Inland river distributed task unloading method based on clustering

Abstract

The invention discloses a clustering-based inland river distributed task unloading method, which comprises the following steps of: randomly dividing and acquiring a plurality of clustering clusters according to a cluster head node and member nodes on the basis of a task unloading clustering strategy; on the basis of a task unloading allocation strategy, dividing a task to be unloaded into sub-task chains according to the link connectivity graph, and allocating member nodes in a cluster to obtain a task distribution sequence; constructing a distributed task unloading network to obtain a plurality of task transmission delays according to the task distribution sequence; constructing a global delay cost target model for distributed task unloading according to the transmission delay of each task; according to a preset to-be-unloaded inland river distributed task, obtaining an optimal clustering cluster corresponding to the minimum global delay cost target model; and realizing inland river distributed task unloading according to the optimal clustering cluster. The problem that when the multiple unmanned ships need to unload the tasks to the edge cloud, due to the influence of resource competition, the task unloading effect is difficult to meet the user requirements, and the user experience is directly influenced is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of inland waterway edge computing, and in particular to a clustering-based distributed task offloading method for inland rivers. Background Art

[0002] Mobile edge computing is an innovative task offloading cloud computing model proposed by the European Telecommunications Standards Institute. It mainly provides high-speed computing units for network edge devices (such as USVs in inland waterway segments). Through this innovative means, services can be closer to users and operators, bringing better application experiences to users, such as reducing data processing latency. Existing edge computing methods in this field include full offloading, partial offloading, dynamic offloading, and task scheduling optimization. Full offloading means offloading the entire computing task to the MEC server for execution; partial offloading means only offloading some computationally intensive tasks, while other tasks are still executed on local devices; dynamic offloading means dynamically adjusting the offloading strategy based on network conditions, computational resource availability, and latency requirements; task scheduling and optimization means using machine learning or optimization algorithms to reasonably allocate computing tasks among edge servers to improve the overall performance of the system.

[0003] Although traditional edge computing methods have advantages in task offloading, they still face high pressure when dealing with a large number of user connections. For example, when multiple unmanned vessels appear in an inland waterway simultaneously and need to offload tasks to the edge cloud, due to the influence of resource competition, the task offloading effect is difficult to meet user requirements and directly affects the user experience. Summary of the Invention

[0004] The present invention provides a clustering-based distributed task offloading method for inland rivers to overcome the above technical problems.

[0005] To achieve the above object, the technical solution of the present invention is as follows:

[0006] A clustering-based distributed task offloading method for inland rivers specifically includes the following steps:

[0007] S1: Using the UAV as a communication relay for heterogeneous distributed USVs to establish a line-of-sight connection between the USV and the edge server TBS, and obtaining a drone-assisted edge network;

[0008] And the heterogeneous distributed USV includes a computationally intensive USV for providing task offloading / communication resources and an assisting USV for providing backup task offloading / communication resources for the computationally intensive USV;

[0009] S2: Using the computationally intensive USV as the cluster head node of the clustering cluster and the assisting USV as the member node of the clustering cluster;

[0010] And based on the UAV-assisted edge network, randomly divide the cluster head nodes and member nodes according to the task offloading clustering strategy and obtain a number of clustering clusters;

[0011] S3: Define a communication link connectivity graph according to the clustering clusters;

[0012] And based on the task offloading allocation strategy, divide the tasks to be offloaded into subtask chains according to the link connectivity graph, and allocate the subtasks in the subtask chains to each member node in the clustering clusters to obtain a task distribution sequence;

[0013] S4: Construct a distributed task offloading network;

[0014] And the distributed task offloading network includes a global task distribution module, an in-cluster task execution module, and a task completion upload module;

[0015] The global task distribution module is used to transmit the subtasks in the task distribution sequence from the cluster head nodes in the clustering clusters to the member nodes through D2D communication to obtain the first task transmission delay;

[0016] The in-cluster task execution module is used to sequentially transmit the subtasks to each member node based on the obtained subtask chain and perform task execution operations, and feedback the completed tasks after the execution operations to the corresponding cluster head nodes for collection to obtain the second task transmission delay;

[0017] The task completion upload module is used to transmit the completed tasks collected by each cluster head node to the edge server TBS to obtain the third task transmission delay;

[0018] S5: Construct a global delay cost objective model for distributed task offloading according to the first task transmission delay, the second task transmission delay, and the third task transmission delay obtained based on the distributed task offloading network;

[0019] S6: Based on the preset distributed tasks to be offloaded in the inland river, obtain the optimal clustering clusters corresponding to minimizing the global delay cost objective model according to step S2;

[0020] And realize the inland river distributed task offloading according to the optimal clustering clusters.

[0021] Further, the S2 specifically includes the following steps:

[0022] S21: Take the computationally intensive USV as the cluster head node of the clustering cluster, denoted as

[0023] where represents the clustering cluster set; L represents the number of computationally intensive USVs in the UAV-assisted edge network;

[0024] will assist the USV as a member node of the cluster, denoted as

[0025] where represents the set of assisting USVs; H represents the number of assisting USVs in the UAV-assisted edge network;

[0026] S22: Based on the task offloading clustering strategy, randomly partition and obtain several clustering clusters according to the cluster head node and member nodes to define the member node set of clustering cluster l and

[0027] where n l represents the number of member nodes in clustering cluster l;

[0028] And the task offloading clustering strategy is specifically as follows:

[0029] Assume that the number of clustering clusters is l and define that there is and can only be one cluster head node in each clustering cluster, and at the same time each assisting USV has and can only join one clustering cluster;

[0030] And the association constraint that each assisting USV has and can only join one clustering cluster is

[0031]

[0032] In the formula: s l,h represents the binary decision variable of assisting USVh as a member node of clustering cluster l, and s l,h =1 indicates that assisting USVh is a member node of clustering cluster l, otherwise s l,h =0;

[0033] Furthermore, define the member node clustering vector of clustering cluster l as S l ={s l,1 , s l,2 ,..., s l,H}.

[0034] Furthermore, the above S2 specifically includes the following steps:

[0035] S31: Define the communication link connectivity graph according to the clustering cluster and

[0036] where V represents the member nodes; E represents the edges between the member nodes, that is, the communication links generated by wireless connections; c represents the allocation constraint of the communication links;

[0037] S32: Based on the task offloading allocation strategy, divide the tasks to be offloaded into sub-task chains according to the connected graph, and allocate them to each member node in the clustering cluster to obtain the task distribution sequence;

[0038] And the task offloading allocation strategy is specifically: according to the number of each member node in the clustering cluster, divide the tasks to be offloaded into sub-tasks corresponding to the number of each member node in the clustering cluster based on the task offloading allocation constraints;

[0039] The task offloading allocation constraints are

[0040]

[0041] In the formula: M represents the task splitter and M = {1, 2,..., m}; k l,m represents the decision variable of the selection state of the splitting layer m in the task splitter and m ∈ M, k l,m = 1 indicates that the splitting layer m is selected, otherwise k l,m = 0;

[0042] Furthermore, define the splitting layer vector of the task splitter used to divide the tasks to be offloaded as k l = {k l,1 , k l,2 ,... k l,m};

[0043] According to the execution order of the tasks to be offloaded, reorganize the divided sub-tasks into sub-task chains;

[0044] S33: According to the communication link relationship of the member nodes in the communication link connected graph in, allocate the sub-task chains to different member nodes in sequence, and then obtain the task distribution sequence of the clustering cluster.

[0045] Furthermore, the expression of the first task transmission delay obtained in S4 is

[0046]

[0047] In the formula: t td,i represents the task allocation delay for transmitting the sub-tasks in the task distribution sequence from the cluster head node in the clustering cluster to the member node through D2D communication; R i represents the D2D data transmission rate between the cluster head node CHi and the member node i; Q l,i represents the task size processed by the member node i;

[0048] The expression of the second task transmission delay obtained is

[0049]

[0050] where: t TE represents an intermediate variable; represents the FP transmission delay from member node i to member node j; represents the intermediate data passed from member node i to member node j after member node i completes local calculation; R i,j represents the transmission rate between member node i and member node j; represents the task execution delay of member node i; C i represents the number of CPU cycles required to process one data sample; f i represents the CPU calculation frequency of member node i;

[0051] The expression for the obtained third task transmission delay is

[0052] t tc,j = t td,i

[0053] where: t tc,j represents the task collection delay for transmitting the completed subtasks from the member nodes in the clustering cluster to the cluster head node through D2D communication.

[0054] Furthermore, the global delay cost objective model for distributed task offloading constructed in S5 includes a global delay cost objective function and objective constraints, and its expression is

[0055]

[0056] where: T represents the global delay cost objective function.

[0057] Furthermore, S6 specifically includes the following steps:

[0058] S61: Construct a member node clustering game mechanism based on graph reinforcement learning and exact potential game;

[0059] And based on the preset inland river distributed tasks to be offloaded, obtain the optimal clustering cluster corresponding to minimizing the global delay cost objective model according to step S2;

[0060] And the constructed member node clustering game mechanism based on graph reinforcement learning and exact potential game specifically includes:

[0061] S611: Given each computationally intensive USV a known clustering strategy s l and s l ∈ S l , then the joint clustering strategy of L clustering clusters can be expressed as s = {s1,..., s l ,... s L};

[0062] S612: Define the utility function of the clustering cluster for the global delay cost objective model, and the utility function U of the clustering cluster l (s l ,s -l ) is expressed as

[0063]

[0064] In the formula: S -l represents other known clustering strategies in the preset clustering strategy pool except for the clustering cluster l; -ψ j (s j ,s -j\l ) represents the utility of the clustering cluster j when the clustering cluster l does not take action; represents the change in the utility of other clustering clusters caused by the clustering strategy adopted by the clustering cluster l; -ψ l (s l ,s -l ) represents the utility of the clustering cluster l when the clustering cluster j does not take action;

[0065] S613: Obtain the exact potential game mechanism EPG based on the potential function, and its expression is

[0066] U l (s l ',s -l )-U l (s l ,s -l )=μ(s l ',s -l )-μ(s l ,s -l ),

[0067]

[0068] In the formula: μ(s l ,s -l ) represents the potential function regarding the clustering strategy s l and the clustering strategy s -l ; U l (s' l ,s -l ) represents the utility function of the clustering strategy s' ; and the clustering strategy s -l ; s' l represents the changed clustering strategy of the clustering cluster l; μ(s' l ,s -l ) represents the potential function regarding the clustering strategy s' l and the clustering strategy s -l ;

[0069] S62: Define the joint strategy of clustering clusters, which is the Nash equilibrium of the clustering game

[0070] To obtain the one for minimizing the utility function U according to each clustering cluster based on the exact potential game mechanism EPG l (s l , s -l ) to reach the maximum value, that is, when the target delay of the global delay cost objective model is the smallest, the Nash equilibrium constraint conditions corresponding to the optimal clustering strategy are obtained;

[0071] And the expression of the Nash equilibrium constraint conditions is

[0072]

[0073] In the formula: Represents the utility when the clustering strategy s l and the clustering strategy s -l reach equilibrium in the game; Represents the utility when only the clustering strategy s -l reaches equilibrium in the game; Represents the optimal solution of the clustering strategy s -l at equilibrium; Represents the optimal solution of the clustering strategy s l at equilibrium;

[0074] S63: Based on the Nash equilibrium constraint conditions, obtain the optimal clustering strategy of the UAV-assisted edge network, and implement the distributed offloading task of the inland waterway according to the optimal clustering clusters.

[0075] Beneficial effects: The present invention provides a clustering-based inland waterway distributed task offloading method. By considering the actual communication condition limitations and actual application requirements of the inland waterway, the potential game theory is used to model the assistance USV clustering optimization problem as a distributed computing offloading game mechanism, aiming to maximize the overall offloading benefit of the constructed distributed task offloading network on the basis of considering the dynamically changing D2D communication quality and clustering cluster topology structure. In order to rationally utilize computing and communication resources, the present invention constructs a global delay minimization problem, that is, the global delay cost objective model of distributed task offloading, and through constructing a member node clustering game mechanism based on graph reinforcement learning and exact potential game, combined with the Nash equilibrium of the clustering game, solves the clustering cluster task allocation problem to achieve global delay minimization, and further obtains the optimal clustering strategy of the UAV-assisted edge network to implement the distributed offloading task of the inland waterway. Brief Description of the Drawings

[0076] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the accompanying drawings required for the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other accompanying drawings can also be obtained based on these drawings.

[0077] Figure 1 It is a flowchart of the distributed task offloading method for inland rivers based on clustering of the present invention;

[0078] Figure 2 It is a schematic diagram of D2D clustering cluster task offloading in the UAV-assisted edge network in this embodiment;

[0079] Figure 3 It is a schematic diagram of the D2D multi-dimensional communication graph within the clustering cluster in this embodiment;

[0080] Figure 4 It is a schematic diagram of the structure of the graph neural network centered on the D2D communication link in this embodiment;

[0081] Figure 5 It is a simulation diagram of the local performance comparison of the three offloading mechanisms in this embodiment under different CC numbers;

[0082] Figure 6 It is a simulation diagram of the local performance comparison of the three offloading mechanisms in this embodiment under different AC numbers;

[0083] Figure 7 It is a simulation diagram of the global performance comparison of the three offloading mechanisms in this embodiment under different AC numbers;

[0084] Figure 8 It is a simulation diagram of the global performance comparison of the three offloading mechanisms in this embodiment under different CC numbers;

[0085] Figure 9 It is a simulation diagram of the performance comparison under different data sizes in this embodiment;

[0086] Figure 10 It is a simulation diagram of the performance comparison under different D2D bandwidths in this embodiment. Detailed implementation manners

[0087] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the scope of protection of the present invention.

[0088] This embodiment provides a clustering-based distributed task offloading method for inland rivers, as Figure 1 shown, which specifically includes the following steps:

[0089] S1: Use the UAV as a communication relay for heterogeneous distributed USVs to establish a line-of-sight connection between the USV and the edge server TBS, and obtain a drone-assisted edge network;

[0090] And the heterogeneous distributed USV includes a computationally intensive USV for providing task offloading / communication resources and an assisting USV for providing backup task offloading / communication resources for the computationally intensive USV;

[0091] Specifically, as Figure 2 shown, the drone-assisted edge network considered in this embodiment consists of multiple distributed USVs with heterogeneous computing / communication resources, multiple UAVs with the same communication resources, and a TBS (acting as an edge server) with sufficient computing resources; each unmanned ship can establish a D2D communication link with neighboring USVs. However, due to factors such as heterogeneous USV communication resources and obstacle occlusion, the USV cannot directly establish a communication link with the TBS. Therefore, the UAV is used as the communication relay between the TBS and the USV, and a line-of-sight connection with the TBS is formed by establishing a wireless cellular link with the USV;

[0092] Specifically, the distributed USV can be divided into two categories, including a computationally intensive client, i.e., a computationally intensive USV (CC), and an assisting client, i.e., an assisting USV (AC). In this embodiment, there are L CCs that can participate in task offloading by contributing their local tasks, and there are H ACs nearby that can help the CCs complete tasks by providing backup computing / communication resources, which can be expressed as Therefore, each CC can cooperate with multiple ACs through D2D communication to form a clustering cluster for task offloading to overcome the bottleneck caused by its limited computing or communication resources, that is, L CCs can be clustered into L learning clusters, and can be expressed as

[0093] S2: Use the computationally intensive USV as the cluster head node of the clustering cluster, use the assisting USV as the member node of the clustering cluster, and based on the drone-assisted edge network, randomly divide the cluster head node and the member node according to the task offloading clustering strategy to obtain several clustering clusters;

[0094] For the convenience of analysis in this embodiment, each clustering cluster only contains one CC, and without loss of generality, assume that the clustering cluster is associated with n l ACs, which can be expressed as In addition, to simplify the analysis and avoid resource sharing conflicts in the AC, this embodiment assumes that each AC can join at most one clustering cluster;

[0095] Specifically, it includes the following steps:

[0096] S21: Take the computationally intensive USV as the cluster head node of the clustering cluster, denoted as

[0097] where represents the set of clustering clusters; L represents the number of computationally intensive USVs in the UAV-assisted edge network;

[0098] Take the assisting USV as the member node of the clustering cluster, denoted as

[0099] where represents the set of assisting USVs; H represents the number of assisting USVs in the UAV-assisted edge network;

[0100] S22: Based on the task offloading clustering strategy, randomly divide and obtain several clustering clusters according to the cluster head node and member nodes to define the member node set of clustering cluster l and

[0101] where n l represents the number of member nodes in clustering cluster l;

[0102] And the task offloading clustering strategy is specifically:

[0103] Assume that the number of clustering clusters is l and Define that there is exactly one cluster head node in each clustering cluster, and at the same time each assisting USV can join at most one clustering cluster;

[0104] And the association constraint that each assisting USV can join at most one clustering cluster is

[0105]

[0106] In the formula: s l,h represents the binary decision variable of assisting USV h as a member node of clustering cluster l, and s l,h =1 means that assisting USV h is a member node of clustering cluster l, otherwise s l,h =0;

[0107] Furthermore, define the member node clustering vector of clustering cluster l as S l ={s l,1 ,s l,2 ,...,s l,H};

[0108] S3: Define the communication link connectivity graph according to the clustering clusters;

[0109] And based on the task offloading assignment strategy, divide the tasks to be offloaded into sub-task chains according to the link connectivity graph, and allocate the sub-tasks in the sub-task chains to each member node in the clustering clusters to obtain the task distribution sequence;

[0110] Specifically, it includes the following steps:

[0111] S31: Define the communication link connectivity graph according to the clustering clusters And

[0112] Where V represents the member nodes; E represents the edges between the member nodes, that is, the communication links generated by wireless connections; c: E → R + Then it represents the communication link allocation constraint. Let n = |V| represent the number of USVs assisting. If any two different vertices in V are connected, it is called a connected graph

[0113] S32: Based on the task offloading assignment strategy, divide the tasks to be offloaded into sub-task chains according to the connectivity graph, and allocate them to each member node in the clustering clusters to obtain the task distribution sequence;

[0114] And the task offloading assignment strategy is specifically: according to the number of each member node in the clustering clusters, divide the tasks to be offloaded into sub-tasks corresponding to the number of each member node in the clustering clusters based on the task offloading assignment constraint;

[0115] In this embodiment, by splitting a complete offloading task into multiple sub-tasks within each clustering cluster and allocating these sub-tasks to them according to the computing capabilities of the members within the clustering cluster, the tasks are completed collaboratively. For simplicity, the tasks considered in this embodiment can be divided into at most the sum of the number of member nodes within the clustering cluster sub-tasks;

[0116] The task offloading assignment constraint is

[0117]

[0118] In the formula: M represents the task splitter and M = {1, 2,..., m}; k l,m Represents the decision variable of the selection state of the splitting layer m in the task splitter and m ∈ M, k l,m = 1 means that the splitting layer m is selected, otherwise k l,m = 0; where the task offloading assignment constraint means that if the UAV as a communication relay has no computing resources, then it will be assigned a virtual task with a computing workload of 0;

[0119] Furthermore, the splitting layer vector of the task splitter used to partition the tasks to be offloaded is defined as k l ={k l,1 , k l,2 ,... k l,m};

[0120] According to the execution order of the tasks to be offloaded, the partitioned subtasks are reorganized into a subtask chain;

[0121] It should be noted that in this embodiment, since the task always starts from the CC, the applicable computing offloading strategy should always allocate the subtask segment at the start of the task to the CC. According to the task allocation strategy k l , each member of the clustering cluster can only hold one subtask. In this way, the member nodes within the clustering cluster can be transformed into a subtask chain according to the order of the subtasks. For simplicity, each member node within the clustering cluster can only forward the execution output of its own model to the previous or next member node in the subtask chain;

[0122] S33: According to the communication link relationship of the member nodes in the communication link connectivity graph , the subtask chain is sequentially allocated to different member nodes, and then the task distribution sequence of the clustering cluster is obtained.

[0123] In this embodiment, in order to better explain the relationship between the clustering strategy based on D2D communication and the task offloading and allocation strategy Figure 3 , an example is given to illustrate the feasible task allocation strategy in the given D2D communication clustering. As Figure 3 shown, there are 4 ACs and 1 CC in the clustering cluster, which means that the task must be divided into 5 subtask segments and then allocated to the member nodes of each clustering cluster. Since the subtask segments can be reorganized into a complete and correct task in a certain order, all cluster members can also form a subtask chain of length 5 in the same order. Then, the subtasks in the subtask chain are sequentially allocated to different member nodes to obtain the task distribution sequence of the clustering cluster;

[0124] S4: Construct a distributed task offloading network;

[0125] And the distributed task offloading network includes a global task distribution module, an intra-cluster task execution module, and a task completion and upload module;

[0126] In this embodiment, the task execution within the distributed task offloading network cluster consists of three main sub - phases: global task distribution, task execution process, and the cluster head collecting tasks, i.e., task completion and upload. When the TC receives the latest global task from the BS, it distributes task segments to other clients according to the given task distribution policy. Then, the CC can start the task execution process; execute according to different client orders and terminate at the last AC. After all clients update their local task segments, the cluster head collects and assembles the completed task segments of other clients;

[0127] In a specific embodiment, the global task distribution module is used to transmit subtasks in the task distribution sequence from the cluster head node in the clustering cluster to the member nodes through D2D communication to obtain the first task transmission delay;

[0128] Specifically, the expression for the obtained first task transmission delay is

[0129]

[0130] In the formula: t td,i represents the task allocation delay for transmitting subtasks in the task distribution sequence from the cluster head node in the clustering cluster to the member nodes through D2D communication; R i represents the corresponding D2D data transmission rate between the cluster head node CHi and the member node i, with the unit of bits per second; Q l,i represents the task size processed by the member node i;

[0131] The in - cluster task execution module is used to sequentially transmit subtasks to each member node based on the obtained subtask chain and perform task execution operations, and feedback the completed tasks after the execution operations to the corresponding cluster head nodes for collection to obtain the second task transmission delay;

[0132] Specifically, the expression for the obtained second task transmission delay is

[0133]

[0134]

[0135] In the formula: t TE represents an intermediate variable; represents the FP transmission delay from the member node i to the member node j; represents the intermediate data passed from the member node i to the member node j after the member node i completes local calculation; R i,j represents the transmission rate between the member node i and the member node j; represents the task execution delay of the member node i; C i represents the number of CPU cycles required to process one data sample; f iDenote the CPU computing frequency of member node i;

[0136] The task completion upload module is used to transmit the completed tasks collected by each cluster head node to the edge server TBS to obtain the third task transmission delay;

[0137] Specifically, after the task execution process ends in this embodiment, the cluster head requests the sub-task segments of other clients in the clustering cluster to be assembled into a complete task. Since the task collection process is almost the same as the in-cluster task distribution process (only in the opposite direction), the model collection delay from client j to client i, that is, the expression of the obtained third task transmission delay is

[0138] t tc,j =t td,i

[0139] In the formula: t tc,j Denote the task collection delay for transmitting the completed sub-tasks from the member nodes in the clustering cluster to the cluster head node through D2D communication.

[0140] This embodiment also includes Definition 1: For any feasible client clustering strategy of a certain cluster in, the corresponding in-cluster D2D communication topology should be a connected graph, denoted by

[0141] Proof 1: According to the above Definition 1 and the task offloading assignment constraints, for any given feasible clustering cluster l in, the corresponding undirected graph can be transformed into an undirected graph with a chain topology Under the feasible task offloading assignment strategy k l , all clients (member nodes) in the clustering cluster are accessed only once. Therefore, it can be concluded that the topological structure obtained by any feasible task offloading assignment strategy is a connected graph. Therefore, it can be further proved that any feasible clustering strategy obtained in the clustering cluster is also a connected graph; in addition, by sequentially merging all the sub-tasks of the clustering cluster l, the complete task of the cluster l is obtained, denoted by where Q l,i (0 ≤ i ≤ n l) is the subtask held by the i-th AC; i = 0 represents the subtask held by the CC. Therefore, for a global task, the task execution method of completing the tasks in order within the cluster and completing the tasks in parallel as a whole is adopted. If the cluster members hold different subtasks, when all clients transmit task segments to the TBS in parallel, it will lead to serious bandwidth conflicts. More importantly, this will also pose a huge challenge to the downlink task broadcast of the TBS. Therefore, considering that the UAV can quickly change its position to keep up with the changing position of the mobile device, this embodiment proposes to use the UAV as the specifically designated cluster head node CH. The cluster head node CH is responsible for the download and upload of global tasks, as well as the allocation and summary of tasks within the cluster; at the same time, the UAV can also adjust its own position to cope with the changing cluster topology and detect the quality of D2D communication to ensure the provision of the best topological link;

[0142] S5: In this embodiment, the global optimization of the USV clustering and task allocation strategy is carried out to minimize the total delay cost of L CCs. This problem can be expressed as a total delay cost minimization problem to construct a global delay cost objective model for distributed task offloading. Specifically, according to the first task transmission delay, the second task transmission delay, and the third task transmission delay obtained based on the distributed task offloading network, a global delay cost objective model for distributed task offloading is constructed;

[0143] Specifically, the constructed global delay cost objective model for distributed task offloading includes a global delay cost objective function and objective constraints, and its expression is

[0144]

[0145] In the formula: T represents the global delay cost objective function;

[0146] S6: According to the preset distributed inland river tasks to be offloaded, based on the optimal clustering cluster corresponding to minimizing the global delay cost objective model obtained in step S2, the distributed inland river task offloading is realized;

[0147] Specifically, it includes the following steps:

[0148] S61: Construct a member node clustering game mechanism based on graph reinforcement learning and exact potential game;

[0149] And according to the preset distributed inland river tasks to be offloaded, based on the optimal clustering cluster corresponding to minimizing the global delay cost objective model obtained in step S2;

[0150] And the constructed member node clustering game mechanism based on graph reinforcement learning and exact potential game specifically includes:

[0151] S611: In this embodiment, the main goal of each learning cluster in each clustering cluster is to find the optimal local policy that minimizes its own latency cost. Therefore, based on this greedy concept, the USV clustering problem can be designed as a multi-player clustering game, where L clusters are the players. In addition, the clustering strategy of each clustering cluster is represented by s l and the clustering strategy space is represented by S l . First, the UAV obtains the network topology of the in-domain data plane, the current link information, and the local computing capabilities of the USVs; then a graph neural network centered on D2D communication links (TRGNN) is constructed, as shown in Figure 4 , including a GNN layer and a DNN layer; and the GNN layer is used to capture the complex features in the topological structure of the unmanned ships in the inland waterway. The GNN layer takes the graph structure of the communication link connectivity graph as the input, shares information between nodes through the message passing mechanism, and finally obtains the embedding representations of each node and edge, and these embedding representations reflect the importance of the nodes in the entire UAV-assisted edge network and the load conditions of each link. The DNN layer is used to capture the D2D communication transmission rate requirements between the source node and the destination node in the communication link connectivity graph . By learning these traffic demand characteristics, the DNN layer can identify their distribution patterns and provide important information for path selection. Through the tight combination of the two modules, the optimization of path selection is jointly completed. The GNN encodes the path through topological information, while the DNN layer generates path selection probabilities according to the time-varying transmission rate. This combination can effectively establish the connection between the topological structure and traffic demand of the communication link connectivity graph in the UAV-assisted edge network, so as to make more effective decisions; according to the obtained embedding representations, the existing shared policy network is used to generate communication links for each computationally intensive (with a heavy computational burden) USV separately, so that the problem scale is reduced from the global path allocation of multiple computationally intensive USVs to the local path allocation of a single computationally intensive USV. The path allocation is to sequentially allocate the subtask chain to different member nodes, so that the learning process is easier to operate and feasible;

[0152] In this embodiment, through the framework of distributed game, the own latency of each computationally intensive USV and the impact of its behavior on the overall network state are considered simultaneously to achieve the exact potential game of computationally intensive USVs. Specifically, for each computationally intensive USV, a known clustering strategy s l and s l ∈S l , then the joint clustering strategy of L clustering clusters can be represented as s = {s1,..., s l ,... s L};

[0153] S612: Define the utility function of the clustering cluster for the global delay cost objective model, and the utility function U of the clustering cluster l (s l , s -l ) is expressed as

[0154]

[0155] In the formula: S -l represents other known clustering strategies in the preset clustering strategy pool except for the clustering cluster l; -ψ j (s j , s -j\l ) represents the utility of the clustering cluster j when the clustering cluster l does not take action; represents the change in the utility of other clustering clusters caused by the clustering strategy adopted by the clustering cluster l; -ψ l (s l , s -l ) represents the utility of the clustering cluster l when the clustering cluster j does not take action;

[0156] S613: Obtain the exact potential game mechanism EPG based on the potential function, and its expression is

[0157] U l (s l ', s -l ) - U l (s l , s -l ) = μ(s l ', s -l ) - μ(s l , s -l ),

[0158]

[0159] In the formula: μ(s l , s -l ) represents the potential function regarding the clustering strategy s l and the clustering strategy s -l ; U l (s' l , s -l ) represents the utility function of the clustering strategy s' l and the clustering strategy s -l ; s' l represents the changed clustering strategy of the clustering cluster l; μ(s' l , s -l ) represents the potential function regarding the clustering strategy s' l and the clustering strategy s -l ;

[0160] This embodiment also includes Theorem 1: The exact potential game mechanism based on the potential function is an EGP.

[0161] The proof is as follows: The potential function is equivalent to the sum of the benefits of all clustering clusters in the entire network. Assume that the decision of group l changes from s l to s l ', then the change in its utility function is as follows:

[0162]

[0163] Although the behavior of clustering cluster l has changed, when clustering cluster l abandons its decision while other users maintain their decisions (i.e., ψ j (s j , s -j\l ) = ψ j (s j , s' -j\l ), the impact of the missing behavior of clustering cluster l on other users is the same. Then, this embodiment can ignore ∑(-ψ j (s j , s -j ) + ψ j (s j , s -j\l )) and reformulate the above formula as:

[0164]

[0165] Based on this, this embodiment can draw the conclusion that when any user unilaterally changes its decision, the change in the personal utility function is the same as the change in the potential function. That is to say, the constructed multi-user computing task offloading game (exact potential game mechanism based on the potential function) is an EPG;

[0166] S62: Define the joint strategy of the clustering cluster, that is, the Nash equilibrium of the clustering game

[0167] Among them, before seeking the Nash equilibrium of the clustering game in this embodiment, it is necessary to verify whether the Nash equilibrium exists. Fortunately, since the exact potential game mechanism based on the potential function is an EPG, it can be guaranteed that there is at least one Nash equilibrium in the joint strategy of the clustering clusters. The original explanation for the existence of at least one Nash equilibrium in the EPG is a well-known technical means in the prior art and will not be elaborated here;

[0168] According to each clustering cluster based on the exact potential game mechanism EPG, obtain the one for making the utility function U l (s l , s -l)When the maximum value is reached, that is, when the target delay of the global delay cost objective model is minimized, the Nash equilibrium constraint conditions corresponding to the optimal clustering strategy are obtained;

[0169] And the expression of the Nash equilibrium constraint condition is

[0170]

[0171] In the formula: represents the utility when the clustering strategies s l and s -l reach equilibrium in the game; represents the utility when only the clustering strategy s -l reaches equilibrium in the game; represents the optimal solution of the clustering strategy s -l at equilibrium; represents the optimal solution of the clustering strategy s l at equilibrium;

[0172] S63: Based on the Nash equilibrium constraint conditions, obtain the optimal clustering strategy of the UAV-assisted edge network, and implement the distributed offloading task of the inland waterway according to the optimal clustering cluster. Among them, in this embodiment, after obtaining the optimal clustering strategy of the UAV-assisted edge network, the existing greedy-based heuristic task allocation algorithm is used for task allocation. Specifically, that is, tasks are allocated in a linear proportion to the computing capabilities of each USV within the cluster. Among them, the method of implementing task allocation by the greedy-based heuristic task allocation algorithm is a well-known technical means in the prior art, and will not be elaborated here too much. In order to effectively solve the task allocation problem in this embodiment, the greedy-based heuristic task allocation algorithm is adopted. The basic idea of this algorithm is to allocate the computing load to a client that is linearly proportional to the computing resources. It should be noted that since the UAVs in the clustering cluster do not undertake computing tasks, a virtual model segment with a computing amount of zero is allocated to them to ensure the feasibility of the task allocation strategy, as shown in Table 1.

[0173] Table 1. Specific process of the greedy-based heuristic task allocation algorithm

[0174]

[0175] In this embodiment, considering the actual communication condition limitations and actual application requirements of inland waterway channels, the potential game theory is used to model the problem of assisting USV clustering optimization as a distributed computing offloading game mechanism. The aim is to maximize the overall offloading benefit of the constructed distributed task offloading network on the basis of considering the dynamically changing D2D communication quality and the clustering cluster topology structure. In order to rationally utilize computing and communication resources, this invention constructs a global delay minimization problem, that is, a global delay cost objective model for distributed task offloading. By constructing a member node clustering game mechanism based on graph reinforcement learning and exact potential game, combined with the Nash equilibrium of the clustering game, the task allocation problem of the clustering cluster is solved to achieve global delay minimization, and then the optimal clustering strategy of the UAV-assisted edge network is obtained to realize the distributed offloading task of the inland waterway channel.

[0176] To verify the effectiveness of the algorithm proposed in this embodiment, computer simulation experiments are carried out using Python. We consider a fan-shaped area centered on the TBS, which represents the active range where the UAV can move freely. There are multiple ACs and CCs in it. Since the UAV acts as a CH, the number of UAVs is the same as that of CCs. The annular radius is 500m, so the coordinates of the TBS are (500, 0). The flight altitude of the UAV and the D2D communication distance threshold are set to 50m respectively. For the D2D communication link connection between client i and client j, the transmission power of client i is set to 24dBm in this embodiment, which is the same for all clients. At the same time, the bandwidth of the D2D communication link connection between client i and client j is set to 5MHz, and the power spectral density of the background noise N0 = -172dBm / Hz. For each computationally intensive USV, the global task data size follows a uniform distribution of [0.2, 5] Mbits. For all users, the number of CPU cycles required per bit follows a uniform distribution of [1500, 2000] cycles / bit. The computing resources of CCs and ACs are set to f c ∈[0.5, 2]×10 9 CPU cycles / s, f a ∈[6, 10]×10 9 CPU cycles / s. For the TBS, its transmission power P is 46dBm, the default communication bandwidth is 5MHz, and the computing resource f tbs =40×10 9 CPU cycles / s. For the proposed graph neural network TRGNN, the default setting is 6 layers. Its global model is initialized using a normal distribution. The learning rate lr of the policy network is set to 0.001, and the ADAM optimization algorithm is used for policy gradient descent. By default, the experiment uses a topology structure of 4 CCs and 8 ACs for simulation.

[0177] Figures 5 to 10The comparison results between the method of this embodiment and other existing algorithms respectively. Figures 5 to 6 This is the comparison between the method of this embodiment and two intra-cluster offloading mechanisms with different numbers of AC and CC. It can be seen that the total delay of the method of this embodiment is lower than that of the other two offloading mechanisms. Figures 7 to 8 This is the comparison between the method of this embodiment and two global offloading mechanisms with different numbers of AC and CC. It can be seen that the method of this embodiment still maintains good performance compared with the traditional offloading mechanism. Figures 9 to 10 This is the comparison between the method of this embodiment and two intra-cluster offloading mechanisms and two global offloading mechanisms under the change of data size and D2D bandwidth. In this case, it can be seen that the method of this embodiment continues to be superior to other offloading mechanisms.

[0178] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A clustering-based inland river distributed task offloading method, characterized in that: The specific steps include: S1: Use UAV as a communication relay for heterogeneous distributed USVs to establish line-of-sight connection between USV and edge server TBS and obtain UAV-assisted edge network; And the heterogeneous distributed USV includes a computing-intensive USV for providing task offloading / communication resources and an assisting USV for providing backup task offloading / communication resources for the computing-intensive USV; S2: The computationally intensive USV is used as the cluster head node of the cluster, and the assisting USV is used as the member node of the cluster; Based on the UAV-assisted edge network, the cluster head nodes and member nodes are randomly divided according to the task offloading clustering strategy and several clusters are obtained; S3: Define the communication link connectivity graph based on the clustering clusters; Based on the task offloading allocation strategy, the task to be offloaded is divided into subtask chains according to the link connectivity graph, and the subtasks in the subtask chain are allocated to each member node in the cluster to obtain the task distribution sequence; S4: Building a distributed task offloading network; And the distributed task offloading network includes a global task distribution module, an intra-cluster task execution module and a task completion upload module; The global task distribution module is used to transmit the subtasks in the task distribution sequence from the cluster head node in the cluster cluster to the member nodes through D2D communication to obtain the first task transmission delay; The intra-cluster task execution module is used to transmit the subtasks to each member node in turn based on the acquired subtask chain and perform task execution operations, and feed back the completed tasks after the execution operations to the corresponding cluster head nodes for collection, and obtain the second task transmission delay; The task completion upload module is used to transmit the completed tasks collected by each cluster head node to the edge server TBS to obtain the third task transmission delay; S5: constructing a global delay cost target model for distributed task offloading according to the first task transmission delay, the second task transmission delay, and the third task transmission delay obtained based on the distributed task offloading network; S6: According to the preset inland river distributed tasks to be offloaded, based on step S2, the optimal clustering corresponding to the target model of minimizing the global delay cost is obtained; And inland river distributed task offloading is achieved based on the optimal clustering cluster.

2. According to claim 1, a clustering-based inland river distributed task offloading method is characterized in that: The S2 specifically includes the following steps: S21: The computationally intensive USV is used as the cluster head node of the cluster, denoted as in represents the set of clusters; L represents the number of computationally intensive USVs in the UAV-assisted edge network; The assisting USV is taken as the member node of the cluster, denoted as in represents the assisting USV set; H represents the number of assisting USVs in the UAV-assisted edge network; S22: Based on the task offloading clustering strategy, randomly divide and obtain several clusters according to the cluster head nodes and member nodes to define the member node set of cluster l and Where n l Represents the number of member nodes in cluster l; And the task offloading clustering strategy is specifically: Assume that the number of clusters is l and It is defined that each cluster has and can only have one cluster head node, and each assisting USV can only join one cluster; And each assisting USV has and can only join one cluster with the associated constraint: Where: s l,h represents the binary decision variable that assists USVh as a member node of cluster l, and s l,h =1 means that the assisting USVh is a member node of cluster l, otherwise s l,h =0; Then the member node clustering vector of cluster l is defined as S l ={s l,1 ,s l,2 ,...,s l,H }.

3. The clustering-based inland river distributed task offloading method according to claim 1 is characterized in that: The S3 specifically includes the following steps: S31: Define the communication link connectivity graph based on clustering and Where V represents the member node; E represents the edge between the member nodes, i.e., the communication link generated by the wireless connection; c represents the allocation constraint of the communication link; S32: Based on the task offloading allocation strategy, the task to be offloaded is divided into subtask chains according to the connectivity graph, and each member node in the cluster is allocated to obtain a task distribution sequence; The task offloading allocation strategy is specifically as follows: according to the number of each member node in the clustering cluster, based on the task offloading allocation constraint, the task to be offloaded is divided into subtasks corresponding to the number of each member node in the clustering cluster; The task offloading allocation constraint is Where: M represents the task splitter and M = {1, 2, ..., m}; k l,m represents the state decision variable selected by the segmentation layer m in the task segmentor and m∈M, k l,m =1 means segmentation layer m is selected, otherwise k l,m =0; Then, the partitioning layer vector of the task partitioner used to partition the offloaded tasks is defined as k l ={k l,1 ,k l,2 ,...k l,m }; According to the execution order of the tasks to be offloaded, the divided subtasks are reorganized into subtask chains; S33: According to the communication link connectivity diagram The communication link relationship of the member nodes in the cluster is used to distribute the subtask chains to different member nodes in turn, thereby obtaining the task distribution sequence of the cluster.

4. The clustering-based inland river distributed task offloading method according to claim 1 is characterized in that: The expression of the first task transmission delay obtained in S4 is: Where: t td,i It represents the task allocation delay of transmitting the subtasks in the task distribution sequence from the cluster head node to the member nodes in the cluster through D2D communication; R i represents the corresponding D2D data transmission rate between the cluster head node CHi and the member node i; Q l,i Indicates the size of the task processed by member node i; The expression of the second task transmission delay obtained is: Where: t TE represents an intermediate variable; represents the FP transmission delay from member node i to member node j; Represents the intermediate data that member node i passes to member node j after completing local calculation; R i,j represents the transmission rate between member node i and member node j; represents the task execution delay of member node i; C i Indicates the number of CPU cycles required to process a data sample; f i Indicates the CPU computing frequency of member node i; The expression of the third task transmission delay obtained is: t tc,j =t td,i Where: t tc,j It represents the task collection delay of transmitting the completed subtask from the member nodes in the cluster to the cluster head node through D2D communication.

5. The clustering-based inland river distributed task offloading method according to claim 4 is characterized in that: The global delay cost objective model of distributed task offloading constructed in S5 includes the global delay cost objective function and objective constraints, and its expression is: Where: T represents the global delay cost objective function.

6. The clustering-based inland river distributed task offloading method according to claim 5 is characterized in that: The S6 specifically comprises the following steps: S61: Constructing a member node clustering game mechanism based on graph reinforcement learning and precise potential game; And according to the preset inland river distributed tasks to be unloaded, based on step S2, the optimal clustering cluster corresponding to the objective model of minimizing the global delay cost is obtained; The member node clustering game mechanism based on graph reinforcement learning and precise potential game is constructed, including: S611: Given each computationally intensive USV a known clustering strategy s l And s l ∈S l , then the joint clustering strategy of L clusters can be expressed as s = {s1,...,s l ,...s L }; S612: Define the utility function of the clustering cluster for the global delay cost target model, and the utility function U of the clustering cluster l (s l ,s -l ) is expressed as Where: S -l Indicates other known clustering strategies in the preset clustering strategy pool except cluster cluster l; -ψ j (s j ,s -j\l ) represents the utility of cluster j when cluster l does not take action; It represents the change in the utility of other clusters due to the clustering strategy adopted by cluster l; -ψ l (s l ,s -l ) represents the utility of cluster l when cluster j does not take action; S613: Obtain the precise potential game mechanism EPG based on the potential function, which is expressed as U l (s′ l ,s -l )-U l (s l ,s -l )=μ(s′ l ,s -l )-μ(s l ,s -l ), Where: μ(s l ,s -l ) represents the clustering strategy s l With clustering strategies -l Potential function; U l (s' l ,s -l ) represents the clustering strategy s' l With clustering strategies -l The utility function of s' l represents the clustering strategy of cluster l; μ(s' l ,s -l ) represents the clustering strategy s' l With clustering strategies -l The potential function of S62: Define the joint strategy of clusters, i.e. the Nash equilibrium of clustering games Based on the precise potential game mechanism EPG of each cluster, the utility function U is obtained. l (s l ,s -l ) reaches the maximum value, that is, when the target delay of the global delay cost target model is minimized, the Nash equilibrium constraint condition corresponding to the optimal clustering strategy is obtained; And the expression of Nash equilibrium constraint is Where: Represents the clustering strategy s l With clustering strategies -l The utility when the game between the two reaches equilibrium; Indicates that only clustering strategy s -l The utility achieved when the game reaches equilibrium; Represents the clustering strategy s when equilibrium is reached -l The optimal solution of Represents the clustering strategy s when equilibrium is reached l The optimal solution of S63: Based on the Nash equilibrium constraints, the optimal clustering strategy of the UAV-assisted edge network is obtained, and the distributed unloading task of the inland waterway is realized according to the optimal clustering cluster.

Images