A space-time perception dynamic virtual network mapping method for heterogeneous network

Abstract

The application belongs to the technical field of tactical communication network, network virtualization and network resource management and control, and particularly relates to a space-time perception dynamic virtual network mapping method for heterogeneous networks. The specific process is as follows: inputting a global graph representation and a virtual network request into a topology perception Transformer encoder of a global intelligent agent to obtain a cross-domain division result and an active backup decision; dividing the virtual request into a plurality of sub-requests and issuing the sub-requests to corresponding domain intelligent agents, while issuing the active backup decision and reliability constraints; the domain intelligent agent extracts time flow features and physical space flow features of the domain resources through a double-flow space-time graph transformer to form a space-time fusion representation in the domain; the domain intelligent agent calculates the mapping scores between each virtual node and the candidate physical node by using the space-time fusion representation in the domain according to the selected node mapping decoding mode, realizes the mapping in the domain, and generates a main path or a backup path by using a robust path selection algorithm.

Description

Technical Field

[0001] This invention belongs to the technical fields of tactical communication networks, network virtualization, and network resource management, and specifically relates to a spatiotemporal-aware dynamic virtual network mapping method for heterogeneous networks. Background Technology

[0002] Tactical communication networks typically include global control nodes, multi-domain controllers, fixed / mobile routing equipment, and tactical access terminals. Due to significant differences in computing power, communication bandwidth, resource costs, and reliability among different nodes, tactical communication networks require unified awareness, collaborative orchestration, and on-demand scheduling of multi-dimensional resources such as links, computing, and storage. With the increasing variety of service types, the expansion of operational areas, and drastic changes in network load, a single tactical physical network often needs to simultaneously carry multiple virtual task requests with different bandwidth, latency, priority, and reliability requirements. Virtual network mapping technology maps virtual nodes and virtual links in virtual requests to available devices and communication links in the physical tactical network, enabling unified orchestration and flexible scheduling of limited battlefield resources.

[0003] Virtual network mapping typically optimizes metrics such as request acceptance rate, resource gain, resource cost, and end-to-end latency, selecting bearer nodes and routing paths under constraints of node capacity, link bandwidth, and service latency. For tactical communication networks with domain-based management, a hierarchical control approach of "global controller - domain controller" may also be adopted. The global controller performs cross-domain partitioning and macro-coordination, while each domain controller performs intra-domain node mapping and link mapping based on the status of routing devices and terminal access nodes within its domain.

[0004] However, link status, routing congestion, available bandwidth, forwarding latency, and device load in heterogeneous networks continuously change over time. When the original mapped path degrades, becomes unavailable, or cross-domain link status deteriorates, the path initially available during mapping may not be able to continuously meet business needs throughout its lifecycle, leading to service interruptions, frequent migrations, and resource reconstruction overhead. Existing dynamic virtual network mapping methods mainly perform passive reconfiguration after link failure, failing to fully utilize the predictability of historical load and link quality trends for forward-looking resource planning. While some methods introduce backup paths, they often adopt a fixed redundancy protection approach, failing to combine the remaining link duration, latency change trends, historical load, and the remaining lifecycle of the request for joint decision-making, which can easily lead to redundant backup resources or a decrease in the access capability of subsequent requests.

[0005] Existing technologies suffer from the following main shortcomings: First, they lack forward-looking utilization of predictable topology and load evolution. Existing solutions primarily focus on current resource availability, rarely incorporating factors such as remaining link duration, link latency change rate, historical load cycles, and terminal access trends into initial mapping decisions. This can easily lead to the selection of paths with currently sufficient resources but poor future stability. Second, reliability assurance methods are simplistic. Relying solely on active backup consumes significant backup resources; relying solely on passive migration for recovery after link failure easily results in service interruptions and recovery delays. Third, the action search space is large, and many actions are ineffective. Candidate routers, edge nodes, terminal access nodes, and candidate paths in terrestrial intelligent networks change rapidly with service load and link status. Without reliability masks and robust path filtering, policy training and online inference can easily generate infeasible or high-risk mapping actions. Fourth, cross-domain partitioning, intra-domain mapping, and online maintenance are disconnected. Traditional solutions typically handle initial deployment and operation and maintenance separately, making it difficult to coordinate benefits, resource costs, latency, and service continuity within a unified decision-making framework. Summary of the Invention

[0006] The technical problem to be solved by this invention is: under the conditions of continuous changes in network topology, link quality, routing congestion and resource status and continuous arrival of service requests, how to use the spatiotemporal evolution characteristics of heterogeneous networks to make joint decisions on cross-domain partitioning, intra-domain mapping, active backup and failure migration of virtual network requests, so as to improve the continuous carrying capacity during the service life cycle and reduce dynamic maintenance overhead and end-to-end latency.

[0007] The technical solution for implementing the present invention is as follows:

[0008] In a first aspect, the present invention provides a spatiotemporally aware dynamic virtual network mapping method for heterogeneous networks, the specific process of which is as follows: Global encoding and decision-making: Inputting global graph representations and virtual network requests into the global agent. In the topology-aware Transformer encoder, cross-domain partitioning results and proactive backup decisions are obtained; Issuing Virtual Sub-Requests: Based on the cross-domain partitioning results, the virtual request is divided into several sub-requests and issued to the corresponding domain agents. At the same time, it issues proactive backup decisions and reliability constraints; Extracting Spatiotemporal Features within the Domain: Domain Agent The temporal and physical spatial features of resources in this domain are extracted by a dual-stream spatiotemporal graph transformer to form a spatiotemporal fusion representation within the domain. Intradomain mapping: Domain agents Based on the selected node mapping decoding mode, the mapping score between each virtual node and the candidate physical node is calculated using the intra-domain spatiotemporal fusion representation. Then, a relevant mask is superimposed on the mapping score to obtain the candidate node mapping score matrix, generating the actual node mapping result. The domain agent determines the source and destination physical nodes of each virtual link based on the node mapping result, and combines this with the global agent. The issued proactive backup decisions and reliability constraints are used to generate primary or backup paths using a robust path selection algorithm.

[0009] Optionally, the specific process of global encoding and decision-making described in this invention is as follows: Global intelligent agent For each virtual node in the virtual network request Extracting local request features For each physical domain in the global graph representation Constructing domain aggregated feature vectors And introduce request lifecycle features As input, the global context representation is jointly determined using the Transformer encoder. Based on global context representation Output cross-domain partitioning results and proactive backup decisions .

[0010] Optionally, the local request feature described in this invention Includes: Normalization calculation requirements Virtual node degree Total bandwidth requirement of adjacent virtual links and the average bandwidth requirement of adjacent virtual links The domain aggregated feature vector Including: Node size within the domain Total remaining computing power Total remaining bandwidth Average latency Maximum available computing power Maximum available bandwidth and computing resource load rate The request lifecycle normalization feature for:

[0011] in, Indicates the lifecycle or remaining service duration of the current virtual network request; This indicates the time scale used for normalization.

[0012] Optionally, in the Transformer encoder, the present invention first extracts the shortest path distance matrix based on the virtual request topology. S and adjacency matrix A Then, the shortest path distance matrix S Mapped to shortest path distance bias Then Adjacency matrix A and non-adjacency relation matrix 1 A Attention scores are obtained by superimposing additive topological biases. for:

[0013] in, Q, K, V These respectively represent the input features , , The query matrix, key matrix, and value matrix obtained by linear mapping The dimension of the key vector. and represents the learnable weights corresponding to adjacent edges and non-adjacent edges, respectively, and Softmax(·) represents the normalization function.

[0014] Optionally, the cross-domain partitioning result described in this invention and proactive backup decisions for:

[0015]

[0016] Among them, among them, and These represent the weight matrix and bias vector of the domain partitioning output head, respectively; and These represent the weight matrix and bias vector of the backup decision output head, respectively; (·) represents the Sigmoid activation function.

[0017] Optionally, the process for obtaining the dynamic topological features of the local domain according to the present invention is as follows: Resource time flow: Domain controllers are based on physical nodes within the domain. Historical state matrix and periodic access state characteristics As input, a time-series encoder GRU(·) is used to extract representations of node resources changing over time. The congestion prediction module Predictor(·) is used to obtain the future congestion tendency of nodes. ; Physical space flow: Taking the physical topology and link states of the local domain as input, it maps the link states to edge weights. For node pairs in the physical topology The higher-order edge biases are obtained by aggregating the edge weights along the shortest path. ; Dual-stream fusion: The congestion tendency obtained from the resource time stream Convert to time bias matrix and obtained from physical space flow and Joint injection The layer attention matrix is ​​used to calculate the current attention score, and the fused domain environment representation is obtained through a transformer. .

[0018] Optionally, the domain mapping described in this invention includes an inference phase, a node mapping phase, and a link mapping phase; During the reasoning phase: Domain agent Intelligent Agent The issued virtual sub-requests, proactive backup decisions, reliability constraints, and domain scope constraints are used together as decision conditions to select the decoding mode of node mapping; During the node mapping phase: Domain agent Based on the selected node mapping decoding mode, the mapping score between each virtual node and the candidate physical node is calculated using the intra-domain spatiotemporal fusion representation. Then, a capacity mask, a domain range mask, a resource occupancy mask, and an LRD-based reliability mask are superimposed on the mapping score to obtain the candidate node mapping score matrix or node action probability distribution. The node with the highest matrix score or the largest probability distribution is selected as the node mapping result. During the link mapping phase: the domain agent determines the source and destination physical nodes of each virtual link based on the node mapping results, and combines the proactive backup decision and reliability constraints issued by the global agent to generate the primary path or backup path using a robust path selection algorithm.

[0019] Optionally, the domain intelligent agent described in this invention global intelligent agent The issued virtual sub-request, proactive backup decision, and reliability constraints serve as decision conditions to select the decoding mode for node mapping. Specifically: Case 1: When the virtual sub-request contains critical links or cross-domain boundary nodes, or the proactive backup decision indicates the need to reserve a backup path, or the reliability constraints show insufficient candidate link LRD and a high risk of subsequent interruption, the domain agent... An autoregressive decoding method with causal masking is adopted; Scenario 2: When the number of virtual sub-request nodes is large, the structure is relatively regular, and no active backup or high reliability risk is triggered, the domain agent... Non-autoregressive decoding is used.

[0020] Optionally, in the link mapping stage of this invention: First, the physical topology is pruned using LRD as a stability threshold, eliminating physical links with insufficient remaining duration, insufficient bandwidth, or that do not meet reliability requirements; then, under the premise of satisfying the maximum end-to-end delay constraint, the maximum LRD bottleneck threshold of feasible paths is approximated by binary search, and K candidate paths are generated by searching in the pruned subgraph using the shortest path algorithm; subsequently, the bandwidth resource consumption, end-to-end delay, LRD bottleneck value, and potential migration overhead of each path are comprehensively compared, and the path with the lowest overall cost and satisfactory stability is selected as the main path; finally, if the backup decision of the global agent indicates that active backup is currently requested, the domain agent continues to search for backup paths that meet the requirements of resource reservation ratio, reliability constraints, and separation of primary and backup paths outside the main path.

[0021] Optionally, the present invention also includes an online maintenance and migration recovery process: after deploying the mapping results, the status of the links and nodes is continuously monitored, and when the primary path reaches the warning conditions, a primary-backup switch is performed, and when the primary-backup path is unavailable, a passive migration is triggered.

[0022] Optionally, the present invention further includes performing value assessment and strategy update: inputting the partitioning results, backup decisions, mapping results, and business operation status into a centralized value assessment network V(s), calculating reliability-driven rewards, and updating the global agent. He Domain Intelligent Agent The strategy parameters.

[0023] What are the technical advantages of this application compared to the closest existing technology? First, compared with the scheme that only performs virtual network mapping based on the current remaining resources, this technical solution incorporates link LRD, latency change rate and historical load trend into the state space, which can avoid short-life links and future congested areas in advance during the initial mapping stage, thereby reducing the reconstruction caused by link disconnection or path degradation during service operation.

[0024] Second, compared with solutions that only use passive migration or active backup, this technical solution can not only establish warm backup paths in advance for high-risk requests and perform switching before the main path reaches the risk threshold, shortening the business recovery time and reducing the probability of complete interruption; at the same time, it can also make backup decisions based on request lifecycle, link LRD, latency change trend and resource status, and supplement recovery capabilities through passive migration in the event of sudden failure. Therefore, it can achieve a more balanced effect between reliability and resource utilization efficiency.

[0025] Third, compared with ordinary deep reinforcement learning or graph attention mapping schemes, this technical solution simultaneously models the physical space structure and resource temporal evolution through a dual-stream spatiotemporal graph transformer, and reduces the invalid action space through reliability masking and robust path selection, which is conducive to improving online inference efficiency and mapping feasibility. Attached Figure Description

[0026] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 This is the overall architecture diagram of a spatiotemporally aware dynamic virtual network mapping system.

[0028] Figure 2 This is a flowchart of robust path selection based on LRD threshold binary search.

[0029] Figure 3 Convergence curves for service completion rates of different algorithms.

[0030] Figure 4 The cumulative distribution of end-to-end average latency for different algorithms.

[0031] Figure 5 A comparison of service stability for different algorithms.

[0032] Figure 6 A comparison of the economic benefits and dynamic costs of different algorithms. Detailed Implementation

[0033] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0034] It should be noted that, in the absence of conflict, the following embodiments and features can be combined with each other; and, based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0035] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this disclosure, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.

[0036] like Figure 1 As shown in the figure, this application embodiment presents a spatiotemporally aware dynamic virtual network mapping process for heterogeneous networks. This process takes a time-varying physical network composed of controllers, routers, terminals, and access devices, as well as online virtual network requests, as input. It explicitly extracts spatiotemporal information such as remaining link duration, link latency change rate, historical load sequences, terminal access changes, and service cycle characteristics. At the global layer, it determines the cross-domain partitioning and backup decision results of virtual requests based on a Transformer. At the intra-domain layer, it completes node mapping, link mapping, and online migration based on a dual-stream spatiotemporal graph transformer. During the runtime phase, it combines active warm standby switching and passive migration recovery to maintain service continuity. The specific process of this invention is as follows: Data Acquisition (Preparation Phase): Collect heterogeneous intelligent network status and service requests: Obtain the global topology status, local status of each domain, link quality, remaining node resources, historical load sequences, and virtual network requests for the current time slot. Construct Spatiotemporal State Features: Update the physical topology based on the collected network data and calculate link transmission capacity, overall latency, remaining link duration (LRD), and link latency change rate. Parse Virtual Network Requests: Extract virtual node resource requirements, virtual link bandwidth requirements, topological relationships, maximum latency constraints, and arrival times.

[0037] Global encoding and decision-making: Inputting global graph representations and virtual network requests into the global agent. In the topology-aware Transformer encoder, cross-domain partitioning results and proactive backup decisions are obtained; Issuing Virtual Sub-Requests: Based on the cross-domain partitioning results, the virtual request is divided into several sub-requests and issued to the corresponding domain agents. At the same time, it issues proactive backup decisions and reliability constraints; Extracting Spatiotemporal Features within the Domain: Domain Agent The temporal and physical spatial features of resources in this domain are extracted by a dual-stream spatiotemporal graph transformer to form a spatiotemporal fusion representation within the domain. Intradomain mapping: Domain agents Based on the selected node mapping decoding mode, the mapping score between each virtual node and the candidate physical node is calculated using the spatiotemporal fusion representation within the domain. Then, a relevant mask is superimposed on the mapping score to obtain the candidate node mapping score matrix, generating the actual node mapping result. The domain agent determines the source and destination physical nodes of each virtual link based on the node mapping result, and combines the proactive backup decision and reliability constraints issued by the global agent to generate the primary path or backup path using a robust path selection algorithm.

[0038] This implementation also includes online maintenance and migration recovery phases, as well as performance value assessments and strategy updates: Online maintenance and migration recovery phase: After deploying the mapping results, continuously monitor the status of links and nodes. When the primary path reaches the warning conditions, perform primary / backup switchover. When the primary / backup path is unavailable, trigger passive migration.

[0039] Execution of value assessment and strategy update: Input the partitioning results, backup decisions, mapping results, and business operation status into the centralized value assessment network V(s), calculate reliability-driven rewards, and update the global agent. He Domain Intelligent Agent The strategy parameters.

[0040] The important processes described above in this invention will be explained in detail below: Request partitioning and backup decisions based on Transformer Set up a global agent in the global control unit It is used to complete the cross-domain partitioning of virtual network requests with limited computing overhead and to determine whether the request needs to establish an active backup path.

[0041] The global intelligent agent A0 adopts a Transformer encoder with embedded graph topology bias to simultaneously perceive the virtual request structure, the carrying capacity of each physical domain, and the request lifecycle characteristics, thus avoiding the introduction of additional cross-domain link splicing costs and subsequent interruption risks by dividing virtual nodes with close topological relationships into different domains.

[0042] At the input layer, for each virtual node in the virtual network request... v Extract local request features. These local request features... Including normalization computation requirements Virtual node degree Total bandwidth requirement of adjacent virtual links and the average bandwidth requirement of adjacent virtual links This is used to characterize the local importance and resource demand intensity of virtual nodes in the request topology.

[0043]

[0044] At the same time, for each physical domain Constructing domain aggregated feature vectors Including the size of nodes within the domain Total remaining computing power Total remaining bandwidth Average latency Maximum available computing power Maximum available bandwidth and computing resource load rate This is used to characterize the overall carrying capacity of the physical domain at the current moment.

[0045]

[0046] Furthermore, the global agent introduces request lifecycle normalization features. Used to assist in backup decisions, among which Indicates the lifecycle or remaining service duration of the current virtual network request; This indicates the timescale used for normalization. For virtual requests with longer lifecycles that span more network state evolution time slices, the backup decision output can provide a more comprehensive risk assessment.

[0047]

[0048] During the Transformer encoding phase, the shortest path distance matrix is ​​first extracted based on the virtual request topology. S and adjacency matrix A Then, the shortest path distance matrix S Mapped to shortest path distance bias Then Adjacency matrix A and non-adjacency relation matrix 1 A Attention scores are obtained by superimposing additive topological biases. Thus, the model can explicitly distinguish between adjacent and non-adjacent nodes and perceive the topological distance between virtual nodes. This allows the global agent to prioritize virtual node relationships with strong business coupling and close topological distance when partitioning requests, reducing unnecessary cross-domain link splicing and providing a more reasonable sub-request structure for subsequent stable mapping within the domain.

[0049]

[0050] in, Q, K, V These represent the query matrix, key matrix, and value matrix obtained by linear mapping of the input features, respectively. The dimension of the key vector. and represents the learnable weights corresponding to adjacent edges and non-adjacent edges, respectively, and Softmax(·) represents the normalization function.

[0051] After topological bias Transformer encoding, the feature sequence of the virtual request node is obtained. Physical domain aggregated feature sequence and request lifecycle characteristics Jointly determine the global context representation Then, the global agent sets the domain partitioning output header and the backup decision output header, with the domain partitioning output header based on... Output the probability distribution of virtual nodes or virtual subtasks assigned to each autonomous system. Backup decision output header is based on Output the probability that the current virtual request enables active backup. .

[0052]

[0053]

[0054] in, and These represent the weight matrix and bias vector of the domain partitioning output head, respectively; and These represent the weight matrix and bias vector of the backup decision output head, respectively; (·) represents the Sigmoid activation function.

[0055] Intradomain mapping of two-stream spatiotemporal graph transformer The domain controller uses a dual-stream spatiotemporal graph transformer to extract the resource temporal stream and physical space stream, obtain the dynamic topology characteristics of the domain, and then completes the mapping of nodes and links within the domain through adaptive dual-mode decoding, reliability masking and robust path selection mechanism.

[0056] Resource time-series data is used to characterize the evolution of node load and resource status over time. Domain controllers are defined by the physical nodes within the domain. Historical state matrix and periodic access state characteristics As input, a time-series encoder GRU(·) is used to extract representations of node resources changing over time. The congestion prediction module Predictor(·) is used to obtain the future congestion tendency of nodes. Where || represents feature concatenation, [ 1] indicates that the last time step is taken as the output.

[0057]

[0058]

[0059] Physical space flow is used to inject the topology and dynamic link attributes of terrestrial intelligent networks into the attention mechanism. Domain controllers take the physical topology and link states of their domain as input and process the links... Construct dynamic edge features, including link bandwidth Overall latency Remaining duration and the rate of change of time delay And map it to edge weights through a fully connected layer. For node pairs The higher-order edge biases are obtained by aggregating the edge weights along the shortest path. Output of physical space flow Spatial bias with shortest path This is used together for subsequent attention fusion. Represents a node arrive Shortest path hop count, This indicates the first position on the shortest path. k One link.

[0060]

[0061]

[0062]

[0063] In dual-stream fusion attention, the domain controller will consider the congestion tendency obtained from the resource time stream. Convert to time bias matrix and obtained from physical space flow and Joint injection The layer attention matrix is ​​used to calculate the current attention score, and the fused domain environment representation is obtained through a transformer. .

[0064]

[0065]

[0066] Intradomain mapping Domain intelligent agent By combining capacity mask, domain range mask, resource occupancy mask, and LRD reliability mask, a node mapping decoding mode is selected and node mapping is performed. Then, a robust path acceleration algorithm based on binary search is used to determine the main path, such as... Figure 2As shown, a warm backup path is selected for requests to enable active backup. This specifically includes the inference phase, node mapping phase, and link mapping phase: During the reasoning phase, the domain agent... Using the spatiotemporal fusion representation within the domain as input, the virtual sub-requests, proactive backup decisions, reliability constraints, and domain-wide constraints issued by the global agent are used together as decision conditions to adaptively select the node mapping decoding mode. Specifically: Case 1: When the virtual sub-request contains critical links or cross-domain boundary nodes, or the proactive backup decision indicates the need to reserve a backup path, or the reliability constraints show that the candidate link LRD is insufficient and the risk of subsequent interruption is high, the domain agent... An autoregressive decoding method with causal masking is adopted; node mapping actions are output step by step according to the virtual node sequence. After each node mapping is completed, the system immediately updates the resource occupancy mask, remaining resource status, and related link availability status, so that subsequent node selection can be aware of the previous mapping results. Scenario 2: When the number of virtual sub-request nodes is large, the structure is relatively regular, and no active backup or high reliability risk is triggered, the domain agent... Non-autoregressive decoding is used; The joint mapping score matrix from virtual nodes to candidate physical nodes is output in one go, and the candidate deployment results are obtained by bipartite graph matching after superimposing constraint masks.

[0067] In the node mapping phase, firstly, the domain agent calculates the mapping score between each virtual node and candidate physical nodes based on the selected node mapping decoding mode and the intra-domain spatiotemporal fusion representation output by the dual-stream spatiotemporal graph transformer. Subsequently, the system superimposes a capacity mask, a domain range mask, a resource occupancy mask, and an LRD-based reliability mask onto the mapping scores to obtain the candidate node mapping score matrix or node action probability distribution, thereby generating the actual node mapping result. Specifically, the capacity mask filters physical nodes with insufficient computational resources; the domain range mask ensures that virtual sub-requests can only be mapped to the target physical domain defined by the global agent; the resource occupancy mask filters physical nodes that are already occupied by the current request or other requests and cannot be reused; and the LRD reliability mask filters candidate nodes with insufficient stability in surrounding links.

[0068] During the link mapping phase, the domain agent determines the source and destination physical nodes of each virtual link based on the node mapping results. Combining the proactive backup decision and reliability constraints issued by the global agent, a robust path selection algorithm is used to generate the primary or backup path. Specifically: the algorithm first prunes the physical topology using LRD as a stability threshold, eliminating physical links with insufficient remaining duration, insufficient bandwidth, or that do not meet reliability requirements. Then, under the premise of satisfying the maximum end-to-end latency constraint, it approximates the maximum LRD bottleneck threshold of feasible paths through binary search and uses the shortest path algorithm to search and generate K candidate paths in the pruned subgraph. Subsequently, it comprehensively compares the bandwidth resource consumption, end-to-end latency, LRD bottleneck value, and potential migration overhead of each path, selecting the path with the lowest overall cost and satisfactory stability as the primary path. If the global agent's backup decision indicates that proactive backup is currently requested, the domain agent continues to search for backup paths that meet the requirements of resource reservation ratio, reliability constraints, and primary / backup path separation, in addition to the primary path.

[0069] Simulation results: To verify the feasibility of this technical solution in dynamic heterogeneous networks, a simulated network with 100 dynamic physical nodes was constructed. Links between nodes were updated according to time slices, and virtual network requests arrived according to a Poisson process, with the request lifecycle following an exponential distribution with a mean of 1800 seconds. This technical solution was compared with dynamic network resource mapping methods such as AC-DMARL, FRMG-SAGIN, and rDQN-SFC, and analyzed from the perspectives of service completion rate, end-to-end latency, service stability, and economic benefits.

[0070] Regarding service completion rate, such as Figure 3 As shown, AC-DMARL and rDQN-SFC, due to the lack of explicit prediction of future link disconnection risks, maintain a service completion rate of approximately 80% to 85% after convergence; FRMG-SAGIN mainly relies on recovery mechanisms after failure, with a service completion rate of approximately 80%; this technical solution, through a dual-stream spatiotemporal graph converter, active backup, and passive migration collaborative mechanism, achieves a service completion rate of over 93%.

[0071] Regarding end-to-end latency, such as Figure 4 As shown, the average latency of FRMG-SAGIN is approximately 275 ms, while the average latency of this technical solution is approximately 285 ms. Although the average latency of this technical solution is slightly higher than that of FRMG-SAGIN, its service completion rate is significantly higher, and the proportion of high-latency tail samples is smaller, indicating that this technical solution can maintain relatively stable end-to-end service quality while maintaining high service carrying capacity.

[0072] Regarding service stability, such as Figure 5As shown, the average number of migrations for rDQN-SFC and AC-DMARL is approximately 18.8 and 16.1, respectively, while the number of migrations for FRMG-SAGIN is approximately 11.0, and the number of migrations for this technical solution is approximately 14.2. Combined with the service completion rate, it can be seen that this technical solution does not simply reduce migrations, but rather suppresses invalid and duplicate migrations while improving the service completion rate. Correspondingly, the average migration times for AC-DMARL, rDQN-SFC, FRMG-SAGIN, and this technical solution are approximately 321.9 ms, 377.6 ms, 150.4 ms, and 182.1 ms, respectively. This technical solution achieves a good balance between migration efficiency and service capacity.

[0073] In terms of economic benefits, such as Figure 6 As shown, AC-DMARL and rDQN-SFC have a benefit-cost ratio of less than 0.40 due to frequent migration and dynamic reconfiguration; this technical solution can still achieve a benefit-cost ratio of about 0.54 even with the introduction of warm backup path reservation, indicating that converting some uncontrollable post-fault recovery overhead into controllable pre-protection overhead is beneficial to reducing long-term dynamic maintenance costs.

[0074] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A spatiotemporally aware dynamic virtual network mapping method for heterogeneous networks, characterized in that, The specific process is as follows: Global encoding and decision-making: Inputting global graph representations and virtual network requests into the global agent. In the topology-aware Transformer encoder, cross-domain partitioning results and proactive backup decisions are obtained; Issuing Virtual Sub-Requests: Based on the cross-domain partitioning results, the virtual request is divided into several sub-requests and issued to the corresponding domain agents. At the same time, it issues proactive backup decisions and reliability constraints; Extracting Spatiotemporal Features within the Domain: Domain Agent The temporal and physical spatial features of resources in this domain are extracted by a dual-stream spatiotemporal graph transformer to form a spatiotemporal fusion representation within the domain. Intradomain mapping: Domain agents Based on the selected node mapping decoding mode, the mapping score between each virtual node and the candidate physical node is calculated using the intra-domain spatiotemporal fusion representation. Then, a relevant mask is superimposed on the mapping score to obtain the candidate node mapping score matrix and generate the actual node mapping result. The domain agent determines the source and destination physical nodes of each virtual link based on the node mapping results, and combines this with the global agent. The issued proactive backup decisions and reliability constraints are used to generate primary or backup paths using a robust path selection algorithm.

2. The spatiotemporally aware dynamic virtual network mapping method for heterogeneous networks according to claim 1, characterized in that, The specific process of global encoding and decision-making is as follows: Global intelligent agent For each virtual node in the virtual network request Extracting local request features For each physical domain in the global graph representation Constructing domain aggregated feature vectors And introduce request lifecycle features As input, the global context representation is jointly determined using the Transformer encoder. Based on global context representation Output cross-domain partitioning results and proactive backup decisions .

3. The spatiotemporally aware dynamic virtual network mapping method for heterogeneous networks according to claim 2, characterized in that, The local request feature Includes: Normalization calculation requirements Virtual node degree Total bandwidth requirement of adjacent virtual links and the average bandwidth requirement of adjacent virtual links ; The domain aggregated feature vector Including: Node size within the domain Total remaining computing power Total remaining bandwidth Average latency Maximum available computing power Maximum available bandwidth and computing resource load rate ; The request lifecycle normalization feature for: in, Indicates the lifecycle or remaining service duration of the current virtual network request; This indicates the time scale used for normalization.

4. The spatiotemporally aware dynamic virtual network mapping method for heterogeneous networks according to claim 3, characterized in that, In the Transformer encoder, the shortest path distance matrix is ​​first extracted based on the virtual request topology. S and adjacency matrix A Then, the shortest path distance matrix S Mapped to shortest path distance bias Then Adjacency matrix A and non-adjacency relation matrix 1 A Attention scores are obtained by superimposing additive topological biases. for: in, Q, K, V These respectively represent the input features , , The query matrix, key matrix, and value matrix obtained by linear mapping The dimension of the key vector. and represents the learnable weights corresponding to adjacent edges and non-adjacent edges, respectively, and Softmax(·) represents the normalization function.

5. The spatiotemporally aware dynamic virtual network mapping method for heterogeneous networks according to claim 4, characterized in that, The cross-domain partitioning results and proactive backup decisions for: Among them, among them, and These represent the weight matrix and bias vector of the domain partitioning output head, respectively; and These represent the weight matrix and bias vector of the backup decision output head, respectively; (·) represents the Sigmoid activation function.

6. The spatiotemporally aware dynamic virtual network mapping method for heterogeneous networks according to claim 1, characterized in that, The process for obtaining the dynamic topological features of this domain is as follows: Resource time flow: Domain controllers are based on physical nodes within the domain. Historical state matrix and periodic access state characteristics As input, a time-series encoder GRU(·) is used to extract representations of node resources changing over time. The congestion prediction module Predictor(·) is used to obtain the future congestion tendency of nodes. ; Physical space flow: Taking the physical topology and link states of the local domain as input, it maps the link states to edge weights. For node pairs in the physical topology The higher-order edge biases are obtained by aggregating the edge weights along the shortest path. ; Dual-stream fusion: The congestion tendency obtained from the resource time stream Convert to time bias matrix and obtained from physical space flow and Joint injection The layer attention matrix is ​​used to calculate the current attention score, and the fused domain environment representation is obtained through a transformer. .

7. The spatiotemporally aware dynamic virtual network mapping method for heterogeneous networks according to claim 6, characterized in that, The domain mapping includes the inference phase, the node mapping phase, and the link mapping phase. During the reasoning phase: Domain agent Intelligent Agent The issued virtual sub-requests, proactive backup decisions, reliability constraints, and domain scope constraints are used together as decision conditions to select the decoding mode of node mapping; During the node mapping phase: Domain agent Based on the selected node mapping decoding mode, the mapping score between each virtual node and the candidate physical node is calculated using the intra-domain spatiotemporal fusion representation. Then, a capacity mask, a domain range mask, a resource occupancy mask, and an LRD-based reliability mask are superimposed on the mapping score to obtain the candidate node mapping score matrix or node action probability distribution. The node with the highest matrix score or the largest probability distribution is selected as the node mapping result. In the link mapping phase: the domain agent determines the source and destination physical nodes of each virtual link based on the node mapping results, and generates the primary path or backup path by combining the proactive backup decision and reliability constraints issued by the global agent and using a robust path selection algorithm.

8. The spatiotemporally aware dynamic virtual network mapping method for heterogeneous networks according to claim 7, characterized in that, The domain intelligent agent global intelligent agent The issued virtual sub-request, proactive backup decision, and reliability constraints serve as decision conditions to select the decoding mode for node mapping. Specifically: Case 1: When the virtual sub-request contains critical links or cross-domain boundary nodes, or the proactive backup decision indicates the need to reserve a backup path, or the reliability constraints show insufficient candidate link LRD and a high risk of subsequent interruption, the domain agent... An autoregressive decoding method with causal masking is adopted; Scenario 2: When the number of virtual sub-request nodes is large, the structure is relatively regular, and no active backup or high reliability risk is triggered, the domain agent... Non-autoregressive decoding is used.

9. The spatiotemporally aware dynamic virtual network mapping method for heterogeneous networks according to claim 7, characterized in that, In the link mapping phase: First, the physical topology is pruned using LRD as a stability threshold, eliminating physical links with insufficient remaining duration, insufficient bandwidth, or that do not meet reliability requirements. Then, under the premise of satisfying the maximum end-to-end delay constraint, the maximum LRD bottleneck threshold of feasible paths is approximated by binary search, and K candidate paths are generated by searching the pruned subgraph using the shortest path algorithm. Subsequently, the bandwidth resource consumption, end-to-end delay, LRD bottleneck value, and potential migration overhead of each path are comprehensively compared, and the path with the lowest overall cost and satisfactory stability is selected as the primary path. Finally, if the backup decision of the global agent indicates that active backup is requested, the domain agent continues to search for backup paths that meet the requirements of resource reservation ratio, reliability constraints, and separation of primary and backup paths, in addition to the primary path.

10. The spatiotemporally aware dynamic virtual network mapping method for heterogeneous networks according to claim 1, characterized in that, This also includes online maintenance and migration recovery processes, as well as performance evaluation and strategy updates; Online maintenance and migration recovery process: After the deployment mapping results are deployed, the status of links and nodes is continuously monitored. When the primary path reaches the warning conditions, the primary and backup switching is performed. When the primary and backup paths are unavailable, passive migration is triggered. Execution of value assessment and strategy update: Input the partitioning results, backup decisions, mapping results, and business operation status into the centralized value assessment network V(s), calculate reliability-driven rewards, and update the global agent. He Domain Intelligent Agent The strategy parameters.

Images