A station building intelligent auxiliary decision generation method based on a knowledge graph

Abstract

This invention discloses a knowledge graph-based intelligent auxiliary decision-making generation method for railway stations, comprising the following steps: S1, collecting multi-source operational data of the railway station and preprocessing it; S2, establishing a node set according to preset entity mapping rules and establishing an edge set based on semantic decision relationships to generate a knowledge graph; S3, generating vector representations of each node in the knowledge graph, selecting a starting node, and performing structural traversal based on multi-hop adjacency relationships to establish a state subgraph; S4, performing strategy path evolution in the state subgraph and modeling the conflict relationships between path sequences; S5, parsing the task dependencies in the paths, assembling node content and associated attributes, generating and issuing auxiliary decision-making instructions; S6, collecting feedback data and incrementally updating node attributes and path evolution strategies. This invention enables the structured generation of auxiliary decision-making instructions for railway stations, improving the stability of the decision-making process.

Description

Technical Field

[0001] This invention relates to the field of station building decision support technology, and in particular to a knowledge graph-based intelligent auxiliary decision generation method for station buildings. Background Technology

[0002] With the continuous expansion of communication station scale and the increasing complexity of the operating environment, the demand for intelligent decision-making in equipment operation monitoring, resource scheduling, operation and maintenance, and risk management is constantly increasing. Existing station management systems mostly rely on rule configuration or human experience for decision support, and usually judge the operating status based on threshold alarms, fixed processes, or simple association rules, making it difficult to uniformly model and analyze the complex relationships between multi-source operating data.

[0003] While some existing technologies incorporate data analysis or intelligent algorithms to assess the operational status of railway stations, most solutions focus on single indicators or localized data processing, lacking the overall ability to express the semantic relationships between station operational elements. This makes it difficult to characterize the multi-layered dependencies between equipment status, operating conditions, and response strategies. When operational scenarios change or abnormal situations overlap, existing technologies often require manual intervention or rule reconfiguration, limiting the efficiency and stability of decision-making responses.

[0004] Furthermore, existing research on decision generation processes largely focuses on output levels, lacking structured constraints and conflict resolution mechanisms for the decision path formation process. In situations involving multiple tasks and strategies, conflicts can easily arise in decision order, resource allocation, or inconsistent execution logic. Consequently, the generated decision results are difficult to directly translate into executable instructions, reducing the practicality of decision support.

[0005] Meanwhile, some existing solutions fail to make full use of feedback information after decision execution, and there is a lack of an effective closed-loop update mechanism between the decision model and the operating environment. This makes it difficult to continuously adjust the decision strategy as the station's operating status changes, and the decision may easily deviate or fail in the long run.

[0006] Therefore, how to provide a knowledge graph-based intelligent auxiliary decision-making generation method for railway stations is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] One objective of this invention is to propose a knowledge graph-based intelligent auxiliary decision generation method for railway stations. This invention fully utilizes knowledge graph modeling, graph embedding representation, path game and strategy evolution algorithms, and instruction-level topology parsing technology to perform semantic modeling and association reasoning on multi-source operational data of railway stations, construct state-aware decision paths, and generate executable auxiliary decision instructions through adversarial reasoning and feedback-driven incremental update mechanisms. The invention details the complete technical process from semantic extraction, path generation, conflict resolution to instruction generation and continuous updating, and possesses the advantages of clear decision logic, strict structural constraints, strong interpretability, and outstanding continuous adaptive update capabilities.

[0008] A knowledge graph-based intelligent auxiliary decision generation method for railway stations according to an embodiment of the present invention includes the following steps: S1. Collect multi-source operational data from the station and preprocess it to construct a standard semantic set; S2. Based on the standard semantic set, establish a node set according to the preset entity mapping rules, and establish an edge set by combining the semantic decision relationship between nodes to generate a knowledge graph; S3. A graph embedding algorithm based on random walk is used to generate vector representations of each node in the knowledge graph. The node that is closest to the semantics of the real-time running state is selected as the starting node by using the vector distance metric method. The structure traversal is performed in combination with multi-hop adjacency relationship to establish the state subgraph. S4. In the state subgraph, the policy path evolution is carried out based on the semantic dependency relationship of the nodes in the graph and the edge attribute weight, and the conflict relationship between the path sequences is modeled. Multi-round adversarial reasoning is carried out through the path policy evolution mechanism. S5. Based on the path sequence after adversarial reasoning, the topological sorting method is used to parse the task dependencies in the path, and combined with the preset instruction template mapping mechanism, the node content and associated attributes in the path are structurally assembled to generate and issue auxiliary decision-making instructions. S6. Collect execution feedback data of auxiliary decision-making instructions, and use a joint mechanism based on dependency propagation and weight adjustment to incrementally update the strategies for node attributes and path evolution in the knowledge graph.

[0009] Optionally, the preprocessing specifically includes: performing data type identification and timestamp alignment operations on the collected multi-source running data; correcting missing values ​​and outliers using correction rules based on statistical thresholds; segmenting text-type running data into semantic units using word segmentation and part-of-speech tagging methods; performing standardized encoding on structured running data according to field mapping rules; and mapping the processed results uniformly to semantic items in the standard semantic set.

[0010] Optionally, S2 specifically includes: S21. Based on the semantic units in the standard semantic set, a rule-driven entity mapping method is adopted to map the semantic units into entity nodes of different types, and to assign a set of node attributes to each node according to the entity type. S22. Perform attribute consistency processing on the attribute set of entity nodes, specifically including: For attribute items in different nodes, semantic similarity is calculated for attribute names to identify attribute pairs with the same semantics, and corresponding attribute items are merged. For attribute items with the same attribute name but different semantics, determine the semantic conflict relationship by combining the entity type of the node and the field content, and rename the attribute name or adjust its attribute category. Set the data type, value range, or set of selectable values ​​for all attribute items to form legality constraint rules and attach them to the corresponding attributes; S23. Based on the semantic association between entity nodes, the relation template matching method is used to identify the semantic decision relation type between nodes, and corresponding relation edges are generated for different relation types. S24. Add edge attribute weights to the generated relation edges. The edge attribute weights are calculated based on the co-occurrence frequency of node pairs in historical semantic data and the degree of dependency of attribute values. They serve as a numerical representation of the strength of decision association between nodes, thus constructing a complete knowledge graph structure.

[0011] Optionally, S3 specifically includes: S31. Based on the edge connection relationships of the knowledge graph, an adjacency list is constructed for each node, and a graph embedding algorithm based on random walk is used to sample the paths of the nodes. The random walk process includes: Starting from any node in the knowledge graph, retrieve the set of candidate adjacent nodes directly connected to the current node from the adjacency list; Based on the weight values ​​of the edges corresponding to the candidate adjacent nodes, the candidate adjacent nodes are normalized, and the normalization result is used as the node transition probability. The next hop node is randomly selected from the set of candidate adjacent nodes according to the node transition probability, and the selected node is added to the current access sequence. Repeat the node selection process until the preset maximum step size is reached to obtain a node visit sequence. Repeat the random walk process for each node to generate multiple node visit sequences starting from that node; S32. The node visit sequence generated by the random walk is windowed to construct node context pairs. Skip-Gram is used to embed the node context pairs. The node vector parameters are updated through the negative sampling mechanism to obtain the node embedding vector of the node structural proximity relationship. S33. Based on the standard semantic set corresponding to the real-time running state, extract the state-related semantic items and numerical items, perform normalization processing on the numerical items, perform vector mapping on the semantic items, and concatenate the two to generate a state semantic vector with the same dimension as the node embedding vector. S34. Using the vector distance metric, calculate the similarity score between the embedding vector of each node and the state semantic vector, and select the nodes that meet the threshold condition as the starting node set based on the similarity ranking results. S35. Using the set of starting nodes as the starting point of traversal, perform a layer-by-layer structural traversal based on the multi-hop adjacency relationship in the knowledge graph, and record the node access path and its corresponding edge relationship during the traversal process to form a node set and a relationship set, and construct a state subgraph.

[0012] Optionally, the embedding training specifically includes: Using the current node in the node access sequence as the center node, select adjacent nodes within a preset window range as context nodes to form a positive sample node pair consisting of the center node and the context node. Based on the Skip-Gram prediction mechanism, the conditional probability of generating context nodes is calculated using the center node vector as input, and the training objective is to maximize the conditional probability. During training, several nodes are selected from non-context nodes as negative sample nodes, and the center node and the negative sample nodes are paired to form a negative sample node pair. A loss function is constructed based on positive sample node pairs and negative sample node pairs, and the parameters of the center node vector and the context node vector are adjusted respectively through gradient update. Repeat the node pair construction and parameter update process until the preset training rounds or convergence conditions are met, and obtain the node embedding vector that reflects the proximity relationship of the node structure.

[0013] Optionally, S4 specifically includes: S41. Perform semantic dependency analysis on the node pairs in the state subgraph. Based on the type mapping relationship and connection order between entity nodes, determine the semantic dependency direction and calculate the dependency strength between node pairs by combining the edge attribute weights. S42. Set the upper limit of path growth depth and the limit of circular access. Starting from the set of starting nodes, adopt the depth-first path expansion strategy to extend the path of node pairs whose dependency strength meets the set threshold, and form a path set. The upper limit of the path growth depth is an integer parameter, determined by the maximum number of nodes that can be added consecutively during the path expansion process. When the number of nodes in the current path reaches this upper limit, the continued expansion of the current path is terminated. The circular access restriction maintains an access count table for each path during the path expansion process. When a node appears a certain number of times in the current path reaches a preset threshold, the subsequent expansion of that node is skipped. S43. Using node resource conflict identification rules and temporal order constraint rules, determine the conflict relationship between any two path sequences in the path set, and mark and classify overlapping nodes or sequential conflict relationships. S44. For the path set that has completed conflict marking and classification, a genetic algorithm is used to perform strategy path evolution, encode the path sequence into genetic individuals, set the path selection probability and structural perturbation control parameters, and perform selection, crossover and mutation operations at the path node sequence level. S45. Based on the evolved path set, a multi-round adversarial reasoning mechanism is adopted, using the result of each round of iteration as the input of the next round, and gradually modifying the path structure according to the change of the conflict marker state until the path set meets the preset conflict resolution rate threshold.

[0014] Optionally, the strategy path evolution specifically includes: Each path in the path set is encoded as a fixed-length node sequence vector, which is composed of unique node identifiers, and retains edge weights and conflict markers in the path as additional bits. Construct a fitness function and calculate the fitness value based on the number of conflict markers between nodes in the path sequence and the cumulative edge weight of the path. The roulette wheel selection algorithm is used to select individual paths from the current path set according to the probability distribution of fitness values ​​to form the parent set; For any two path sequences in the parent set, randomly set the intersection position within the matching length range, swap the subsequences after the intersection, form a new path sequence, and re-encode the node order. Set mutation probability for the generated path sequence, and randomly select mutation positions from the remaining nodes after excluding the start node, end node and positions with conflict markers; At the selected mutation location, based on the adjacency relationship in the knowledge graph, a replacement node is selected from the set of candidate nodes that are connected to the nodes before and after the location by edges, and the replacement node is inserted into the path sequence to complete the path mutation operation; The crossover and mutation-derived path sets are used as the next generation of candidate paths. The fitness calculation and individual selection process is repeated. The results are sorted according to fitness values, and the path sequences that meet the preset retention conditions are retained. Set the maximum number of generations and the minimum number of conflicts thresholds. When either condition is met, terminate the evolution process and output the set of path sequences with the best fitness in the current generation.

[0015] Optionally, S5 specifically includes: S51. Analyze the node sequence in each path sequence after adversarial reasoning, read the entity type and operation attribute identifier corresponding to the node, select the entity node with the executable operation identifier as the task node, and record the edge connection relationship and its direction information between the task nodes. S52. Construct a directed acyclic structure based on the directional information between task nodes. Use the Kahn topological sorting algorithm to select the task node with an in-degree of zero in each round of sorting and add it to the execution sequence. Remove the outgoing edge corresponding to the task node and update the in-degree value of the remaining task nodes until the sorting of all task nodes is completed. S53. Process the task nodes one by one according to the execution sequence, and perform the following operations on each task node: Read the attribute value representing the operation category in the task node to determine the corresponding instruction operation type; Read the attribute values ​​representing control parameters from the task node, and write the attribute values ​​into the instruction parameter positions according to the parameter names; Read the attribute values ​​representing dependencies from the outgoing edges of the task node, and write the dependencies into the precondition flag of the current instruction; S54. Sequentially concatenate the instruction contents generated by each task node according to the execution sequence. During the concatenation process, align the precondition identifiers of adjacent instructions with the trigger identifiers of subsequent instructions to form auxiliary decision-making instructions. S55. Based on the precondition identifier and trigger identifier in the auxiliary decision instruction sequence, construct an instruction dependency directed graph, and use a depth-first search algorithm to perform loop detection on the instruction dependency directed graph.

[0016] Optionally, S6 specifically includes: S61. Establish instruction execution identifiers for the issued auxiliary decision-making instructions, and collect corresponding execution status data for each instruction. The execution status data includes at least an execution success identifier, an execution failure identifier, an execution interruption identifier, and the execution order completion status. S62. Based on the instruction execution identifier, the execution status data is back-linked to the task node that generated the instruction and its corresponding path sequence, establishing a mapping relationship between instruction, task node, and path sequence, and the status of the instruction is determined based on the execution status data: When the instruction execution status is failure or interruption, extract the node attribute values ​​of the corresponding task node and the edge attribute weights of its incoming and outgoing edges. When the instruction execution status is successful, extract the edge attribute weights of the corresponding task node to form a set of edge weights to be updated; S63. For the set of edge weights to be updated, a weight update algorithm based on Bellman relaxation rules is adopted. The edge attribute weights are relaxed and updated edge by edge according to the instruction execution status. The edge weights corresponding to instructions that fail or are interrupted are decayed and updated, while the edge weights corresponding to instructions that are successfully executed are enhanced. S64. Based on the updated edge attribute weights, perform dependency propagation along the edge connection direction in the knowledge graph, pass the edge weight changes to adjacent nodes layer by layer, and synchronously update the node attribute values ​​representing the dependency strength in the adjacent nodes. S65. Based on the updated node attributes and edge attributes, read the updated edge attribute weights from the knowledge graph, replace the weight parameters of the corresponding edges during the path strategy evolution process, and complete the incremental update.

[0017] The beneficial effects of this invention are: First, this invention uses unified semantic parsing and knowledge graph modeling of multi-source operational data of the station to express equipment status, operating conditions and disposal relationships in a structured form, thus avoiding the problems of data fragmentation and semantic inconsistency in the prior art.

[0018] Secondly, this invention introduces path generation and strategy evolution mechanisms on the basis of knowledge graphs, performs structured search, conflict identification and multi-round reasoning on decision paths that match the current running state, and transforms complex decision results into auxiliary decision instructions with clear execution order and dependencies through topological constraints and instruction-level parsing.

[0019] Finally, by continuously collecting and backtracking the results of the execution of auxiliary decision-making instructions, and combining dependency propagation and weight update mechanisms, this invention incrementally updates the node attributes and path-related parameters in the knowledge graph, enabling the decision generation process to be continuously adjusted as the station's operating status changes. Attached Figure Description

[0020] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of a knowledge graph-based intelligent auxiliary decision generation method for station buildings proposed in this invention; Figure 2 This is a schematic diagram of the state subgraph construction and node matching process of a knowledge graph-based intelligent auxiliary decision generation method for station buildings proposed in this invention. Figure 3 This is a schematic diagram of the auxiliary decision-making instruction generation process of the intelligent auxiliary decision-making generation method for station buildings based on knowledge graph proposed in this invention. Detailed Implementation

[0021] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0022] refer to Figure 1-3 A knowledge graph-based intelligent auxiliary decision generation method for railway stations includes the following steps: S1. Collect multi-source operational data from the station and preprocess it to construct a standard semantic set; S2. Based on the standard semantic set, establish a node set according to the preset entity mapping rules, and establish an edge set by combining the semantic decision relationship between nodes to generate a knowledge graph; S3. A graph embedding algorithm based on random walk is used to generate vector representations of each node in the knowledge graph. The node that is closest to the semantics of the real-time running state is selected as the starting node by using the vector distance metric method. The structure traversal is performed in combination with multi-hop adjacency relationship to establish the state subgraph. S4. In the state subgraph, the policy path evolution is carried out based on the semantic dependency relationship of the nodes in the graph and the edge attribute weight, and the conflict relationship between the path sequences is modeled. Multi-round adversarial reasoning is carried out through the path policy evolution mechanism. S5. Based on the path sequence after adversarial reasoning, the topological sorting method is used to parse the task dependencies in the path, and combined with the preset instruction template mapping mechanism, the node content and associated attributes in the path are structurally assembled to generate and issue auxiliary decision-making instructions. S6. Collect execution feedback data of auxiliary decision-making instructions, and use a joint mechanism based on dependency propagation and weight adjustment to incrementally update the strategies for node attributes and path evolution in the knowledge graph.

[0023] In this embodiment, the preprocessing specifically includes: performing data type identification and timestamp alignment operations on the collected multi-source running data; correcting missing values ​​and outliers using correction rules based on statistical thresholds; segmenting text-type running data into semantic units using word segmentation and part-of-speech tagging methods; performing standardized encoding on structured running data according to field mapping rules; and mapping the processed results uniformly to semantic items in the standard semantic set.

[0024] In this embodiment, S2 specifically includes: S21. Based on semantic units in the standard semantic set, a rule-driven entity mapping method is used to map semantic units into entity nodes of different types, and a set of node attributes is assigned to each node according to the entity type. The entity mapping method includes: Extract field names and values ​​from semantic units, perform string matching on field names, and identify feature fields that represent entity categories; Among the identified fields, the entity type to which the field belongs is determined based on the field's location and the context structure within the semantic unit; The field values ​​are bound to the attribute names represented by the field names and combined with the identified entity types to form entity nodes and their attribute sets. By setting the number prefix and field encoding method, a unique identifier for the entity node is generated; S22. Perform attribute consistency processing on the attribute set of entity nodes, specifically including: For attribute items in different nodes, semantic similarity is calculated for attribute names to identify attribute pairs with the same semantics, and corresponding attribute items are merged. For attribute items with the same attribute name but different semantics, determine the semantic conflict relationship by combining the entity type of the node and the field content, and rename the attribute name or adjust its attribute category. Set the data type, value range, or set of selectable values ​​for all attribute items to form legality constraint rules and attach them to the corresponding attributes; S23. Based on the semantic association between entity nodes, the relation template matching method is used to identify the semantic decision relation type between nodes, and corresponding relation edges are generated for different relation types. S24. Add edge attribute weights to the generated relation edges. The edge attribute weights are calculated based on the co-occurrence frequency of node pairs in historical semantic data and the degree of dependency of attribute values. They serve as a numerical representation of the strength of decision association between nodes, thus constructing a complete knowledge graph structure.

[0025] In this embodiment, S3 specifically includes: S31. Based on the edge connection relationships of the knowledge graph, an adjacency list is constructed for each node, and a graph embedding algorithm based on random walk is used to sample the paths of the nodes. The random walk process includes: Starting from any node in the knowledge graph, retrieve the set of candidate adjacent nodes directly connected to the current node from the adjacency list; Based on the weight values ​​of the edges corresponding to the candidate adjacent nodes, the candidate adjacent nodes are normalized, and the normalization result is used as the node transition probability. The next hop node is randomly selected from the set of candidate adjacent nodes according to the node transition probability, and the selected node is added to the current access sequence. Repeat the node selection process until the preset maximum step size is reached to obtain a node visit sequence. Repeat the random walk process for each node to generate multiple node visit sequences starting from that node; S32. The node visit sequence generated by the random walk is windowed to construct node context pairs. Skip-Gram is used to embed the node context pairs. The node vector parameters are updated through the negative sampling mechanism to obtain the node embedding vector of the node structural proximity relationship. S33. Based on the standard semantic set corresponding to the real-time running state, extract the state-related semantic items and numerical items, perform normalization processing on the numerical items, perform vector mapping on the semantic items, and concatenate the two to generate a state semantic vector with the same dimension as the node embedding vector. S34. Using the vector distance metric, calculate the similarity score between the embedding vector of each node and the state semantic vector, and select the nodes that meet the threshold condition as the starting node set based on the similarity ranking results. S35. Using the set of starting nodes as the starting point of traversal, perform a layer-by-layer structural traversal based on the multi-hop adjacency relationship in the knowledge graph, and record the node access path and its corresponding edge relationship during the traversal process to form a node set and a relationship set, and construct a state subgraph.

[0026] In this embodiment, the embedding training specifically includes: Using the current node in the node access sequence as the center node, select adjacent nodes within a preset window range as context nodes to form a positive sample node pair consisting of the center node and the context node. Based on the Skip-Gram prediction mechanism, the conditional probability of generating context nodes is calculated using the center node vector as input, and the training objective is to maximize the conditional probability. During training, several nodes are selected from non-context nodes as negative sample nodes, and the center node and the negative sample nodes are paired to form a negative sample node pair. A loss function is constructed based on positive sample node pairs and negative sample node pairs, and the parameters of the center node vector and the context node vector are adjusted respectively through gradient update. Repeat the node pair construction and parameter update process until the preset training rounds or convergence conditions are met, and obtain the node embedding vector that reflects the proximity relationship of the node structure.

[0027] In this embodiment, S4 specifically includes: S41. Perform semantic dependency analysis on the node pairs in the state subgraph. Based on the type mapping relationship and connection order between entity nodes, determine the semantic dependency direction and calculate the dependency strength between node pairs by combining the edge attribute weights. S42. Set the upper limit of path growth depth and the limit of circular access. Starting from the set of starting nodes, adopt the depth-first path expansion strategy to extend the path of node pairs whose dependency strength meets the set threshold, and form a path set. The upper limit of the path growth depth is an integer parameter, determined by the maximum number of nodes that can be added consecutively during the path expansion process. When the number of nodes in the current path reaches this upper limit, the continued expansion of the current path is terminated. The circular access restriction maintains an access count table for each path during the path expansion process. When a node appears a certain number of times in the current path reaches a preset threshold, the subsequent expansion of that node is skipped. S43. Using node resource conflict identification rules and temporal order constraint rules, determine the conflict relationship between any two path sequences in the path set, and mark and classify overlapping nodes or sequential conflict relationships. S44. For the path set that has completed conflict marking and classification, a genetic algorithm is used to perform strategy path evolution, encode the path sequence into genetic individuals, set the path selection probability and structural perturbation control parameters, and perform selection, crossover and mutation operations at the path node sequence level. S45. Based on the evolved path set, a multi-round adversarial reasoning mechanism is adopted, using the result of each round of iteration as the input of the next round, and gradually modifying the path structure according to the change of the conflict marker state until the path set meets the preset conflict resolution rate threshold.

[0028] In this embodiment, the strategy path evolution specifically includes: Each path in the path set is encoded as a fixed-length node sequence vector, which is composed of unique node identifiers, and retains edge weights and conflict markers in the path as additional bits. Construct a fitness function and calculate the fitness value based on the number of conflict markers between nodes in the path sequence and the cumulative edge weight of the path. The roulette wheel selection algorithm is used to select individual paths from the current path set according to the probability distribution of fitness values ​​to form the parent set; For any two path sequences in the parent set, randomly set the intersection position within the matching length range, swap the subsequences after the intersection, form a new path sequence, and re-encode the node order. Set mutation probability for the generated path sequence, and randomly select mutation positions from the remaining nodes after excluding the start node, end node and positions with conflict markers; At the selected mutation location, based on the adjacency relationship in the knowledge graph, a replacement node is selected from the set of candidate nodes that are connected to the nodes before and after the location by edges, and the replacement node is inserted into the path sequence to complete the path mutation operation; The crossover and mutation-derived path sets are used as the next generation of candidate paths. The fitness calculation and individual selection process is repeated. The results are sorted according to fitness values, and the path sequences that meet the preset retention conditions are retained. Set the maximum number of generations and the minimum number of conflicts thresholds. When either condition is met, terminate the evolution process and output the set of path sequences with the best fitness in the current generation.

[0029] In this embodiment, S5 specifically includes: S51. Analyze the node sequence in each path sequence after adversarial reasoning, read the entity type and operation attribute identifier corresponding to the node, select the entity node with the executable operation identifier as the task node, and record the edge connection relationship and its direction information between the task nodes. S52. Construct a directed acyclic structure based on the directional information between task nodes. Use the Kahn topological sorting algorithm to select the task node with an in-degree of zero in each round of sorting and add it to the execution sequence. Remove the outgoing edge corresponding to the task node and update the in-degree value of the remaining task nodes until the sorting of all task nodes is completed. S53. Process the task nodes one by one according to the execution sequence, and perform the following operations on each task node: Read the attribute value representing the operation category in the task node to determine the corresponding instruction operation type; Read the attribute values ​​representing control parameters from the task node, and write the attribute values ​​into the instruction parameter positions according to the parameter names; Read the attribute values ​​representing dependencies from the outgoing edges of the task node, and write the dependencies into the precondition flag of the current instruction; S54. Sequentially concatenate the instruction contents generated by each task node according to the execution sequence. During the concatenation process, align the precondition identifiers of adjacent instructions with the trigger identifiers of subsequent instructions to form auxiliary decision-making instructions. S55. Based on the precondition identifiers and trigger identifiers in the auxiliary decision-making instruction sequence, a directed graph of instruction dependencies is constructed, and a depth-first search algorithm is used to perform loop detection on the directed graph of instruction dependencies, specifically including: When a node that is already in the recursive stack is visited again during the depth-first search, a circular dependency is determined to exist. Extract the instruction subsequence corresponding to the loop dependency, and locate the precondition identifier and trigger identifier fields of each instruction in the loop; The instructions to be processed are determined based on the number of conflicts and the number of dependent edges within the loop, and the instructions to be processed are either eliminated or their preconditions are cleared. For the instruction sequence after elimination or clearing, the instruction dependency directed graph is reconstructed and loop detection is repeatedly performed until no loop dependency is detected. The auxiliary decision instruction sequence that passes the loop detection is then output for issuance.

[0030] In this embodiment, S6 specifically includes: S61. Establish instruction execution identifiers for the issued auxiliary decision-making instructions, and collect corresponding execution status data for each instruction. The execution status data includes at least an execution success identifier, an execution failure identifier, an execution interruption identifier, and the execution order completion status. S62. Based on the instruction execution identifier, the execution status data is back-linked to the task node that generated the instruction and its corresponding path sequence, establishing a mapping relationship between instruction, task node, and path sequence, and the status of the instruction is determined based on the execution status data: When the instruction execution status is failure or interruption, extract the node attribute values ​​of the corresponding task node and the edge attribute weights of its incoming and outgoing edges. When the instruction execution status is successful, extract the edge attribute weights of the corresponding task node to form a set of edge weights to be updated; S63. For the set of edge weights to be updated, a weight update algorithm based on Bellman relaxation rules is adopted. The edge attribute weights are relaxed and updated edge by edge according to the instruction execution status. The edge weights corresponding to instructions that fail or are interrupted are decayed and updated, while the edge weights corresponding to instructions that are successfully executed are enhanced. S64. Based on the updated edge attribute weights, perform dependency propagation along the edge connection direction in the knowledge graph, pass the edge weight changes to adjacent nodes layer by layer, and synchronously update the node attribute values ​​representing the dependency strength in the adjacent nodes. S65. Based on the updated node attributes and edge attributes, read the updated edge attribute weights from the knowledge graph, replace the weight parameters of the corresponding edges during the path strategy evolution process, and complete the incremental update.

[0031] Example 1: To verify the feasibility of this invention in practice, it was applied to a station operation management scenario characterized by multi-device collaborative operation. Such stations involve various equipment states, operating conditions, and operational behaviors during operation, with complex relationships between different data sources. Traditional rule-based or human experience-based decision-making methods struggle to form stable and consistent auxiliary decisions under complex conditions.

[0032] In this application scenario, the system first performs unified preprocessing on the multi-source operational data collected during the station's operation, transforming the data from different sources into a standard semantic expression form, and constructing a knowledge graph structure based on semantic units to describe equipment status, operational behavior, and their semantic dependencies. When the station's operational status changes, the system selects matching nodes from the knowledge graph based on the real-time status and constructs a state subgraph, generating multiple candidate decision paths in the subgraph.

[0033] To address potential execution order and resource consumption conflicts among candidate paths, the system identifies and marks these conflicts. Through path evolution and multi-round adversarial reasoning, it gradually filters out stable paths with clear dependencies and fewer conflicts. Subsequently, the system parses the nodes in the stable paths, identifying nodes with operational attributes as task nodes. Based on the directional relationships between nodes, it performs topological sorting to generate a task execution sequence that satisfies execution constraints. Building upon this, the system transforms the task node attributes into structured auxiliary decision-making instructions and verifies the dependencies between these instructions to ensure the executability of the instruction sequence.

[0034] During instruction execution, the system continuously collects execution feedback information and backtracks the feedback results to the corresponding nodes and paths in the knowledge graph. It incrementally updates the edge attribute weights in the graph and synchronously corrects the relevant node attributes through the dependency propagation mechanism, so that the subsequent decision-making process can be continuously adjusted based on the updated graph state.

[0035] To verify the feasibility of this invention in practice, the method of this invention was compared with the traditional rule-based decision-making method under the same operating conditions, and the following statistical results were obtained.

[0036] Table 1. Comparison of Application Effects of Knowledge Graph-Based Intelligent Decision-Making Aid Methods for Station Buildings

[0037] As can be seen from the results in Table 1, the method of the present invention is superior to the traditional method in terms of decision path stability, conflict control capability, and execution consistency, which verifies the feasibility and practicality of the present invention in the intelligent auxiliary decision-making scenario of station buildings.

[0038] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A knowledge graph-based intelligent auxiliary decision-making generation method for railway stations, characterized in that, Includes the following steps: S1. Collect multi-source operational data from the station and preprocess it to construct a standard semantic set; S2. Based on the standard semantic set, establish a node set according to the preset entity mapping rules, and establish an edge set by combining the semantic decision relationship between nodes to generate a knowledge graph; S3. A graph embedding algorithm based on random walk is used to generate vector representations of each node in the knowledge graph. The node that is closest to the semantics of the real-time running state is selected as the starting node by using the vector distance metric method. The structure traversal is performed in combination with multi-hop adjacency relationship to establish the state subgraph. S4. In the state subgraph, the policy path evolution is carried out based on the semantic dependency relationship of the nodes in the graph and the edge attribute weight, and the conflict relationship between the path sequences is modeled. Multi-round adversarial reasoning is carried out through the path policy evolution mechanism. S5. Based on the path sequence after adversarial reasoning, the topological sorting method is used to parse the task dependencies in the path, and combined with the preset instruction template mapping mechanism, the node content and associated attributes in the path are structurally assembled to generate and issue auxiliary decision-making instructions. S6. Collect execution feedback data of auxiliary decision-making instructions, and use a joint mechanism based on dependency propagation and weight adjustment to incrementally update the strategies for node attributes and path evolution in the knowledge graph.

2. The knowledge graph-based intelligent auxiliary decision-making generation method for station buildings according to claim 1, characterized in that, The preprocessing specifically includes: performing data type identification and timestamp alignment operations on the collected multi-source running data; correcting missing values ​​and outliers using correction rules based on statistical thresholds; segmenting text-type running data into semantic units using word segmentation and part-of-speech tagging methods; performing standardized encoding on structured running data according to field mapping rules; and mapping the processed results uniformly to semantic items in the standard semantic set.

3. The knowledge graph-based intelligent auxiliary decision-making generation method for station buildings according to claim 1, characterized in that, S2 specifically includes: S21. Based on the semantic units in the standard semantic set, a rule-driven entity mapping method is adopted to map the semantic units into entity nodes of different types, and to assign a set of node attributes to each node according to the entity type. S22. Perform attribute consistency processing on the attribute set of entity nodes, specifically including: For attribute items in different nodes, semantic similarity is calculated for attribute names to identify attribute pairs with the same semantics, and corresponding attribute items are merged. For attribute items with the same attribute name but different semantics, determine the semantic conflict relationship by combining the entity type of the node and the field content, and rename the attribute name or adjust its attribute category. Set the data type, value range, or set of selectable values ​​for all attribute items to form legality constraint rules and attach them to the corresponding attributes; S23. Based on the semantic association between entity nodes, the relation template matching method is used to identify the semantic decision relation type between nodes, and corresponding relation edges are generated for different relation types. S24. Add edge attribute weights to the generated relation edges. The edge attribute weights are calculated based on the co-occurrence frequency of node pairs in historical semantic data and the degree of dependency of attribute values. They serve as a numerical representation of the strength of decision association between nodes, thus constructing a complete knowledge graph structure.

4. The knowledge graph-based intelligent auxiliary decision-making generation method for station buildings according to claim 1, characterized in that, S3 specifically includes: S31. Based on the edge connection relationships of the knowledge graph, an adjacency list is constructed for each node, and a graph embedding algorithm based on random walk is used to sample the paths of the nodes. The random walk process includes: Starting from any node in the knowledge graph, retrieve the set of candidate adjacent nodes directly connected to the current node from the adjacency list; Based on the weight values ​​of the edges corresponding to the candidate adjacent nodes, the candidate adjacent nodes are normalized, and the normalization result is used as the node transition probability. The next hop node is randomly selected from the set of candidate adjacent nodes according to the node transition probability, and the selected node is added to the current access sequence. Repeat the node selection process until the preset maximum step size is reached to obtain a node visit sequence. Repeat the random walk process for each node to generate multiple node visit sequences starting from that node; S32. The node visit sequence generated by the random walk is windowed to construct node context pairs. Skip-Gram is used to embed the node context pairs. The node vector parameters are updated through the negative sampling mechanism to obtain the node embedding vector of the node structural proximity relationship. S33. Based on the standard semantic set corresponding to the real-time running state, extract the state-related semantic items and numerical items, perform normalization processing on the numerical items, perform vector mapping on the semantic items, and concatenate the two to generate a state semantic vector with the same dimension as the node embedding vector. S34. Using the vector distance metric, calculate the similarity score between the embedding vector of each node and the state semantic vector, and select the nodes that meet the threshold condition as the starting node set based on the similarity ranking results. S35. Using the set of starting nodes as the starting point of traversal, perform a layer-by-layer structural traversal based on the multi-hop adjacency relationship in the knowledge graph, and record the node access path and its corresponding edge relationship during the traversal process to form a node set and a relationship set, and construct a state subgraph.

5. The knowledge graph-based intelligent auxiliary decision-making generation method for station buildings according to claim 4, characterized in that, The embedding training specifically includes: Using the current node in the node access sequence as the center node, select adjacent nodes within a preset window range as context nodes to form a positive sample node pair consisting of the center node and the context node. Based on the Skip-Gram prediction mechanism, the conditional probability of generating context nodes is calculated using the center node vector as input, and the training objective is to maximize the conditional probability. During training, several nodes are selected from non-context nodes as negative sample nodes, and the center node and the negative sample nodes are paired to form a negative sample node pair. A loss function is constructed based on positive sample node pairs and negative sample node pairs, and the parameters of the center node vector and the context node vector are adjusted respectively through gradient update. Repeat the node pair construction and parameter update process until the preset training rounds or convergence conditions are met, and obtain the node embedding vector that reflects the proximity relationship of the node structure.

6. The knowledge graph-based intelligent auxiliary decision-making generation method for station buildings according to claim 1, characterized in that, S4 specifically includes: S41. Perform semantic dependency analysis on the node pairs in the state subgraph. Based on the type mapping relationship and connection order between entity nodes, determine the semantic dependency direction and calculate the dependency strength between node pairs by combining the edge attribute weights. S42. Set the upper limit of path growth depth and the limit of circular access. Starting from the set of starting nodes, adopt the depth-first path expansion strategy to extend the path of node pairs whose dependency strength meets the set threshold, and form a path set. The upper limit of the path growth depth is an integer parameter, determined by the maximum number of nodes that can be added consecutively during the path expansion process. When the number of nodes in the current path reaches this upper limit, the continued expansion of the current path is terminated. The circular access restriction maintains an access count table for each path during the path expansion process. When a node appears a certain number of times in the current path reaches a preset threshold, the subsequent expansion of that node is skipped. S43. Using node resource conflict identification rules and temporal order constraint rules, determine the conflict relationship between any two path sequences in the path set, and mark and classify overlapping nodes or sequential conflict relationships. S44. For the path set that has completed conflict marking and classification, a genetic algorithm is used to perform strategy path evolution, encode the path sequence into genetic individuals, set the path selection probability and structural perturbation control parameters, and perform selection, crossover and mutation operations at the path node sequence level. S45. Based on the evolved path set, a multi-round adversarial reasoning mechanism is adopted, using the result of each round of iteration as the input of the next round, and gradually modifying the path structure according to the change of the conflict marker state until the path set meets the preset conflict resolution rate threshold.

7. The knowledge graph-based intelligent auxiliary decision-making generation method for station buildings according to claim 6, characterized in that, The strategy path evolution specifically includes: Each path in the path set is encoded as a fixed-length node sequence vector, which is composed of unique node identifiers, and retains edge weights and conflict markers in the path as additional bits. Construct a fitness function and calculate the fitness value based on the number of conflict markers between nodes in the path sequence and the cumulative edge weight of the path. The roulette wheel selection algorithm is used to select individual paths from the current path set according to the probability distribution of fitness values ​​to form the parent set; For any two path sequences in the parent set, randomly set the intersection position within the matching length range, swap the subsequences after the intersection, form a new path sequence, and re-encode the node order. Set mutation probability for the generated path sequence, and randomly select mutation positions from the remaining nodes after excluding the start node, end node and positions with conflict markers; At the selected mutation location, based on the adjacency relationship in the knowledge graph, a replacement node is selected from the set of candidate nodes that are connected to the nodes before and after the location by edges, and the replacement node is inserted into the path sequence to complete the path mutation operation; The crossover and mutation-derived path sets are used as the next generation of candidate paths. The fitness calculation and individual selection process is repeated. The results are sorted according to fitness values, and the path sequences that meet the preset retention conditions are retained. Set the maximum number of generations and the minimum number of conflicts thresholds. When either condition is met, terminate the evolution process and output the set of path sequences with the best fitness in the current generation.

8. The knowledge graph-based intelligent auxiliary decision-making generation method for station buildings according to claim 1, characterized in that, S5 specifically includes: S51. Analyze the node sequence in each path sequence after adversarial reasoning, read the entity type and operation attribute identifier corresponding to the node, select the entity node with the executable operation identifier as the task node, and record the edge connection relationship and its direction information between the task nodes. S52. Construct a directed acyclic structure based on the directional information between task nodes. Use the Kahn topological sorting algorithm to select the task node with an in-degree of zero in each round of sorting and add it to the execution sequence. Remove the outgoing edge corresponding to the task node and update the in-degree value of the remaining task nodes until the sorting of all task nodes is completed. S53. Process the task nodes one by one according to the execution sequence, and perform the following operations on each task node: Read the attribute value representing the operation category in the task node to determine the corresponding instruction operation type; Read the attribute values ​​representing control parameters from the task node, and write the attribute values ​​into the instruction parameter positions according to the parameter names; Read the attribute values ​​representing dependencies from the outgoing edges of the task node, and write the dependencies into the precondition flag of the current instruction; S54. Sequentially concatenate the instruction contents generated by each task node according to the execution sequence. During the concatenation process, align the precondition identifiers of adjacent instructions with the trigger identifiers of subsequent instructions to form auxiliary decision-making instructions. S55. Based on the precondition identifier and trigger identifier in the auxiliary decision instruction sequence, construct an instruction dependency directed graph, and use a depth-first search algorithm to perform loop detection on the instruction dependency directed graph.

9. The knowledge graph-based intelligent auxiliary decision-making generation method for station buildings according to claim 1, characterized in that, S6 specifically includes: S61. Establish instruction execution identifiers for the issued auxiliary decision-making instructions, and collect corresponding execution status data for each instruction. The execution status data includes at least an execution success identifier, an execution failure identifier, an execution interruption identifier, and the execution order completion status. S62. Based on the instruction execution identifier, the execution status data is back-linked to the task node that generated the instruction and its corresponding path sequence, establishing a mapping relationship between instruction, task node, and path sequence, and the status of the instruction is determined based on the execution status data: When the instruction execution status is failure or interruption, extract the node attribute values ​​of the corresponding task node and the edge attribute weights of its incoming and outgoing edges. When the instruction execution status is successful, extract the edge attribute weights of the corresponding task node to form a set of edge weights to be updated; S63. For the set of edge weights to be updated, a weight update algorithm based on Bellman relaxation rules is adopted. The edge attribute weights are relaxed and updated edge by edge according to the instruction execution status. The edge weights corresponding to instructions that fail or are interrupted are decayed and updated, while the edge weights corresponding to instructions that are successfully executed are enhanced. S64. Based on the updated edge attribute weights, perform dependency propagation along the edge connection direction in the knowledge graph, pass the edge weight changes to adjacent nodes layer by layer, and synchronously update the node attribute values ​​representing the dependency strength in the adjacent nodes. S65. Based on the updated node attributes and edge attributes, read the updated edge attribute weights from the knowledge graph, replace the weight parameters of the corresponding edges during the path strategy evolution process, and complete the incremental update.

Images