A digital twin edge collaborative quasi-static deployment method and system based on vehicle mobility locality characteristics

Abstract

The present invention discloses a method for quasi-static deployment of digital twins based on the local characteristics of vehicle mobility. The method includes the following steps: obtaining a gridded map of the area to be deployed, performing a first clustering of the grids in the gridded map to obtain K final first clusters; constructing a target model to obtain an optimization target; selecting one final first cluster as the final second cluster; performing a second clustering of the grids in the final second cluster to obtain multiple final third clusters; generating a server allocation plan corresponding to each final third cluster using simulated annealing and particle swarm optimization algorithms based on the optimization target; selecting another final first cluster as the final second cluster; repeating until all final first clusters are selected; and integrating all server allocation plans to obtain a final server allocation plan. This method can effectively deploy digital twins with low resource overhead and configuration cost.

Description

Technical Field

[0001] The present invention relates to the field of digital twins in Internet of Vehicles (IoV), and more specifically, to a quasi-static deployment method for digital twins based on the local characteristics of vehicle mobility. Background Art

[0002] The Internet of Vehicles (IoV) forms a highly intelligent and automated transportation network by interconnecting and sharing information across multiple systems, including vehicles, transportation infrastructure, smart devices, and cloud platforms. Digital twins are a core enabling technology for the digital and intelligent development of IoV smart transportation. Numerous IoV digital twins will be deployed on network edge servers, forming IoV edge digital twins. Digital twins use modeling tools to construct high-fidelity digital models of physical entities in the IoV, such as vehicles, pedestrians, transportation infrastructure, and communication networks. These digital twins maintain real-time interaction with the physical entities in the real environment, while using virtual models to simulate the real-world behavior of the physical entities. Assisted by central cloud servers, IoV edge digital twins collaborate on autonomous vehicle task calculations, providing digital twin services such as early warning, navigation, traffic flow monitoring, and network optimization, furthering the intelligent connectivity and digitalization of vehicles.

[0003] Because vehicles are constantly in motion and the road conditions they encounter are constantly changing, digital twins must also constantly adapt to these changes. However, due to the large size of digital twins, frequent switching back and forth causes task migration and network scheduling issues, which in turn increases configuration costs. Therefore, digital twins are often deployed across multiple edge servers to reduce switching costs. However, traditional deployment methods are not efficient for deploying digital twins, resulting in high resource consumption and configuration costs.

[0004] The prior art discloses a method for deploying edge digital twins based on transfer learning, including the following steps: using the KM algorithm to establish a match between the edge servers of a new factory and the edge servers of an existing factory, and using the simulated annealing algorithm to establish a match between the edge servers in the new factory and the IIoT device m; each edge server in the new factory is matched with an edge server in an existing factory; the source domain model in the edge server of the existing factory is transferred to the corresponding edge server in the new factory; the edge server in the new factory migrates the received source domain model to the matching IIoT device m, building a source domain model for IIoT device m; within the new factory, the edge server uses the source domain model and combines it with the data uploaded by IIoT device m to build a digital twin for the matching IIoT device m. This method cannot effectively deploy IoV digital twins based on the characteristics of vehicle operation. Summary of the Invention

[0005] In response to the defects of digital twins in the existing technology, such as low efficiency, large resource overhead and high configuration cost, the present invention provides a quasi-static deployment method of digital twins based on the local characteristics of vehicle mobility. This method can effectively deploy digital twins with low resource overhead and low configuration cost.

[0006] The primary purpose of the present invention is to solve the above technical problems, and the technical solutions of the present invention are as follows:

[0007] A quasi-static deployment method for digital twin edge collaboration based on the local characteristics of vehicle movement, comprising:

[0008] S1: Obtain a rasterized map of the area to be deployed, perform a first clustering on the grids in the rasterized map, and obtain K final first clusters;

[0009] S2: Build an optimization target model for the IoV digital twin and obtain the optimization target for digital twin deployment.

[0010] S3: Select a final first cluster as the final second cluster;

[0011] S4: performing a second clustering on the grids in the final second cluster to obtain multiple final third clusters;

[0012] S5: Generate a server allocation plan corresponding to each final third cluster using simulated annealing and particle swarm optimization algorithms according to the digital twin deployment optimization goal;

[0013] S6: Select another final first cluster as the final second cluster; repeat steps S4 to S5 until all final first clusters are selected;

[0014] S7: Integrate all server allocation plans to obtain the final server allocation plan.

[0015] Furthermore, the first clustering includes:

[0016] S101: Obtain a rasterized map and use it as a cluster to form a cluster set;

[0017] S102: Selecting the cluster with the largest diameter in the cluster set as the second cluster;

[0018] S103: Calculating the average dissimilarity between each grid in the second cluster and other grids in the second cluster as a first average dissimilarity; and sorting the first average dissimilarity to obtain the grid with the largest first average dissimilarity as a new first grid;

[0019] S104: Calculating the dissimilarity between each grid in the second cluster and the first grid as a second dissimilarity;

[0020] S105: comparing the first average dissimilarity and the second dissimilarity of each grid in the second cluster; if the first average dissimilarity is greater than or equal to the second dissimilarity, adding the grid to the fourth cluster; if the first average dissimilarity is less than the second dissimilarity, adding the grid to the third cluster;

[0021] S106: replacing the second cluster in the cluster set with the third cluster, and adding the fourth cluster as a new first cluster to the cluster set;

[0022] S107: Create two empty clusters as the new third cluster and the new fourth cluster respectively; repeat steps S102 to S106 until the number of clusters in the cluster set is K, and take all clusters in the cluster set as the final first cluster.

[0023] Furthermore, the calculation formula of the diameter of the cluster is as follows:

[0024]

[0025] DIS(i,j)=ε1D nom (i,j)+ε2TR mom (i,j)

[0026] ε1 and ε2 represent weight factors, ε1+ε2=1, D nom (i,j) represents the geographical distance from grid i to grid j, TR mom (i, j) represents the number of switching digital twins required for the vehicle to move between grid i and grid j, and i and j represent sequence numbers.

[0027] Furthermore, the dissimilarity is the Euclidean distance between two grids.

[0028] Furthermore, the second clustering includes:

[0029] S301: Selecting grids in the final first cluster to construct a grid sparse graph, forming a grid sparse graph matrix; wherein the nodes of the sparse graph represent grids; and the edges of the sparse graph represent the traffic intensity distance between grids;

[0030] S302: Inputting the grid sparse graph matrix and the corresponding grid attributes into a deep feature model to obtain grid deep features;

[0031] S303: performing soft clustering on the grid deep-level features to obtain multiple final third clusters.

[0032] Furthermore, the deep feature model includes: a self-loop layer, a linear layer, a first GCN unit, and a second GCN unit;

[0033] The grid sparse graph matrix is ​​input into the input end of the ring layer, and the grid attributes are input into the input end of the first GCN unit; the output end of the second GCN unit outputs the classification result;

[0034] The output end of the self-loop layer is connected to the input end of the linear layer, the output end of the linear layer is connected to the input end of the first GCN unit and the input end of the second GCN unit, and the output end of the first GCN unit is connected to the input end of the second GCN unit.

[0035] Furthermore, the first GCN unit includes a first hidden layer, a first fully connected layer, and a first activation layer connected in sequence;

[0036] The grid attributes and the output end of the linear layer are input to the input end of the first hidden layer, and the output end of the first activation layer is connected to the input end of the second GCN unit.

[0037] Furthermore, the second GCN unit includes a second hidden layer, a second fully connected layer, and a second activation layer connected in sequence;

[0038] The output end of the first activation layer and the output end of the linear layer are connected to the input end of the second hidden layer, and the second activation layer outputs the classification result.

[0039] Furthermore, the digital twin deployment optimization objectives are as follows:

[0040]

[0041] M is the unit deployment benefit, is the coefficient that balances the load balancing LO(A, L) and the virtual-real interaction delay DE(A, L), ρ represents the cost of a single switch, and μ is the cost of a single deployment; i,k =1 means that the i-th grid is assigned to the k-th final third cluster, i, k, q represent the sequence number, l k,q Indicates that the kth final third cluster is assigned to the qth server, Q is the total number of servers, K is the total number of final third clusters, N is the total number of grids, and w i (t) represents the load of the i-th grid from t-1 to t, b k,q represents the bandwidth resource allocated by the qth server to the kth final third cluster, B q represents the total bandwidth resources of the qth server, F q represents the total computing power resources of the qth server, f i represents the computing power resources required for the i-th grid, and A and L represent the obtained deployment utility optimization solution.

[0042] A digital twin edge collaborative quasi-static deployment system based on the local characteristics of vehicle movement, comprising:

[0043] The first clustering module: obtains a rasterized map of the area to be deployed, performs a first clustering on the grids in the rasterized map, and obtains K final first clusters;

[0044] Model building module: Builds the IoV digital twin optimization target model and obtains the digital twin deployment optimization target;

[0045] Cluster selection module: selects a final first cluster as the final second cluster;

[0046] A second clustering module: performing a second clustering on the grids in the final second cluster to obtain a plurality of final third clusters;

[0047] Allocation plan generation module: Based on the digital twin deployment optimization goal, the simulated annealing and particle swarm optimization algorithms are used to generate the server allocation plan corresponding to each final third cluster;

[0048] The third clustering module: select another final first cluster as the final second cluster; return to the second clustering module until all final first clusters are selected;

[0049] Allocation plan integration module: integrates all server allocation plans to obtain the final server allocation plan.

[0050] Compared with the prior art, the present invention has the following beneficial effects:

[0051] The present invention decomposes the global map data into multiple local areas through the first clustering, further simplifying and optimizing the data processing process; through the second clustering, the complex spatial relationship between geographical locations is effectively mined; the use of simulated annealing and particle swarm optimization algorithms can comprehensively consider the optimization problem composed of virtual and real interaction delay, load balancing, number of switching times and deployment cost. At the same time, the global search capability of simulated annealing and the fast local search advantage of particle swarm are combined to obtain the optimal solution. This hybrid optimization algorithm effectively avoids the problem of falling into the local optimal solution and realizes the optimal configuration of resources. In summary, the deployment method of the present invention can effectively deploy digital twins with low resource overhead and low configuration cost. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 A flowchart of a digital twin edge collaborative quasi-static deployment method based on the local characteristics of vehicle movement provided in Example 1.

[0053] Figure 2 This is a flowchart of the first clustering provided in Example 1.

[0054] Figure 3 This is a flowchart of the second clustering provided in Example 1.

[0055] Figure 4 This is a structural diagram of the deep feature model provided in Example 1.

[0056] Figure 5 This is a structural diagram of the first GCN unit provided in Example 1.

[0057] Figure 6 This is a structural diagram of the second GCN unit provided in Example 1.

[0058] Figure 7 Schematic diagram of the principle of the deep feature model provided in Example 1.

[0059] Figure 8 This is a schematic diagram of the principle of the second clustering provided in Example 1.

[0060] Figure 9 This is a schematic diagram of the server allocation solution application provided in Example 1. DETAILED DESCRIPTION

[0061] The accompanying drawings are for illustrative purposes only and are not to be construed as limiting this patent;

[0062] In order to better illustrate this embodiment, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product size;

[0063] It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0064] The technical solution of the present invention is further described below with reference to the accompanying drawings and embodiments.

[0065] Example 1

[0066] like Figure 1 As shown, a digital twin edge collaborative quasi-static deployment method based on the local characteristics of vehicle movement includes:

[0067] S1: Obtain a rasterized map of the area to be deployed, perform a first clustering on the grids in the rasterized map, and obtain K final first clusters;

[0068] S2: Build an optimization target model for the IoV digital twin and obtain the optimization target for digital twin deployment.

[0069] S3: Select a final first cluster as the final second cluster;

[0070] S4: performing a second clustering on the grids in the final second cluster to obtain multiple final third clusters;

[0071] S5: Generate a server allocation plan corresponding to each final third cluster using simulated annealing and particle swarm optimization algorithms according to the digital twin deployment optimization goal;

[0072] S6: Select another final first cluster as the final second cluster; repeat steps S4 to S5 until all final first clusters are selected;

[0073] S7: Integrate all server allocation plans to obtain the final server allocation plan.

[0074] Furthermore, if Figure 2 As shown, the first clustering includes:

[0075] S101: Obtain a rasterized map and use it as a cluster to form a cluster set;

[0076] S102: Selecting the cluster with the largest diameter in the cluster set as the second cluster;

[0077] S103: Calculating the average dissimilarity between each grid in the second cluster and other grids in the second cluster as a first average dissimilarity; and sorting the first average dissimilarity to obtain the grid with the largest first average dissimilarity as a new first grid;

[0078] S104: Calculating the dissimilarity between each grid in the second cluster and the first grid as a second dissimilarity;

[0079] S105: comparing the first average dissimilarity and the second dissimilarity of each grid in the second cluster; if the first average dissimilarity is greater than or equal to the second dissimilarity, adding the grid to the fourth cluster; if the first average dissimilarity is less than the second dissimilarity, adding the grid to the third cluster;

[0080] S106: replacing the second cluster in the cluster set with the third cluster, and adding the fourth cluster as a new first cluster to the cluster set;

[0081] S107: Create two empty clusters as the new third cluster and the new fourth cluster respectively; repeat steps S102 to S106 until the number of clusters in the cluster set is K, and take all clusters in the cluster set as the final first cluster.

[0082] It should be noted that the grid represents a map grid, a grid-based map representation method. The clusters in the first clustering process represent a collection of grids. The first clustering process decomposes the global map into multiple local regions, transforming a global problem into multiple local problems.

[0083] It's important to note that the global map's data complexity is high, and a single data processing model is difficult to adapt to diverse analytical needs, such as traffic flow control and accident prevention. This limits the system's response speed and accuracy. Decomposing the global map using hierarchical recursive clustering can effectively address this problem. Dividing the global map into multiple, relatively independently managed local regions not only reduces the complexity of individual processing tasks but also improves data processing efficiency. Each local region can be analyzed and managed specifically based on its unique geographic and traffic characteristics, resulting in a more precise response to local conditions.

[0084] Furthermore, the calculation formula of the diameter of the cluster is as follows:

[0085]

[0086] DIS(i,j)=ε1D nom (i,j)+ε2TR mom (i,j)

[0087] ε1 and ε2 represent weight factors, ε1+ε2=1, D nom (i,j) represents the geographical distance from grid i to grid j, TR mom (i, j) represents the number of switching digital twins required for the vehicle to move between grid i and grid j, and i and j represent sequence numbers.

[0088] Furthermore, the dissimilarity is the Euclidean distance between two grids.

[0089] Furthermore, if Figure 3 As shown, the second clustering includes:

[0090] S301: Selecting grids in the final first cluster to construct a grid sparse graph, forming a grid sparse graph matrix; wherein the nodes of the sparse graph represent grids; and the edges of the sparse graph represent the traffic intensity distance between grids;

[0091] S302: Inputting the grid sparse graph matrix and the corresponding grid attributes into a deep feature model to obtain grid deep features;

[0092] S303: performing soft clustering on the grid deep-level features to obtain multiple final third clusters.

[0093] In a specific embodiment, the soft clustering adopts the soft-kmeans algorithm.

[0094] Furthermore, if Figure 4 As shown, the deep feature model includes: a self-loop layer, a linear layer, a first GCN unit, and a second GCN unit;

[0095] The grid sparse graph matrix is ​​input into the input end of the ring layer, and the grid attributes are input into the input end of the first GCN unit; the output end of the second GCN unit outputs the classification result;

[0096] The output end of the self-loop layer is connected to the input end of the linear layer, the output end of the linear layer is connected to the input end of the first GCN unit and the input end of the second GCN unit, and the output end of the first GCN unit is connected to the input end of the second GCN unit.

[0097] Furthermore, if Figure 5 As shown, the first GCN unit includes a first hidden layer, a first fully connected layer, and a first activation layer connected in sequence;

[0098] The grid attributes and the output end of the linear layer are input to the input end of the first hidden layer, and the output end of the first activation layer is connected to the input end of the second GCN unit.

[0099] Furthermore, if Figure 6 As shown, the second GCN unit includes a second hidden layer, a second fully connected layer, and a second activation layer connected in sequence;

[0100] The output end of the first activation layer and the output end of the linear layer are connected to the input end of the second hidden layer, and the second activation layer outputs the classification result.

[0101] It should be noted that the application of GCN units not only improves the accuracy of data analysis but also makes local regional features more distinct, facilitating accurate clustering analysis. A topological graph is constructed with grids as vertices and traffic flow intensity distances as edges. This topological graph is then fed into the GCN neural network for feature extraction, mining deeper relationships between grids to obtain node embedding representations. Traditional clustering often fails to fully exploit the complex spatial relationships and topological structures between geographic locations, resulting in limited accuracy in traffic flow analysis. These methods typically focus on the independent attributes of points rather than their interactions, making the analysis results unable to reflect actual traffic conditions and the dynamic changes in the geographical environment. To address this, a graph convolutional network (GCN) is used to process local regional data. By leveraging the topological relationships between grids, GCN can effectively capture and learn traffic flows and spatial dependencies between grids, thereby improving the accuracy and efficiency of data analysis and obtaining a reasonably distributed clustering result.

[0102] like Figure 7 As shown, the GCN unit can be expressed by the following formula:

[0103]

[0104]

[0105] is a matrix with self-connection, I N is the identity matrix, for The degree matrix of W (0) 、W (1) denote the weight matrices of the first and second layers in the normalization layer, respectively. σ(·) and ReLU(·) are activation functions. Each GCN unit multiplies the adjacency matrix M by the feature matrix H to obtain a summary of the features of each vertex's neighbors. This is then multiplied by a parameter matrix W(l), and a nonlinear transformation is performed with the activation function σ to obtain the matrix H(l+1) that aggregates the features of adjacent vertices. This matrix is ​​the desired clustering matrix A.

[0106] The training uses modularity as the loss function. The specific formula is as follows:

[0107]

[0108] M∈R N×N is the adjacency matrix of the graph, which represents the actual connections between nodes in the graph, M ij represents the actual connection weight between nodes i and j. represents the connection probability between nodes in a random graph, Indicates the difference between actual connections and random connections. m is the total number of edges in the graph, i.e. m = ∑M ij . k is the node degree vector, k i =∑ j M ij , represents the degree of node i, that is, the sum of the number of connected edges. Z∈R N×k is the community distribution matrix of the node, that is, the output feature of GCN, Z i represents the distribution of node i in k clusters. Tr(·) is a trace operation used to extract the overall characteristics of the matrix.

[0109] Construct the difference matrix between actual connections and random connections This difference matrix is ​​weightedly projected through the node's community distribution matrix Z, that is, calculated Thus, the connection quality contribution of the node in each community is obtained. Through the trace operation Tr(·), the sum of the connection differences within all communities is extracted to reflect the overall difference between the current community division and random connection. Finally, through the normalization term The results are normalized to ensure that the value of modularity Q is within a reasonable range. The training goal is to maximize modularity Q so that the actual connection density within the graph is as high as possible, while the connections between communities are as sparse as possible, thereby obtaining a reasonable partitioning structure.

[0110] The present invention performs training with the goal of maximizing the modularity function, which is a criterion for measuring whether the community division is reasonable.

[0111] Furthermore, the digital twin deployment optimization objectives are as follows:

[0112]

[0113]

[0114] M is the unit deployment benefit, is the coefficient that balances the load balancing LO(A, L) and the virtual-real interaction delay DE(A, L), ρ represents the cost of a single switch, and μ is the cost of a single deployment; i,k =1 means that the i-th grid is assigned to the k-th final third cluster, i, k, q represent the sequence number, l k,q Indicates that the kth final third cluster is assigned to the qth server, Q is the total number of servers, K is the total number of final third clusters, N is the total number of grids, and w i (t) represents the load of the i-th grid from t-1 to t, b k,q represents the bandwidth resource allocated by the qth server to the kth final third cluster, B q represents the total bandwidth resources of the qth server, F q represents the total computing power resources of the qth server, f i represents the computing power resources required for the i-th grid, and A and L represent the obtained deployment utility optimization solution.

[0115] It should be noted that constraint C1 specifies unique grid affiliation, meaning any grid can only be assigned to one cluster. Constraint C2 specifies unique matching of the final third cluster, meaning any final third cluster can only be matched to one edge server. Constraint C3 indicates that an edge server can be matched to multiple final third clusters, provided its bandwidth and computing power resources allow. Constraints C4 and C5 represent the bandwidth and computing power resource constraints of edge server g, respectively. Constraint C6 specifies the type of optimization solution, meaning that all optimization solutions are integers of 0 or 1.

[0116] It should be noted that the optimization objectives of digital twin deployment are divided into virtual-real interaction delay, load balancing, number of switching times, and deployment cost;

[0117] It should be noted that the virtual-real interaction delay can be deduced from this, assuming that the computing capacity of the edge server meets the maximum workload, thus ignoring the processing delay and focusing on the transmission delay and computing delay. i The distance to the edge server q is d i,q , then the grid g within the cluster i The channel gain with edge server q is:

[0118]

[0119] Where β is the fixed transmission loss.

[0120] B q It represents the bandwidth resource set provided by BS or RSU for virtual-real interaction within the map range. Assuming that the entire map can be covered, b is used to represent the bandwidth required for communication with the edge server q. The grid g within the cluster i The communication rate of uploading data to the edge server q is:

[0121]

[0122] where p q represents the transmission power of the edge server q, and σ represents the noise power.

[0123] In-cluster grid g i The amount of data is D i , then the grid g within the cluster i The transmission delay of data uploaded to edge server q is:

[0124]

[0125] where w i (t) is the grid g i The load of the grid g i Number of virtual-real interactions between vehicle users

[0126] The total transmission delay is the sum of the transmission delays of uploading the grid data in each cluster in the map to the edge server:

[0127]

[0128] where l k,q Indicates the placement of the digital twin of the grid vehicle within the cluster, a i,k Represents the raster clustering result.

[0129] The computing power required per unit load is f i , the computing power of edge server q is F q , the computing delay measured by the edge server computing resources is:

[0130]

[0131] The total computing latency is the sum of the computing latency measured by the computing resources of each edge server in the map:

[0132]

[0133] In summary, the virtual-reality interaction delay of the system can be expressed as:

[0134] DE(A,L)=T up +T per

[0135] It should be noted that load balancing is introduced from this.

[0136] The load balancing index of the system based on Jain fairness is as follows. The goal is to evaluate the load balancing in the system by considering the placement and connectivity of the digital twins of grid vehicles in different clusters:

[0137]

[0138] in

[0139]

[0140] |Q| represents the cardinality of set Q, i.e. the number of edge servers Q; U Q The total load on the edge server Q is a weighted sum of the connectivity of the vehicle digital twins in each grid within the cluster on that server. By optimizing this index, the system configuration can be adjusted to make resource allocation more equitable, thereby improving the performance and efficiency of the entire system.

[0141] It should be noted that the number of switching times is derived from this.

[0142] In the trace, the total workload of all clusters is constant for a given duration, so the maintenance cost can be simply expressed as the migration cost, i.e., the number of switches. The goal is stated as:

[0143]

[0144] in

[0145]

[0146] tr k,k′ (t) represents cluster C k To cluster C k′ The number of switches between time t and time t-1, tr i,j (t) represents the grid g in this time period i To grid g jThe number of switching times.

[0147] It should be noted that the deployment cost is derived from this.

[0148] Considering the peak total load of each cluster in the time dimension, the maximum workload of each cluster in a given time period is:

[0149]

[0150] Bandwidth and computing power are considered as part of the cost because they are used in cluster C at the edge server. k In the process of communication, it involves the use of communication and computing resources.

[0151] Use edge server q in cluster C k The bandwidth and computing power generated in are expressed as:

[0152]

[0153] The total deployment cost is the total cost of bandwidth and computing power required for the peak total load of each cluster in the time dimension:

[0154]

[0155] In a specific embodiment, the generation process of the simulated annealing and particle swarm optimization algorithm includes:

[0156] S501: Define the objective function f(x), set the particle swarm size N, determine the maximum number of iterations max_iter, and initialize the particle position x i and speed v i Set the inertia weight w, learning factors c1 and c2, initial temperature T0, cooling rate α, and minimum temperature T min .

[0157] S502: For each particle i, calculate the fitness f(x i ).

[0158] If this fitness f(x i ) is better than the individual optimal fitness f(p besti ), then update the individual's optimal position to the current position p besti is x i ;

[0159] If the individual's optimal fitness f(p besti ) is better than the global optimal fitness f(g best ), then update the global optimal position g best For p vesti ;

[0160] S503: Update the particle's velocity v i and position x i .

[0161] S504: Generate new solution x new By perturbing x i , calculate the fitness of the new solution f(x new ). If f(x new ) is better than f(x i ), directly accept the new solution, otherwise calculate the energy difference ΔE=f(x new )-f(x i ) and accept the new solution with probability P = exp(-ΔE / T).

[0162] S505: If the temperature T is lower than the minimum temperature T min , then terminate the iteration and return the global optimal solution g best And its objective function value f(g best ).

[0163] It's important to note that traditional matching algorithms don't effectively handle the matching between the final third cluster and the edge server. Using simulated annealing and particle swarm optimization to find the optimal cluster-edge server match can optimize resource allocation and improve service response speed and efficiency. This optimization algorithm combines the advantages of global and local search, effectively avoiding local optimal solutions and finding more reasonable matching solutions.

[0164] In one specific embodiment, simulated annealing and particle swarm optimization algorithms are used to generate a server allocation plan corresponding to each final third cluster. The final third cluster can only be matched with one server, but a server can be matched with multiple clusters if bandwidth resources and computing power allow. The clustering matrix, along with simulated annealing and particle swarm optimization algorithms, continuously search for the optimal solution, achieving a stable and ideal match between clusters and servers. Digital twins are placed on the corresponding servers, and vehicle users can interact with and synchronize updates with the digital twins on the servers through nearby road test units.

[0165] A digital twin edge collaborative quasi-static deployment system based on the local characteristics of vehicle movement, comprising:

[0166] The first clustering module: obtains a rasterized map of the area to be deployed, performs a first clustering on the grids in the rasterized map, and obtains K final first clusters;

[0167] Model building module: Builds the IoV digital twin optimization target model and obtains the digital twin deployment optimization target;

[0168] Cluster selection module: selects a final first cluster as the final second cluster;

[0169] A second clustering module: performing a second clustering on the grids in the final second cluster to obtain a plurality of final third clusters;

[0170] Allocation plan generation module: Based on the digital twin deployment optimization goal, the simulated annealing and particle swarm optimization algorithms are used to generate the server allocation plan corresponding to each final third cluster;

[0171] The third clustering module: select another final first cluster as the final second cluster; return to the second clustering module until all final first clusters are selected;

[0172] Allocation plan integration module: integrates all server allocation plans to obtain the final server allocation plan.

[0173] like Figure 8 、 Figure 9 As shown in the figure, the global map is decomposed into multiple local regions after the first clustering step. For each local region, a topological graph is constructed with grid cells as vertices and the traffic flow intensity distances between grid cells as edges. This graph is then fed into the GCN neural network model for training, aiming to uncover deeper relationships between grid cells. After training the neural network model, a feature matrix of aggregated adjacent nodes is obtained. The learned node embedding representations are clustered using the Soft-Kmeans algorithm to obtain clustering results and a cluster matrix. Using simulated annealing and particle swarm optimization algorithms, an optimal solution for the deployment of clusters and edge servers is obtained—the best matching result, which maximizes the deployment utility. Each cluster is associated with its corresponding edge server. The cluster places the grid vehicle digital twin on the corresponding edge server. Grid vehicle users in the cluster interact with and synchronize updates with the digital twin on the edge server through adjacent road test units.

[0174] The same or similar reference numerals correspond to the same or similar components;

[0175] The terms used in the drawings to describe positional relationships are for illustrative purposes only and should not be construed as limiting this patent;

[0176] Obviously, the above embodiments of the present invention are merely examples for the purpose of clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. Those skilled in the art will appreciate that other variations or modifications can be made based on the above description. It is not necessary and impossible to enumerate all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the claims of the present invention.

Claims

1. A digital twin edge collaborative quasi-static deployment method based on the local characteristics of vehicle movement, characterized by: include: S1: Obtain a rasterized map of the area to be deployed, perform a first clustering on the grids in the rasterized map, and obtain K final first clusters; S2: Build an optimization target model for the IoV digital twin and obtain the optimization target for digital twin deployment. S3: Select a final first cluster as the final second cluster; S4: performing a second clustering on the grids in the final second cluster to obtain multiple final third clusters; S5: Generate a server allocation plan corresponding to each final third cluster using simulated annealing and particle swarm optimization algorithms according to the digital twin deployment optimization goal; S6: Select another final first cluster as the final second cluster; repeat steps S4 to S5 until all final first clusters are selected; S7: Integrate all server allocation plans to obtain the final server allocation plan; The first clustering includes: S101: Obtain a rasterized map and use it as a cluster to form a cluster set; S102: Selecting the cluster with the largest diameter in the cluster set as the second cluster; S103: Calculating the average dissimilarity between each grid in the second cluster and other grids in the second cluster as a first average dissimilarity; and sorting the first average dissimilarity to obtain the grid with the largest first average dissimilarity as a new first grid; S104: Calculating the dissimilarity between each grid in the second cluster and the first grid as a second dissimilarity; S105: comparing the first average dissimilarity and the second dissimilarity of each grid in the second cluster; if the first average dissimilarity is greater than or equal to the second dissimilarity, adding the grid to the fourth cluster; if the first average dissimilarity is less than the second dissimilarity, adding the grid to the third cluster; S106: replacing the second cluster in the cluster set with the third cluster, and adding the fourth cluster as a new first cluster to the cluster set; S107: Create two empty clusters as the new third cluster and the new fourth cluster respectively; repeat steps S102 to S106 until the number of clusters in the cluster set is K, and use all clusters in the cluster set as the final first cluster; The second clustering includes: S301: Selecting grids in the final first cluster to construct a grid sparse graph, forming a grid sparse graph matrix; wherein the nodes of the sparse graph represent grids; and the edges of the sparse graph represent the traffic intensity distance between grids; S302: Inputting the grid sparse graph matrix and the corresponding grid attributes into a deep feature model to obtain grid deep features; S303: performing soft clustering on the grid deep-level features to obtain multiple final third clusters.

2. According to claim 1, a digital twin edge collaborative quasi-static deployment method based on the local characteristics of vehicle movement is characterized by: The calculation formula of the diameter of the cluster is as follows: and represents the weight factor, + =1, represents the geographical distance from grid i to grid j, It represents the number of switching digital twins required for the vehicle to move between grid i and grid j, where i and j represent sequence numbers.

3. The method for quasi-static deployment of digital twin edge collaboration based on the local characteristics of vehicle movement according to claim 1 is characterized in that: The dissimilarity is the Euclidean distance between two grids.

4. The method for quasi-static deployment of digital twin edge collaboration based on the local characteristics of vehicle movement according to claim 1 is characterized in that: The deep feature model includes: a self-loop layer, a linear layer, a first GCN unit, and a second GCN unit; The grid sparse graph matrix is ​​input into the input end of the ring layer, and the grid attributes are input into the input end of the first GCN unit; the output end of the second GCN unit outputs the classification result; The output end of the self-loop layer is connected to the input end of the linear layer, the output end of the linear layer is connected to the input end of the first GCN unit and the input end of the second GCN unit, and the output end of the first GCN unit is connected to the input end of the second GCN unit.

5. The method for quasi-static deployment of digital twin edge collaboration based on the local characteristics of vehicle movement according to claim 4 is characterized in that: The first GCN unit includes a first hidden layer, a first fully connected layer, and a first activation layer connected in sequence; The grid attributes and the output end of the linear layer are input to the input end of the first hidden layer, and the output end of the first activation layer is connected to the input end of the second GCN unit.

6. The method for quasi-static deployment of digital twin edge collaboration based on the local characteristics of vehicle movement according to claim 5 is characterized in that: The second GCN unit includes a second hidden layer, a second fully connected layer, and a second activation layer connected in sequence; The output end of the first activation layer and the output end of the linear layer are connected to the input end of the second hidden layer, and the second activation layer outputs the classification result.

7. The method for quasi-static deployment of digital twin edge collaboration based on the local characteristics of vehicle movement according to claim 1 is characterized in that: The digital twin deployment optimization objectives are as follows: M is the unit deployment benefit, ∂ is the coefficient that balances the load balancing LO(A,L) and the virtual-real interaction delay DE(A,L), represents the cost of a single switch, μ is the cost of a single deployment; Indicates that the i-th grid is assigned to the k-th final third cluster, i, k, q represent serial numbers, Indicates that the kth final third cluster is assigned to the qth server, Q is the total number of servers, K is the total number of final third clusters, and N is the total number of grids. represents the load of the i-th grid from t-1 to t, represents the bandwidth resource allocated by the qth server to the kth final third cluster, represents the total bandwidth resources of the qth server, Represents the total computing power resources of the qth server, represents the computing power resources required for the i-th grid, and A and L represent the obtained deployment utility optimization solution.

8. A digital twin edge collaborative quasi-static deployment system based on the local characteristics of vehicle movement, applied to the deployment method according to any one of claims 1 to 7, characterized in that: include: The first clustering module: obtains a rasterized map of the area to be deployed, performs a first clustering on the grids in the rasterized map, and obtains K final first clusters; Model building module: Builds the IoV digital twin optimization target model and obtains the digital twin deployment optimization target; Cluster selection module: selects a final first cluster as the final second cluster; A second clustering module: performing a second clustering on the grids in the final second cluster to obtain a plurality of final third clusters; Allocation plan generation module: Based on the digital twin deployment optimization goal, the simulated annealing and particle swarm optimization algorithms are used to generate the server allocation plan corresponding to each final third cluster; The third clustering module: select another final first cluster as the final second cluster; Return to the second clustering module until all final first clusters are selected; Allocation plan integration module: integrates all server allocation plans to obtain the final server allocation plan.

Images