Method for intelligent routing optimization based on machine learning in software-defined networks

Abstract

The application discloses a method for intelligent routing optimization based on machine learning in a software-defined network, and relates to the technical field of communication networks, comprising the following steps: obtaining link state encoding and alleyway topology data of a target object; performing topology skeleton mapping on the alleyway topology data to generate a twin topology base map; performing time series difference operation on the link state encoding to generate a sequence of encoding difference values, and performing abnormality measurement on the sequence of encoding difference values to generate a sequence of abnormality evidence. According to the method, communication abnormality is searched in the twin topology base map, a block probability map is generated by forward accumulation and backward smoothing, and is updated to a reachability update twin graph, so that noise misjudgment is inhibited, reachability is continuously evolved with abnormality, an action mask is generated by double-layer mask synthesis of the reachability update twin graph and the block probability map, and an actionable candidate solution is constructed, thereby, unreachable and high-risk actions are eliminated in advance, and invalid exploration is reduced.

Description

Technical Field

[0001] This invention relates to the field of communication network technology, specifically to a machine learning-based intelligent routing optimization method in software-defined networks. Background Technology

[0002] With the development of the Industrial Internet, smart mines, and emergency communication technologies, the ability to ensure communication networks in complex environments has become one of the important research directions in the field of network technology. When sudden accidents such as collapses and roof falls occur in underground mines, the original communication infrastructure is easily damaged, the network topology changes drastically, and the link status exhibits highly unstable characteristics. Against this background, software-defined networking decouples the control plane from the forwarding plane, enabling the network to have centralized control, global awareness, and flexible scheduling capabilities, providing a new technical foundation for the reconstruction and dynamic optimization of communication networks in complex scenarios.

[0003] Existing routing optimization schemes are mainly based on centralized routing control schemes in software-defined networks. Specifically, the controller periodically acquires link status information and network topology information, calculates the shortest path or a path that meets specific constraints, and then sends the calculation results to the forwarding devices. These schemes have certain technical effects, such as improving link utilization, reducing average latency, or reducing the probability of congestion. However, these schemes assume that the network topology is logically complete and reachable, and only reflect changes in network quality through link weights or status indicators. In the scenario of underground mine collapse rescue networks, the collapse causes physical inaccessibility. For example, after a collapse in a mine tunnel, the corresponding area is physically impassable, but logical connections may still exist in the network topology. The above schemes will still include this area in the candidate path calculation range, thus generating invalid routes that are "logically reachable but physically unreachable," resulting in insufficient accuracy of the routing optimization scheme. Summary of the Invention

[0004] The purpose of this invention is to provide a machine learning-based intelligent routing optimization method for software-defined networks to solve the problems mentioned in the background section.

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

[0006] In a first aspect, this invention discloses a machine learning-based intelligent routing optimization method for software-defined networks, applied to intelligent routing optimization in underground mine collapse rescue networks, comprising the following steps:

[0007] Obtain the link-state code and tunnel topology data of the target object;

[0008] Perform topological skeleton mapping on the tunnel topology data to generate a twin topological base map;

[0009] The link-state coding is subjected to time-series differential operation to generate a sequence of coded differential values, and anomaly measurement is performed on the sequence of coded differential values ​​to generate an anomaly evidence sequence.

[0010] The twin topology base map is subjected to communication anomaly retrieval, and the path consistency inversion processing of the communication anomaly retrieval results in the neighborhood is performed according to the anomaly evidence sequence and the link state code to generate a blocking probability map. Then, the twin topology base map is updated according to the blocking probability map to generate a reachability update twin map.

[0011] Perform a two-layer mask synthesis process on the reachability update twin graph and the blocking probability graph to generate an action mask, and construct a candidate solution for possible actions using the action mask as a constraint;

[0012] Based on the abnormal evidence sequence, the blocking probability graph, and the reachability update twin graph, the joint cost solution and mapping relationship are established for the candidate solutions of the possible actions, generating a joint cost function. Based on the joint cost function, the candidate solutions of the possible actions are value-selected to generate a route optimization scheme.

[0013] Secondly, this invention discloses a machine learning-based intelligent routing optimization system for software-defined networks, comprising:

[0014] The data acquisition module is used to acquire the link status code and tunnel topology data of the target object;

[0015] The topology processing module is used to perform topology skeleton mapping on the tunnel topology data and generate a twin topology base map;

[0016] An anomaly analysis module is used to perform time-series differential operations on the link state coding to generate a sequence of coded differential values, and to perform anomaly measurement on the sequence of coded differential values ​​to generate an anomaly evidence sequence.

[0017] The update module is used to perform communication anomaly retrieval on the twin topology base map, and perform path consistency inversion processing on the communication anomaly retrieval results in the neighborhood according to the anomaly evidence sequence and the link state code to generate a blocking probability map, and then update the twin topology base map according to the blocking probability map to generate a reachability updated twin map.

[0018] The candidate solution construction module is used to perform a two-layer mask synthesis process on the reachability update twin graph and the blocking probability graph to generate an action mask, and construct feasible action candidate solutions with the action mask as a constraint.

[0019] The scheme generation module is used to solve the joint cost and establish the mapping relationship for the candidate solutions of the possible actions based on the abnormal evidence sequence, the blocking probability graph and the reachability update twin graph, generate a joint cost function, and perform value screening on the candidate solutions of the possible actions based on the joint cost function to generate a route optimization scheme.

[0020] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0021] 1. This scheme performs communication anomaly retrieval on the twin topology base map and generates a blocking probability map through forward accumulation and backward smoothing, which is then updated to the reachability update twin map. This suppresses noise misjudgment and makes reachability continuously evolve with anomalies. The two-layer mask of the reachability update twin map and the blocking probability map is synthesized to generate an action mask and construct candidate solutions for possible actions. This eliminates unreachable and high-risk actions in advance, reduces invalid exploration, and improves the accuracy of the subsequent route optimization scheme.

[0022] 2. This scheme establishes an evidence aggregation mapping based on the communication anomaly retrieval results, achieving unified alignment between anomaly landing points and twin structural units, avoiding drift of anomaly evidence in structural positioning. Based on the evidence aggregation mapping, corresponding evidence items are extracted from the anomaly evidence sequence to form an evidence set, and the evidence strength value is calculated, thereby quantifying dispersed anomalies into structural unit-level strength characterizations. Based on the encoded time series obtained by link state encoding time organization, the evidence set is differentially analyzed to obtain the evidence consistency value, identifying the difference between short-term fluctuations and continuous anomalies, reducing misjudgments caused by transient jitter. The evidence strength value and evidence consistency value are combined to generate a blocking probability map, which facilitates the hierarchical expression of neighborhood blocking risks. Attached Figure Description

[0023] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings, the same reference numerals are used to refer to the same parts. Wherein:

[0024] Figure 1 This is a flowchart illustrating the steps of the intelligent routing optimization method based on machine learning in software-defined networks according to the present invention.

[0025] Figure 2 This is a schematic diagram of the process for generating a blocking probability map provided by the present invention;

[0026] Figure 3 A schematic diagram of the process for generating an action mask provided by the present invention;

[0027] Figure 4 A schematic diagram of the process for constructing feasible action candidate solutions provided by the present invention;

[0028] Figure 5A schematic diagram illustrating the module functions of the machine learning-based intelligent routing optimization system in software-defined networks provided by this invention. Detailed Implementation

[0029] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0030] Application Overview:

[0031] In traditional software-defined network (SDN) routing optimization schemes, when applied to underground mine collapse rescue network scenarios, the collapse renders some tunnels physically inaccessible, yet logical connections still exist in the network topology. The controller struggles to effectively distinguish between logically reachable and physically inaccessible states, thus including physically inaccessible areas in the candidate path calculation, resulting in invalid routes that are logically reachable but physically unreachable. This problem stems from existing technologies reflecting network quality changes solely through link weights or state indicators, without considering the actual impact of physical blockages on the network topology. Furthermore, this issue reduces the accuracy of routing optimization schemes, thereby affecting key performance indicators such as network communication reliability and real-time performance.

[0032] For example, after a collapse in an underground mine, tunnel B is completely blocked and physically impassable. However, the network nodes in this area maintain power and communication, and the link status information periodically acquired by the controller shows that the logical connection in this area is normal. In this situation, when the controller calculates the route based on the shortest path algorithm, it uses tunnel B as a valid path node and generates a routing scheme that includes this area. When data packets are forwarded to this area, due to the physical blockage, the data packets cannot continue to be transmitted, causing communication interruption. The rescue command center cannot obtain sensor data in this area, affecting the timeliness of rescue decisions. In this scenario, the contradiction between physical inaccessibility and logical accessibility directly leads to the failure of the routing scheme. Furthermore, network resources are ineffectively occupied, and the results of communication anomaly retrieval cannot accurately reflect the physical spatial status.

[0033] If the above problems are not addressed, invalid routing will lead to increased packet loss rate, higher probability of network congestion, and a continuous decline in communication reliability. In particular, the consequences of this technology are especially serious in emergency rescue scenarios, where communication interruptions may hinder the transmission of critical information and adversely affect the safety of rescue personnel and the efficiency of accident handling. Therefore, there is an urgent need for a technical solution that can identify physical spatial obstructions and dynamically update network topology to improve the adaptability of routing optimization.

[0034] After introducing the basic concept of the present invention, the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0035] Example 1:

[0036] Please see Figure 1 A machine learning-based intelligent routing optimization method in software-defined networking is applied to intelligent routing optimization in underground mine collapse rescue networks, including the following steps:

[0037] Obtain the link-state code and tunnel topology data of the target object;

[0038] Perform topological skeleton mapping on the tunnel topology data to generate a twin topological base map;

[0039] Perform time-series differential operations on the link-state coding to generate a sequence of coded differential values, and perform anomaly measurement on the coded differential value sequence to generate an anomaly evidence sequence;

[0040] Communication anomaly retrieval is performed on the twin topology base map, and path consistency inversion processing is performed on the communication anomaly retrieval results in the neighborhood based on the anomaly evidence sequence and link state coding to generate a blocking probability map. Then, the twin topology base map is updated based on the blocking probability map to generate a reachability updated twin map.

[0041] Perform a two-layer mask synthesis process on the reachability update twin graph and the blocking probability graph to generate an action mask, and construct candidate solutions for possible actions using the action mask as a constraint;

[0042] Based on the abnormal evidence sequence, the blocking probability graph, and the reachability update twin graph, the joint cost of the candidate solutions for possible actions is solved and the mapping relationship is established to generate a joint cost function. Based on the joint cost function, the candidate solutions for possible actions are value-selected to generate a route optimization scheme.

[0043] Link state coding refers to a fixed-length bit sequence or fixed-length byte sequence data object formed by any network device in the underground mine collapse rescue network sampling, normalizing, and integrating the operating status of at least one link under its jurisdiction in one go within the sampling period.

[0044] Tunnel topology data refers to graph-structured data objects used to characterize the spatial connectivity of tunnels within an underground mine rescue area;

[0045] Twin topology basemaps refer to graph-structured data objects used to represent the spatial connectivity structure and deployable location constraints of underground mine tunnels;

[0046] The encoded differential value sequence refers to the numerical sequence formed by arranging the link state codes obtained at consecutive sampling times of the same target link in chronological order after calculating the differential measurement between adjacent time times;

[0047] An anomaly evidence sequence refers to a set of anomaly evidence values ​​arranged in chronological order, formed by link state codes obtained at continuous sampling times based on the same link or the same pair of communication endpoints, after time-series organization, differential measurement and anomaly measurement.

[0048] Communication anomaly retrieval results refer to a set of structured data obtained at the structural unit granularity of the twin topology base map, which characterizes the distribution of communication anomalies in the topological space by performing structured alignment and spatial mapping processing on the anomaly evidence sequence and link state code, using the twin topology base map as the structural carrier.

[0049] A blocking probability graph is a structured probability mapping data graph constructed by a software-defined network controller based on network observation results of an underground mine collapse rescue communication network. It is used to quantitatively characterize the probability of communication blocking or impending communication blocking occurring in each structural unit of the twin topology base map.

[0050] Reachability update twin graph refers to a graph data object formed by quantifying and updating the communicability status of each structural unit in the twin topology base graph, which is used to characterize the joint state of roadway structure and communication reachability at the current moment.

[0051] Action mask is an action constraint data structure used in machine learning-driven route optimization to impose prior feasibility restrictions and risk weight modulation on candidate route actions.

[0052] A feasible action candidate solution refers to a set of candidate routing actions that are topologically reachable, risk-acceptable, and action-permitted.

[0053] The joint cost function is a composite function used to uniformly characterize the overall cost level of candidate solutions in terms of communication stability, physical reachability, consistency of blocking risk, and continuity of topology changes.

[0054] A route optimization scheme refers to a set of structured route decision data that can be directly issued and executed after joint cost calculation and screening of candidate solutions for feasible actions.

[0055] This scheme generates a twin topological base map by mapping the tunnel topology data into a topological skeleton, which compresses the tunnel connection relationship into a computable structural carrier, reduces the computational redundancy caused by full map expansion and maintains consistency in subsequent retrieval and positioning. After performing temporal differential operation on the link state code to generate a sequence of coded differential values ​​and performing anomaly measurement on the coded differential value sequence to generate anomaly evidence sequence, the link state mutation is quantified into a comparable anomaly intensity sequence, reducing the dependence on fixed thresholds and field statistics and improving the stability of anomaly discrimination.

[0056] Communication anomaly retrieval is performed on the twin topology base map. Based on the anomaly evidence sequence and link state coding, the path consistency inversion of the communication anomaly retrieval results is performed in the neighborhood to generate a blocking probability map. The twin topology base map is then updated accordingly to generate an reachability update twin map. This allows communication anomalies to be mapped into a probabilistic expression of road blockage and written back into the twin structure. This enables reachability to be updated synchronously with changes in the field to reduce erroneous route selection. A two-layer mask synthesis is performed on the reachability update twin map and the blocking probability map to generate an action mask. After constructing candidate solutions for actionable actions by constraining the action mask, unreachable actions are suppressed in the generation stage and high-risk actions are weighted and constrained, reducing invalid exploration and invalid rerouting from the source.

[0057] Based on the abnormal evidence sequence, the blocking probability map, and the reachability update twin graph, the joint cost solution of the candidate solutions for actionable actions is established to generate a mapping relationship and a joint cost function. Based on this value screening, a route optimization scheme is generated, so that the candidate solutions for actionable actions can achieve a unified quantitative ranking of stability and risk under the same cost framework. The primary and backup paths and switching rules that can be directly issued are output to improve the continuity of rescue communication.

[0058] The above describes a complete scheme for intelligent routing optimization based on machine learning in software-defined networks. The following section describes how to obtain the link-state coding and lane topology data of the target object, specifically including:

[0059] The controller sends a status reporting trigger command to the switching or relay device to which the target object belongs and receives the status code object returned by it to obtain the link status code of the target object. The link status code includes, but is not limited to, the status expression of whether the link is connected, the status expression of changes in link forwarding behavior, and the status expression of the trend of link load change.

[0060] The roadway topology data of the target object is obtained through the mine information management data interface; the roadway topology data includes, but is not limited to, the expression of the connection relationship between roadway nodes, the expression of the adjacency relationship between roadway segments, and the description of the overall structural connectivity of the roadway.

[0061] The above describes how to obtain the link-state code and tunnel topology data of the target object. The following describes how to perform topology skeleton mapping on the tunnel topology data to generate a twin topology base map, specifically including:

[0062] The tunnel topology data is compressed and mapped to generate a topology skeleton sequence, and the topology skeleton sequence is indexed and labeled to generate a deployment skeleton index.

[0063] Based on the deployed skeleton index, the tunnel topology data and topology skeleton sequence are structured and encapsulated into a graph structure to generate a twin topology base map.

[0064] Among them, the topology skeleton sequence refers to a linearized structural data sequence used to express in compressed form the core structural units and their connection relationships in the roadway topology that play a decisive role in communication deployment and path connectivity;

[0065] Deploying a skeleton index refers to the unique set of structural index identifiers assigned to each skeleton structural unit based on the topological skeleton sequence obtained by compressing the tunnel topological data.

[0066] The above content will be described in detail below:

[0067] The tunnel topology data is compressed and mapped using a topology skeleton to generate a topology skeleton sequence. This sequence is then indexed and annotated to generate a deployment skeleton index.

[0068] The connection relationship expansion operation is performed on the roadway topology data, and all connection relationships in the roadway topology are integrated into a "connection pair set" to form a connection relationship set. Then, the node degree statistics operation is performed on the connection relationship set to obtain the number of connections for each node, thus forming a node degree set. Subsequently, the skeleton segmentation operation is performed on the connection relationship set based on the node degree set. The calculation process of the skeleton segmentation operation is as follows: the node degree is not equal to two as the segmentation boundary, and the continuous connection pairs in the connection relationship set whose first and last ends are the segmentation boundaries are concatenated in the topological adjacent order to obtain a continuous segment set. Then, each continuous segment set is integrated to form a skeleton segment set.

[0069] To achieve compression mapping, an isomorphic merge operation is performed on the skeleton segment set. The calculation process of the isomorphic merge operation is to extract the segment length value and endpoint metric pair of each skeleton segment to form a segment feature pair, and to perform equivalent merge on the segment feature pairs of all skeleton segments. Skeleton segments with the same segment feature pair are grouped into the same merge group, and the representative segment of the merge group replaces each skeleton segment in the group, thereby obtaining the compressed skeleton segment set.

[0070] The compressed skeleton segment set is subjected to sequential encoding operation. The calculation process of sequential encoding operation is to construct a segment adjacency table based on the endpoint adjacency relationship of the compressed skeleton segment set, and generate a segment number table based on the segment adjacency table. For each skeleton segment, a correspondence relationship of "segment number - adjacent number set" is constructed with its adjacent segment number, thereby forming a topological skeleton sequence.

[0071] The topological skeleton sequence is subjected to indexing and annotation operations. The calculation process of the indexing and annotation operations is as follows: for each entry in the topological skeleton sequence, extract its segment number and the set of adjacent numbers, generate segment index codes and integrate them to form segment index entries. At the same time, for the endpoint positions of each segment entry, generate endpoint index codes and integrate them to form endpoint index entries. The segment index code is generated by concatenating the segment number with a fixed-length check bit to obtain a unique code. The endpoint index code is generated by concatenating the segment number to which the endpoint belongs with the endpoint orientation identifier to obtain a unique code.

[0072] Entries with the same index code are merged into a single entry, and the merge count is summed to obtain the occurrence value. The occurrence value and the index code are written together into the index entry to characterize the index stability, thereby generating a deployment skeleton index.

[0073] Based on the deployed skeleton index, the tunnel topology data and topology skeleton sequence are subjected to structured mapping and encapsulated into a graph structure to generate a twin topology base map:

[0074] The skeleton unit extraction operation is performed on the topological skeleton sequence to form a skeleton unit set. The calculation process of the skeleton unit extraction operation is to read skeleton connection units and skeleton node units from each sequence position of the topological skeleton sequence, and combine the read skeleton connection units and skeleton node units to obtain the skeleton unit set.

[0075] Perform index alignment operation on the deployment skeleton index to form an index alignment table. The calculation process of the index alignment operation is to read the unique index code and adjacent relationship identifier of each index entry from the deployment skeleton index, and align them with the unique index code as the key and the adjacent relationship identifier as the value to obtain the index alignment table.

[0076] The original topology merge mapping operation is performed on the roadway topology data to form a merge correspondence table. The calculation process of the original topology merge mapping operation is as follows: read the roadway connection relationship from the roadway topology data and represent each connection relationship as an original connection unit identifier pair. Then, based on the adjacency relationship of the skeleton units in the topology skeleton sequence, determine the merge landing point of the original connection unit identifier pair in the skeleton structure. Establish a correspondence between each original connection unit identifier and the corresponding skeleton unit index code to form a merge correspondence table.

[0077] The graph structure assembly operation is performed to generate a twin topology base map. The calculation process of the graph structure assembly operation is as follows: the index code corresponding to each skeleton node unit in the skeleton unit set is used as the graph node identifier set, the index code corresponding to each skeleton connection unit in the skeleton unit set is used as the graph edge identifier set, and the two end graph node identifiers connected by each graph edge identifier are determined according to the adjacent relationship identifier in the index alignment table. The graph node identifier set and the graph edge identifier set and their connection relationship are integrated to form the graph structure body. The merge correspondence table is used as the mapping appendix so that any graph node identifier or graph edge identifier in the graph structure can trace its corresponding original connection unit identifier pair set through the merge correspondence table. This completes the structured mapping and encapsulation of the roadway topology data and the topology skeleton sequence under the deployment of skeleton index constraints, and generates a twin topology base map.

[0078] Perform time-series differential operations on the link-state coding to generate a sequence of coded differential values, and then perform anomaly measurement on the coded differential value sequence to generate an anomaly evidence sequence:

[0079] The link state encoding of consecutively arriving data is processed by time organization. Each arriving encoded object is formed into an encoded time sequence according to its arrival order. Then, a time difference operation is performed on two adjacent encoded objects in the encoded time sequence. The calculation process of the time difference operation is as follows: the two adjacent encoded objects are regarded as the preceding encoded object and the following encoded object, respectively. The preceding encoded object and the following encoded object are regarded as binary bit sequences of equal length. The bit sequence is XORed bit by bit to obtain the difference bit sequence. The number of bits with a value of 1 in the difference bit sequence is counted to obtain the difference count. The difference count is divided by the total number of bits in the binary bit sequence to obtain the normalized difference value. The normalized difference values ​​obtained for each pair of adjacent encoded objects are sorted in time order to generate the encoded difference value sequence.

[0080] Anomaly measurement operations are performed on the coded difference value sequence. The calculation process of the anomaly measurement operation is as follows: sum the total difference of all difference values ​​in the coded difference value sequence to obtain the total difference, and then divide the total difference by the number of difference values ​​to obtain the difference mean; calculate the absolute difference between each difference value in the coded difference value sequence and the difference mean, and integrate the absolute differences sequentially to form a deviation sequence; sum the deviation sequence to obtain the total deviation, and then divide the total deviation by the number of deviations to obtain the deviation mean; divide each absolute difference in the deviation sequence by the deviation mean to obtain the normalized deviation value, and use the normalized deviation value as the anomaly evidence value corresponding to that difference value. Then, integrate the anomaly evidence values ​​in chronological order to generate an anomaly evidence sequence.

[0081] Communication anomaly retrieval on twin topology base map:

[0082] The structural unit enumeration operation is performed on the twin topology base map to obtain a set of structural units. Each structural unit in the set is assigned a retrieval identifier and integrated to form a retrieval entry. Then, a time-aligned writing operation is performed on the abnormal evidence sequence. The calculation process of the time-aligned writing operation is as follows: the evidence items in the abnormal evidence sequence are integrated sequentially according to the time order of the encoded time sequence to form an evidence time sequence table. A windowed aggregation operation is performed on the evidence time sequence table to form an evidence aggregation list. The calculation process of the windowed aggregation operation is as follows: the evidence time sequence table is segmented and summed according to the order of adjacent evidence items to obtain the total number of evidence segments. The total number of evidence segments is divided by the number of evidence items in that segment to obtain the average value of the evidence segments. The average values ​​of each evidence segment are aggregated in time order to form an evidence aggregation list.

[0083] An association retrieval operation is performed on the search entries and the evidence aggregation list to generate anomaly hit entries. The calculation process of the association retrieval operation is as follows: for each search entry, its search identifier is read sequentially, and the corresponding average value of the evidence segment is determined in the evidence aggregation list using the search identifier. Then, the average value of the evidence segment is written into the outlier field of the search entry to form an anomaly assigned entry. A sorting and taking-first-out operation is performed on all anomaly assigned entries to form a set of anomaly candidate entries. The calculation process of the sorting and taking-first-out operation is as follows: a sorting sequence is generated according to the outlier field in the anomaly assigned entries from largest to smallest, and several entries are selected sequentially from the sorting sequence to form a set of anomaly candidate entries. The set of anomaly candidate entries and its structural unit identifier and adjacent relationship identifier in the twin topology base map are encapsulated together to obtain the communication anomaly retrieval result.

[0084] This scheme performs topological skeleton compression mapping on the roadway topology data to generate a topological skeleton sequence, converging the roadway connectivity from a full structure to a skeleton-connected expression. This reduces the scale of subsequent calculations and decreases the frequency of invalid recalculations during landslide changes. The topological skeleton sequence is indexed and labeled to generate a deployment skeleton index, allowing deployable locations to be determined with a unified index and quickly located. This reduces the search space for relay deployment and routing and decreases the probability of unreachable locations being selected. Based on the deployment skeleton index, the roadway topology data and the topological skeleton sequence are structured and encapsulated into a graph structure to generate a twin topology base map. This ensures a one-to-one correspondence between the roadway physical structure and the network decision carrier, maintaining topological consistency. This improves the matching accuracy of reachability constraint write-back and route distribution and reduces erroneous route changes caused by structural mismatch.

[0085] The above describes the topological skeleton mapping of the tunnel topology data to generate a twin topological base map. The following describes the path consistency inversion processing of communication anomaly retrieval results within the neighborhood based on anomaly evidence sequences and link-state codes, generating a blocking probability map. Please refer to [link to relevant documentation]. Figure 2 , Figure 2 This is a schematic diagram of the process for generating a blocking probability map provided in an embodiment of this application. Generating the blocking probability map specifically includes:

[0086] An evidence aggregation mapping is established for the communication anomaly retrieval results; wherein, the evidence aggregation mapping is determined by the correspondence between the deployed skeleton index and the twin topology base map;

[0087] Based on the evidence aggregation mapping, corresponding evidence items are extracted from the abnormal evidence sequence to form an evidence set, and the evidence strength value is calculated based on the evidence set;

[0088] The evidence set is subjected to pairwise difference operations in chronological order based on the encoded time series to obtain the evidence consistency value; the encoded time series is obtained by performing time organization operations on the link state encoding.

[0089] The evidence strength value and evidence consistency value are combined to generate a blocking probability map.

[0090] Among them, evidence aggregation mapping refers to a mapping data structure established by the software-defined network controller when performing communication anomaly inversion processing, based on the existing correspondence between the deployment skeleton index and the twin topology base map, to merge communication anomaly evidence from the network observation space to the twin topology structure space.

[0091] The evidence set refers to the ordered data set of a group of abnormal evidence data items corresponding to the same twin topology unit, obtained by the controller from the abnormal evidence sequence by filtering and merging them according to the evidence set mapping during the execution of communication abnormal path consistency inversion process;

[0092] The strength of evidence value refers to quantitative data used to characterize the degree of communication anomalies carried by a twin topology unit within the current time frame;

[0093] Encoding time series refers to an ordered encoding set formed by the controller organizing the continuously acquired link status codes in chronological order. This ordered encoding set is used to reflect the changes in the link operation status of the same communication link or the same twin topology unit at different points in time.

[0094] Evidence consistency value refers to the data quantification result used to quantify whether the same communication anomaly exhibits "continuous, stable, and non-random" characteristics in the time dimension.

[0095] The above content will be described in detail below:

[0096] An evidence aggregation mapping is established for the communication anomaly retrieval results; wherein, the evidence aggregation mapping is determined by the correspondence between the deployed skeleton index and the twin topology base map:

[0097] Perform correspondence parsing on the deployed skeleton index, extract the index code and its adjacent relationship identifier for each index entry, and locate the twin structural unit identifier corresponding to the index entry in the twin topology base map using the adjacent relationship identifier as the retrieval key, construct the correspondence between "index code and twin structural unit identifier", and thus generate the skeleton base map correspondence table;

[0098] Read the abnormal records in the communication anomaly retrieval results one by one, extract the abnormality identifier of the abnormal record and its associated skeleton index code, and retrieve the corresponding twin structural unit identifier in the skeleton to bottom map correspondence table using the skeleton index code. Then integrate the correspondence of "abnormality identifier - twin structural unit identifier" to generate an evidence landing point table.

[0099] The evidence landing point table is aggregated according to the twin structural unit identifier. The calculation process of the aggregation operation is as follows: using the twin structural unit identifier as the grouping key, the set of abnormal identifiers with the same twin structural unit identifier is written into the same grouping entry and the number of abnormal identifiers in the grouping entry is counted as the aggregation count. The "twin structural unit identifier - abnormal identifier set - aggregation count" are integrated to generate the evidence aggregation mapping.

[0100] Based on the evidence set mapping, corresponding evidence items are extracted from the abnormal evidence sequence to form an evidence set, and the evidence strength value is calculated based on the evidence set:

[0101] Read each mapping entry of the evidence collection mapping one by one. Each mapping entry indicates at least one unit identifier and one anomaly identifier. Use the unit identifier as the collection key and use the anomaly identifier to perform location reading in the anomaly evidence sequence to extract the corresponding anomaly evidence value.

[0102] When the same unit identifier corresponds to multiple anomaly identifiers in the evidence aggregation mapping, the multiple anomaly evidence values ​​read will be written sequentially into the aggregation list corresponding to that unit identifier in the order of reading. After all mapping entries have been read, the aggregation lists corresponding to all unit identifiers will be encapsulated into an evidence set, and the evidence strength value will be calculated based on the evidence set. The specific calculation formula is as follows:

[0103] ;

[0104] In the formula, Indicates unit identifier The number of evidence items contained in the evidence set. Indicates unit identifier The first in the evidence set The abnormal evidence values ​​for each evidence item; all of the above data have been normalized during the calculation.

[0105] The evidence set is subjected to pairwise difference operations in chronological order based on the encoded time series to obtain the evidence consistency value; the encoded time series is obtained by performing time organization operations on the link state encoding.

[0106] The strength of evidence value and the consistency of evidence value are combined to generate a blocking probability map. The specific calculation formula is as follows:

[0107] ;

[0108] In the formula, Indicates the cell identifier in the blocking probability graph The blocking probability value, Indicates unit identifier Consistency value of evidence This represents the set of structural units involved in the calculation. Represents a very small positive number. Indicates unit identifier Strength of evidence value Indicates unit identifier The evidence consistency values ​​are all normalized during the calculation process.

[0109] Update the twin topology base map based on the blocking probability map, and generate an updated twin map based on reachability:

[0110] Each structural unit identifier in the blocking probability map is matched one-to-one with the structural unit identifier in the twin topology base map to obtain an alignment mapping. Based on the alignment mapping, the blocking probability value of each structural unit in the blocking probability map is written into the corresponding structural unit in the twin topology base map to form a blocking representation after writing.

[0111] The blocking representation after writing is processed by adjacent smoothing synthesis to form a smoothed blocking representation. The calculation process of adjacent smoothing synthesis is as follows: for each structural unit in the twin topology base map, its adjacent structural unit set is determined. The blocking label value of the structural unit is read and the blocking label value of each structural unit in the adjacent structural unit set is read. The blocking label values ​​of each structural unit in the adjacent structural unit set are summed to obtain the total adjacent blocking. The total adjacent blocking is then divided by the number of elements in the adjacent structural unit set to obtain the average adjacent blocking. Subsequently, the blocking label value of the structural unit is summed with the average adjacent blocking and divided by two to obtain the smoothed blocking value. The blocking label value of the structural unit is replaced with the smoothed blocking value to complete the smoothed blocking representation.

[0112] An update encapsulation operation is performed on the twin topology base graph to generate a reachability-updated twin graph. The calculation process of the update encapsulation operation is to merge and encapsulate the set of structural units, the set of adjacency relationships, and the smoothed blocking representation of the twin topology base graph to form a unified graph structure object, and output the unified graph structure object as the reachability-updated twin graph.

[0113] This scheme establishes an evidence aggregation mapping, which can accurately locate communication anomaly retrieval results to structural units according to the correspondence between the deployed skeleton index and the twin topology base map, reducing anomaly source drift and mislocation. It extracts evidence items to form an evidence set and calculates evidence strength values, thereby aggregating discrete anomaly evidence into comparable quantified strengths, suppressing false blocking judgments caused by single-point noise. Based on the coded time series, it obtains the evidence consistency value by the adjacent difference of the evidence set, which can distinguish between instantaneous jitter and continuous anomalies, improving the stability of identifying real landslide blockages. It combines the evidence strength value and the continuous consistency value to generate a blocking probability map, forming a neighborhood continuous blocking probability space representation, which facilitates subsequent reachability updates and constrained routing, and reduces invalid rerouting and control surface oscillations.

[0114] The above describes the path consistency inversion processing of communication anomaly retrieval results within the neighborhood based on the anomaly evidence sequence and link state coding, generating a blocking probability map. The following describes the two-layer mask synthesis processing performed on the reachability update twin graph and the blocking probability map to generate an action mask. Please refer to [reference needed]. Figure 3 , Figure 3 This is a schematic diagram of the process for generating an action mask provided in an embodiment of this application. Generating the action mask specifically includes:

[0115] Perform an intersection operation on the topological skeleton sequence and the reachability update twin graph to generate a set of reachable skeleton units. Extract the structural units and adjacency relationships corresponding to the set of reachable skeleton units from the twin topological base graph and assemble them into a reachability constraint subgraph.

[0116] The blockage probability graph is complemented and mapped to generate risk availability values. These risk availability values ​​are then written into the reachability constraint subgraph to form a risk labeling subgraph. Membership mapping is then performed on the risk labeling subgraph to generate structural feasibility marker values.

[0117] The risk availability value is mapped to generate risk feasibility weights, and the structural feasibility label value and the risk feasibility weights are multiplied to generate an action mask.

[0118] The reachable skeleton unit set refers to the set of skeleton units that, at the current moment, remain physically connected or in a risk-allowed state after being determined by the reachability update twin graph, among all skeleton units described by the topological skeleton sequence.

[0119] The reachability constraint subgraph refers to a restricted topology graph structure generated by the controller in a software-defined network, based on a twin topology base graph as the basic structural carrier, under the joint constraints of introducing a topology skeleton sequence and a reachability update twin graph, through structural unit intersection filtering and adjacency relationship pruning, which only contains the communication nodes and communication connection relationships corresponding to the current physically reachable roadway structure;

[0120] Risk availability refers to data that characterizes the relative feasibility of a structural unit being used for routing or deployment actions at the current moment;

[0121] Risk-labeled subgraphs refer to subgraph data structures that use reachability-constrained subgraphs as the basic topology, and without changing the set of nodes and the set of adjacent relationships, write the risk availability value obtained by the blocking probability graph through complement mapping into each structural unit, thus forming a subgraph data structure that simultaneously contains "structural reachability attributes" and "risk availability attributes".

[0122] The structurally feasible marker value refers to a type of binary structural judgment data formed by the software-defined network controller during the generation of action masks. It is used to characterize whether a candidate action satisfies the most basic constraint condition of "physically reachable and structurally existent" at the topological level.

[0123] Risk feasibility weight refers to the numerical weight obtained by the controller after taking the complement of the blocking probabilities of each structural unit in the blocking probability diagram, which is used to characterize the "risk acceptability" of candidate actions in the current roadway environment.

[0124] The above content will be described in detail below:

[0125] Perform an intersection operation on the topological skeleton sequence and the reachability update twin graph to generate a set of reachable skeleton elements. Then, extract the structural elements and adjacency relationships corresponding to the reachable skeleton element set from the twin topological base graph and assemble them into a reachability constraint subgraph.

[0126] The skeleton unit reading operation is performed on the topological skeleton sequence to form a skeleton unit list. The connected unit reading operation is performed on the reachability update twin graph to form a connected unit list. Then, the intersection operation is performed on the skeleton unit list and the connected unit list. The calculation process of the intersection operation is as follows: the skeleton unit identifier is extracted item by item from the skeleton unit list, and the same identifier is matched in the connected unit list. The successfully matched skeleton unit identifiers are integrated to finally obtain the reachable skeleton unit set.

[0127] The twin topology base map is read and structural element extraction operation is performed. The calculation process of structural element extraction operation is as follows: structural element identifiers and their adjacent relationship identifiers are read one by one from the twin topology base map to form structural element pairs. The reachable skeleton element set is used as the filtering basis. Set membership matching is performed on the structural element identifiers in each structural element pair. Subgraph node set is formed according to the structural element pairs with successful membership matching. Adjacency relationship assembly operation is performed on the adjacent relationship identifiers in the structural element pairs written into the subgraph node set. The calculation process of adjacent relationship assembly operation is as follows: when the two structural element identifiers pointed to by the adjacent relationship identifier exist in the subgraph node set at the same time, the adjacent relationship identifiers are integrated to form the subgraph edge set. The subgraph node set and the subgraph edge set are encapsulated into a reachability constraint subgraph.

[0128] The blockage probability graph is complemented to generate risk availability values, which are then written into the reachability constraint subgraph to form a risk labeling subgraph. Furthermore, the risk labeling subgraph is subjected to membership mapping to generate structural feasibility marker values.

[0129] Extract the set of blocking probability values ​​corresponding to all structural units in the blocking probability map, and extract the maximum probability value from the set of blocking probability values. Then, perform a normalization operation on the blocking probability value of each structural unit. The normalization operation is calculated by dividing the blocking probability value of the structural unit by the maximum probability value to obtain the normalized blocking value. Then, perform a complement operation on the normalized blocking value. The complement operation is calculated by subtracting the normalized blocking value from one to obtain the risk availability value corresponding to the structural unit.

[0130] Read the reachability constraint subgraph and extract its structural unit set and structural unit identifier set. Then, perform identifier matching on each structural unit identifier in the structural unit identifier set to obtain the corresponding risk availability value. Write the obtained risk availability value into the label position of the corresponding structural unit in the reachability constraint subgraph to complete the risk label writing and encapsulation of the reachability constraint subgraph, thereby forming a risk label subgraph.

[0131] Read the risk-marked subgraph and extract its structural unit identifier set to form a feasible structure set. Perform candidate action pointing mapping operation on the risk-marked subgraph to obtain a candidate action identifier set. The calculation process of the candidate action pointing mapping operation is to generate a candidate action identifier for each pair of adjacent structural units in the risk-marked subgraph according to their adjacency relationship, and pair the candidate action identifier with its pointing structural unit identifier and write it into the mapping entry to form a candidate action identifier set. Then, perform a membership mapping operation on the candidate action identifier set and the feasible structure set. The calculation process of the membership mapping operation is to search in the feasible structure set for whether there is a structural unit identifier that is the same as the candidate action identifier set. If it is found, output the structural feasibility mark value corresponding to the candidate action as one; if it is not found, output the structural feasibility mark value corresponding to the candidate action as zero, thereby generating a structural feasibility mark value.

[0132] The risk availability value is mapped to generate a risk feasibility weight, specifically by performing normalization processing on the risk availability value;

[0133] An action mask is generated by multiplying the structural feasibility marker value and the risk feasibility weight;

[0134] This scheme eliminates physically unreachable units, compresses the search space, and reduces invalid recalculations by intersecting the topological skeleton sequence with the reachability update twin graph and assembling the reachability constraint subgraph. It also obtains the risk availability by complementing the blocking probability graph and writes it into the subgraph to form a risk labeling subgraph. The membership mapping then obtains the structural feasibility label value, which pre-constrains the collapse risk and reduces incorrect path selection. The risk availability value is mapped to the risk feasibility weight and multiplied with the structural feasibility label value to synthesize the action mask. This achieves hard reachability + soft risk synergistic constraints, suppresses high-risk actions, accelerates convergence, and improves stability.

[0135] The above describes the two-layer mask synthesis process performed on the reachability update twin graph and the blocking probability graph to generate an action mask. The following section describes the construction of candidate action solutions using the action mask as constraints. Please refer to [link / reference]. Figure 4 , Figure 4 This is a flowchart illustrating the process of constructing a candidate solution for a feasible action, as provided in an embodiment of this application. The specific steps of constructing a candidate solution for a feasible action include:

[0136] The deployment skeleton index is subjected to position filtering and sequence encoding to generate deployment sequence codes. Based on the deployment sequence codes, structural unit mapping is performed in the reachability constraint subgraph and written according to the adjacency relationship to generate path sequence codes.

[0137] The deployment sequence code and the path sequence code are concatenated and encapsulated to generate candidate decodes. The path sequence codes corresponding to the candidate decodes are then segmented and processed to generate a set of candidate actions.

[0138] Delete candidate actions with an action mask of zero, and reorganize the remaining candidate actions according to their original adjacency relationships to generate candidate solutions for actionable actions.

[0139] Among them, deployment sequence encoding refers to the serialization encoding result formed by the controller selecting several deployment index identifiers from the deployment skeleton index in sequence during a candidate solution construction process, and arranging the deployment index identifiers in order according to their logical relationship in the same candidate solution.

[0140] Path sequence coding refers to the serialized encoded data formed by the software-defined network controller, under reachability constraints, based on the acquired reachability constraint subgraph and deployment sequence coding, to uniformly express the access order of structural units in the twin topology of candidate communication paths;

[0141] Candidate decoding refers to composite sequence data generated by the controller, which is used to simultaneously represent deployment selection information and path selection information in a candidate communication scheme, and ensures that the two have a one-to-one structural relationship in the same encoding object;

[0142] The candidate action set refers to the set of action units formed by the controller according to the path adjacency structure relationship, which decomposes the path sequence code into several "adjacent structural unit transfer pairs" under the candidate solution code determined by the given deployment sequence code and path sequence code, and assigns a unique action identifier to each adjacent structural unit transfer pair.

[0143] The above content will be described in detail below:

[0144] Position filtering and sequence encoding are performed on the deployment skeleton index to generate deployment sequence codes. Based on the deployment sequence codes, structural unit mapping is performed in the reachability constraint subgraph and written according to the adjacency relationship to generate path sequence codes.

[0145] The location screening operation is performed on the deployment skeleton index. The calculation process of the location screening operation is to read the index code and adjacent relationship identifier of all index entries in the deployment skeleton index one by one, and combine the index entries with adjacent relationship identifiers to form a candidate index set. At the same time, the candidate index set is deduplicated to ensure that only one candidate index is retained for the same index code.

[0146] The sequential encoding operation is performed on the candidate index set. The calculation process of the sequential encoding operation is as follows: the index codes in the candidate index set are sorted lexicographically to obtain the index sorting sequence, and then each index code is written in the order of the index sorting sequence to form an index sequence. This index sequence is then used as the deployment sequence encoding output.

[0147] Each index code in the deployment sequence encoding is written as a location key into the location key set. Then, the location key set is used as the mapping input to perform a structural unit mapping operation on the reachability constraint subgraph. The calculation process of the structural unit mapping operation is to retrieve the corresponding structural unit identifier for each location key in the reachability constraint subgraph and combine the retrieved structural unit identifiers to form a path structural unit set.

[0148] The adjacency relationship writing operation is performed on the path structure unit set. The calculation process of the adjacency relationship writing operation is as follows: based on the existing adjacency relationships in the reachability constraint subgraph, the structure unit identifiers in the path structure unit set are arranged in order according to the adjacency relationship, and the arranged structure unit identifiers are combined in sequence. The combination result is output as the path sequence code.

[0149] The deployment sequence encoding and path sequence encoding are concatenated and encapsulated to generate candidate decoding. Then, the path sequence encoding corresponding to the candidate decoding is processed into segmented actions to generate a candidate action set.

[0150] The length identifier is calculated for the deployment sequence code. The length identifier is calculated by counting the deployment entries written sequentially in the deployment sequence code to obtain the number of deployment entries and using it as the deployment length value.

[0151] The path sequence encoding is performed to calculate the length identifier. The length identifier calculation is obtained by counting the path entries written sequentially in the path sequence encoding one by one, and the number of path entries is used as the path length value.

[0152] The deployment length value is written into the encapsulation header as the first position identifier, the path length value is written into the encapsulation header as the second position identifier, the deployment sequence code is written into the first half of the encapsulation body in its original order, the path sequence code is written into the second half of the encapsulation body in its original order, and the deployment length value is used as the boundary marker in the encapsulation body to determine the end position of the deployment entry, thereby generating candidate solution codes.

[0153] The encapsulation header in the candidate deencoding is read to obtain the deployment length value and determine the latter half of the encapsulation body as the path sequence encoding corresponding to the candidate deencoding. Adjacent segment decomposition is performed on the path sequence encoding. Adjacent segment decomposition is obtained by reading the path entries in sequence and forming a path segment by two adjacent path entries. Each path segment in the path segment sequence represents a jump from the previous entry to the next destination. Then, action encoding is performed on the path segment sequence. The calculation process of action encoding is as follows: extract the previous destination identifier and the next destination identifier for each path segment, concatenate the two in a fixed order to form an action pair identifier, integrate the action pair identifiers to form an action entry, and write the action entries continuously in the order of the path segment sequence to form an action sequence. The action sequence is encapsulated into a set structure and output as a candidate action set.

[0154] Delete candidate actions with an action mask of zero, and reorganize the remaining candidate actions according to their original adjacency relationships to generate candidate solutions for actionable actions.

[0155] This scheme generates deployment sequence codes by sequentially encoding the deployment skeleton index after screening. Based on the deployment sequence codes, it maps them to the reachability constraint subgraph and writes them into the generated path sequence codes according to the adjacency relationship. This limits the deployment and route selection range, ensures path reachability, and reduces invalid searches. The deployment sequence codes and path sequence codes are concatenated to generate candidate solution codes. The path sequence codes are segmented into actions to generate action sets, turning the paths into verifiable actions and improving scoring accuracy. Actions with zero action mask are deleted and recombined according to the adjacency relationship to obtain actionable candidate solutions. This eliminates inactive actions, reduces recalculation and misselection, and improves stability.

[0156] The above describes the construction of actionable candidate solutions using action masks as constraints. The following describes the joint cost calculation and mapping relationship establishment for the actionable candidate solutions based on anomaly evidence sequences, blocking probability graphs, and reachability update twin graphs, generating a joint cost function. Specifically, this includes:

[0157] For the candidate solutions of the actionable actions, perform cover matching in the reachability update twin graph to obtain a set of cover structure units;

[0158] Based on the set of covered structural units, perform probability extraction on the blocking probability map and calculate the arithmetic mean to generate the risk mean;

[0159] Based on the set of covered structural units, evidence is extracted from the sequence of abnormal evidence and the arithmetic mean is calculated to generate the evidence mean.

[0160] Based on the set of covered structural units, perform update extraction in the reachability update twin graph and calculate the arithmetic mean to generate the change mean;

[0161] A weighted composite operation is performed on the risk mean, evidence mean, and change mean to generate a joint cost value, and a mapping relationship is established from the candidate solutions of actionable actions to the joint cost value to obtain the joint cost function.

[0162] Among them, the set of covered structural units refers to the set of all structural units actually traversed, occupied, or associated by a certain action candidate solution in the reachability update twin graph;

[0163] Risk mean refers to the risk quantification result obtained by extracting the blocking probability value of the corresponding structural unit from the blocking probability graph based on the coverage relationship of the candidate action in the reachability update twin graph during the joint cost function solution process, and performing an arithmetic mean operation on the extracted blocking probability value;

[0164] The evidence mean refers to the numerical result obtained by performing an arithmetic average of the abnormal evidence scores corresponding to the set of structural units in the abnormal evidence sequence within the range of the set of structural units covered by a single candidate action.

[0165] The mean change refers to the average change of numerical labels reflecting the degree of structural state change in the reachability update twin graph within the corresponding set of covered structural units for a candidate solution of a certain action.

[0166] The joint value refers to the data quantification result of the software-defined network controller for each actionable candidate solution, based on the range of structural units it covers in the current rescue communication environment, comprehensively reflecting the overall status of the candidate solution in terms of the degree of blocking risk, the intensity of communication anomalies, and the magnitude of twin structure changes.

[0167] The above content will be described in detail below:

[0168] For each action candidate solution, perform cover matching in the reachability update twin graph to obtain a set of cover structure elements:

[0169] Action segment extraction is performed on the candidate solutions of actionable actions to obtain the action segment sequence. The calculation process of the action segment extraction operation is to extract the path selection expression from the candidate solutions of actionable actions and integrate them into the action segment sequence according to the order of adjacent connection segments. Then, the structural unit positioning operation is performed on the action segment sequence. The calculation process of the structural unit positioning operation is to read the pointing identifier of each action segment and perform identifier matching in the reachability update twin graph. The matched structural unit identifier is used as the structural unit positioning result of the action segment and written into the positioning result sequence. The positioning result sequence is output as the covering structural unit set.

[0170] According to the set of covered structural units Perform probability extraction and calculate the arithmetic mean to generate the risk mean in the blocking probability graph. The specific calculation formula is as follows:

[0171] ;

[0172] In the formula, This indicates the number of elements in the set of covering structural units. Represents the blocking probability diagram and the covering structure unit The corresponding blocking probability values, all of which have been normalized during the calculation process;

[0173] According to the method of generating risk mean, evidence mean is generated by extracting evidence from the abnormal evidence sequence based on the set of covered structural units and calculating the arithmetic mean. Then, change mean is generated by extracting update from the reachability update twin graph based on the set of covered structural units and calculating the arithmetic mean.

[0174] For the mean risk Mean of evidence with the mean of change Perform a weighted composition operation to generate a joint cost. The specific calculation formula is as follows:

[0175] ;

[0176] In the formula, This represents the risk weighting coefficient. This represents the evidence weighting coefficient. The above data has been normalized during the calculation, and the weighting coefficients represent the changes.

[0177] A one-to-one correspondence is established between the candidate solutions of the possible actions and the joint cost value to obtain the joint cost function.

[0178] This scheme can accurately limit the scope of candidate solutions in the current physical reachable structure by performing coverage matching on the reachability update twin graph, thereby avoiding structural conflicts between candidate solutions and updated unreachable areas and improving the consistency between the route calculation results and the actual roadway reachability status.

[0179] By performing probability extraction and calculating the risk mean on the set of coverage structural units in the blocking probability graph, the path risk of candidate solutions is quantified into a comparable numerical expression, thereby reducing the possibility of incorrect path selection even when passing through high blocking probability areas. By performing evidence extraction and calculating the evidence mean on the set of coverage structural units in the anomalous evidence sequence, the sensitivity of candidate solutions to communication anomalies is reflected, thereby suppressing repeated route selection in areas where communication conditions are continuously deteriorating. By performing update extraction and calculating the change mean on the set of coverage structural units in the reachability update twin graph, the adaptability of candidate solutions to twin structure changes is quantified, thereby reducing invalid rerouting in areas where structures change frequently. By performing a weighted synthesis operation on the risk mean, evidence mean, and change mean and establishing a mapping relationship, the comprehensive cost of candidate solutions is expressed in a unified function form, thereby providing a stable, comparable, and multi-source data-consistent evaluation basis for subsequent decision-making.

[0180] The above describes the joint cost calculation and mapping relationship establishment for candidate action solutions based on anomaly evidence sequences, blocking probability maps, and reachability update twin graphs, generating a joint cost function. The following describes the value screening of candidate action solutions based on the joint cost function to generate route optimization schemes, specifically including:

[0181] The candidate solutions for actionable actions are rank-fused and encoded according to the joint cost function to generate a set of candidate deployment paths. Then, action segments are constructed and filtered item by item in the set of candidate deployment paths according to the action mask to generate a set of candidate actions.

[0182] Based on the abnormal evidence sequence, the coded difference value sequence, the blocking probability map and the joint cost function, the candidate action set is calculated and summarized to generate a reward component table;

[0183] The reward subscale is summed and sorted to generate a strategy value table. The abnormal evidence sequence and the coded difference value sequence are normalized and synthesized to generate change level values.

[0184] The strategy value table, deployment path candidate set, and change level value are structured and combined to generate a route optimization scheme.

[0185] Among them, the deployment path candidate set refers to a set of structured candidate solution data generated by the software-defined network controller after value screening of the candidate solutions of feasible actions under the constraints of the joint cost function;

[0186] The candidate action set refers to the action-level candidate set data formed by structurally decomposing the path selection expression of each deployment path candidate solution in the deployment path candidate set, and then filtering and reorganizing the decomposition results item by item in combination with action mask, which is used for subsequent strategy benefit calculation.

[0187] The reward subscale is a structured data table generated by the software-defined network controller during the routing optimization decision-making process. It is used to perform multi-dimensional quantitative evaluation of each candidate action in the candidate action set.

[0188] The strategy value table is a structured data table generated by the controller after calculating the reward components of the candidate action set. It is used to quantify and organize the comprehensive strategy value of each action candidate solution at the current moment.

[0189] Change level value refers to a type of continuous control data generated by the controller during the route optimization process. It is used to characterize the magnitude of structural adjustments allowed by the route optimization scheme at the current moment, based on the degree of change in network operating status.

[0190] The above content will be described in detail below:

[0191] Based on the joint cost function, rank fusion encoding is performed on the candidate solutions of actionable actions to generate a candidate set of deployment paths. Then, action segments are constructed and filtered item by item in the candidate set of deployment paths according to the action mask to generate a candidate action set.

[0192] A descending order sorting operation is performed on the joint cost values ​​corresponding to the candidate solutions of the action to obtain a cost sorting sequence. At the same time, the path selection expression of each candidate solution of the action is read and the coverage is calculated to obtain the coverage value. The calculation process of the coverage value is to remove duplicate structural unit identifiers contained in the path selection expression and count the number of duplicate identifiers. The number of identifiers is used as the coverage value. A descending order sorting operation is performed on the coverage values ​​of all candidate solutions to obtain a coverage sorting sequence.

[0193] For each action, candidate solutions are made to determine the cost noun in the cost ranking sequence and the coverage noun in the coverage ranking sequence, and rank fusion encoding is performed to obtain the fusion rank value. The fusion rank value is calculated by summing the cost noun and the coverage noun and dividing by two to obtain the average rank, and the average rank is used as the fusion rank value.

[0194] The fusion rank values ​​are sorted from smallest to largest to obtain a fusion sorting table. Then, the action candidates with truncated length values ​​are selected from the fusion sorting table in order of sorting to decapsulate and generate a deployment path candidate set.

[0195] Specifically, the truncation length value is obtained by calculating the difference between the maximum and minimum fusion ranks of the fusion rank value sequence to obtain the rank span, then summing the fusion rank value sequence and dividing by the number of sequence items to obtain the rank mean, then dividing the rank span by the rank mean to obtain the normalized dispersion, and finally multiplying the normalized dispersion by the total number of candidate solutions for possible actions and rounding down to the nearest integer.

[0196] Read the deployment path candidate set and extract the path selection expression of each candidate item by item. Divide the path selection expression into action segment sequence according to the adjacent connection relationship. Each action segment consists of "starting structural unit identifier and target structural unit identifier".

[0197] For each action segment in the action segment sequence, an action segment identifier is constructed, and the corresponding action mask value is obtained by retrieving the action mask using the action segment identifier. Then, a zero-value filtering operation is performed on the action mask value to delete the action segments with an action mask value of zero and retain the action segments with an action mask value greater than zero, thereby obtaining the filtered action segment sequence. The filtered action segment sequence is integrated with the corresponding candidate identifier to form candidate action entries, and all candidate action entries are summarized and encapsulated to generate a candidate action set.

[0198] The candidate action set is calculated and summarized based on the abnormal evidence sequence, coded difference value sequence, blocking probability map, and joint cost function to generate a reward component table:

[0199] The stationary component of evidence is calculated based on the abnormal evidence sequence. The calculation process is as follows: take the two adjacent items of the abnormal evidence sequence in chronological order, perform difference and take the absolute value to obtain the evidence difference set. Summate the evidence difference set to obtain the total evidence difference. Divide the total evidence difference by the number of elements in the evidence difference set to obtain the evidence difference mean. At the same time, take the maximum value of the abnormal evidence sequence to obtain the evidence maximum value. Divide the evidence difference mean by the evidence maximum value to obtain the evidence normalized fluctuation value. Subtract the evidence normalized fluctuation value from one to obtain the stationary component of evidence.

[0200] The differential stationary component is calculated based on the encoded differential value sequence. The calculation process of the differential stationary component is as follows: take two adjacent items in the encoded differential value sequence in chronological order, perform difference and take the absolute value to obtain the differential fluctuation set. Summate the differential fluctuation set to obtain the total differential fluctuation. Divide the total differential fluctuation by the number of elements in the differential fluctuation set to obtain the differential fluctuation mean. At the same time, take the maximum value of the encoded differential value sequence to obtain the maximum differential value. Divide the differential fluctuation mean by the differential maximum value to obtain the differential normalized fluctuation value. Subtract the differential normalized fluctuation value from one to obtain the differential stationary component.

[0201] The risk aversion component is calculated based on the blocking probability map. The calculation process of the risk aversion component is as follows: Read the blocking probability value of each item in the set of covered structural units in the blocking probability map and sum them to obtain the total risk. Then, divide the total risk by the number of elements in the set of covered structural units to obtain the risk mean. At the same time, take the maximum value of all blocking probability values ​​in the blocking probability map to obtain the maximum risk value. Divide the risk mean by the maximum risk value to obtain the risk normalization value. Subtract the risk normalization value from one to obtain the risk aversion component.

[0202] The calibration consistency component is calculated based on the joint cost function. The calculation process of the calibration consistency component is as follows: read the joint cost value corresponding to the candidate in the joint cost function, take the maximum value of all joint cost values ​​in the joint cost function to obtain the maximum cost value, divide the joint cost value of the candidate by the maximum cost value to obtain the cost normalization value, and subtract the cost normalization value to obtain the calibration consistency component.

[0203] Each candidate's identifier is encapsulated along with the evidence-stationary component, the difference-stationary component, the risk-averse component, and the calibration-consistent component to generate a reward component table.

[0204] Calculate the dispersion of the reward component table The specific calculation formula is as follows:

[0205] ;

[0206] In the formula, This indicates the number of candidates in the reward component table. Indicates the first The evidence for each candidate is stable and substantial. Indicates the first The difference-stationary components of each candidate Indicates the first Risk aversion factor for each candidate Indicates the first Calibration consistency of each candidate component Indicates the first The evidence for each candidate is stable and substantial. Indicates the first The difference-stationary components of each candidate Indicates the first Risk aversion factor for each candidate Indicates the first The calibration consistency components of each candidate are all normalized during the calculation.

[0207] The minimum dispersion is taken, and then the minimum dispersion is divided by the dispersion to obtain the adaptive weight. The reward subtable is weighted and summed according to the adaptive weight, and the weighted sum is multiplied by the corresponding feasibility value to obtain the strategy reward value. The strategy reward values ​​are then integrated to generate the strategy value table.

[0208] The feasibility value is obtained by multiplying the action masks corresponding to the candidate action set:

[0209] Read the action segment sequence corresponding to each candidate in the candidate action set, and generate an action identifier corresponding to each action segment in the action segment sequence. Then, retrieve the mask value corresponding to each action segment in the action mask one by one according to the action identifier, form a mask value sequence, and perform a chain multiplication operation on the mask value sequence to calculate the feasibility value.

[0210] The abnormal evidence sequence and the coded difference value sequence are normalized and synthesized to generate a change level value. The specific calculation formula is as follows:

[0211] ;

[0212] In the formula, Indicates the weighting coefficient. Indicates the sequence length. The sequence of anomalous evidence is in the first... The value of anomalous evidence at each moment. The sequence of anomalous evidence is in the first... The value of anomalous evidence at each moment. The sequence of anomalous evidence is in the first... The value of anomalous evidence at each moment. Indicates the encoded difference value sequence at the th The encoded difference value at each time step. Indicates the encoded difference value sequence at the th The encoded difference value at each time step. Indicates the encoded difference value sequence at the th The encoded difference values ​​at each time point; all of the above data have been normalized during the calculation.

[0213] Sort the strategy value table in descending order to obtain the value ranking sequence;

[0214] Based on the first item in the value ranking sequence, extract the deployment selection expression and path selection expression from the deployment path candidate set, and generate the corresponding deployment selection result and main path set:

[0215] The first item read operation is performed on the value ranking sequence to obtain the first item identifier. The calculation process of the first item read operation is to take the first position of the value ranking sequence as the first item and read the candidate item identifier recorded in the first item to obtain the first item identifier. Then, the candidate matching operation is performed in the deployment path candidate set to locate the target candidate that matches the first item identifier. The calculation process of the candidate matching operation is to read the candidate item identifier of each candidate in the deployment path candidate set one by one and compare it with the first item identifier. The matching candidate is marked as the target candidate and output.

[0216] The deployment selection expression is read from the target candidates and the deployment expression encapsulation operation is performed to generate the deployment selection result. The path selection expression is read from the target candidates and the path expression encapsulation operation is performed to generate the main path set.

[0217] Calculate the difference between the remaining items in the value ranking sequence and the main path set, and sort the remaining items according to the difference. Select the top-ranked item A to generate the backup path set: where A is a positive integer between 10 and 15.

[0218] Extract the main path coverage set from the main path set. The calculation process of the main path coverage set is to read the structural unit identifiers that each main path in the main path set passes through in its path order and perform a union summation to obtain the main path coverage set.

[0219] For each remaining item in the value ranking sequence, extract the candidate path coverage set of that remaining item. The calculation process of the candidate path coverage set is to read the structural unit identifiers that the path selection expression corresponding to that remaining item passes through in the path order and perform union summation to obtain the candidate path coverage set.

[0220] Calculate the intersection and union of the two sets respectively. The intersection is calculated by performing an intersection operation on the main path coverage set and the candidate path coverage set and counting the number of intersection elements. The union is calculated by performing a union operation on the main path coverage set and the candidate path coverage set and counting the number of union elements. Divide the intersection by the union to get the overlap ratio, and subtract the overlap ratio to get the difference.

[0221] Sort the differences in descending order to obtain the difference ranking sequence. Select the first A items from the difference ranking sequence starting from the first item. Read the path selection expression of the remaining items corresponding to the selected items in the value ranking sequence. Encapsulate each path selection expression into a set structure to generate a backup path set.

[0222] The deployment selection results, the primary path set, and the backup path set are combined in a structured manner to generate a route optimization scheme.

[0223] This scheme generates a deployment path candidate set by performing rank fusion encoding on the candidate solutions of feasible actions according to the joint cost function, so that deployment selection and path selection are uniformly measured in the same value space, thereby reducing the value distortion caused by the separation of deployment and routing decisions and improving the overall stability of candidate solutions. By combining action masks to construct action segments item by item in the deployment path candidate set and filtering them, unreachable or high-risk actions can be directly eliminated in the data processing stage, reducing the subsequent computational complexity and avoiding invalid exploration. By using abnormal evidence sequences, encoded difference value sequences, blocking probability maps and joint cost functions to calculate and summarize the reward components of the candidate action set, the quality of candidate solutions can be quantified from multiple dimensions such as communication stability, topology risk and structural changes, improving the ability to perceive sudden collapse and link fluctuations.

[0224] By summarizing and sorting the reward component table, a strategy value table is generated, providing a clear data ranking basis for selecting the optimal solution. Simultaneously, by normalizing and synthesizing the abnormal evidence sequence and the coded difference value sequence, a change level value is generated, enabling adaptive control of the rerouting amplitude. Finally, by structurally combining the strategy value table, the deployment path candidate set, and the change level value, a routing optimization scheme can be generated. This can significantly reduce the frequency of invalid rerouting and improve the continuity and reliability of rescue communications while ensuring communication connectivity.

[0225] The above describes the value screening of candidate solutions for feasible actions based on the joint cost function to generate route optimization schemes. The following describes the summarization and sorting of policy returns on the reward component table to generate a policy value table, specifically including:

[0226] The dispersion of the reward component table is calculated, and the minimum dispersion is selected. Then, the adaptive weight is calculated based on the dispersion and the minimum dispersion.

[0227] The reward subscale is weighted and summed according to the adaptive weights, and the weighted sum is multiplied by the corresponding feasibility value to obtain the strategy reward value. The strategy reward values ​​are then integrated to generate a strategy value table.

[0228] The feasibility value is obtained by multiplying the action mask corresponding to the candidate action set.

[0229] Among them, dispersion refers to the quantitative result of the degree of dispersion of the numerical distribution of multiple reward components corresponding to the same candidate in the reward component table within the same time section;

[0230] Adaptive weights refer to a set of weight parameters automatically calculated by the controller based on the dispersion between the reward components within the reward component table when performing strategy benefit aggregation processing on the reward component table.

[0231] The weighted summation result refers to the single numerical summary result formed by the controller performing a weighted summation operation on each reward component corresponding to each candidate action in the reward component table under the adaptive weight constraints calculated.

[0232] Feasibility value refers to numerical data used to quantify the overall executability of a candidate action under the current network state and physical reachability constraints;

[0233] The strategy benefit value refers to the comprehensive decision score data calculated by the controller for each candidate strategy corresponding to a candidate action. It is used to measure the stability consistency, risk avoidance, and execution feasibility of the candidate strategy under the current network state and reachability constraints.

[0234] This part has already been described in detail above, so I will not repeat it here.

[0235] This scheme generates adaptive weights by calculating the dispersion of the reward component table and taking the minimum dispersion. It can automatically adjust the influence ratio of each reward component according to the fluctuation differences of different candidate actions in the multi-dimensional reward components, avoiding policy bias caused by fixed weights and improving the adaptability of policy evaluation to complex network state changes. By using adaptive weights to perform weighted summation on the reward component table and multiplying the weighted result by the corresponding feasibility value to generate the policy reward value, while comprehensively considering stability, risk avoidance and structural feasibility, a quantitative constraint on path executability is introduced, thereby effectively suppressing the overestimation of infeasible or high-risk policies. By sorting the policy reward values ​​in descending order to generate a policy value table, a clear and comparable priority basis is provided for subsequent routing decisions, enabling the controller to quickly select the optimal or suboptimal policy, reducing invalid rerouting and improving the overall stability and continuity of the underground mine rescue communication network in dynamic collapse scenarios.

[0236] The above describes the process of summarizing and sorting the reward component table to generate a strategy value table. The following describes the structured combination of the strategy value table, deployment path candidate set, and change level values ​​to generate a route optimization scheme, specifically including:

[0237] The strategy value table is sorted in descending order to obtain the value ranking sequence. Based on the first item in the value ranking sequence, the deployment selection expression and path selection expression are extracted from the deployment path candidate set, and the deployment selection result and main path set are generated accordingly.

[0238] Calculate the difference between the remaining items in the value ranking sequence and the main path set, and sort the remaining items according to the difference. Select the top-ranked item A to generate the backup path set.

[0239] The deployment selection results, the primary path set, and the backup path set are combined in a structured manner to generate a route optimization scheme.

[0240] Among them, the value ranking sequence refers to the ordered sequence data formed by the software-defined network controller after performing a unified sorting process on each candidate item according to the corresponding policy benefit value based on the generated policy value table.

[0241] Deployment selection results refer to a type of structured data generated by the software-defined network controller after processing the policy value table, deployment path candidate set, and change level values.

[0242] The main path set refers to the path data set formed by the controller after sorting the strategy value table, selecting the path selection expression corresponding to the candidate with the highest strategy value from the deployment path candidate set, and encapsulating the path selection expression in a structured manner.

[0243] The degree of difference refers to the quantitative value of the structural difference between the set of structural units covered by the candidate path in the reachability constraint subgraph and the set of structural units covered by the main path set.

[0244] The backup path set refers to an ordered set of multiple path selection expressions that are significantly different from the main path set in terms of topological coverage structure, but still satisfy the reachability constraint subgraph and action mask constraints, obtained by the controller from the deployment path candidate set, given that the main path set has been determined.

[0245] This part has already been described in detail above, so I will not repeat it here.

[0246] This solution sorts the strategy value table in descending order to obtain a value ranking sequence. Based on the first item in the value ranking sequence, it extracts the deployment selection expression and path selection expression from the deployment path candidate set to generate deployment selection results and a main path set. It directly selects the candidate solution with the best overall stability, risk avoidance, and feasibility as the main execution plan, reducing unnecessary trial switching. It calculates the difference between the remaining items in the value ranking sequence and the main path set, and selects the top-ranked A items according to the difference to generate a backup path set. The difference metric ensures that the backup path is structurally sufficiently distinguishable from the main path, thereby providing effective alternative capabilities when the main path fails and avoiding simultaneous obstruction of the main and backup paths. The deployment selection results, the main path set, and the backup path set are structurally combined to generate a routing optimization plan. The deployment decision and routing decision are uniformly encapsulated into a single, deployable result, improving the overall consistency and execution stability of routing adjustments.

[0247] Example 2:

[0248] Please see Figure 5 Intelligent routing optimization systems based on machine learning in software-defined networks include:

[0249] The data acquisition module is used to acquire the link status code and tunnel topology data of the target object;

[0250] The topology processing module is used to perform topological skeleton mapping on the tunnel topology data and generate a twin topology base map.

[0251] The anomaly analysis module is used to perform time-series differential operations on link-state codes to generate a sequence of coded differential values, and to perform anomaly measurement on the coded differential value sequence to generate an anomaly evidence sequence.

[0252] The update module is used to perform communication anomaly retrieval on the twin topology base map, and perform path consistency inversion processing on the communication anomaly retrieval results in the neighborhood based on the anomaly evidence sequence and link state coding to generate a blocking probability map. Then, the twin topology base map is updated based on the blocking probability map to generate a reachability updated twin map.

[0253] The candidate solution construction module is used to perform two-layer mask synthesis processing on the reachability update twin graph and the blocking probability graph to generate an action mask, and construct feasible action candidate solutions with the action mask as a constraint.

[0254] The scheme generation module is used to solve the joint cost and establish the mapping relationship for the candidate solutions of feasible actions based on the abnormal evidence sequence, the blocking probability map and the reachability update twin graph, generate the joint cost function, and perform value screening on the candidate solutions of feasible actions based on the joint cost function to generate the route optimization scheme.

[0255] This embodiment has the same technical effects as Embodiment 1.

[0256] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. The data mentioned in this application have undergone normalization and other preprocessing to unify dimensions during formula calculations.

[0257] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A machine learning-based intelligent routing optimization method for software-defined networks, applied to intelligent routing optimization in underground mine collapse rescue networks, characterized in that... Includes the following steps: Obtain the link-state code and tunnel topology data of the target object; Perform topological skeleton mapping on the tunnel topology data to generate a twin topological base map; The link-state coding is subjected to time-series differential operation to generate a sequence of coded differential values, and anomaly measurement is performed on the sequence of coded differential values ​​to generate an anomaly evidence sequence. The twin topology base map is subjected to communication anomaly retrieval, and the path consistency inversion processing of the communication anomaly retrieval results in the neighborhood is performed according to the anomaly evidence sequence and the link state code to generate a blocking probability map. Then, the twin topology base map is updated according to the blocking probability map to generate a reachability update twin map. Perform a two-layer mask synthesis process on the reachability update twin graph and the blocking probability graph to generate an action mask, and construct a candidate solution for possible actions using the action mask as a constraint; Based on the abnormal evidence sequence, the blocking probability graph, and the reachability update twin graph, the joint cost solution and mapping relationship are established for the candidate solutions of the possible actions, generating a joint cost function. Based on the joint cost function, the candidate solutions of the possible actions are value-selected to generate a route optimization scheme.

2. The intelligent routing optimization method based on machine learning in software-defined networks according to claim 1, characterized in that: The process of mapping the tunnel topology data into a topological skeleton to generate a twin topological base map specifically includes: The tunnel topology data is compressed and mapped to generate a topology skeleton sequence, and the topology skeleton sequence is indexed and labeled to generate a deployment skeleton index. The tunnel topology data and the topology skeleton sequence are structured and encapsulated into a graph structure based on the deployment skeleton index to generate a twin topology base map.

3. The intelligent routing optimization method based on machine learning in software-defined networks according to claim 2, characterized in that: Based on the abnormal evidence sequence and the link state code, path consistency inversion processing is performed on the communication anomaly retrieval results within the neighborhood to generate a blocking probability map, specifically including: An evidence aggregation mapping is established for the communication anomaly retrieval results; wherein, the evidence aggregation mapping is determined by the correspondence between the deployment skeleton index and the twin topology base map; According to the evidence aggregation mapping, corresponding evidence items are extracted from the abnormal evidence sequence to form an evidence set, and the evidence strength value is calculated based on the evidence set; Based on the encoded time series, the evidence set is subjected to a difference operation on each pair of adjacent items in chronological order to obtain the evidence consistency value; wherein, the encoded time series is obtained by performing a time organization operation on the link state encoding; The evidence strength value and the evidence consistency value are combined to generate a blocking probability map.

4. The intelligent routing optimization method based on machine learning in software-defined networks according to claim 2, characterized in that: Performing a two-layer mask synthesis process on the reachability update twin graph and the blocking probability graph to generate an action mask specifically includes: Perform an intersection operation on the topological skeleton sequence and the reachability update twin graph to generate a set of reachable skeleton units. Extract the structural units and adjacency relationships corresponding to the set of reachable skeleton units from the twin topological base graph and assemble them into a reachability constraint subgraph. The blocking probability graph is complemented and mapped to generate a risk availability value. The risk availability value is then written into the reachability constraint subgraph to form a risk labeling subgraph. The risk labeling subgraph is then subjected to membership mapping to generate a structural feasibility label value. The risk availability value is mapped to generate a risk feasibility weight, and the structural feasibility label value and the risk feasibility weight are multiplied to generate an action mask.

5. The intelligent routing optimization method based on machine learning in software-defined networks according to claim 4, characterized in that: Constructing candidate solutions for actionable actions using the action mask as a constraint specifically includes: The deployment skeleton index is subjected to position filtering and sequence encoding to generate deployment sequence encoding. Based on the deployment sequence encoding, structural unit mapping is performed in the reachability constraint subgraph and written according to the adjacent relationship to generate path sequence encoding. The deployment sequence code and the path sequence code are concatenated and encapsulated to generate candidate decodes, and the path sequence code corresponding to the candidate decodes is segmented into actionable parts to generate a candidate action set. Delete the candidate actions whose action mask is zero, and reorganize the remaining candidate actions according to their original adjacency relationship to generate a candidate solution for possible actions.

6. The intelligent routing optimization method based on machine learning in software-defined networks according to claim 5, characterized in that: Based on the abnormal evidence sequence, the blocking probability graph, and the reachability update twin graph, a joint cost calculation and mapping relationship are established for the candidate solutions of the possible actions, generating a joint cost function specifically including: The candidate solutions for the possible actions are subjected to cover matching in the reachability update twin graph to obtain a set of cover structure units; Based on the set of coverage structure units, probability extraction is performed on the blocking probability map, and the arithmetic mean is calculated to generate the risk mean. Based on the set of covered structural units, evidence extraction is performed in the abnormal evidence sequence, and the arithmetic mean is calculated to generate the evidence mean. Based on the set of covered structural units, update extraction is performed in the reachability update twin graph, and the arithmetic mean is calculated to generate the change mean; A weighted composite operation is performed on the mean risk, the mean evidence, and the mean change to generate a joint cost value, and a mapping relationship is established between the candidate solutions of the actionable actions and the joint cost value to obtain the joint cost function.

7. The intelligent routing optimization method based on machine learning in software-defined networks according to claim 6, characterized in that: The specific steps for generating a route optimization scheme include: performing value screening on the candidate solutions of the feasible actions based on the joint cost function; and generating a route optimization scheme based on the following. The candidate solutions for action are subjected to rank fusion encoding according to the joint cost function to generate a candidate set of deployment paths. Then, action segments are constructed and filtered item by item in the candidate set of deployment paths according to the action mask to generate a candidate action set. The candidate action set is calculated and summarized based on the abnormal evidence sequence, the encoded difference value sequence, the blocking probability map, and the joint cost function to generate a reward component table; The reward subtable is subjected to strategy benefit aggregation and sorting processing to generate a strategy value table, and the abnormal evidence sequence and the coded difference value sequence are normalized and synthesized to generate change level value. The strategy value table, the deployment path candidate set, and the change level value are combined in a structured manner to generate a route optimization scheme.

8. The intelligent routing optimization method based on machine learning in software-defined networks according to claim 7, characterized in that: The strategy benefit aggregation and sorting process is performed on the reward subtable to generate the strategy value table, specifically including: The dispersion of the reward component table is calculated, and the minimum dispersion is selected. Then, adaptive weights are calculated based on the dispersion and the minimum dispersion. The reward subtable is weighted and summed according to the adaptive weights, and the weighted summation result is multiplied by the corresponding feasibility value to obtain the strategy benefit value. The strategy benefit values ​​are then integrated to generate a strategy value table. The feasibility value is obtained by multiplying the action mask corresponding to the candidate action set.

9. The intelligent routing optimization method based on machine learning in software-defined networks according to claim 7, characterized in that: The strategy value table, the deployment path candidate set, and the change level value are structurally combined to generate a route optimization scheme, specifically including: The strategy value table is sorted in descending order to obtain a value ranking sequence. Based on the first item in the value ranking sequence, the deployment selection expression and path selection expression are extracted from the deployment path candidate set to generate the deployment selection result and main path set accordingly. Calculate the difference between the remaining items in the value ranking sequence and the main path set, and sort the remaining items according to the difference, selecting the top-ranked item A to generate a backup path set; The deployment selection results, the primary path set, and the backup path set are combined in a structured manner to generate a routing optimization scheme.

10. A machine learning-based intelligent routing optimization system for software-defined networks, characterized in that, include: The data acquisition module is used to acquire the link status code and tunnel topology data of the target object; The topology processing module is used to perform topology skeleton mapping on the tunnel topology data and generate a twin topology base map; An anomaly analysis module is used to perform time-series differential operations on the link state coding to generate a sequence of coded differential values, and to perform anomaly measurement on the sequence of coded differential values ​​to generate an anomaly evidence sequence. The update module is used to perform communication anomaly retrieval on the twin topology base map, and perform path consistency inversion processing on the communication anomaly retrieval results in the neighborhood according to the anomaly evidence sequence and the link state code to generate a blocking probability map, and then update the twin topology base map according to the blocking probability map to generate a reachability updated twin map. The candidate solution construction module is used to perform a two-layer mask synthesis process on the reachability update twin graph and the blocking probability graph to generate an action mask, and construct feasible action candidate solutions with the action mask as a constraint. The scheme generation module is used to solve the joint cost and establish the mapping relationship for the candidate solutions of the possible actions based on the abnormal evidence sequence, the blocking probability graph and the reachability update twin graph, generate a joint cost function, and perform value screening on the candidate solutions of the possible actions based on the joint cost function to generate a route optimization scheme.

Images