A collaborative management method for network slice resources for edge computing

Abstract

This invention discloses a network slice resource collaborative management method for edge computing, comprising the following steps: Step 1: Generating slice collaborative sample units for the network slices to be managed; Step 2: Constructing a slice resource collaborative heterogeneous graph based on the slice collaborative sample units; Step 3: Generating a slice coupling representation sequence using an improved HGNN model; Step 4: Performing a splitting process on the slice coupling representation sequence according to the slice instance identifier to generate a candidate collaborative action set; Step 5: Injecting the candidate collaborative action set item by item into the shadow slice body generated by the current real running state of each slice instance, and performing counterfactual inference processing to form an action result trajectory set; Step 6: Performing business constraint verification processing on each trajectory in the action result trajectory set to generate a stable collaborative action sequence. This invention achieves intelligent collaborative management of edge slices through an improved HGNN model.

Description

Technical Field

[0001] This invention relates to the field of resource management technology, and in particular to a collaborative management method for network slice resources oriented towards edge computing. Background Technology

[0002] With the continuous development of edge computing and network slicing technologies, an increasing number of low-latency, high-reliability, and differentiated services are being deployed to edge nodes closer to the terminal. Network slicing, by organizing and isolating wireless access, transmission, and computing resources on demand, can provide differentiated carrying environments for different services; edge computing, by pushing service processing capabilities down to the network edge, improves service response speed and reduces the burden on the core network. Against this backdrop, how to collaboratively manage multiple types of network slices in edge computing scenarios has gradually become an important research direction in the field of network resource scheduling and intelligent orchestration.

[0003] In existing technologies, network slice management methods for edge computing mostly adopt resource allocation methods based on static rules, local monitoring indicators, or single prediction results. These methods typically adjust the deployment of slice instances, reselect links, or migrate resources based on the node load, link utilization, or service demand intensity at the current moment. While these approaches can achieve slice resource scheduling to a certain extent, they mostly focus on local state analysis of a single node, a single link, or a single moment. They lack a unified characterization of the coupling effects between slice service processing, edge node bearing relationships, wireless access relationships, link forwarding relationships, and terminal area migration relationships, making it difficult to accurately reflect the dynamic collaborative characteristics of network slices in edge computing scenarios.

[0004] Therefore, how to provide a collaborative management method for network slice resources oriented towards edge computing is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a collaborative management method for network slice resources in edge computing. By constructing a heterogeneous collaborative graph of slice resources and combining it with an improved HGNN model, candidate collaborative action generation, shadow slice counterfactual inference, and business constraint verification, this invention achieves a refined representation of network slice resource relationships in edge computing scenarios, accurate location of conflicting chains, and effective screening of stable collaborative actions, thereby improving the accuracy, stability, and foresight of collaborative management of network slice resources.

[0006] A network slice resource collaborative management method for edge computing according to an embodiment of the present invention includes the following steps: Step 1: Within the current management cycle, perform unified time base alignment and sliding window segmentation on the network slices to be managed, and then perform merging, concatenation, and splicing processes to generate slice collaborative sample units; Step 2: Generate various types of nodes, heterogeneous association edges, temporal association edges, and functional chain association edges based on slice collaborative sample units, and construct a slice resource collaborative heterogeneous graph; Step 3: Input the slice resource collaborative heterogeneous map into the improved HGNN model, and generate the slice coupling representation sequence through the node mapping module, channel aggregation module, chain segment gating module and temporal sequence fusion module; Step 4: Perform splitting processing on the slice coupling representation sequence according to the slice instance identifier to form a chain segment conflict value sequence and generate a candidate cooperative action set; Step 5: Inject the candidate collaborative action set into the shadow slice body generated by the current real running state of each slice instance, and perform counterfactual inference processing to form the action result trajectory set; Step 6: Perform business constraint verification on each trajectory in the action result trajectory set to generate a stable collaborative action sequence.

[0007] Optionally, step one specifically includes: The service request records, edge node operation records, wireless access records, link forwarding records, and terminal area residency records corresponding to the network slices under management within the current management cycle are processed using a unified time reference to generate time-aligned records. The generated time-aligned records are processed by sliding window segmentation according to the preset decision window length and sliding step size to generate original window segments; The business request records in each original window segment are merged according to the slice instance identifier to form a slice business window segment; The edge node running records in each original window segment are merged according to the edge node identifier to form a node running window segment; The wireless access records and link forwarding records in each original window segment are concatenated according to the data flow path to form an access transmission window segment. The terminal region dwell records in each original window segment are processed by trajectory splicing according to the order of terminal movement to form a region migration window segment; Perform positional binding processing on slice business window segments, node operation window segments, access transmission window segments, and regional migration window segments within the same decision window to generate slice collaboration sample units.

[0008] Optionally, step two specifically involves: The business processing requests, session persistence requests, and forwarding processing requests in each slice collaborative sample unit are sorted according to the order of their occurrence within the same decision window of the same slice instance, forming a slice function chain sequence. The edge node occupancy relationship, wireless access occupancy relationship, link forwarding occupancy relationship and area residency relationship corresponding to each processing segment in the slice function chain sequence are associated and attached to generate resource association fragments; For the association relationships in each resource-related segment that have the same slice instance identifier and are within the same decision window, perform node extraction processing to generate slice nodes, edge nodes, access nodes, link nodes, and region nodes; Perform edge processing on the bearer relationship between slice nodes and edge nodes, the access relationship between slice nodes and access nodes, the transmission relationship between slice nodes and link nodes, and the dwell relationship between slice nodes and region nodes to generate heterogeneous association edges. Perform temporal edge processing on region nodes belonging to the same slice instance within adjacent decision windows, and perform sequential edge processing on adjacent functional chain processing segments within the same slice instance to generate temporal and functional chain related edges. By combining slice nodes, edge nodes, access nodes, link nodes, regional nodes, as well as heterogeneous associated edges, temporal associated edges, and functional chain associated edges, a slice resource collaborative heterogeneous graph is generated.

[0009] Optionally, the improved HGNN model is specifically as follows: The slice resource collaborative heterogeneous map is input into the node mapping module, and slice nodes, edge nodes, access nodes, link nodes and regional nodes are divided into corresponding node groups according to node type; Perform intra-type linear mapping processing on the node features within each node group, and unify the feature length of nodes of different categories within their respective mapping channels; After completing the intra-type linear mapping process, each node is grouped and written into a unified feature space to obtain a unified node representation. The unified node representation is input into the channel aggregation module to obtain the channel aggregation representation; The channel aggregation representation is input into the chain segment gating module, the functional chain position index of the corresponding slice node is extracted, and the functional chain position index is matched with the bearer channel representation, access channel representation, transmission channel representation, resident channel representation, timing channel representation and functional chain channel representation one by one. Channel representations that match the current function chain position are retained, while channel representations that do not match the current function chain position are eliminated. The channel representations after retention and elimination are recombined to generate a gated aggregate representation. The gated aggregation representation is input into the temporal sequence fusion module. Cross-window propagation processing is performed on the regional node representations belonging to the same slice instance in adjacent decision windows along the regional temporal association edge to generate temporal enhancement representations. The cross-window propagation processing is to pass the regional migration changes and resource squeezing changes in the previous decision window to the current decision window. Along the functional chain association edge, perform sequential propagation processing on the representation of adjacent functional chain processing segments within the same slice instance to generate chain segment coupling representation. The sequential propagation processing is to continue to pass the load changes, link changes and region changes of the previous segment of the functional chain to the next segment. Cross-type splicing processing is performed on the slice node representation, edge node representation, link node representation and region node representation in the chain segment coupling representation to form a slice coupling representation sequence.

[0010] Optionally, the sub-channel aggregation module specifically comprises: The edge relationships in the slice resource collaborative heterogeneous graph are processed by edge type, according to the bearer edge, access edge, transmission edge, dwell edge, timing edge, and functional chain edge. For each edge type, perform independent neighborhood retrieval processing on the adjacent nodes. Propagate the edge node information, access node information, link node information, region node information, previous decision window region node information, and adjacent functional chain processing segment information connected to the slice node to the slice node along the corresponding edge type. The propagation results of each edge type are processed by intra-channel aggregation to generate bearer channel representation, access channel representation, transmission channel representation, resident channel representation, timing channel representation and functional chain channel representation. These are then written into the same node according to the node position to obtain the sub-channel aggregated representation.

[0011] Optionally, step four specifically involves: The slice coupling representation sequence is split according to the slice instance identifier and arranged in order according to the functional chain position index to generate independent coupling representations corresponding to each slice instance to be regulated. For each chain segment representation in each independent coupling representation, the edge bearing strength, link congestion propagation strength and regional migration propagation strength are extracted respectively to generate a chain segment conflict value sequence; The chain segment conflict value sequence is sorted according to its size, and the chain segment with the highest conflict value is selected as the chain segment with the most concentrated resource conflict in the current slice instance. The chain segment with the most concentrated resource conflict is used as the starting point for action generation. The current bearer edge node, current forwarding link path, current regional dwell position and adjacent chain segments before and after the current functional chain corresponding to the chain segment position are read. Replacement and attachment processing, location exchange judgment processing, path replacement processing, candidate cache position replacement processing and candidate anchor point switching processing are performed respectively to generate candidate collaborative action items. All candidate collaborative action items corresponding to the same slice instance are classified according to their corresponding function chain position index. Candidate collaborative action items that act on the same chain segment position are grouped into the same action group. Candidate collaborative action items in each action group are sorted according to action category. Each action group is then grouped from front to back according to function chain position to form a set of candidate collaborative actions corresponding to the same slice instance.

[0012] Optionally, step five specifically includes: Based on the current actual running status of each slice instance, a corresponding shadow slice body is generated; The candidate collaborative action set is injected into the corresponding shadow slice one by one. For each candidate collaborative action after injection, the shadow slice is processed by counterfactual inference within multiple consecutive decision windows. The processing delay, transmission jitter, packet loss, node occupancy, link congestion and regional migration and diffusion caused by the action are recorded window by window. The various changes obtained from the same candidate collaborative action in multiple consecutive decision windows are spliced ​​together in chronological order to generate the action result trajectory. The action result trajectories corresponding to all candidate collaborative actions under the same slice instance are aggregated to form an action result trajectory set.

[0013] Optionally, the business constraint verification process specifically includes: The latency results in each trajectory of the action result trajectory set are compared with the latency upper limit of the corresponding slice instance. The jitter results in each trajectory are compared with the jitter upper limit of the corresponding slice instance. The packet loss results in each trajectory are compared with the reliability lower limit of the corresponding slice instance. The node occupancy results in each trajectory are compared with the range that the target edge node can bear. Candidate collaborative actions that do not meet any business constraints or resource carrying capacity constraints are eliminated, generating a subset of action trajectories; For candidate collaborative actions belonging to the same target edge region in the subset of action trajectories, perform summary processing, count the number of new slices arriving, the increase in new link occupancy, and the increase in new node load caused in the corresponding region, and weight them according to preset weights to obtain the corresponding regional collaborative pressure value. Candidate collaborative actions with regional collaborative pressure values ​​exceeding a preset upper limit are eliminated. The remaining candidate collaborative actions are sorted from low to high to obtain a stable sequence of collaborative actions.

[0014] The beneficial effects of this invention are: This invention addresses the shortcomings of existing technologies, such as insufficient characterization of network slice collaboration relationships, coarse conflict localization, limited candidate action generation, and inadequate action evaluation, achieving more refined slice resource management tailored to edge computing scenarios. By performing unified time base alignment, sliding window segmentation, merging, concatenation, and splicing processing on service request records, edge node operation records, wireless access records, link forwarding records, and terminal area residency records, it organizes previously scattered multi-source operation information into slice collaboration sample units that can be analyzed in a single decision window, thereby improving the consistency and completeness of subsequent resource relationship modeling. Furthermore, this invention constructs a heterogeneous graph of slice resource collaboration, incorporating slice nodes, edge nodes, access nodes, link nodes, and area nodes, as well as heterogeneous association edges, temporal association edges, and functional chain association edges into a unified graph structure. This allows for the synchronous expression of the coupling relationships of network slices in terms of bearer, access, transmission, residency, cross-window migration, and functional chain order, overcoming the one-sidedness of existing technologies that rely solely on local states for single-point scheduling.

[0015] Building upon this foundation, the improved HGNN model utilizes its node mapping module, channel aggregation module, chain segment gating module, and temporal sequence fusion module to perform targeted propagation and filtering of multiple types of nodes and edge relationships. This allows for the retention of information relevant to the current chain segment deployment decision while eliminating irrelevant propagation results that do not match the current functional chain position. Consequently, the generated slice coupling representation sequence more accurately reflects the chain segment-level resource conflict state. Based on the slice coupling representation sequence, a chain segment conflict value sequence is further formed, generating a set of candidate cooperative actions. This enables resource adjustments to move beyond coarse-grained overall migration and instead develop more targeted replacement attachments, position exchanges, path replacements, candidate cache position replacements, and candidate anchor point switching schemes around the chain segment positions where resource conflicts are most concentrated. Furthermore, by performing counterfactual reasoning through the shadow slice, the changes in processing latency, transmission jitter, packet loss, node occupancy, link congestion, and regional migration and diffusion caused by candidate collaborative actions within multiple consecutive decision windows are assessed in advance. Combined with business constraint verification, layer-by-layer elimination and sorting are completed, effectively reducing the probability of actions that do not meet business constraints or resource carrying capacity constraints entering the actual management process, and improving the reliability of stable collaborative action sequences. Therefore, this invention can improve the accuracy, foresight, and stability of network slice resource collaborative management in edge computing scenarios, and enhance the carrying capacity and collaborative control capabilities of slice services in complex dynamic environments. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of a network slice resource collaborative management method for edge computing proposed in this invention; Figure 2 This is a schematic diagram of the slice collaborative sample unit generation steps in a network slice resource collaborative management method for edge computing proposed in this invention. Figure 3 This is a flowchart illustrating the process of generating a set of candidate collaborative actions for a network slice resource collaborative management method for edge computing proposed in this invention. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0018] refer to Figures 1-3 A collaborative management method for network slice resources for edge computing includes the following steps: Step 1: Within the current management cycle, perform unified time base alignment and sliding window segmentation on the network slices to be managed, and then perform merging, concatenation, and splicing processes to generate slice collaborative sample units; Step 2: Generate various types of nodes, heterogeneous association edges, temporal association edges, and functional chain association edges based on slice collaborative sample units, and construct a slice resource collaborative heterogeneous graph; Step 3: Input the slice resource collaborative heterogeneous map into the improved HGNN model, and generate the slice coupling representation sequence through the node mapping module, channel aggregation module, chain segment gating module and temporal sequence fusion module; Step 4: Perform splitting processing on the slice coupling representation sequence according to the slice instance identifier to form a chain segment conflict value sequence and generate a candidate cooperative action set; Step 5: Inject the candidate collaborative action set into the shadow slice body generated by the current real running state of each slice instance, and perform counterfactual inference processing to form the action result trajectory set; Step 6: Perform business constraint verification on each trajectory in the action result trajectory set to generate a stable collaborative action sequence.

[0019] This step constructs a network slice resource collaborative management link for edge computing scenarios, enabling unified organization, correlation modeling, deep representation, action inference, and constraint screening of multi-source heterogeneous operational information, thereby improving the accuracy and stability of network slice resource regulation. By performing unified time base alignment, sliding window segmentation, and merging, concatenation, and splicing processing on the network slices to be managed within the current management cycle, dispersed service requests, node operation, wireless access, link forwarding, and regional residency information can be converted into structurally unified slice collaborative sample units, reducing the impact of multi-source data temporal misalignment and semantic fragmentation on subsequent analysis. Based on the slice collaborative sample units, a slice resource collaborative heterogeneous graph is constructed, which can uniformly express the bearer relationships, access relationships, transmission relationships, residency relationships, cross-window temporal relationships, and functional chain sequence relationships among slice nodes, edge nodes, access nodes, link nodes, and regional nodes, enhancing the ability to characterize the complex coupling relationships of network slices. Furthermore, by combining the improved HGNN model to generate slice coupling representation sequences, resource conflict propagation features and chain segment dependency features can be extracted more fully, making the chain segment conflict value sequence more accurately reflect the location of the most concentrated resource conflicts in the current slice instance. A candidate set of collaborative actions is generated based on the chain segment conflict value sequence, and each action is injected into the shadow slice body for counterfactual inference processing. This allows for the pre-assessment of the impact of different collaborative actions on latency, jitter, packet loss, node occupancy, link congestion, and regional migration and diffusion before actual adjustments, reducing the risk of service fluctuations caused by blind adjustments. Furthermore, by performing business constraint verification on each trajectory in the action result trajectory set, candidate actions that do not meet business and resource carrying capacity constraints can be effectively eliminated. This ensures that the final generated stable collaborative action sequence better meets the requirements for the continuous and stable operation of network slices in edge computing scenarios. This improves the foresight, precision, and reliability of network slice resource collaborative management.

[0020] In this embodiment, step one specifically includes: Read the business request records, edge node operation records, wireless access records, link forwarding records, and terminal area residency records within the current management cycle, extract the original timestamp field corresponding to each record, and convert the original timestamp field in each record into a unified time unit to generate a unified timestamp record; Using a preset reference time as a unified time starting point, time offset calculation is performed on each record in the unified timestamp record to obtain the time offset value of each record relative to the unified time starting point. Based on the time offset value, time recalibration processing is performed on each record to generate time-aligned records. The time-aligned records are processed by sliding window segmentation according to the preset decision window length and preset sliding step size. Records falling within the same window start and end interval are grouped into the same original window segment, generating a sequence of original window segments arranged in chronological order. The business request records in each original window segment are classified according to the slice instance identifier, and the business request records belonging to the same slice instance identifier are sorted according to the order of their recording time to form slice business window segments. The edge node running records in each original window segment are classified according to the edge node identifier, and the running records belonging to the same edge node identifier are sorted according to the order of recording time to form node running window segments; For each original window segment, the wireless access record and link forwarding record are associated and matched according to the same data flow identifier to obtain the path correspondence between the access record and the link forwarding record. Then, the concatenation process is performed according to the access node order and link forwarding order of the data flow to form the access transmission window segment. The terminal region dwell records in each original window segment are classified according to the terminal identifier, and the regional dwell records belonging to the same terminal identifier are sorted according to the order of dwell time to form a region migration window segment. Using the same decision window as the binding unit, slice service window segments, node operation window segments, access transmission window segments, and area migration window segments are aligned and bound according to a common slice instance identifier, a common time window number, and a common service bearing relationship. This allows the service processing, node bearing, access transmission, and area migration processes corresponding to the same slice instance within the same decision window to be written into the same binding structure, generating slice collaboration sample units.

[0021] This step, by performing unified time base conversion, time offset calculation, and time recalibration on multi-source heterogeneous records within the current management cycle, aligns service request information, node operation information, wireless access information, link forwarding information, and regional residency information—which originally had different collection frequencies and inconsistent recording times—to the same time reference system, reducing correlation deviations caused by inconsistent time granularity. Combined with sliding window segmentation, classification and sorting of similar records, access and link path matching and concatenation, and terminal residency trajectory processing, the scattered raw operational data can be transformed into windowed fragments with continuous temporal semantics and clear service carrying relationships. Furthermore, by forming slice collaborative sample units through positional binding, the service processing, node carrying, access transmission, and regional migration processes of the same slice instance within the same decision window are uniformly encapsulated in the same structure, thus providing a consistent, temporally consistent, and semantically consistent input foundation for subsequent heterogeneous map construction. This improves the organizational integrity of the multi-dimensional operational status of network slices and the reliability of subsequent collaborative analysis, avoiding the problems of fragmented and difficult-to-synchronize multi-source records in existing technologies.

[0022] In this embodiment, step two specifically includes: Read the business processing requests, session persistence requests, and forwarding processing requests from each slice collaborative sample unit, and perform sorting processing according to the order in which the requests occur within the same decision window of the same slice instance, forming a slice function chain sequence; For each processing segment in the slice function chain sequence, extract its corresponding edge node bearer location, wireless access location, link forwarding path, and terminal area camping location, and perform alignment and attachment processing on the edge node bearer location, wireless access location, link forwarding path, and terminal area camping location corresponding to the same processing segment to generate resource association fragments that correspond one-to-one with each processing segment; Each resource-associated segment is grouped according to the slice instance identifier and decision window number. Within each group, slice node extraction is performed on the slice instance identifier, edge node extraction is performed on the edge node bearing location, access node extraction is performed on the wireless access location, link node extraction is performed on each path location in the link forwarding path, and area node extraction is performed on the terminal area camping location, generating slice nodes, edge nodes, access nodes, link nodes, and area nodes belonging to the same slice instance and the same decision window. Within the same group, perform edge connection processing on the carrying relationship between each slice node and its corresponding edge node to generate slice carrying edges; Perform edge processing on the access relationship between each slice node and its corresponding access node to generate slice access edges; Perform transmission relationship edge processing on each slice node and its corresponding link node to generate slice transmission edges; Perform dwell relationship processing on each slice node and its corresponding region node to generate slice dwell edges; Combine slice carrying edges, slice access edges, slice transmission edges, and slice residing edges to generate a heterogeneous set of related edges; For adjacent decision windows, perform temporal pairing processing on the regional nodes that belong to the same slice instance and have corresponding continuous regional migration processes, and perform directed edge processing on the paired regional nodes to generate regional temporal association edges. For adjacent processing segments within the same slice instance in the slice function chain segment sequence, perform sequential pairing processing, and perform directed edge processing on the node corresponding to the previous processing segment and the node corresponding to the next processing segment to generate function chain association edges. The slice nodes, edge nodes, access nodes, link nodes, regional nodes, heterogeneous associated edge sets, regional temporal associated edges, and functional chain associated edges corresponding to the same slice instance are uniformly combined to form a slice resource collaborative heterogeneous graph containing multiple types of nodes and multiple types of edges.

[0023] This step organizes the service processing requests, session persistence requests, and forwarding processing requests in each slice collaborative sample unit into a slice functional chain sequence according to their chronological order within the same slice instance and decision window. This transforms the actual network slice processing flow from discrete records into a chain-like expression with sequential dependencies, providing a clear process foundation for subsequent resource correlation analysis. Furthermore, it performs alignment and attachment processing on the edge node bearer location, radio access location, link forwarding path, and terminal area camping location corresponding to each processing segment. Combined with the extraction of slice nodes, edge nodes, access nodes, link nodes, and area nodes, and the processing of multiple relationship edges, it unifies the slice service processing process with edge bearer relationships, access relationships, transmission relationships, and camping relationships into a unified graph structure. This transforms the originally dispersed resource occupancy and service operation states into a correlated network that can be jointly analyzed. Furthermore, by introducing regional temporal correlation edges and functional chain correlation edges, it is possible not only to characterize the continuous regional migration process of the same slice instance between adjacent decision windows, but also to express the sequential dependency relationship between adjacent processing segments within the same slice instance. This allows the constructed slice resource collaborative heterogeneous graph to simultaneously possess spatial correlation characterization and temporal evolution characterization capabilities. Consequently, it can more comprehensively reflect the resource collaboration relationships, migration and transmission relationships, and functional chain dependencies of network slices in edge computing scenarios, overcoming the problem of incomplete relationship expression caused by static analysis based only on local nodes or single links in existing technologies. It also provides a structurally clear and semantically complete data foundation for subsequent deep coupling feature extraction and resource conflict identification based on the improved HGNN model.

[0024] In this embodiment, the improved HGNN model is specifically as follows: The slice resource collaborative heterogeneous map is input into the node mapping module, and slice nodes, edge nodes, access nodes, link nodes and regional nodes are divided into corresponding node groups according to node type; Perform intra-type linear mapping processing on the node features within each node group, and unify the feature length of nodes of different categories within their respective mapping channels; After completing the intra-type linear mapping process, each node is grouped and written into a unified feature space to obtain a unified node representation. The unified node representation is input into the channel aggregation module, and edge type-based traffic processing is performed on the edge relationships in the slice resource collaborative heterogeneous graph according to the bearer edge, access edge, transmission edge, dwell edge, time sequence edge and functional chain edge; For each edge type, perform independent neighborhood retrieval processing on the adjacent nodes. Propagate the edge node information, access node information, link node information, region node information, previous decision window region node information, and adjacent functional chain processing segment information connected to the slice node to the slice node along the corresponding edge type. The propagation results of each edge type are processed by intra-channel aggregation to generate bearer channel representation, access channel representation, transmission channel representation, resident channel representation, time-series channel representation and functional chain channel representation. These are then written into the same node container according to the node position to obtain the sub-channel aggregation representation. The channel aggregation representation is input into the chain segment gating module, the functional chain position index of the corresponding slice node is extracted, and the functional chain position index is matched with the bearer channel representation, access channel representation, transmission channel representation, resident channel representation, timing channel representation and functional chain channel representation one by one. Channel representations that match the current function chain position are retained, while channel representations that do not match the current function chain position are eliminated. The channel representations after retention and elimination are recombined to generate a gated aggregate representation. The gated aggregation representation is input into the temporal sequence fusion module. Cross-window propagation processing is performed on the regional node representations belonging to the same slice instance in adjacent decision windows along the regional temporal association edge to generate temporal enhancement representations. The cross-window propagation processing is to pass the regional migration changes and resource squeezing changes in the previous decision window to the current decision window. Along the functional chain association edge, perform sequential propagation processing on the representation of adjacent functional chain processing segments within the same slice instance to generate chain segment coupling representation. The sequential propagation processing is to continue to pass the load changes, link changes and region changes of the previous segment of the functional chain to the next segment. Cross-type splicing processing is performed on the slice node representation, edge node representation, link node representation and region node representation in the chain segment coupling representation to form a slice coupling representation sequence.

[0025] The improved HGNN model proposed in this step shares similarities with the traditional HGNN model in that both use graph-structured data as their core processing object and achieve the representation learning of complex relationships through the information propagation mechanism between nodes and edges. Traditional HGNN models typically first map the original features of different nodes into a uniformly computeable feature space, and then perform neighborhood information aggregation based on the connection relationships between nodes in the graph, thus obtaining node representation results that include the influence of adjacency relationships. The improved HGNN model in this step also follows this basic technical route: first, it performs feature mapping on slice nodes, edge nodes, access nodes, link nodes, and region nodes, and then performs propagation and aggregation processing based on the edge relationships in the graph, ultimately forming a slice-coupled representation sequence that reflects the graph structure relationships. From the perspective of processing objectives, both attempt to use graph neural networks to model the dependencies, transitive relationships, and influence relationships between multiple nodes, in order to solve the problem that traditional vector-based processing methods are unable to express complex topological relationships. From the perspective of the internal mechanism of the model, both retain the basic processing framework of "node feature input - neighborhood information propagation - channel aggregation - high-level feature output", and both rely on the structural connection relationship in the graph to determine how information flows and converges. Therefore, the improved HGNN model proposed in this step is not a completely new structure that deviates from the principle of traditional HGNN. Instead, it is a targeted enhancement for the scenario of collaborative management of network slice resources based on the graph structure modeling, feature mapping and neighborhood propagation of traditional HGNN. Its underlying layer still belongs to the category of graph neural networks that use graph structure information to complete high-order association modeling.

[0026] The difference lies in the fact that traditional HGNN models typically focus on uniformly modeling high-order connections in general graph or hypergraph structures. Their propagation process mostly employs relatively universal neighborhood aggregation methods, offering limited adaptability to scenario features such as differences in node categories, edge relationships, and business processing order. The improved HGNN model proposed in this step, however, does not homogenize all nodes and edges. Instead, it constructs a more targeted processing link based on the business characteristics of the heterogeneous graph collaborative network slice resources. The model first divides slice nodes, edge nodes, access nodes, link nodes, and region nodes into different node groups according to node category in the node mapping module, performing intra-type linear mapping processing for different node categories to preserve the semantic differences between different node types from the source. Then, in the channel aggregation module, it performs edge type splitting processing according to bearer edges, access edges, transmission edges, dwell edges, time-series edges, and functional chain edges, allowing neighborhood information from different sources to propagate along their corresponding channels and aggregate independently, avoiding information interference caused by the mixed propagation of multiple relationships in traditional HGNNs. Finally, a chain segment gating module is introduced... The functional chain position index is matched one by one with the channel representation. Matching channels are retained, while mismatched channels are eliminated, enabling the model to filter out irrelevant propagation information based on the current chain segment's position in the business process. Finally, in the temporal sequence fusion module, not only is cross-window propagation processing performed along the regional temporal correlation edges, but sequential propagation processing is also performed along the functional chain correlation edges. This propagates the regional migration and resource compression changes from the previous decision window to the current window, while simultaneously propagating the load changes, link changes, and regional changes of the preceding functional chain segment to the next chain segment. Therefore, this improved model is significantly different from the traditional HGNN model in terms of node classification processing, edge-based channel propagation, chain segment position gating, and dual cross-window and sequential fusion.

[0027] The improvements are beneficial because they better reflect the realities of network slice resource collaborative management scenarios, which involve multiple types of nodes, multiple types of edges, and sequential dependencies in business processing chains. This enhances the accuracy and usability of slice coupling representation sequences. By performing intra-type linear mapping on different types of nodes through the node mapping module, semantic confusion between slice nodes, edge nodes, access nodes, link nodes, and region nodes before unified mapping is avoided, improving the consistency of subsequent propagation calculations. The channel aggregation module performs edge type splitting and independent aggregation on bearer edges, access edges, transmission edges, residing edges, time-series edges, and functional chain edges, enabling edge bearer information, link transmission information, region residing information, and functional chain sequence information to propagate independently along their respective channels, reducing representation distortion caused by heterogeneous relationship mixing. Finally, the chain segment gating module introduces a functional chain position index. Furthermore, by performing position matching processing, propagation results irrelevant to the current chain segment decision are eliminated, retaining only the key channel representations matching the current functional chain position, thereby enhancing the model's ability to locate chain segment-level resource conflicts. Through a temporal-sequential fusion module, cross-window propagation processing and sequential propagation processing are simultaneously performed, unifying cross-decision window regional migration and resource squeezing changes, as well as changes in bearer capacity, links, and regions between preceding and following functional chain segments, into the current chain segment representation. This ensures that the output slice coupling representation sequence includes not only spatial resource relationships but also temporal evolutionary relationships and sequential dependencies within the functional chain. Based on these improvements, the model can provide a more reliable input foundation for subsequent chain segment conflict value sequence generation, candidate collaborative action set generation, and counterfactual inference, thereby improving the accuracy and stability of the entire network slice resource collaborative management method in conflict identification, action selection, and business constraint adaptation.

[0028] In this embodiment, step four specifically includes: The slice coupling representation sequence is split according to the slice instance identifier and arranged in order according to the functional chain position index to generate independent coupling representations corresponding to each slice instance to be regulated. For each chain segment representation in each independent coupling representation, the edge bearing strength, link congestion propagation strength and regional migration propagation strength are extracted respectively to generate a chain segment conflict value sequence; Then, sort the chain segment conflict values ​​in the chain segment conflict value sequence according to their size, and select the chain segment with the highest chain segment conflict value as the chain segment position with the most concentrated resource conflict in the current slice instance; The action generation starts from the chain segment position where resource conflicts are most concentrated, and reads the current bearer edge node, current forwarding link path, current regional dwell position and the positions of adjacent chain segments before and after the current functional chain corresponding to the chain segment position; Perform alternative edge node retrieval processing around the current bearer edge node to generate a candidate edge node set; perform alternative link path retrieval processing around the current forwarding link path to generate a candidate link path set; perform neighboring target area retrieval processing around the current regional camp location to generate a candidate target area set. For each candidate edge node in the candidate edge node set, a replacement attachment process is performed one by one. The replacement attachment process is to switch the position of the chain segment with the most concentrated resource conflict from the current bearing edge node to the corresponding candidate edge node, and generate an edge node remapping action. For the chain segment where the resource conflict is most concentrated, perform position exchange judgment processing on the preceding and following adjacent chain segments in the functional chain. When the current chain segment is adjusted to the position of the preceding adjacent chain segment, a local forward movement of the functional chain is generated. When the current chain segment is adjusted to the position of the following adjacent chain segment, a local backward movement of the functional chain is generated. For each candidate link path in the candidate link path set, a path replacement process is performed one by one. The path replacement process is to switch the current forwarding link path corresponding to the link segment position with the most concentrated resource conflict to the corresponding candidate link path, and generate a link forwarding path reorganization action. Perform candidate cache location replacement processing on the cache residence location corresponding to the chain segment location with the most concentrated resource conflicts, switch the current cache residence location in the target region candidate set to the target cache location, and generate a cache residence location relocation action; Perform candidate anchor point switching processing on the session anchor point corresponding to the chain segment position with the most concentrated resource conflicts, switch the current session anchor point to the target anchor point, and generate a session anchor point switching action; Edge node remapping action, local forward movement of function chain, local backward movement of function chain, link forwarding path reorganization action, cache dwell location relocation action, and session anchor point switching action are selected as candidate collaborative action items. All candidate collaborative action items corresponding to the same slice instance are classified according to their corresponding function chain position index. Candidate collaborative action items that act on the same chain segment position are grouped into the same action group. Candidate collaborative action items in each action group are sorted according to action category. Each action group is then grouped from front to back according to function chain position to form a set of candidate collaborative actions corresponding to the same slice instance.

[0029] This step splits the slice coupling representation sequence according to the slice instance identifier and forms independent coupling representations by combining them with the functional chain position index. This separates the multi-slice information, which was originally mixed in the overall representation, into chain-segment level analysis objects oriented towards a single slice instance, thereby improving the targeting of subsequent resource conflict identification. Further extraction and synthesis of edge bearing strength, link congestion propagation strength, and regional migration propagation strength from each chain segment representation allows for the unified mapping of edge node pressure, link congestion propagation trends, and regional migration impacts onto the chain segment conflict value sequence. This ensures accurate location of the chain segment with the highest concentration of resource conflicts, avoiding the problem of coarse conflict identification caused by relying on overall load averages or single resource indicators in existing technologies. Based on this, using the chain segment with the highest concentration of conflicts as the starting point for action generation, candidate sets for edge nodes, link paths, and target regions are constructed around the current bearing edge node, the current forwarding link path, and the current regional dwell location. This ensures that the subsequently generated collaborative actions directly correspond to the key influencing factors of the current conflicting chain segment, enhancing the matching between action design and conflict sources. Furthermore, through replacement and attachment processing, location exchange determination processing, path replacement processing, candidate cache location replacement processing, and candidate anchor point switching processing, edge node remapping actions, functional chain local forward movement actions, functional chain local backward movement actions, link forwarding path reorganization actions, cache dwell location relocation actions, and session anchor point switching actions can be formed. This expands the resource adjustment method from the traditional single migration or overall relocation to a set of multi-path, multi-level collaborative actions oriented towards chain segment locations. Finally, all candidate collaborative action items corresponding to the same slice instance are classified and grouped according to the functional chain location index and action category. This can organize multiple adjustment schemes at the same chain segment location into a set of candidate collaborative actions in an orderly manner, providing a clear and uniformly granular action input foundation for subsequent counterfactual inference and business constraint verification. This can improve the accuracy of network slice resource conflict location, the richness of candidate action generation, and the effectiveness of subsequent action screening, thereby enhancing the refinement level and dynamic adjustment capability of network slice resource collaborative management in edge computing scenarios.

[0030] In this embodiment, step five specifically includes: Based on the current actual running status of each slice instance, a corresponding shadow slice body is generated; The candidate collaborative action set is injected into the corresponding shadow slice one by one. For each candidate collaborative action after injection, the shadow slice is processed by counterfactual inference within multiple consecutive decision windows. The processing delay, transmission jitter, packet loss, node occupancy, link congestion and regional migration and diffusion caused by the action are recorded window by window. The various changes obtained from the same candidate collaborative action in multiple consecutive decision windows are spliced ​​together in chronological order to generate the action result trajectory. The action result trajectories corresponding to all candidate collaborative actions under the same slice instance are aggregated to form an action result trajectory set.

[0031] This step generates a one-to-one shadow slice based on the current real-world operating state of each slice instance. This provides a virtual verification platform for candidate collaborative actions that closely mirrors the actual network slice operating environment without directly interfering with real-world business operations. This avoids the uncertainties associated with direct adjustments based on static estimations or empirical rules in existing technologies. Furthermore, the candidate collaborative action set is injected item by item into the corresponding shadow slice, and counterfactual analysis is performed within multiple consecutive decision windows. This allows for observation of the continuous impact of different collaborative actions on processing latency, transmission jitter, packet loss, node occupancy, link congestion, and regional migration and diffusion before the actions are actually implemented. This expands action evaluation beyond a single moment or single indicator to a dynamic impact analysis across multiple windows and time series. By recording the various changes caused by action execution window by window, the potential business performance fluctuations, resource occupancy shifts, and migration and diffusion trends that candidate collaborative actions may trigger during subsequent operations can be more comprehensively revealed, thereby improving the ability to identify potential risks in advance. Then, the various change results obtained from the same candidate collaborative action within multiple consecutive decision windows are spliced ​​together in chronological order to form an action result trajectory. This transforms the discrete multi-window inference results into a time-series trajectory expression with continuous evolutionary characteristics, making it easier to compare and filter the long-term impact differences between different candidate actions. Finally, the action result trajectories corresponding to all candidate collaborative actions under the same slice instance are aggregated to form an action result trajectory set. This provides a unified, complete, and comparable trajectory input basis for subsequent business constraint verification, ensuring that the determination of stable collaborative action sequences is based on continuous inference results, rather than on local experience judgments. This improves the foresight, systematicity, and credibility of network slice collaborative action evaluation, reduces the blindness and trial-and-error costs in the resource adjustment process, and enhances the stability and reliability of network slice resource collaborative management in edge computing scenarios.

[0032] In this embodiment, the business constraint verification process specifically includes: The latency results in each trajectory of the action result trajectory set are compared with the latency upper limit of the corresponding slice instance. The jitter results in each trajectory are compared with the jitter upper limit of the corresponding slice instance. The packet loss results in each trajectory are compared with the reliability lower limit of the corresponding slice instance. The node occupancy results in each trajectory are compared with the range that the target edge node can bear. Candidate collaborative actions that do not meet any business constraints or resource carrying capacity constraints are eliminated, generating a subset of action trajectories; For candidate collaborative actions belonging to the same target edge region in the subset of action trajectories, perform summary processing, count the number of new slices arriving, the increase in new link occupancy, and the increase in new node load caused in the corresponding region, and weight them according to preset weights to obtain the corresponding regional collaborative pressure value. Candidate collaborative actions with regional collaborative pressure values ​​exceeding a preset upper limit are eliminated. The remaining candidate collaborative actions are sorted from low to high to obtain a stable sequence of collaborative actions.

[0033] This step compares each trajectory in the action result trajectory set with the corresponding slice instance's latency upper limit, jitter upper limit, reliability lower limit, and the target edge node's carrying capacity. This allows for simultaneous verification of the feasibility of candidate collaborative actions from both business performance and resource carrying capacity constraints before actual execution. This avoids the misjudgment problems caused by prior art that relies solely on a single performance indicator or local resource status for action selection. Candidate collaborative actions that do not meet any business or resource carrying capacity constraints are directly eliminated, preventing invalid actions that may cause excessive business latency, worsened transmission jitter, increased packet loss, or node overload. This reduces the probability of unstable adjustment actions entering the subsequent decision-making process. Further, candidate collaborative actions belonging to the same target edge region within the subset of actionable action trajectories are aggregated and processed. The number of newly arrived slices, the increase in new link occupancy, and the increase in new node load caused by these actions within the corresponding region are statistically analyzed. These are then weighted according to preset weights to obtain the regional collaborative pressure value. This extends the collaborative impact of a single candidate action on a local region to a comprehensive assessment of the overall load-bearing pressure of the target edge region. Action selection no longer focuses solely on whether a single slice meets constraints but also considers its impact on regional resource competition and collaborative stability. Candidate collaborative actions with regional collaborative pressure values ​​exceeding a preset upper limit are further eliminated to prevent multiple seemingly locally feasible actions from overlapping in the same region, causing new congestion and load aggregation. Finally, the remaining candidate collaborative actions are sorted from low to high. Under the premise of meeting business constraints, resource carrying capacity constraints, and regional collaborative pressure constraints, actions with less system disturbance and better resource collaborative effects are prioritized, resulting in a more stable and implementable sequence of collaborative actions. This improves the comprehensiveness, accuracy, and regional adaptability of action selection in network slice resource collaborative management, enhancing the stability and reliability of multi-slice collaborative control in edge computing scenarios.

[0034] Example 1: To verify the feasibility of this invention in practice, it was applied to a port-based equipment manufacturing park in a coastal city. The park deployed various edge services for industrial production, machine vision quality inspection, mobile inspection terminals, and warehousing and logistics scheduling. The park's network adopted a multi-access edge computing architecture, deploying edge nodes in the production workshop, warehousing area, final assembly and testing area, and the park's integrated dispatch center. Differentiated services were provided through network slicing. Industrial control services have extremely high requirements for latency and continuity; machine vision quality inspection services have high requirements for uplink bandwidth and edge computing power; mobile inspection terminal services exhibit significant regional migration characteristics; and warehousing and logistics scheduling services are simultaneously affected by access-side load and link forwarding pressure. In actual operation, the park frequently encountered the following problem during high-load periods: multiple service slices competed for edge nodes and access resources close to the production line at the same time. While some slices appeared to have spare node resources, due to terminal regional migration, link forwarding fluctuations, and the coupling effects of front-end and back-end functional chains, sudden increases in latency, amplified jitter, and increased packet loss still occurred. Existing methods typically make single-point adjustments based on node load or link occupancy. While this can complete slice migration or path switching in a short time, it is difficult to simultaneously consider the coupling effects between slice functional chain order, regional movement trends, and various resource relationships. This results in inaccurate resource conflict localization and often leads to problems such as new chain segments becoming congested again after migration, local links being squeezed, and frequent slice reswitching, affecting the continuity of production in the park and the business stability of inspection terminals.

[0035] In this embodiment, the park deploys the method of the present invention in an edge scheduling and control platform to perform unified and collaborative management of industrial control slices, visual quality inspection slices, mobile inspection slices, and warehouse scheduling slices. The platform continuously collects various business request records, edge node operation records, wireless access records, link forwarding records, and terminal area residency records, aligns them according to a unified time base, and then segments the multi-source data using a fixed decision window and sliding step size. After merging, concatenation, and trajectory stitching, slice collaborative sample units under the same decision window are formed. After this processing, data that was originally scattered across different logs, different sampling frequencies, and different device sides are organized into the same analysis unit. This allows for a complete mapping of the edge bearer status, access location, link forwarding path, and terminal residency area of ​​a visual quality inspection slice at a certain moment. It also allows for the unified mapping of access switching and link changes caused by mobile inspection slices traversing multiple production areas to the same decision window.

[0036] The scheduling platform further constructs a heterogeneous graph of slice resource collaboration based on slice collaborative sample units. In this graph, business processing requests, session persistence requests, and forwarding processing requests for the same slice instance are expanded into a sequence of functional chain segments with a sequential order. Each chain segment is associated with the corresponding edge node, access location, link path, and regional residency information. For the complete processing process within the park, from production line cameras to visual analysis services to quality inspection feedback services, the system can place its front-end upload link, mid-end edge analysis nodes, and back-end result feedback link into the same graph structure, while retaining the related regional residency relationships and cross-window migration relationships. In this way, resource coupling relationships that were previously difficult to reflect in a single load table can be synchronously expressed in the heterogeneous graph.

[0037] Subsequently, the platform inputs the heterogeneous map of slice resource collaboration into the improved HGNN model. The model first performs a unified mapping of slice nodes, edge nodes, access nodes, link nodes, and region nodes, and then receives the propagation results of bearer edges, access edges, transmission edges, dwell edges, temporal edges, and functional chain edges through the channel aggregation module. Due to the obvious functional chain sequence characteristics of park operations, for example, the front-end control command issuance of industrial control slices is most sensitive to latency, the mid-end image analysis of visual quality inspection slices is most sensitive to computing power, and the front and rear segments of mobile inspection slices are closely related to changes in regional dwell time. Therefore, the chain segment gating module retains only the propagation results related to the current chain segment decision based on the functional chain position index, eliminating mismatched channel representations. The temporal sequence fusion module then propagates the regional migration changes and resource squeezing changes in the previous decision window to the current window, while continuing to propagate the bearer changes and link changes in the front end of the functional chain to subsequent chain segments, ultimately generating a slice coupling representation sequence. Through this processing, the platform no longer just knows "which node is busy", but can identify "which segment of which slice is most likely to cause a conflict due to the superposition of edge carrying, link congestion and regional migration".

[0038] In this implementation scenario, the system splits the slice coupling representation sequence according to the slice instance identifier and locates the chain segment with the most concentrated resource conflicts in each slice based on the chain segment conflict value sequence. For industrial control slices, conflicts are mostly concentrated between the control command issuance chain segment and the result feedback chain segment; for visual quality inspection slices, conflicts are mostly concentrated in the image processing chain segment; for mobile inspection slices, conflicts are often concentrated in the access and link connection chain segments when area switching is frequent. The platform generates candidate collaborative action items such as edge node replacement, chain segment position exchange, link path replacement, cache position adjustment, and anchor point switching around these conflicting chain segments, and injects these actions one by one into the shadow slice body in the digital twin environment. The shadow slice body performs counterfactual inference in multiple subsequent consecutive decision windows, outputting changes in processing latency, transmission jitter, packet loss, node occupancy, link congestion, and area migration and diffusion, respectively. Only action trajectories that simultaneously meet the upper limits of latency, jitter, reliability, and node carrying capacity will be retained. These trajectories are then combined with the number of newly arrived slices within the target area, the increase in new link occupancy, and the increase in new node load to calculate the regional collaborative pressure value, from which stable collaborative action sequences are selected. Finally, the platform distributes these stable collaborative action sequences to the resource management plane to complete slice resource collaborative management.

[0039] To verify the application effect of this invention in the park, continuous observation was conducted during peak production hours. The observation area included three business-intensive areas: the production workshop, the warehousing area, and the final assembly and testing area. A total of 48 slices were compared, including 12 industrial control slices, 10 visual quality inspection slices, 14 mobile inspection slices, and 12 warehouse scheduling slices. A total of 8 edge nodes were deployed, with 32 access points and 24 main forwarding links. Compared with the first scheme, which uses an edge node migration method based on a fixed threshold, the present invention uses the network slice resource collaborative management method for edge computing described in the above claims.

[0040] Table 1. Comparison of the Collaborative Management Effects of Multi-Service Edge Slices in the Park ; As shown in Table 1, under the same business load conditions, the problem with the traditional fixed threshold migration method is that it can only make adjustments based on the local node load or local link status. When multiple slices in the park are gathered near the production line at the same time, although some node surfaces still have remaining resources, the propagation of functional chain front and back ends, regional migration trends, and link forwarding compression are not jointly considered, ultimately leading to significant latency amplification and increased back-cutting in industrial control and mobile inspection services. The method of this invention performs better in all indicators, especially in terms of resource conflict location accuracy, shadow slice volume inference deviation, number of repeated adjustments, and total service interruption duration. This indicates that the invention can not only more accurately find the conflict chain segments that truly need adjustment, but also eliminate actions that are likely to cause subsequent instability in advance through action result trajectory set and business constraint verification processing, thereby reducing ineffective adjustments and rollback adjustments.

[0041] This embodiment demonstrates that, addressing the challenges of fragmented multi-source data, complex resource relationships, difficulty in locating conflicting chains, lack of continuous window verification for candidate actions, and insufficient screening of business constraints among network slices in edge computing scenarios, this invention achieves refined collaborative management of network slice resources through slice collaborative sample unit construction, slice resource collaborative heterogeneous graph generation, improved HGNN model coupling representation, candidate collaborative action formation, shadow slice volume counterfactual inference, and business constraint verification processing. In real-world campus application scenarios, it achieves lower latency, lower jitter, lower packet loss rate, more balanced node load, and fewer repeated adjustments, verifying the invention's good engineering applicability and significant technical benefits.

[0042] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for edge computing oriented network slice resource collaborative management, characterized in that, Includes the following steps: Step 1: Within the current management cycle, perform unified time base alignment and sliding window segmentation on the network slices to be managed, and then perform merging, concatenation, and splicing processes to generate slice collaborative sample units; Step 2: Generate various types of nodes, heterogeneous association edges, temporal association edges, and functional chain association edges based on slice collaborative sample units, and construct a slice resource collaborative heterogeneous graph; Step 3: Input the slice resource collaborative heterogeneous map into the improved HGNN model, and generate the slice coupling representation sequence through the node mapping module, channel aggregation module, chain segment gating module and temporal sequence fusion module; Step 4: Perform splitting processing on the slice coupling representation sequence according to the slice instance identifier to form a chain segment conflict value sequence and generate a candidate cooperative action set; Step 5: Inject the candidate collaborative action set into the shadow slice body generated by the current real running state of each slice instance, and perform counterfactual inference processing to form the action result trajectory set; Step 6: Perform business constraint verification on each trajectory in the action result trajectory set to generate a stable collaborative action sequence. 2.The method of claim 1, wherein, Step one specifically involves: The service request records, edge node operation records, wireless access records, link forwarding records, and terminal area residency records corresponding to the network slices under management within the current management cycle are processed with a unified time base alignment to generate time-aligned records. The generated time-aligned records are processed by sliding window segmentation according to the preset decision window length and sliding step size to generate original window segments; The business request records in each original window segment are merged according to the slice instance identifier to form a slice business window segment; The edge node running records in each original window segment are merged according to the edge node identifier to form a node running window segment; The wireless access records and link forwarding records in each original window segment are concatenated according to the data flow path to form an access transmission window segment. The terminal region dwell records in each original window segment are processed by trajectory splicing according to the order of terminal movement to form a region migration window segment; Perform positional binding processing on slice business window segments, node operation window segments, access transmission window segments, and regional migration window segments within the same decision window to generate slice collaboration sample units.

3. The network slice resource collaborative management method for edge computing according to claim 1, characterized in that, Step two specifically involves: The business processing requests, session persistence requests, and forwarding processing requests in each slice collaborative sample unit are sorted according to the order of their occurrence within the same decision window of the same slice instance, forming a slice function chain sequence. The edge node occupancy relationship, wireless access occupancy relationship, link forwarding occupancy relationship and area residency relationship corresponding to each processing segment in the slice function chain sequence are associated and attached to generate resource association fragments; For the association relationships in each resource-related segment that have the same slice instance identifier and are within the same decision window, perform node extraction processing to generate slice nodes, edge nodes, access nodes, link nodes, and region nodes; Perform edge processing on the bearer relationship between slice nodes and edge nodes, the access relationship between slice nodes and access nodes, the transmission relationship between slice nodes and link nodes, and the dwell relationship between slice nodes and region nodes to generate heterogeneous association edges. Perform temporal edge processing on region nodes belonging to the same slice instance within adjacent decision windows, and perform sequential edge processing on adjacent functional chain processing segments within the same slice instance to generate temporal and functional chain related edges. By combining slice nodes, edge nodes, access nodes, link nodes, regional nodes, as well as heterogeneous associated edges, temporal associated edges, and functional chain associated edges, a slice resource collaborative heterogeneous graph is generated.

4. The network slice resource collaborative management method for edge computing according to claim 1, characterized in that, The improved HGNN model is specifically as follows: The slice resource collaborative heterogeneous map is input into the node mapping module, and slice nodes, edge nodes, access nodes, link nodes and regional nodes are divided into corresponding node groups according to node type; Perform intra-type linear mapping processing on the node features within each node group, and unify the feature length of nodes of different categories within their respective mapping channels; After completing the intra-type linear mapping process, each node is grouped and written into a unified feature space to obtain a unified node representation. The unified node representation is input into the channel aggregation module to obtain the channel aggregation representation; The channel aggregation representation is input into the chain segment gating module, the functional chain position index of the corresponding slice node is extracted, and the functional chain position index is matched with the bearer channel representation, access channel representation, transmission channel representation, resident channel representation, timing channel representation and functional chain channel representation one by one. Channel representations that match the current function chain position are retained, while channel representations that do not match the current function chain position are eliminated. The channel representations after retention and elimination are recombined to generate a gated aggregate representation. The gated aggregation representation is input into the temporal sequence fusion module. Cross-window propagation processing is performed on the regional node representations belonging to the same slice instance in adjacent decision windows along the regional temporal association edge to generate temporal enhancement representations. The cross-window propagation processing is to pass the regional migration changes and resource squeezing changes in the previous decision window to the current decision window. Along the functional chain association edge, perform sequential propagation processing on the representation of adjacent functional chain processing segments within the same slice instance to generate chain segment coupling representation. The sequential propagation processing is to continue to pass the load changes, link changes and region changes of the previous segment of the functional chain to the next segment. Cross-type splicing processing is performed on the slice node representation, edge node representation, link node representation and region node representation in the chain segment coupling representation to form a slice coupling representation sequence.

5. A network slice resource collaborative management method for edge computing according to claim 4, characterized in that, The channel aggregation module specifically comprises: Edge type-based traffic splitting is performed on the edge relationships in the slice resource collaborative heterogeneous graph according to the bearer edge, access edge, transmission edge, dwell edge, timing edge, and functional chain edge; For each edge type, perform independent neighborhood retrieval processing on the adjacent nodes. Propagate the edge node information, access node information, link node information, region node information, previous decision window region node information, and adjacent functional chain processing segment information connected to the slice node to the slice node along the corresponding edge type. The propagation results of each edge type are processed by intra-channel aggregation to generate bearer channel representation, access channel representation, transmission channel representation, resident channel representation, timing channel representation and functional chain channel representation. These are then written into the same node according to the node position to obtain the sub-channel aggregated representation.

6. The network slice resource collaborative management method for edge computing according to claim 1, characterized in that, Step four specifically involves: The slice coupling representation sequence is split according to the slice instance identifier and arranged in order according to the functional chain position index to generate independent coupling representations corresponding to each slice instance to be regulated. For each chain segment representation in each independent coupling representation, the edge bearing strength, link congestion propagation strength and regional migration propagation strength are extracted respectively to generate a chain segment conflict value sequence; The chain segment conflict value sequence is sorted according to its size, and the chain segment with the highest conflict value is selected as the chain segment with the most concentrated resource conflict in the current slice instance. The chain segment with the most concentrated resource conflict is used as the starting point for action generation. The current bearer edge node, current forwarding link path, current regional dwell position and adjacent chain segments before and after the current functional chain corresponding to the chain segment position are read. Replacement and attachment processing, location exchange judgment processing, path replacement processing, candidate cache position replacement processing and candidate anchor point switching processing are performed respectively to generate candidate collaborative action items. All candidate collaborative action items corresponding to the same slice instance are classified according to their corresponding function chain position index. Candidate collaborative action items that act on the same chain segment position are grouped into the same action group. Candidate collaborative action items in each action group are sorted according to action category. Each action group is then grouped from front to back according to function chain position to form a set of candidate collaborative actions corresponding to the same slice instance.

7. The network slice resource collaborative management method for edge computing according to claim 1, characterized in that, Step five specifically involves: Based on the current actual running status of each slice instance, a corresponding shadow slice body is generated; The candidate collaborative action set is injected into the corresponding shadow slice one by one. For each candidate collaborative action after injection, the shadow slice is processed by counterfactual inference within multiple consecutive decision windows. The processing delay, transmission jitter, packet loss, node occupancy, link congestion and regional migration and diffusion caused by the action are recorded window by window. The various changes obtained from the same candidate collaborative action in multiple consecutive decision windows are spliced ​​together in chronological order to generate the action result trajectory. The action result trajectories corresponding to all candidate collaborative actions under the same slice instance are aggregated to form an action result trajectory set.

8. The network slice resource collaborative management method for edge computing according to claim 1, characterized in that, The business constraint verification process is specifically as follows: The latency results in each trajectory of the action result trajectory set are compared with the latency upper limit of the corresponding slice instance. The jitter results in each trajectory are compared with the jitter upper limit of the corresponding slice instance. The packet loss results in each trajectory are compared with the reliability lower limit of the corresponding slice instance. The node occupancy results in each trajectory are compared with the range that the target edge node can bear. Candidate collaborative actions that do not meet any business constraints or resource carrying capacity constraints are eliminated, generating a subset of action trajectories; For candidate collaborative actions belonging to the same target edge region in the subset of action trajectories, perform summary processing, count the number of new slices arriving, the increase in new link occupancy, and the increase in new node load caused in the corresponding region, and weight them according to preset weights to obtain the corresponding regional collaborative pressure value. Candidate collaborative actions with regional collaborative pressure values ​​exceeding a preset upper limit are eliminated. The remaining candidate collaborative actions are sorted from low to high to obtain a stable sequence of collaborative actions.

Images