An intelligent power 5G slice scheduling method

Abstract

The application discloses a power 5G slice intelligent scheduling method, and belongs to the field of power communication and 5G network slice scheduling. In order to solve the problem that when multiple power businesses share a communication module, it is difficult to timely adapt to the business priority, link state and safety isolation requirements, the application generates and issues a slice scheduling instruction through business flow identification, link slice state collection, multi-layer heterogeneous state graph construction, graph embedding and time sequence coding, constrained executor-evaluator scheduling and deterministic safety gating, thereby achieving the technical effects of reducing key business resource preemption, improving link switching timeliness and slice resource utilization.

Description

Technical Field

[0001] This invention relates to the field of power communication and 5G network slicing scheduling, and in particular to a smart scheduling method for power 5G slicing. Background Technology

[0002] With the development of power distribution automation, electricity consumption information collection, load control, and on-site operation and maintenance services, power 5G user premises equipment and public-private integrated communication modules need to simultaneously support multiple types of services, such as concentrators, distribution terminals, load control terminals, and handheld terminals. These services have different requirements in terms of latency, bandwidth, reliability, and security isolation, and the same communication module often needs to select access and schedule resources among public network candidate links, private network candidate links, and different network slices.

[0003] Existing communication modules mostly select access based on fixed link priorities, static network parameters, or manually configured rules. When base station load, wireless signal quality, slice capacity, service queues, and control command status change rapidly, it is difficult to make timely scheduling decisions based on power service priorities. Especially when control services and ordinary data acquisition services are transmitted simultaneously, problems such as ordinary services occupying critical link resources, low-latency service handover delays, and uneven utilization of slice resources are likely to occur.

[0004] In addition, while some existing solutions can predict service traffic or network status, they lack a unified state representation that models services, terminal interfaces, candidate links, network slices, base station cells, and security domains. They also lack mechanisms for security gating, action masking, risk calibration, and event preemption processing after the output of the learning-based scheduling model, making it difficult to balance dynamic resource utilization with the safe and reliable transmission of power services. Summary of the Invention

[0005] One objective of this invention is to propose an intelligent scheduling method for 5G power slicing. Addressing the problem that existing technologies using fixed link priorities or static network parameters cannot adapt in a timely manner to changes in various power services, base station load, link quality, and sudden control commands, this invention proposes a technical solution based on service state vectors, link slice state vectors, multi-layer heterogeneous state diagrams, joint scheduling state characteristics, constrained executor-evaluator scheduling models, and deterministic security gating rules. This invention achieves the technical effects of reducing preemption of critical service resources, improving the timeliness of link switching, and enhancing the utilization rate of slice resources.

[0006] This invention provides a power 5G slicing intelligent scheduling method, applied to communication equipment including power 5G user premises equipment or public-private integrated communication modules, comprising: S1, identifying power service flows accessing the communication equipment and generating service state vectors; S2, collecting the operating status of candidate communication links and candidate network slices and generating link slice state vectors; S3, constructing a multi-layer heterogeneous state graph based on the service state vectors and the link slice state vectors; S4, performing graph embedding processing on the multi-layer heterogeneous state graph and performing time-series encoding processing on link quality, service load, and queue changes within multiple consecutive scheduling cycles to generate joint scheduling state features; S5, inputting the joint scheduling state features and burst control event features into a trained constrained executor-evaluator scheduling model to generate candidate scheduling actions; S6, performing action masking and action correction on the candidate scheduling actions according to deterministic security gating rules to generate a final scheduling instruction and writing the final scheduling instruction into a local scheduling table.

[0007] Optionally, S1 includes:

[0008] Extract service type, service direction, message length, queue backlog, latency budget, reliability requirements, bandwidth requirements, security isolation level, and control command identifier from service messages, terminal access identifiers, and local service configuration tables;

[0009] The characteristics of control-type services, ordinary data collection services, handheld terminal services, and sudden control events are determined based on the service type, the terminal access identifier, the local service configuration table, and the control instruction identifier.

[0010] The deadline margin is calculated based on the timestamp of the message entering the queue, the delay budget, and the start time of the current scheduling cycle;

[0011] The service state vector is generated by combining the service type, service direction, message length, queue backlog, latency budget, reliability requirements, bandwidth requirements, security isolation level, control command identifier, deadline margin, and sudden control event characteristics in a preset field order.

[0012] Optionally, S2 includes:

[0013] Reference signal received power, signal-to-interference-plus-noise ratio, block error rate, modulation and coding scheme, physical resource block occupancy rate, round-trip time, packet loss rate, and cell load of public network candidate links and private network candidate links are collected respectively.

[0014] Collect the remaining service level, available resources, isolation level, access domain identifier, reachability of user plane function network elements, reachability of edge computing nodes, and alarm status of candidate network slices;

[0015] The collected operating status is time-aligned according to a preset sampling period, and the link slice status vector is generated according to the preset unit conversion rules corresponding to each field.

[0016] Optionally, S3 includes:

[0017] Establish a service node for each service flow, a terminal interface node for each terminal interface, a candidate link node for each public network candidate link or private network candidate link, a slice node for each candidate network slice, a base station node for each base station cell, and a security domain node for each security isolation domain.

[0018] Establish an access edge between the service node and the terminal interface node, establish a candidate bearer edge between the service node and the slice node, establish a wireless access edge between the candidate link node and the base station node, establish an isolation edge between the slice node and the security domain node, and establish a competition edge between service nodes that share the same candidate link or the same candidate network slice.

[0019] The business state vector, the link slice state vector, and the edge type encoding are written into the corresponding node and the corresponding edge to obtain the multi-layer heterogeneous state graph.

[0020] Optionally, S4 includes:

[0021] Type embedding parameters are set for different types of nodes and different types of edges in the multi-layer heterogeneous state graph;

[0022] Graph neural networks are used to pass messages based on the features of neighboring nodes, edge type encoding, and edge weights, generating business node embeddings, candidate link node embeddings, and slice node embeddings.

[0023] According to the scheduling cycle order, the business nodes, candidate link nodes, and slice nodes of multiple cycles are embedded into the time encoder to generate trend state features;

[0024] The joint scheduling state feature is obtained by concatenating the graph embedding feature of the current period, the trend state feature, and the sudden control event feature.

[0025] Optionally, S5 includes:

[0026] The joint scheduling state features are input into the executor network of the constrained executor-evaluator scheduling model to generate the action probability distribution and action parameters of each service flow in slice selection, link selection, queue weight, resource reservation ratio, handover hold time and backup link strategy;

[0027] The joint scheduling state features, the action probability distribution, and the action parameters are input into the evaluator network to obtain the state action value corresponding to the candidate scheduling action.

[0028] A set of candidate scheduling actions is generated based on the action probability distribution, the action parameters, the state action value, and the slice capacity constraint, isolation constraint, and switching count constraint.

[0029] Candidate scheduling actions whose state action values ​​satisfy preset selection rules are selected from the candidate scheduling action set and used as input for security gating processing.

[0030] Optionally, S6 includes:

[0031] The deterministic security gating rules are composed of the whitelist of power control services, the first latency threshold, the first reliability threshold, the isolation rules between private and public networks, the slice capacity limit, the handover jitter limit, and the control command preemption rules.

[0032] Based on the joint scheduling state characteristics and the candidate scheduling actions selected from the candidate scheduling action set, the prediction delay and prediction reliability of power control services are determined by the performance prediction output of the constrained executor-evaluator scheduling model.

[0033] When a candidate scheduling action causes the predicted latency to exceed the first latency threshold, the predicted reliability to be less than the first reliability threshold, the candidate network slice to exceed the slice capacity limit, the candidate link to violate the private network and public network isolation rules, or the number of handovers to exceed the handover jitter limit, an action mask is executed on the corresponding candidate scheduling action.

[0034] When there are unmasked candidate scheduling actions, a new candidate scheduling action is selected from the unmasked candidate scheduling actions according to the order of state action value. The queue weight, resource reservation ratio, slice selection or link selection are then adjusted according to the control instruction preemption rules to generate the final scheduling instruction.

[0035] When there are no candidate scheduling actions that are not masked, a safety backup scheduling instruction that maintains the current slice and the current link is generated as the final scheduling instruction, and the final scheduling instruction is written into the local scheduling table;

[0036] Furthermore, after the candidate scheduling action selected from the candidate scheduling action set is output and before entering step S6, a confidence risk calibration process is also included: using the service queue length of the most recent multiple scheduling cycles, reference signal received power, signal-to-interference-plus-noise ratio, block error rate, modulation and coding scheme, physical resource block occupancy rate, round-trip delay, packet loss rate, cell load, user plane functional network element reachability, edge computing node reachability, and historical action results as calibration samples, the delay error distribution, reliability error distribution, bandwidth satisfaction error distribution, and handover failure error distribution are calculated respectively.

[0037] The corresponding quantile value is determined from each error distribution based on the preset target coverage rate;

[0038] The prediction results of candidate services, candidate network slices and candidate link combinations are calibrated based on the quantile values, and a confidence risk vector composed of the upper bound of latency risk, the lower bound of reliability risk, the probability of bandwidth satisfaction and the risk of handover failure is output.

[0039] Furthermore, after the confidence risk vector is output, the process also includes slice action conflict graph processing: taking the candidate scheduling actions selected from the candidate scheduling action set, the remaining resources of the slice, the service priority, the security isolation level, the queue weight, the access domain identifier, and the confidence risk vector as input, action nodes, service nodes, slice nodes, link nodes, and security domain nodes are established.

[0040] Establish resource mutual exclusion edges between action nodes and slice nodes, establish isolation constraint edges between business nodes and security domain nodes, establish preemption edges between business nodes corresponding to control-type business and business nodes corresponding to ordinary collection-type business, and establish handover jitter edges between link nodes.

[0041] The conflict type, conflict service, and conflict resource root cause are determined based on the resource mutual exclusion edge, isolation constraint edge, preemption edge, and switching jitter edge, and the conflict type, conflict service, and conflict resource root cause are input into step S6;

[0042] Furthermore, it also includes event preemption gating processing: when the sudden control event characteristic characterization determined according to claim 2 indicates the existence of control-type service messages, and the upper bound of the latency risk of the corresponding candidate service, candidate network slice and candidate link combination represented by the confidence risk vector exceeds the first latency threshold or the lower bound of the reliability risk is less than the first reliability threshold, the action mask is executed for the new resource occupation action of ordinary collection-type services, the public network switching action of handheld terminal services and the action of violating security isolation level according to the conflict type, conflict service and conflict resource root cause;

[0043] Generate reserved slice activation instructions, private network link maintenance instructions, backup link activation instructions, or queue weight enhancement instructions for control-type services, and write the generated instructions into the local scheduling table. The local scheduling table is used for subsequent packet classification, link dialing, single network slice selection auxiliary information selection, service quality flow identifier mapping, and queue scheduling.

[0044] The beneficial effects of this invention are:

[0045] 1. This invention describes service requirements, wireless links, network slices, and security isolation status simultaneously through service state vectors and link slice state vectors, enabling scheduling decisions to consider service priority, link quality, slice capacity, and public-private network isolation relationships at the same time, reducing scheduling lag caused by static access selection.

[0046] 2. This invention generates joint scheduling state features through multi-layer heterogeneous state diagrams and time-series coding, enabling the bearer, contention, reachability, and isolation relationships between service flows, terminal interfaces, candidate links, network slices, base station cells, and security domains to enter the constrained scheduling model, thereby improving the state adaptability of scheduling actions in multi-service concurrent scenarios.

[0047] 3. This invention uses deterministic security gating, confidence risk calibration, slice action conflict graph, and event preemption gating to mask, correct, reselect, or provide security fallback for candidate scheduling actions. This ensures that control services can receive priority transmission guarantees when there are risks to latency or reliability, and reduces the impact of ordinary data acquisition services on critical control services. Attached Figure Description

[0048] 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:

[0049] Figure 1 This is a flowchart of the intelligent scheduling method for 5G power slicing according to the present invention.

[0050] Figure 2 This is a flowchart of step S6 of the present invention, which involves confidence calibration, conflict analysis, deterministic gating, and event preemption processing. Detailed Implementation

[0051] 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.

[0052] refer to Figures 1-2A power 5G slicing intelligent scheduling method includes: S1, identifying power service flows accessing the communication equipment and generating service state vectors; S2, collecting the operating status of candidate communication links and candidate network slices and generating link slice state vectors; S3, constructing a multi-layer heterogeneous state graph based on the service state vectors and the link slice state vectors; S4, performing graph embedding processing on the multi-layer heterogeneous state graph and performing time-series encoding processing on link quality, service load, and queue changes within multiple consecutive scheduling cycles to generate joint scheduling state features; S5, inputting the joint scheduling state features and burst control event features into a trained constrained executor-evaluator scheduling model to generate candidate scheduling actions; S6, performing action masking and action correction on the candidate scheduling actions according to deterministic security gating rules to generate a final scheduling instruction and writing the final scheduling instruction into a local scheduling table.

[0053] In this specific embodiment, S1 includes:

[0054] Used to perform consistent identification of power service flows accessing communication equipment and generate service state vectors within the local scheduling cycle;

[0055] The communication device defines a service flow using a 5-tuple <SrcIP, DstIP, SrcPort, DstPort, Proto> and maintains an independent transmission queue and queue statistics for each service flow;

[0056] The communication device performs field extraction on each service packet entering the queue. The field sources include the service packet, the terminal access identifier, and the local service configuration table. The terminal access identifier is the user identifier UEID that can be reliably obtained from the communication module side and is used as the primary key of the local service configuration table.

[0057] The local service configuration table is stored in non-volatile memory and loaded into memory during device manufacturing or engineering deployment. Table entries include fields such as Service Type (BT), Service Direction (BD), and latency budget. Reliability requirements Bandwidth requirements Security isolation level Control command identifier matching rules and the fixed attributes of the service associated with UEID;

[0058] The service type (BT) is categorized into one of three types—control, general acquisition, or handheld terminal—based on enumeration encoding, with each encoding value uniquely determined within the device. The service direction (BD) is encoded using uplink and downlink binary codes, with each encoding value uniquely determined within the device. Security isolation level... Classification and coding by isolation domain, and consistent referencing for subsequent isolation gating;

[0059] The message length is taken as the number of bytes of the IP layer payload and recorded as . The queue backlog is taken as the number of bytes currently waiting to be sent in the service flow's sending queue and recorded as . ;

[0060] The control command identifier is composed of the message field and Jointly determined, among which Defined as a ternary rule of "matching field position + matching mask + target value set", the communication device retrieves the value according to the mask after parsing the target field and compares it with the target value set. If a match is found, the control command flag is set. ,otherwise ;

[0061] Communication equipment is based on BT, UEID, local service configuration table and Generate business category identifiers and synchronously output sudden control event characteristics. Wherein BT is configured as a control class and Time setting In other cases, This ensures that the triggering conditions for "arrival of control service messages" are deterministic and reproducible within the device;

[0062] The communication device writes an enqueue timestamp to each message entering the queue and records it as such. The start time of the scheduling cycle is given by the local high-precision clock and denoted as . The latency budget is given by the local business configuration table and denoted as... And calculate the deadline margin accordingly. To characterize the remaining window before the current message times out, the cutoff time slack satisfies the formula:

[0063] ;

[0064] in Indicates the remaining time before the deadline. Indicates the message enqueue timestamp. This indicates the latency budget for this service flow. Indicates the start time of the current scheduling cycle;

[0065] To ensure the comparability of service state vectors across different scheduling cycles and service flows, the communication device normalizes continuous value fields using a fixed-dimensional conversion rule. This fixed-dimensional conversion rule is stored as a constant in the device and is not adaptively updated during runtime. and Each component is scaled proportionally to a preset upper limit constant, and any portion exceeding the upper limit is truncated to the upper limit. With interval The probability of success for the target within is directly encoded. Write directly using discrete encoding;

[0066] Finally, the communication equipment strictly follows the preset field order. , The business state vector is generated by concatenation, and the business state vector and the corresponding business flow identifier are written into the local state cache for use in the subsequent step S3 graph construction.

[0067] In this specific embodiment, S2 includes:

[0068] Used to collect the running status of candidate communication links and candidate network slices and generate link slice status vectors under the same scheduling cycle benchmark as step S1.

[0069] Communication equipment divides available access bearers into a set of public network candidate links. With private network candidate link set Each candidate link It corresponds to a specific set of cellular access configuration and dialing parameters and has a unique link identifier within the device. ;

[0070] The communication equipment uses the measurement reporting interface of the cellular modem during a preset sampling period. The preset sampling period is used to retrieve the radio and link layer metrics for each candidate link. Set it to 100ms in the device firmware and keep it unchanged during runtime;

[0071] For each candidate link The fields collected include the reference signal received power. Signal-to-interference-plus-noise ratio Block error rate Modulation coding method Physical resource block occupancy rate Round-trip delay Packet loss rate and cell load ;

[0072] in and The results are directly provided by the baseband measurement and link adaptation module. The address of the user plane probe on this link is obtained by the device performing fixed-length ICMP or UDP echo probes and using the most recent one. The median of the sub-probes was used as From the recent The percentage of lost probe messages is calculated to determine the cell load. Obtained by parsing fields broadcast by the base station or measured and reported by the network side;

[0073] Communication equipment simultaneously maintains a set of candidate network slices Each candidate network slice Identify by slice Uniquely identifiable and related to the access domain identifier The status fields of candidate network slices are bound and sampled at the same period by the network-side management interface through the device's built-in slice status agent. Pull and cache locally, the status field includes service level remaining quantity. Available resources Isolation level Access domain identifier User plane function network element reachability Edge computing node reachability and alarm status ;

[0074] in The ratio is calculated by converting the difference between "daily used quota" and "daily total quota". It is expressed as the proportion of the remaining PRB or remaining bandwidth of the slice issued by the network side to the total number of slices. With respect to the security isolation level in step S1 Use the same level of coding system to ensure consistent gating later. The device performs application-layer heartbeats on the user plane probe address bound to the slice and continuously Second successful placement Otherwise set The device performs a TCP handshake on the edge service probe address bound to the slice and uses it continuously. Second successful placement Otherwise set Network-side alarm codes are mapped to binary alarm flags to ensure that the field dimensions are fixed.

[0075] To achieve time alignment of indicators from different sources, the communication device writes a unified sampling timestamp for each sampling moment. And for each field, the "most recent time no later than" rule is applied. The effective observations are aligned, and the time alignment rule satisfies the following conditions. and efficient ,in Indicates alignment to the sampling time. field values, Indicates the field in the original time The observed values, Indicates the most recent valid observation timestamp used for alignment. This indicates taking the maximum value from the set of times that meet the conditions;

[0076] When in window If no valid observations are found, this field is set to the default code, and a default flag is simultaneously set in the vector. The window length... Set the time to 500ms in the device firmware and make it consistent across all fields;

[0077] In terms of dimension conversion, the communication equipment normalizes all continuous value fields to a preset dimension conversion rule. The interval is then truncated, where according to Linear mapping, according to Linear mapping, According to the upper limit constant Scaling Scaled to the maximum encoding value of 28. and Write directly as a ratio value, discrete field and Write using enumeration encoding, binary field and by Write;

[0078] Finally, the communication device will sort each candidate link according to the field order. of , With each candidate network slice of Concatenate to generate link slice state vectors, and use and The index is written to the local state cache for later use when S3 builds a multi-layered heterogeneous state graph.

[0079] In this specific embodiment, S3 includes:

[0080] Used to organize service state vectors and link slice state vectors into a multi-layer heterogeneous state diagram according to the bearer and constraint relationship of "service-interface-link-slice-base station-security domain" within the current scheduling cycle;

[0081] Communication equipment at the start of the scheduling cycle Trigger the graph construction process and use the scheduling cycle number Identify the current graph batch, which includes the set of business flows. Taken from the set of service flow identifiers and terminal interface sets that have been established in step S1 and contain pending messages in the queue. The candidate link set is the set of interface identifiers within the communication device that can be used for cellular access, and each interface identifier is uniquely determined within the device. Take the set of public network candidate links maintained in step S2. With private network candidate link set Union of candidate network slices Take the candidate network slice set and base station cell set maintained in step S2. Take step S2 for each candidate link The set of parsed and deduplicated cell identifiers, with each cell identifier uniquely represented by CellID, is the set of security isolation domains. The local isolation domain mapping table is fixed and the isolation domain identifier is uniquely represented by DID;

[0082] The multi-layered heterogeneous state diagram constructed by communication equipment is denoted as follows: ,in Indicates the first The multi-layered heterogeneous state diagram corresponding to each scheduling cycle. Represents a set of nodes. Denotes the set of edges. This represents the set of node features, where each node feature is a fixed-length vector. It represents a set of edge features, and each edge feature contains an edge type code and an edge weight;

[0083] In terms of node construction, the communication equipment configures each service flow. A business node is established and the business state vector of the business flow generated in step S1 is written into the features of the business node. The order of the fields of the business state vector is consistent with that in step S1 to ensure that the input dimension of subsequent graph embedding is fixed.

[0084] Communication equipment for each terminal interface Establish a terminal interface node and write the interface type code and interface availability flag as node characteristics. The interface availability flag is determined by the interface dialing status and is based on... express;

[0085] The communication equipment for each candidate link Establish a candidate link node and apply step S2 to the collected and normalized data of the candidate link. Write the candidate link node characteristics in a fixed field order;

[0086] Communication equipment for each Candidate network slices Create a slice node and apply step S2 to the slice after collection and normalization. Write the slice node features according to a fixed field order;

[0087] The communication equipment assigns a CellID to each base station cell. A base station node is established and its aggregated load characteristics are written as node characteristics. The aggregated load characteristics are taken from all candidate links connected to this CellID. The arithmetic mean of the aggregated load characteristics is equal to that of the candidate link when there is only one candidate link connected to the CellID. ;

[0088] Communication equipment for each security isolation domain A security domain node is established, and the isolation domain level code and domain availability flag are written as node characteristics. The domain availability flag is determined by the local security policy state and is based on… express;

[0089] In terms of edge construction, the communication device binds each service flow according to the local interface binding table. Bind to a unique terminal interface An access edge is established between the business node and the terminal interface node to represent the relationship of "business access via interface";

[0090] The communication equipment generates a set of allowed services to slices based on the local slice availability table and the security isolation domain mapping table. The local slice availability table, indexed by service type and security isolation level, provides a set of allowed slice identifiers and is permanently stored within the device, not adaptively changing during operation. This allows for the management of each service flow. Candidate bearer edges are established between the business node and all allowed slice nodes to represent the relationship that "the business can be carried by this slice";

[0091] The communication equipment based on each candidate link The resolved CellID A radio access edge is established between the candidate link node and the corresponding base station node to represent the relationship that "the link is attached to the cell";

[0092] Communication equipment according to each slice isolation level The unique security isolation domain is determined by the local isolation domain mapping table. And establish isolation edges between slice nodes and security domain nodes to represent the relationship that "slices belong to the isolation domain";

[0093] Communication equipment can handle any two different service flows and When performing shared resource determination, if the local slice availability table determines that there is at least one shared allowed slice or the local interface binding table determines that the two are bound to the same terminal interface, a competitive edge is established between the corresponding business nodes to represent the relationship of "shared resources leading to potential competition".

[0094] Regarding edge feature writing, the communication device writes the edge type code ET and edge weight for each edge. The edge type encoding ET uses a one-hot vector of length 5 and corresponds to the access edge, candidate bearer edge, wireless access edge, isolation edge, and contention edge in a fixed order, with edge weights... The access edge, candidate bearer edge, wireless access edge, and isolation edge are all set to 1 to indicate that the relationship exists and its strength is fixed. The edge weight of the contention edge is set accordingly. Set as the normalized queue backlog in both business nodes. The larger one is used to reflect the intensity of competition and to ensure that the calculation process is reproducible;

[0095] After the drawing is completed, the communication equipment will Write the node type index and edge type index into the local state cache in the form of an adjacency list, so that the subsequent step S4 can perform graph embedding processing according to the node type and edge type.

[0096] In this specific embodiment, S4 includes:

[0097] Used for multi-layer heterogeneous state diagrams The execution graph embedding is performed and the state changes of multiple consecutive scheduling cycles are temporally encoded to generate joint scheduling state features for subsequent scheduling decisions;

[0098] The communication device maintains a set of node type embedding parameters for each node type in its local state cache. ,in This indicates the node type, and its values ​​are limited to service nodes, terminal interface nodes, candidate link nodes, slice nodes, base station nodes, and security domain nodes. It is a trainable vector of length 8 and is stored in the device memory after the model training is completed;

[0099] The communication device also maintains a set of edge type transformation parameters for each edge type. ,in The edge type is indicated and corresponds one-to-one with the edge type code ET in step S3, and is limited to access edge, candidate bearer edge, wireless access edge, isolation edge, and contention edge. This represents the layer index of the graph embedding network, with values ​​of 0 and 1. For dimension The trainable matrix is ​​then stored in the device memory after the model training is complete.

[0100] Since different types of nodes have different original feature dimensions, the communication device configures a set of input projection parameters for each node type. Used to generate the original node feature vector of this type of node Projected onto a fixed dimension of 64, where This represents the set of nodes that are identified by their node identifier and belong to step S3. Represents a node During the scheduling period The original node feature vectors are written in step S3. For dimension Trainable matrix and Let be the original feature dimension of this node type, and let be the initial node representation after projection. It is then compressed back to 64 dimensions via a linear layer to ensure that all nodes have the same dimension when entering the graph neural network. Represents a node Node type, This indicates vector concatenation;

[0101] Graph embedding networks employ a two-layer heterogeneous message passing structure and... All nodes perform updates, in the... layer to the first During layer updates, the communication device updates each node. Aggregate neighbor node information based on its inbound edge set, and combine edge type and edge weight. The edge weights are weighted. Consistent with step S3, and reflecting the competition intensity for competing edges, while setting the value to 1 for the remaining edges, the updated relation satisfies the formula:

[0102] ;

[0103] in Represents a node During the scheduling period The Layer node embedding vector, Indicates the corresponding first Layer node embedding vector, Indicates the first The self-loop transformation matrix of the layer has a dimension of This represents the set of edges constructed in step S3. Indicates from node Pointing to node One of the edges, This indicates the edge weight. This indicates the edge type index corresponding to the edge and is consistent with the edge type code ET. This indicates that the edge type is in the first... The layer corresponds to the edge type transformation matrix with dimension . This represents a non-linear activation function with ReLU enabled.

[0104] After completing the two-layer update, the communication device extracts the service node embedding, candidate link node embedding, and slice node embedding from the final layer output according to node type and records them as follows: ,in This indicates the business node identifier corresponding to the business flow and originates from the business flow set. Indicates the candidate link node identifier and comes from the candidate link set. Indicates the slice node identifier and that it comes from the candidate network slice set. The candidate link node embeddings and slice node embeddings are then subjected to type-based average pooling to obtain the global link features for the current period. Global features of slices ,in For all The arithmetic mean vector with dimension . For all The arithmetic mean vector with a dimension of 64;

[0105] To incorporate the changing trends across multiple consecutive scheduling cycles, the communication device stores the most recent data in a circular buffer. Each scheduling cycle and When the buffer has less than 8 periods, a zero vector is used to fill in the missing periods while maintaining a constant sequence length. The resulting sequence is then concatenated in chronological order. Input a timing encoder to output trend state features ,in Indicates the first The link slice features of each cycle are concatenated with a dimension of 128. The time encoder uses a single-layer gated recurrent unit (GRU) with an input dimension of 128 and a hidden state dimension of 64. The hidden state of the last time step is taken as the trend state feature. And the dimension is 64;

[0106] Finally, the communication equipment performs each service flow... Construct its joint scheduling state feature vector:

[0107] ;

[0108] in Used to characterize the current topology embedding of business nodes, the global state of candidate links and slices, and cross-cycle trend information. The sudden control event characteristics determined in step S1 are set to either 0 or 1, and all business flows are... Stacked in order of service flow identifiers, they form a joint scheduling state feature tensor. Write to the local state cache for direct reading and use in subsequent steps S5.

[0109] In this specific embodiment, S5 includes:

[0110] Used to transform the joint scheduling state feature tensor Input the trained constrained executor-evaluator scheduling model to generate candidate scheduling actions and complete the construction and screening of a set of candidate actions based on constraints;

[0111] The communication device stores the network parameters of the constrained executor-evaluator scheduling model in its local memory and in each scheduling cycle number. Upon arrival, a forward inference is performed once, where for each business flow... Read its joint scheduling state feature vector Its composition is consistent with step S4 and includes the characteristics of a sudden control event. The constrained executor-evaluator scheduling model consists of an executor network and an evaluator network, employing a two-way structure with shared input. The executor network comprises a shared feature extraction trunk and a multi-head output layer. The shared feature extraction trunk is a two-layer fully connected network with layer widths of 256 and 128 respectively. The first layer has an input dimension of... The dimension is 257 in this embodiment. ReLU activation is used between the two layers, and batch normalization is not used to ensure device-side inference determinism. The trunk output is denoted as And the dimension is 128;

[0112] In the output layer of the executor network, set the slice selection header, link selection header, queue weight header, resource reservation ratio header, handover hold time header, and backup link policy header. The slice selection header outputs the candidate network slice set. Unnormalized scoring The link selection header outputs a set of candidate communication links. Unnormalized scoring The queue weight head outputs the normalized queue weight. Convert the queue scheduling weight to an integer before writing the action parameters. To ensure that the local scheduling table fields are definite integers, the resource reservation ratio header outputs the resource reservation ratio. It is used for subsequent slice capacity constraint calculations, and the switch hold time header outputs the switch hold time. And the values ​​are limited to a set. The indicated number of scheduling cycles is used to ensure controllable handover jitter, and the backup link policy header outputs the backup policy. And the value is limited to These indicate that backup is disabled, cold backup is enabled, and hot backup is enabled, respectively, and the subsequent local dialing strategy will be consistent.

[0113] The enforcer network converts the unnormalized scores of the slice selection head into an action probability distribution for slice selection using softmax. The transformation satisfies the formula:

[0114] ;

[0115] in Indicates business flow During the scheduling period Select candidate network slices The probability value, This indicates that the executor selects the network slice head for candidate network slices. The unnormalized score of the output, This represents the set of candidate network slices maintained in step S2. express any candidate network slice index in the database. Represents an exponential function;

[0116] Link selection header for candidate communication link set The action probability distribution for link selection is generated using the same softmax method as described above. and Together they constitute the action probability distribution and action parameters of candidate scheduling actions;

[0117] The evaluator network outputs state-action values ​​and consists of two fully connected layers with layer widths of 256 and 128, respectively. Its input is the joint scheduling state feature vector. With candidate scheduling action encoding vector The concatenated vector, where One-hot encoding by slice selection, one-hot encoding by link selection, one-hot encoding by backup strategy, and continuous parameters and and The values ​​are obtained by concatenating fields in a fixed order to ensure a fixed dimension for the action representation. The evaluator network outputs scalar state-action values. And serve as the basis for ranking and filtering candidate actions;

[0118] The communication device constructs a set of candidate scheduling actions based on the output of the executor network. The construction method is from Select the three candidate network slices with the highest probability and from The two candidate communication links with the highest probability are selected and combined using Cartesian methods to obtain at most six discrete "slice selection-link selection" combinations. The same set of continuous action parameters is then attached to each discrete combination. Candidate scheduling actions are generated;

[0119] In terms of constraint handling, communication equipment is constructed... Simultaneously apply slice capacity constraints, isolation constraints, and switchover count constraints, where the slice capacity constraint is based on the available resources of the slice collected in step S2. The upper bound is set, and the slice reservation ratio accumulation table is maintained within this scheduling cycle. For each candidate scheduling action, if its selected slice is Then determine And after passing the judgment Updated to This ensures that the reserved actions of multiple service flows within the same period do not exceed the available resources of the slice;

[0120] Isolation constraints use the security isolation level in step S1 Isolation level of slices in step S2 Compare with the same coding system and require And at the same time, candidate network slices are required. With business flow Multilayer heterogeneous state diagram constructed in step S3 There are candidate bearer edges to ensure that actions are generated only from the set of allowed bearers;

[0121] The handover count constraint is implemented through a local handover counter. accomplish, Record business flow The number of slice or link changes that occurred in the last 20 scheduling cycles is updated in a circular queue. If a candidate scheduling action causes a change in slice or link selection relative to a scheduling instruction that has taken effect in the previous cycle, the predicted handover count is incremented by one, and a further requirement is made. Only then can the candidate scheduling action be retained;

[0122] In the candidate scheduling action set After the construction is completed, the communication device will assign each candidate scheduling action to a specific action. The input evaluator network obtains state-action values ​​and sorts them from highest to lowest. Then, according to a preset selection rule, the candidate scheduling action with the highest state-action value that satisfies all the above constraints is selected as the business flow. Output the candidate scheduling actions and summarize the candidate scheduling actions of all business flows into a candidate scheduling action set. Write to the local state cache as input for S6 security gating processing.

[0123] In this specific embodiment, S6 includes:

[0124] The set of candidate scheduling actions used to output step S5 and write it to the local state cache. Perform confidence risk calibration, action conflict graph root cause analysis, and deterministic safety gating and event preemption gating to generate the final scheduling instruction and write it to the local scheduling table;

[0125] Communication devices embed a set of deterministic security gating rules in their firmware. , Whitelist of power control businesses First delay threshold First reliability threshold Private network and public network isolation rules Maximum slice size Switch jitter limit and control command preemption rules Composed of components that do not adaptively update during runtime, among which by For key storage control type services, the preemptible service items are those whose BT is consistent with step S1. Fixed Fixed Fixed requirements when business flow Security isolation level satisfy Only the private network candidate link set can be selected at this time. And only the slice isolation level can be selected. Candidate network slices A fixed value of 0.95 is used to limit the cumulative reservation ratio of the same slice within a single cycle to no more than the available resource quantity of the slice. of , A fixed value of 2 is used to limit the number of slice or link switching operations in the most recent 20 scheduling cycles to no more than 2. Fixed rules when control command identifier And the characteristics of sudden control events The queue scheduling weight will be an integer. Revised to 10 and the resource reservation ratio The value should be revised to be no less than 0.20, and public network handover for this service flow should be prohibited simultaneously.

[0126] Before entering the deterministic safety gating, the communication equipment considers the set of candidate scheduling actions. Perform confidence risk calibration to obtain the confidence risk vector. The calibration sample window used for confidence risk calibration is fixed and taken from the nearest sample. Each scheduling cycle and the sample fields are fixed and include the business queue length, , And historical action results, the communication equipment for each candidate scheduling action The predicted latency is obtained by calling the performance prediction output of the constrained executor-evaluator scheduling model. Prediction reliability Predicted bandwidth satisfaction probability And predict the probability of handover failure The performance prediction output is implemented by a performance prediction network embedded in the device, which is a two-layer fully connected network with a defined structure. The input is the joint scheduling state feature vector from step S4. With the candidate scheduling action encoding vector of step S5 The concatenated vectors have a first layer width of 128 and use ReLU activation. The second layer has an output width of 4, and the delay output is constrained to a positive number using softplus, and the reliability and probability outputs are constrained to a positive number using sigmoid. interval;

[0127] The communication equipment calculates the error distribution between the predicted and actual observed values ​​of historical actions performed within the calibration sample window and applies it to the target coverage. The corresponding quantile values ​​are calculated, and the error distribution is fixed using a delay positive error distribution, a reliability negative error distribution, a bandwidth satisfaction negative error distribution, and a handover failure positive error distribution. This calibrates the prediction results of candidate scheduling actions into a confidence risk vector, which satisfies the formula:

[0128] ;

[0129] in Indicates business flow During the scheduling period For candidate scheduling actions The confidence risk vector, This indicates the upper bound of the calibrated latency risk. This indicates the lower bound of the reliability risk after calibration. This indicates that the calibrated bandwidth meets the lower probability bound. This indicates the upper bound of the risk of handover failure after calibration. This represents the prediction latency of the performance prediction network output. This indicates the reliability of the prediction output of the performance prediction network. This represents the probability that the predicted bandwidth output by the performance prediction network will meet the requirement. This represents the predicted handover failure probability output by the performance prediction network. Indicates the target coverage rate The following are the quantile values ​​obtained from the distribution of positive time delay errors. Indicates the target coverage rate The following are the quantiles obtained from the negative reliability error distribution. Indicates the target coverage rate The quantile values ​​are obtained from bandwidths that satisfy a negative error distribution. Indicates the target coverage rate The following are the quantile values ​​obtained from the positive error distribution of switching failures;

[0130] After the confidence risk vector is generated, the communication device... The process of performing slice action conflict graph processing outputs the conflict type, conflicting business, and conflicting resource root cause, which serve as additional inputs for subsequent gating. The conflict graph is denoted as... And includes a set of action nodes Business node set Slice node set Link node set With security domain node set The communication equipment will schedule each candidate action. Establish as an action node and connect it to its corresponding business flow via an index. Candidate network slices Candidate communication links With the security isolation domain DID, and establish resource mutual exclusion edges between action nodes and slice nodes to express "the cumulative reservation ratio of the same slice exceeds the limit". The mutual exclusion relationship is expressed by establishing isolation constraint edges between business nodes and security domain nodes. To address the violation relationship, a preemptive edge is established between the business nodes corresponding to control-type services and the business nodes corresponding to ordinary data collection services to express the "new resource consumption by ordinary services leading to control services" violation. or The influence relationship is explained by establishing handover jitter edges between link nodes to represent "the number of handovers exceeds..." The relationship between "";

[0131] Communication equipment Execute the determined traversal rules and output the conflict type. Conflicting business sets Root causes of conflict resources When resource mutual exclusion edges exist, the setting is... and Record the corresponding slice identifier The accumulated value of exceeding the limit is set when there is an isolation constraint edge. and Record the violation isolation domain identifier (DID) and the violation business identifier, and reset them when there is a preemption edge. pre and Record ordinary business identifiers that lead to risk exceeding limits, and set them when there is switching jitter edge. and Record the corresponding service identifier and the switching counter value;

[0132] During the deterministic security gating phase, the communication equipment gates each service flow. The candidate scheduling actions are valued according to the state and action given in S5. Traverse and perform motion mask judgment and motion correction from high to low, with motion mask judgment based on... and If the candidate scheduling action is executed jointly and any condition is met, a mask flag is set for that action. The conditions include The selected slices caused the slice reservation ratio to accumulate to more than [a certain percentage]. Action selection link violation The action caused the switching counter to exceed Conflict types Root causes of conflict resources This indicates that the action occurred on a conflicting link or conflicting slice;

[0133] When there are candidate scheduling actions that are not masked, the communication device takes the first one after sorting. The candidate scheduling actions are used as the gating mechanism for the current service flow, and action corrections are performed. These action corrections include at least the following: right and The determination of the correction and when And when the conflict slice identifier hits the slice selected by the action, it will Revised to ensure that the accumulation does not exceed The maximum allowed value and keep slice selection and link selection unchanged;

[0134] When no candidate scheduling action is not masked, the communication device generates a safety backup scheduling instruction and sets it as the final scheduling instruction. The safety backup scheduling instruction is fixed to maintain the slice selection that was effective in the previous scheduling cycle. Link selection And will Revised to be no less than 5 and The value was increased to 8 to extend the hold time and thus suppress jitter.

[0135] During the event preemption and gating phase, after completing the aforementioned masking and correction, the communication equipment re-checks the characteristics of the sudden control event. With control command identifier ,when and Furthermore, the corresponding control-type business still satisfies the requirements under the current gate access action. or At that time, the communication equipment based on and For new resource occupancy actions of ordinary data collection services, execute action mask and forcibly correct their resource reservation ratio to 0; for public network switching actions of handheld terminal services, execute action mask and forcibly maintain their previous cycle link selection; execute action mask for all actions that violate security isolation levels.

[0136] Simultaneously, the communication equipment generates and immediately writes a set of preemption guarantee instructions for control-related services into the local scheduling table. , The fixed instruction includes at least one of the following: reserved slice activation instruction, private network link maintenance instruction, backup link activation instruction, and queue weight increase instruction, and the selection of which is determined by [the specific instruction]. Confirmed, when Indicate slice capacity root cause of master time enable reserved slice and will Revised to 0.30, when Indicates link jitter root cause is when the main timer is active. Corrected to 8 and maintains private network link, when Indicates the upper bound of the risk of root cause of link quality failure due to primary or switchover failure. When the value is greater than 0.10, the backup policy will be adjusted. Corrected to hot backup and forced selection of private network candidate links;

[0137] Finally, the communication device writes the final scheduling instruction obtained through deterministic security gating and event preemption gating into the local scheduling table Tab according to a fixed field order. The fixed fields include at least the service flow identifier, slice selection identifier, link selection identifier, and queue scheduling weight integer. Resource reservation ratio Switching hold time Backup strategy Effective period serial number The local scheduling table, along with the mask reason code, is directly queried and executed during subsequent message classification, link dialing, single network slice selection auxiliary information selection, quality of service flow identifier mapping, and queue scheduling processes.

[0138] 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.

[0139] This invention connects service identification, link slice status acquisition, multi-layer heterogeneous graph modeling, joint feature encoding, constrained scheduling model and security gating rules into a closed scheduling chain, enabling the power 5G communication module to form an executable local scheduling table when the public network, private network and network slice status change, thereby improving the resource competition problem between critical services and ordinary services in multi-service co-carrying scenarios.

[0140] This invention introduces confidence risk calibration, slice action conflict graph and event preemption gating after the candidate scheduling action is output. This allows scheduling actions to be constrained by risk boundaries, conflict root causes and safety bottom line instructions even when prediction is uncertain, action conflicts or sudden control commands arrive. Therefore, it can better meet the real-time, reliability and safety isolation requirements of power control services.

Claims

1. A power 5G slicing intelligent scheduling method, characterized in that, include: S1. Identify the power service flow of the access communication equipment and generate a service state vector; S2. Collect the operating status of candidate communication links and candidate network slices and generate a link slice state vector; S3. Construct a multi-layer heterogeneous state graph based on the service state vector and the link slice state vector; S4. Perform graph embedding processing on the multi-layer heterogeneous state graph and perform time-series coding processing on the link quality, service load and queue changes in multiple consecutive scheduling cycles to generate joint scheduling state features; S5. Input the joint scheduling state characteristics and sudden control event characteristics into the trained constrained executor-evaluator scheduling model to generate candidate scheduling actions; S6. Perform action masking and action correction on the candidate scheduling actions according to the deterministic safety gating rules to generate the final scheduling instruction and write the final scheduling instruction into the local scheduling table.

2. The intelligent scheduling method for 5G power slicing according to claim 1, characterized in that, S1 includes: extracting service type, service direction, message length, queue backlog, latency budget, reliability requirements, bandwidth requirements, security isolation level, and control command identifier from service messages, terminal access identifiers, and local service configuration tables; determining control-type services, ordinary collection-type services, handheld terminal services, and sudden control event characteristics based on the service type, terminal access identifier, local service configuration table, and control command identifier; calculating the deadline margin based on the timestamp of the message entering the queue, the latency budget, and the start time of the current scheduling cycle; and combining the service type, service direction, message length, queue backlog, latency budget, reliability requirements, bandwidth requirements, security isolation level, control command identifier, deadline margin, and sudden control event characteristics according to a preset field order to generate the service state vector.

3. The intelligent scheduling method for 5G power slicing according to claim 1, characterized in that, S2 include: The system collects reference signal received power, signal-to-interference-plus-noise ratio, block error rate, modulation and coding scheme, physical resource block occupancy rate, round-trip time, packet loss rate, and cell load for public network candidate links and private network candidate links, respectively. It also collects the remaining service level, available resources, isolation level, access domain identifier, reachability of user plane function network elements, reachability of edge computing nodes, and alarm status of candidate network slices. The collected operating status is time-aligned according to a preset sampling period, and the link slice status vector is generated according to the preset unit conversion rules corresponding to each field.

4. The intelligent scheduling method for 5G power slicing according to claim 1, characterized in that, S3 includes: Establish a service node for each service flow, a terminal interface node for each terminal interface, a candidate link node for each public network candidate link or private network candidate link, a slice node for each candidate network slice, a base station node for each base station cell, and a security domain node for each security isolation domain. An access edge is established between the service node and the terminal interface node; a candidate bearer edge is established between the service node and the slice node; a wireless access edge is established between the candidate link node and the base station node; an isolation edge is established between the slice node and the security domain node; and a competition edge is established between service nodes sharing the same candidate link or the same candidate network slice. The service state vector, the link slice state vector, and the edge type encoding are written into the corresponding node and the corresponding edge to obtain the multi-layer heterogeneous state diagram.

5. The intelligent scheduling method for 5G power slicing according to claim 1, characterized in that, S4 includes: setting type embedding parameters for different types of nodes and different types of edges in the multi-layer heterogeneous state graph; using a graph neural network to perform message passing based on the features of adjacent nodes, edge type encoding, and edge weights to generate service node embeddings, candidate link node embeddings, and slice node embeddings; inputting the service node embeddings, candidate link node embeddings, and slice node embeddings of multiple cycles into a time-series encoder according to the scheduling cycle order to generate trend state features; and concatenating the graph embedding features of the current cycle, the trend state features, and the sudden control event features to obtain the joint scheduling state features.

6. The intelligent scheduling method for power 5G slicing according to claim 1, characterized in that, S5 includes: The joint scheduling state characteristics are input into the executor network of the constrained executor-evaluator scheduling model to generate the action probability distribution and action parameters of each service flow in slice selection, link selection, queue weight, resource reservation ratio, handover hold time and backup link strategy; the joint scheduling state characteristics, the action probability distribution and the action parameters are input into the evaluator network to obtain the state action value corresponding to the candidate scheduling action; a candidate scheduling action set is generated according to the action probability distribution, the action parameters, the state action value and the slice capacity constraint, isolation constraint and handover number constraint. Candidate scheduling actions whose state action values ​​satisfy preset selection rules are selected from the candidate scheduling action set and used as input for security gating processing.

7. The intelligent scheduling method for 5G power slicing according to claim 6, characterized in that, S6 include: The deterministic security gating rules are composed of a whitelist of power control services, a first latency threshold, a first reliability threshold, private network and public network isolation rules, slice capacity limits, handover jitter limits, and control command preemption rules. Based on the joint scheduling state characteristics and candidate scheduling actions selected from the candidate scheduling action set, the predicted latency and predicted reliability of power control services are determined by the performance prediction output of the constrained executor-evaluator scheduling model. When a candidate scheduling action causes the predicted latency to exceed the first latency threshold, the predicted reliability to be less than the first reliability threshold, the candidate network slice to exceed the slice capacity limit, the candidate link to violate the private network and public network isolation rules, or the number of handovers to exceed the handover jitter limit, an action mask is applied to the corresponding candidate scheduling action. When there are unmasked candidate scheduling actions, a new candidate scheduling action is selected from the unmasked candidate scheduling actions according to the order of state action value. The queue weight, resource reservation ratio, slice selection or link selection are then adjusted according to the control instruction preemption rules to generate the final scheduling instruction. When there are no unmasked candidate scheduling actions, a safety-backup scheduling instruction that maintains the current slice and the current link is generated as the final scheduling instruction, and the final scheduling instruction is written into the local scheduling table.

8. The intelligent scheduling method for 5G power slicing according to claim 6, characterized in that, After the candidate scheduling action selected from the candidate scheduling action set is output and before proceeding to step S6, a confidence risk calibration process is also included: using the service queue length, reference signal received power, signal-to-interference-plus-noise ratio, block error rate, modulation and coding scheme, physical resource block occupancy rate, round-trip time, packet loss rate, cell load, user plane functional network element reachability, edge computing node reachability, and historical action results of the most recent multiple scheduling cycles as calibration samples, the delay error distribution, reliability error distribution, bandwidth satisfaction error distribution, and handover failure error distribution are calculated respectively; the corresponding quantile value is determined from each error distribution according to the preset target coverage rate; the prediction results of candidate services, candidate network slices, and candidate link combinations are calibrated according to the quantile value, and a confidence risk vector composed of the upper bound of delay risk, the lower bound of reliability risk, the bandwidth satisfaction probability, and the handover failure risk is output.

9. The intelligent scheduling method for power 5G slicing according to claim 8, characterized in that, After the confidence risk vector is output, the slice action conflict graph processing is also included: taking the candidate scheduling actions selected from the candidate scheduling action set, the remaining resources of the slice, the service priority, the security isolation level, the queue weight, the access domain identifier and the confidence risk vector as input, action nodes, service nodes, slice nodes, link nodes and security domain nodes are established. Establish resource mutual exclusion edges between action nodes and slice nodes, establish isolation constraint edges between service nodes and security domain nodes, establish preemption edges between service nodes corresponding to control-type services and service nodes corresponding to ordinary acquisition-type services, and establish handover jitter edges between link nodes; determine the conflict type, conflict service, and conflict resource root cause based on the resource mutual exclusion edges, isolation constraint edges, preemption edges, and handover jitter edges, and input the conflict type, conflict service, and conflict resource root cause into step S6.

10. The intelligent scheduling method for 5G power slicing according to claim 9, characterized in that, Step S6 also includes event preemption gating processing: when the sudden control event characteristics indicate the presence of control-type service messages, and the confidence risk vector indicates that the upper bound of the latency risk of the corresponding candidate service, candidate network slice, and candidate link combination exceeds the first latency threshold or the lower bound of the reliability risk is less than the first reliability threshold, the action mask is executed for the new resource occupation action of ordinary collection-type services, the public network switching action of handheld terminal services, and the action of violating security isolation level, based on the conflict type, conflict service, and conflict resource root cause. Generate reserved slice activation instructions, private network link maintenance instructions, backup link activation instructions, or queue weight enhancement instructions for control-type services, and write the generated instructions into the local scheduling table. The local scheduling table is used for subsequent packet classification, link dialing, single network slice selection auxiliary information selection, service quality flow identifier mapping, and queue scheduling.

Images