then the downlink transmission power allocation is optimized.

[0100] Step 6: The UAV executes the optimal joint decision to support the offloading request of the IoT device, and at the same time executes the optimal downlink power allocation decision to improve the downlink throughput of the system, and step 6 is specifically:

[0101] Step 6.1: The UAV adjusts its flight trajectory according to the flight trajectory planning decision obtained in step 4, and processes the offloading task request of the IoT device according to the task offloading decision and the computing resource allocation decision in step 4.

[0102] Step 6.2: The UAV transmits the task result back to the IoT device according to the optimal downlink power allocation decision obtained in step 5, and improves the downlink throughput of the system by maximizing the minimum downlink transmission power.

[0103] The technical solutions of the present application will be described in detail below in combination with specific experimental platforms and experimental results.

[0104] The simulation of the present application is realized based on the Matlab and PyCharm platforms, and a 1000m*1000m IoT device working area is considered, in which 3 UAV edge servers are deployed, the initial positions of 100 IoT devices are subject to uniform distribution, 30 anchor nodes are fixed in the considered area in a uniform random distribution manner, the node Gaussian Markov random motion model is used to simulate the mobility of the IoT device, and the UAV flies at a fixed height and provides computing support for the IoT device. The computing frequency of the IoT device is uniformly distributed in [0.1, 0.5]GHz, the data size of the task is uniformly distributed in [100, 1000]KB, the required computing resource per unit task data is uniformly distributed in [500, 1000]cycles / bit, the maximum computing resource capacity of the UAV is set to 20GHz, and the flight speed is uniformly distributed in [10, 15]m / s. The experiment simulates 6000 time slots, in each time period, the IoT device generates a computing task with a probability of 0.99, the UAV will hover to the specified position according to the action generated in the last time period, continue to fly after receiving the offloading task, and process the computing task, therefore, the positions of the IoT device and the UAV are dynamically changed in different time periods. The parameter list is shown in Table 1: k k In [100, 1000]KB, the required computing resource per unit task data is uniformly distributed in [500, 1000]cycles / bit, the maximum computing resource capacity of the UAV is set to 20GHz, and the flight speed is uniformly distributed in [10, 15]m / s. The experiment simulates 6000 time slots, in each time period, the IoT device generates a computing task with a probability of 0.99, the UAV will hover to the specified position according to the action generated in the last time period, continue to fly after receiving the offloading task, and process the computing task, therefore, the positions of the IoT device and the UAV are dynamically changed in different time periods. The parameter list is shown in Table 1:

[0105] Table 1 Parameter settings ​

[0106]

[0107] where K is the number of IoT devices, M is the number of UAVs, N is the number of anchor nodes, φ1 and φ3 are constant parameters related to the hardware of IoT devices, η is the discount rate of rewards, θ is the Actor network parameter, μ is the Critic network parameter, is the maximum computing resource of the UAV v m , is the maximum downlink transmission power of the UAV v m , is the noise power.

[0108] At the same time, the application sets a comparative algorithm in the simulation implementation. For the positioning algorithm, in order to evaluate the root mean squared error (RMSE) of the positioning position and the actual position and the positioning energy consumption of the IoT device, the least squares method, the single-device positioning method based on SDR, the multi-device positioning method based on SDR and the joint positioning method proposed in the patent are compared on the Matlab platform.

[0109] Figure 4 It is shown that when the measurement noise σ 2 =1, the root mean squared error of the positioning position and the actual position changes with the average number of anchor nodes around the IoT device. The RMSE obtained by the joint positioning method is lower than that of other comparative algorithms, because the joint positioning method has the advantages of the single-device positioning method based on SDR and the multi-device positioning method based on SDR, and can adaptively adjust the algorithm according to the distribution of anchor nodes.

[0110] Figure 5 It is shown that the average positioning energy consumption of the IoT device changes with the average number of anchor nodes around the IoT device. The positioning energy consumption of our method is similar to that of the least squares method, and is much lower than that of other comparative algorithms, because the joint positioning method uses the least squares principle, and when the number of anchor nodes increases, the joint positioning method will gradually tend to use the single-device positioning method based on SDR, thereby realizing energy saving.

[0111] For the joint optimization algorithm of task offloading, UAV computing resource allocation and trajectory planning, in order to evaluate the convergence, the weighted sum of total energy consumption and delay, the task uncompleted ratio, the average offloaded task data size, the proportion of tasks that are not offloaded after state coordination to the coordinated tasks, and the UAV computing resource utilization, the deep deterministic policy gradient algorithm with punishment mechanism (PF-DDPG), the deep Q network algorithm with punishment mechanism (PF-DQN), the greedy algorithm (Greedy) and the multi-agent state coordination normalized deep reinforcement learning algorithm (MASC-DDPG) proposed in the patent are compared on the PyCharm platform.

[0112] Figure 6 The convergence performance of each algorithm is shown, which is represented by the change of the total reward of the UAV agent. From the figure, it can be seen that with the increase of time slots, the reward of the Greedy method does not change much, and the rewards of other deep reinforcement learning methods gradually increase and eventually tend to be stable, and the reward after training is greater than that obtained by the Greedy method. In addition, compared with the other two deep reinforcement learning algorithms, our algorithm converges faster and obtains higher total reward, because the MASC-DDPG method considers the state coordination normalization of multiple UAV agents, and the speed of the DQN method decreases significantly when the number of action quantization increases.

[0113] Figure 7 The change of the weighted sum of total energy consumption and delay with training time is shown. It can be seen that with the increase of time slots, the weighted sum of total energy consumption and delay of the Greedy method does not change much, and the weighted sum of total energy consumption and delay of other deep reinforcement learning methods gradually decreases and eventually tends to be stable, and is less than that of the Greedy method after training. In addition, compared with the other two deep reinforcement learning algorithms, the weighted sum of total energy consumption and delay of our algorithm is smaller.

[0114] Figure 8 The change of the task uncompleted ratio is shown. From the figure, it can be seen that with the increase of time slots, the task uncompleted ratio gradually decreases and converges to a minimum value. The MASC-DDPG algorithm can achieve a smaller uncompleted ratio than other algorithms, and the amplitude fluctuation is smaller, and the performance is more stable.

[0115] Figure 9a and 9b The average offloaded task data size of IoT devices to UAV is shown under different training rounds and different task data sizes. From the figure, it can be seen that when the training round is 50, the average offloaded task data size of the Greedy algorithm is the largest, and the offloading ratio of other deep reinforcement learning algorithms is smaller. Figure 9a It can be concluded that when the training round is 50, the average offloaded task data size of the Greedy algorithm is the largest, and the offloading ratio of other deep reinforcement learning algorithms is smaller. From the figure, it can be seen that when the task data size is 1000, the average offloaded task data size of the Greedy algorithm is the largest, and the offloading ratio of other deep reinforcement learning algorithms is smaller.Figure 9b It can be concluded that when the number of training rounds reaches 6000, the offloading task amount of the deep reinforcement learning-based algorithm is higher than that of the Greedy algorithm. This is because in the initial stage, the training effect has not yet appeared, and when the training reaches a stable state, the deep reinforcement learning-based algorithm will choose to offload tasks as much as possible according to the reward.

[0116] Figure 10 It is shown that after the coordination state, the proportion of tasks not offloaded in the action generated by the MASC-DDPG algorithm accounts for the coordination tasks. It can be concluded that the initial non-offloading proportion is about 0.47, which gradually decreases with the training and finally converges to about 0.05. After analysis, it can be known that after adding the penalty factor for non-offloading action in the reward function, the network will continuously tend to offload tasks during training, so that the situation of tasks not offloaded in the action after the state coordination is significantly reduced.

[0117] Figure 11 It is shown that the change of the UAV computing resource utilization rate with the number of IoT devices. It can be seen from the figure that with the increase of the number of IoT devices, the computing resource utilization rate gradually increases and tends to be stable. Because the computing resources of the UAV are limited, when the number of IoT devices increases, the total number of tasks increases, and the computing resources allocated to the tasks by the UAV will eventually approach its maximum computing resource capacity, at which time the resource utilization rate cannot continue to rise. It can also be seen from the figure that our computing resource utilization rate is the highest.

[0118] Considering the hardware limitations and mobility of IoT devices, we first proposed a joint positioning algorithm based on semi-positive relaxation and arrival time. Then, for the complex Internet of Things scene of multiple IoT device numbers and multiple UAVs, the MASC-DDPG algorithm is proposed, which considers the state coordination normalization of multiple UAV agents to reduce state conflict and action conflict, and improve the convergence speed of network training. The experimental results show that the scheme proposed in the patent can accurately position the IoT devices in real time, and compared with the comparative algorithm, it realizes the lowest positioning error and the lowest positioning energy consumption. The scheme proposed in the patent shortens the convergence time of network training, effectively reduces the weighted sum of total energy consumption and delay, improves the average offloaded task data amount, and improves the computing resource utilization rate of the UAV, providing dynamic and reliable computing support for IoT devices.

[0119] The above only describes the preferred embodiments of the present application and is not intended to limit the idea of the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A joint positioning and computing support method for Internet of Things (IoT) devices, characterized in that, Includes the following steps: Step 1: Collect distance information between IoT devices and anchor nodes within communication range, as well as the remaining battery power information of the anchor nodes, to obtain the set of available anchor nodes for each IoT device; Step 2: For cases where the number of available anchor nodes is less than 3, construct a set of auxiliary IoT devices based on the communication range of the target IoT device, and calculate the distance information between the target IoT device and all auxiliary IoT devices; Step 3: The IoT device sends the collected information to the UAV. The UAV constructs a positioning model based on the location and distance information. It then transforms the positioning problem into a semidefinite programming problem using weighted least squares and semidefinite relaxation. Finally, a convex optimization tool is used to solve for the location coordinates of the target IoT device. Specifically: Step 3.1: For each IoT device, the localization problem is defined as estimating the location of the target IoT device based on the measured distance containing measurement noise and NLOS scene noise, given the locations of available anchor nodes and auxiliary IoT devices. The localization model for the IoT device is constructed as follows: in, Indicates IoT devices To anchor node Or auxiliary IoT devices Observation distance, Indicates IoT devices To anchor node Or auxiliary IoT devices The actual distance It follows a pattern with a mean of 0 and a variance of . Gaussian distribution measurement noise, It is NLOS scene noise, and ; For the set of available anchor nodes, To assist in the aggregation of IoT devices; using aggregation Stores distance information between IoT devices, and aggregates them. Stores distance information between IoT devices and anchor nodes; Step 3.2: Squaring both sides of the location model, omitting terms much smaller than the remaining terms. At the same time, The above positioning model can be transformed into: Step 3.3: Set the coordinates of the IoT device as follows: The coordinates of the anchor node are ,in For IoT device collection, Given the set of anchor nodes, based on the positioning model obtained in step 3.2, the unknown parameters are... and Using weighted least squares to estimate, the problem is transformed into the following nonlinear weighted least squares problem: Among them, weight parameters , This represents the variance corresponding to the measurement noise, when From collection hour, ,when From collection hour, ; Step 3.4: Introduce auxiliary variables ,Right now and The location problem can be transformed as follows: Step 3.5: Using SDR to relax constraints and combining it with the Schur complement theorem, the non-convex problem in Step 3.4 is transformed into an SDP problem, which is then solved using convex optimization tools. Through continuous iteration, relatively stable IoT device positioning coordinates are obtained. ; Step 4: The UAV edge server dynamically plans the flight trajectory and allocates computing resources to handle offloaded task requests. The goal is to minimize the total system energy consumption and latency. A deep reinforcement learning method is used to obtain joint decisions on computing task offloading, UAV trajectory planning, and UAV computing resource allocation. Step 5: The UAV edge server obtains the optimal downlink power allocation decision through a successive convex approximation method, with the goal of maximizing the minimum downlink throughput; Step 6: The UAV performs optimal joint decision-making to support the offloading requests of IoT devices, while performing optimal downlink power allocation decision-making to improve the downlink throughput of the system.

2. The joint positioning and computing support method for IoT devices as described in claim 1, characterized in that, Step 1 is as follows: Step 1.1: Construct a set of candidate anchor nodes for IoT device positioning, including all anchor nodes within communication range; Step 1.2: The IoT device sends a location request signal to the candidate anchor node, which contains the IoT device's unique identifier and location request information; Step 1.3: After the location request signal arrives at the candidate anchor node, the anchor node calculates the Euclidean distance between the corresponding IoT device and the anchor node based on the sending timestamp and arrival timestamp of the request, and returns its own identification information, the timestamp of the request arrival, the distance information to the IoT device, its own location information, and the remaining battery information to the IoT device. Step 1.4: For candidate anchor nodes in the set, the IoT device filters anchor nodes whose remaining power is greater than a threshold based on the collected anchor node remaining power information, forming a set of available anchor nodes, and calculates the number of anchor nodes in the set of available anchor nodes.

3. The joint positioning and computing support method for IoT devices as described in claim 1, characterized in that, Step 2 is as follows: Step 2.1: Build an auxiliary IoT device set for the IoT device, which includes all IoT devices within communication range; Step 2.2: The IoT device sends a location request signal to the auxiliary IoT device, which contains the unique identifier of the IoT device and the location request information; Step 2.3: After the location request signal arrives at the auxiliary IoT device, the auxiliary IoT device calculates the Euclidean distance between itself and the corresponding IoT device based on the sending and arrival timestamps of the request, and then exchanges its own identification information, the request arrival timestamp information, and the distance between itself and the IoT device. The distance information is returned to the target IoT device.

4. The joint positioning and computing support method for IoT devices as described in claim 1, characterized in that, Step 4 is as follows: Step 4.1: Define the weighted sum of total latency and energy consumption of the UAV computing support system, and then jointly optimize the offloading decision, UAV path planning decision, and UAV computing resource allocation decision with the goal of minimizing the total system energy consumption and latency; Step 4.2: Deploy a deep reinforcement learning module as an agent in each UAV edge server, and define a Markov decision process corresponding to the joint optimization problem, including Agent, State, Action, and Reward, to provide a foundation for neural network learning and training on the UAV edge server. Step 4.3: Each UAV edge server's agent selects an action through the Actor policy network. ; Step 4.4: The Agent executes the action obtained in Step 4.

3. And observe the rewards and the next state ; Step 4.5: The Agent will transfer the experience tuples The experience is stored in the experience buffer, and mini-batch experiences in the experience buffer are sampled to update the neural network to accelerate the training process. The main Actor network is updated by minimizing the policy objective function, and the main Critic network is updated by minimizing the loss function. Step 4.6: The Agent updates the target Actor network and the target Critic network in each time slot using a soft update strategy; Step 4.7: To address the issues of action conflicts caused by conflicting Agent states and magnitude differences between input state elements when IoT devices are located in multiple UAV coverage areas, each Agent communicates after generating an action. For cases of conflicting actions, a corresponding penalty is added to the reward to achieve state coordination and normalization. Step 4.8: The agent continuously repeats steps 4.3-4.6 for trial and error and learning, and finally arrives at the optimal joint decision for task unloading, UAV trajectory planning and computing resource allocation.

5. The joint positioning and computing support method for IoT devices as described in claim 4, characterized in that, Step 4.1 aims to minimize the total system energy consumption and delay, as shown in the following expression: Where t is used to indicate the time slot. Used to indicate UAVs, where n is the total time slot length and M is the number of UAVs. To calculate the unloading ratio variable, Here are the coordinates of the UAV position. To calculate resource allocation variables, It is to reduce latency Energy consumption Maintain the same order of magnitude; UAV The computational latency and the sum of the transmission and computational latency corresponding to the task requests being processed. This represents the total energy consumption of UAV flight and computing, as well as the transmission and computing energy consumption corresponding to the task requests it processes.

6. The joint positioning and computing support method for IoT devices as described in claim 1, characterized in that, Step 5 specifically involves: Step 5.1: Define the minimum average downlink transmission rate of the UAV calculation support system to optimize downlink power allocation and thus improve downlink throughput; in, Assign variables to downlink transmission power. This indicates the proportion of IoT device tasks that are offloaded to UAV edge servers for processing. For IoT devices With UAV edge server Downlink transmission rate between; Step 5.2: By utilizing the properties of the logarithmic function in downlink transmission power and the first-order Taylor expansion, the problem of maximizing the minimum average downlink transmission rate is transformed into a convex problem. Using the successive convex approximation algorithm SCA, the optimal power value is found in each iteration until the difference between the optimal values ​​of two consecutive iterations is less than a certain threshold, thereby optimizing the maximum downlink throughput and thus optimizing the downlink transmission power allocation.

7. The joint positioning and computing support method for IoT devices as described in claim 1, characterized in that, Step 6 specifically involves: Step 6.1: The UAV adjusts its own flight trajectory based on the flight trajectory planning decision obtained in Step 4, and processes the unloading task request of the IoT device based on the task unloading decision and computing resource allocation decision in Step 4. Step 6.2: The UAV transmits the task results back to the IoT device based on the optimal downlink power allocation decision obtained in Step 5, thereby improving the downlink throughput of the system by maximizing the minimum downlink transmission power.

Images