Intelligent matching and decision method for urban rail transit accidents based on knowledge graph

Abstract

The application discloses a kind of city rail transit accident intelligent matching and decision-making method based on knowledge graph, comprising: constructing accident case knowledge graph and emergency plan knowledge graph;The historical accident disposal measures in accident case knowledge graph are carried out semantic analysis and standardization processing;Obtain the new accident feature vector corresponding to historical accident case characteristic space;Based on new accident feature vector, hierarchical case retrieval is carried out in accident case knowledge graph, and multiple similar historical accident cases are obtained;From accident case knowledge graph and emergency plan knowledge graph, corresponding accident disposal measures, standard measures and plan basis are obtained;Target accident information, similar historical accident case information and corresponding measure basis chain are input into large language model, and city rail transit accident emergency decision scheme is generated.The application can integrate accident case knowledge and emergency plan knowledge, realize the accurate matching of similar cases and generate intelligent decision scheme.

Description

Technical Field

[0001] This invention relates to the field of urban rail transit, specifically to a knowledge graph-based intelligent matching and decision-making method for urban rail transit accidents. Background Technology

[0002] As urban rail transit networks continue to expand, their operation involves multiple subsystems, including rolling stock, power supply, signaling, communication, tracks, and station organization. The operating environment is characterized by strong equipment interrelationships, complex business processes, and diverse types of emergencies. In actual operation, various incidents may occur, such as train equipment failures, power supply equipment anomalies, signaling system failures, passenger emergencies, and personnel entering the tracks. To address these incidents, operations dispatchers typically need to quickly identify the type and key characteristics of the incident based on the on-site description, search for similar events in historical accident cases, and develop corresponding response measures in conjunction with emergency plans to ensure operational safety and minimize the impact of the incident.

[0003] Currently, existing methods for retrieving urban rail transit accident cases typically rely on keyword matching or simple text similarity calculations. However, due to the often different expressions used in accident reports and on-site descriptions, the same fault phenomenon may be described in different language forms. For example, phrases like "doors cannot be opened," "doors cannot be closed," and "doors are stuck" may all correspond to the same type of door fault. Traditional keyword matching methods struggle to accurately identify the semantic relationships between different expressions, easily leading to incomplete similar case retrieval or insufficient matching accuracy. Furthermore, existing methods often fail to fully consider the hierarchical relationships between accident location, accident cause category, and specific fault type, making it difficult to achieve refined matching based on accident characteristics. On the other hand, urban rail transit emergency plans typically describe responsible entities, duties, and handling procedures in the form of normative clauses, while actual handling measures in historical accident cases are often scenario-based, experience-based, and fragmented in their expression, lacking an effective semantic association mechanism between the two. Although knowledge graph technology can provide a structured representation of the relationships between accidents, measures, departments, responsibilities, and contingency plan clauses, existing knowledge graph-based retrieval methods typically only return relevant nodes or relationship paths, making it difficult to further combine historical case experience to form emergency decision-making solutions with practical guidance.

[0004] Therefore, to solve the above problems, a knowledge graph-based intelligent matching and decision-making method for urban rail transit accidents is needed. This method can integrate accident case knowledge and emergency plan knowledge to achieve accurate matching of similar cases and generate intelligent decision-making solutions, thereby improving the efficiency of information utilization and decision support capabilities in the emergency response process. Summary of the Invention

[0005] In view of this, the purpose of this invention is to overcome the deficiencies in the prior art and provide an intelligent matching and decision-making method for urban rail transit accidents based on knowledge graphs. This method can integrate accident case knowledge and emergency plan knowledge to achieve accurate matching of similar cases and generate intelligent decision-making solutions, thereby improving the efficiency of information utilization and decision support capabilities in the emergency response process.

[0006] The present invention provides a knowledge graph-based intelligent matching and decision-making method for urban rail transit accidents, comprising:

[0007] Obtain historical accident data and emergency response plan text data of urban rail transit, and construct accident case knowledge graph and emergency response plan knowledge graph respectively;

[0008] Semantic analysis and standardization are performed on historical accident handling measures in the accident case knowledge graph to form standard measure nodes. Semantic mapping relationships are established between standard measure nodes and emergency plan basis nodes in the emergency plan knowledge graph to form a link between accident handling measures and emergency plan basis.

[0009] Obtain the description text of the urban rail transit accident to be processed, and perform semantic feature extraction and standardization on the accident description text to obtain a new accident feature vector corresponding to the feature space of historical accident cases;

[0010] Based on the new accident feature vector, hierarchical case retrieval is performed in the accident case knowledge graph. Candidate cases are screened through macro-level accident category features, and the similarity between candidate cases and target accidents is calculated based on fine-grained accident features to obtain multiple similar historical accident cases.

[0011] Based on similar historical accident cases, corresponding accident handling measures, standard measures, and emergency response plan basis are obtained from the accident case knowledge graph and the emergency plan knowledge graph;

[0012] The target accident information, similar historical accident case information, and corresponding measures are input into the big language model, which then performs semantic recombination of the information to generate an emergency decision-making plan for urban rail transit accidents.

[0013] Furthermore, an accident case knowledge graph is constructed using the following method:

[0014] Data cleaning and field extraction were performed on historical urban rail transit accident cases to obtain accident name, location, type, cause, result, response measures, and loss information.

[0015] A large language model is used to perform multi-round semantic parsing on accident report texts to extract basic accident information, accident classification information, accident causal information, and response measure information.

[0016] The extracted results are normalized according to the preset data format, and accident components or fault types with different expressions but the same semantics are mapped to a unified entity, resulting in an accident case knowledge graph including accident case nodes, accident attribute nodes, and handling measure nodes.

[0017] Furthermore, a link is established between accident response measures and the basis of the contingency plan, specifically including:

[0018] The source measures text corresponding to historical accidents are extracted from the accident case knowledge graph. The semantic coding model is used to convert each source measure text into a semantic vector. Based on the semantic similarity, the source measure text is clustered and a standard measure node with a unified semantic description is generated according to the clustering results.

[0019] Based on the implementing entity, handling stage, action content, and plan chapter information corresponding to the standard measures, the standard measure nodes are mapped to the plan clause nodes, work nodes, or responsibility nodes in the emergency plan knowledge graph, forming a cross-knowledge graph mapping chain that includes source measure nodes, standard measure nodes, plan basis nodes, and basis level nodes.

[0020] Furthermore, cluster analysis was performed on the source measures texts, specifically including:

[0021] The BERT model is used to encode the text of each source measure to obtain the corresponding high-dimensional semantic vector;

[0022] Agglomerative hierarchical clustering is performed on the source measure text using the cosine similarity between semantic vectors as the clustering distance.

[0023] Clustering thresholds are determined based on the silhouette coefficient, Davies-Bouldin index, Calinski-Harabasz index, and standard measure mapping accuracy to obtain a set of standard measures that meet the semantic consistency requirements.

[0024] Furthermore, semantic feature extraction and standardization processing are performed on the accident description text, specifically including:

[0025] Input the text describing the accident to be processed into the large language model, constrain the output structured fields of the large language model through the preset prompt template, and extract four types of retrieval fields: the category of the accident location, the specific location of the accident, the category of the accident cause, and the specific cause of the accident.

[0026] Normalization is performed on the same accident entity corresponding to different natural language expressions to generate a new accident retrieval vector with the same field structure as the feature matrix of historical accident cases.

[0027] Furthermore, the hierarchical case retrieval specifically includes:

[0028] The first layer of retrieval involves a hard screening of historical accident cases in the accident case knowledge graph based on the type of location and cause of the target accident, to obtain a set of first candidate cases.

[0029] The second layer of retrieval calculates the comprehensive similarity based on the matching relationship between the specific location and cause of the accident in the historical accident cases in the first candidate case set and the corresponding fields of the target accident.

[0030] The third-level retrieval, when the number of cases in the first candidate case set is less than the preset return number, further expands the retrieval to at least one historical accident case that matches a macro category.

[0031] Output the Top-K similar historical accident cases according to the retrieval hierarchy priority and comprehensive similarity ranking rules.

[0032] Furthermore, the overall similarity is calculated using the following method:

[0033] The weights of different fine-grained accident characteristics are determined using the analytic hierarchy process (AHP). The specific location of the accident corresponds to the first weight, the specific cause of the accident corresponds to the second weight, and the second weight is greater than the first weight.

[0034] Calculate the Jaccard similarity between the target accident and historical accident cases in terms of both the specific location of the accident and the specific cause of the accident.

[0035] The similarity scores of each dimension are weighted and summed based on the first and second weights to obtain a comprehensive matching score for historical accident cases.

[0036] Furthermore, the accident case knowledge graph and emergency response plan knowledge graph are stored in a graph database;

[0037] The graph database is used to store accident case nodes, accident attribute nodes, measure nodes, standard measure nodes, contingency plan basis nodes, and basis level nodes, and establishes the association path between nodes through occurrence relationship, action relationship, mapping relationship, basis relationship, and attribution relationship;

[0038] By querying the graph database, we can obtain historical cases and complete action chains corresponding to the target accident.

[0039] Furthermore, an emergency decision-making plan for urban rail transit accidents is generated, specifically including:

[0040] The target accident description text, new accident retrieval vector, Top-K similar historical accident cases, and corresponding measures reference chains are encapsulated into structured input data;

[0041] Construct integrated prompts that include character settings, input data, task constraints, and output format;

[0042] Using a large language model, based on fusion prompts, the handling measures in multiple similar historical accident cases are merged, analyzed, and correlated.

[0043] Output a structured emergency decision-making text organized according to case analysis, similar cases, recommended measures, and standard criteria.

[0044] Furthermore, when generating emergency decision-making text, only suggested measures are allowed to be generated based on the input historical accident case information, standard measure information, and emergency plan basis nodes; each suggested measure is associated with the corresponding historical accident source and emergency plan basis; and accident feature information, similar case information, knowledge graph association paths, and emergency decision-making suggestions are output through the presentation layer.

[0045] The beneficial effects of this invention are as follows: This invention discloses an intelligent matching and decision-making method for urban rail transit accidents based on knowledge graphs. By constructing accident case knowledge graphs and emergency plan knowledge graphs, it structurally associates historical accident information, response measures, and plan basis. It also utilizes a large language model to extract features and standardize the semantics of new accident descriptions. Combined with case reasoning and hierarchical similarity matching, it obtains similar accident cases and further generates emergency decision-making schemes based on the measure mapping chain. This invention can improve the ability to identify synonymous heterogeneous accidents, increase the accuracy of case matching, achieve traceable association between response measures and plan basis, and enhance the reliability and interpretability of emergency decision-making results. Attached Figure Description

[0046] The present invention will be further described below with reference to the accompanying drawings and embodiments:

[0047] Figure 1 This is a schematic diagram of the overall technical roadmap;

[0048] Figure 2 A schematic diagram illustrating the construction of an emergency response plan knowledge graph and an accident case knowledge graph;

[0049] Figure 3 A semantic mapping path diagram for source measures, standard measures, and emergency plans;

[0050] Figure 4 A flowchart for intelligent retrieval and decision generation based on CBR and large models;

[0051] Figure 5 A schematic diagram of the three-tier architecture and four functional modules of the matching and decision-making system;

[0052] Figure 6 Generate case illustrations for decision-making recommendations. Detailed Implementation

[0053] The present invention will be further described below with reference to the accompanying drawings, as shown in the figures:

[0054] This embodiment discloses an intelligent matching and decision-making method for urban rail transit accidents based on knowledge graphs, including the following steps:

[0055] Obtain historical accident data and emergency response plan text data of urban rail transit, and construct accident case knowledge graph and emergency response plan knowledge graph respectively;

[0056] Semantic analysis and standardization are performed on historical accident handling measures in the accident case knowledge graph to form standard measure nodes. Semantic mapping relationships are established between standard measure nodes and emergency plan basis nodes in the emergency plan knowledge graph to form a link between accident handling measures and emergency plan basis.

[0057] Obtain the description text of the urban rail transit accident to be processed, and perform semantic feature extraction and standardization on the accident description text to obtain a new accident feature vector corresponding to the feature space of historical accident cases;

[0058] Based on the new accident feature vector, hierarchical case retrieval is performed in the accident case knowledge graph. Candidate cases are screened through macro-level accident category features, and the similarity between candidate cases and target accidents is calculated based on fine-grained accident features to obtain multiple similar historical accident cases.

[0059] Based on similar historical accident cases, corresponding accident handling measures, standard measures, and emergency response plan basis are obtained from the accident case knowledge graph and the emergency plan knowledge graph;

[0060] The target accident information, similar historical accident case information, and corresponding measures are input into the big language model, which then performs semantic recombination of the information to generate an emergency decision-making plan for urban rail transit accidents.

[0061] In this embodiment, the data sources for the emergency response plan knowledge graph include national-level emergency response plans for urban rail transit operations and emergency response plan texts from some cities. The plan texts undergo format standardization, chapter numbering standardization, text extraction, title hierarchy restoration, and sentence segmentation to form a corpus suitable for entity recognition and relation extraction.

[0062] To address the relatively stable structure and clearly defined responsibilities and tasks in emergency response plans, a BERT-BiLSTM-CRF model was adopted for entity recognition. Entity categories include four types: departments, responsibilities, tasks, and others. Departments represent responsible entities such as operating units, transportation authorities, public security organs, and fire and rescue agencies; responsibilities represent duties such as "reporting information" and "coordinating on-site"; tasks represent executable tasks such as "organizing passenger evacuation" and "carrying out equipment repair"; and others represent accident scenarios, triggering conditions, status descriptions, and supplementary semantic information.

[0063] Relation extraction employs a combination of dependency parsing and semantic role labeling to identify semantic components such as action initiator, action object, condition, purpose, cause, and method. Extracted relations are then uniformly categorized into relation types such as execution, action on, condition, responsibility, based on, due to, method, purpose, inclusion, and involvement. This ultimately forms an emergency response plan knowledge graph G_p, which includes department nodes, responsibility nodes, work nodes, condition nodes, plan clause nodes, and their hierarchical relationships.

[0064] The data source for the accident case knowledge graph is historical operational accident cases of urban rail transit. The historical accident case database includes several urban rail transit operational accident cases. After cleaning and screening, each case is extracted with fields such as accident name, city of occurrence, management unit, accident location, accident type, specific accident location, accident cause type, specific accident cause, accident result, response measures, accident type, date of occurrence, number of casualties, and economic losses.

[0065] To address the characteristics of accident report texts—loose structure, strong narrative, and frequent cross-sentence expressions of response measures—a large language model was employed for multi-round cue word extraction. The first round extracted basic factual information, including accident name, time, location, number of casualties, and economic losses. The second round extracted categorization information, including management unit, incident location, cause, result, and type. The third round extracted causal attributes and response measures, including specific cause descriptions, specific result descriptions, and a list of response measures. The model output was limited to a preset JSON schema to ensure field integrity and format stability.

[0066] The extracted results undergo entity normalization, unifying expressions such as "door cannot be opened," "door cannot be closed," and "door stuck" into "door malfunction," and unifying expressions such as "signal abnormality" and "no signal display" into "signal malfunction." Fine-grained malfunctions are then categorized into corresponding macro-level cause categories, such as "train-related," "power supply-related," and "signal-related." This ultimately forms an accident case knowledge graph G_c, which includes accident case nodes, accident attribute nodes, specific measure nodes, and their relationships.

[0067] In this embodiment, after obtaining the emergency response plan knowledge graph G_p and the accident case knowledge graph G_c, a semantic mapping relationship between the two graphs is further constructed. Here, "source measures" come from the specific measure nodes in the accident case knowledge graph G_c, that is, the actual handling actions extracted from the accident report; "standard measures" are intermediate nodes formed after semantic clustering and standardized naming of the source measures; and "emergency response plan basis" comes from the clause nodes, work nodes, or responsibility nodes in the emergency response plan knowledge graph G_p.

[0068] The specific mapping process is as follows: First, extract the specific measures text associated with each accident case from the accident case knowledge graph G_c to form a set of source measures; second, use the BERT model to encode each source measure into a high-dimensional semantic vector, and perform agglomerative hierarchical clustering using cosine similarity as the distance metric; third, manually verify and standardize the naming of each cluster to form standard measures. For example, "publishing announcements via Weibo," "broadcasting notifications to passengers," and "publishing reminders on electronic displays within the station" are grouped into "publishing relevant information through multiple channels."

[0069] Once standard measures are formulated, they are mapped to corresponding clauses or work nodes in the emergency response plan knowledge graph G_p, based on their action meaning, responsible party, handling stage, and plan chapter level. For example, the standard measure "release relevant information through multiple channels" maps to clauses in the plan such as "timely release of event and operational adjustment information." This forms a complete link: source measure node, mapping to, standard measure node, basis, emergency response plan basis node, attribution, and basis level node. This link allows actual accident handling experience to be traced back to the plan's standard clauses, and also allows plan clauses to be linked back to specific implementation methods in historical cases.

[0070] A comprehensive optimization strategy was adopted for selecting the clustering threshold. The silhouette coefficient, Davies-Bouldin index, and Calinski-Harabasz index were calculated under multiple candidate thresholds, and a comprehensive judgment was made by combining the proportion of single-sample clusters, semantic consistency of clusters, workload of manual verification, and the Precision, Recall, and F1 scores of the mapping from source measures to standard measures. In this embodiment, 0.17 was selected as the clustering threshold, which achieves a good balance between semantic granularity, mapping quality, and manual verification cost.

[0071] All of the above knowledge is stored in the Neo4j graph database. Neo4j is used to store nodes such as accident cases, accident attributes, specific measures, standard measures, contingency plans, and basis levels, and enables cross-graph path retrieval through relationships such as "occurred at", "belongs to", "taken", "mapped to", "based on", and "belongs to".

[0072] In this embodiment, before proceeding to CBR (Case-Based Reasoning) similarity retrieval, the semantic features of the new accident natural language description (the text describing the urban rail transit accident to be processed) input by the user are extracted and standardized using a large language model. The reason for this setting is that subsequent hard matching and Jaccard similarity calculation rely on stable structured fields. If the ambiguous expressions in the original text are used directly, it is easy to cause missed detections or low similarity.

[0073] After receiving a new accident description, the large language model outputs at least four key fields for case retrieval: accident location category, specific accident location, accident cause category, and specific accident cause.

[0074] The model output process simultaneously standardizes terminology; for example, "door cannot be opened / door cannot be closed / door stuck" is standardized to "door malfunction," and "signal abnormality / no signal display" is standardized to "signal malfunction." The large language model uses DeepSeek-Chat. To ensure output stability, output fields are constrained using JSON Schema, prohibiting the model from adding undefined fields or generating content without source evidence.

[0075] In the case-based reasoning retrieval framework, the structured representation of cases is the foundation for similarity calculation. Four feature dimensions are selected for case retrieval: incident location category, specific incident location, incident cause category, and specific incident cause, corresponding to the previous features. The meanings of the four retrieval features are shown in Table 1 below.

[0076] Table 1

[0077]

[0078] The n cases in the historical accident case database are used to construct a feature matrix M, with the number of feature dimensions d=4. The matrix can be represented as: M= [m ij ] (n×4) ;

[0079] In this matrix, each row corresponds to a historical accident case, and each column corresponds to a retrieval feature dimension. Column 1 represents the accident location category, column 2 represents the specific location of the accident, column 3 represents the accident cause category, and column 4 represents the specific cause of the accident. All matrix elements are stored as standardized text values ​​and are not numerically encoded.

[0080] For newly occurring incidents, four retrieval fields are extracted and standardized using prompt words from a large language model to form a new case retrieval vector V_new:

[0081] V_new = [P_cat_new, P_sub_new, C_cat_new, C_sub_new];

[0082] Wherein, P_cat_new represents the incident location category, P_sub_new represents the specific incident location, C_cat_new represents the incident cause category, and C_sub_new represents the specific incident cause. V_new has the same field structure and feature space as the historical case feature matrix M, and can be directly used for subsequent hierarchical retrieval and similarity calculation.

[0083] In this embodiment, the present invention employs the analytic hierarchy process (AHP) to determine the weights of two feature dimensions: the specific location of the incident and the specific cause of the incident. The incident location category and the incident cause category are used for hard matching in hierarchical retrieval and do not participate in weight calculation. This avoids the redundant effect of macro-level category weights and hard screening conditions, while ensuring basic consistency among candidate cases in terms of accident system category and cause category.

[0084] Based on the degree of impact on emergency decision-making, the specific location of the incident is used to depict the precise system location where the accident occurred, while the specific cause of the incident is used to depict the triggering mechanism and failure type. Relatively speaking, the specific cause of the incident has a stronger direct impact on the selection of emergency response measures, and therefore is given higher weight.

[0085] Normalizing the second-order judgment matrix yields a weight of 0.4 for the specific location of the incident and a weight of 0.6 for the specific cause of the incident.

[0086] For a second-order judgment matrix, λ max =2, CI=0, CR=0, the matrix has complete consistency, therefore the weight allocation result is reasonable and reliable.

[0087] For the two textual features, "specific location of the incident" and "specific cause of the incident," the Jaccard coefficient is used to calculate attribute similarity. Let the set of words for a new case on a certain feature be denoted as . The set of words in historical cases that share the same feature is Then the Jaccard similarity of this feature is:

[0088] ;

[0089] The overall similarity is weighted and summed according to the above weights, as shown in the formula:

[0090] .

[0091] In this invention, "target case" refers to a new accident that the user currently inputs and requires emergency decision support; "historical case" refers to accident case nodes already stored in the accident case knowledge graph G_c. Hierarchical retrieval directly operates on case nodes and their standardized attributes in the knowledge graph, rather than retrieving them separately from the knowledge graph.

[0092] The first layer is a hard screening: based on two features, the incident location category and the incident cause category, historical cases that simultaneously satisfy both criteria are screened in Neo4j to form a candidate set C1.

[0093] C1 = {V_i | P_cat_i = P_cat_new and C_cat_i = C_cat_new};

[0094] The second layer is a weighted similarity ranking: for each historical case in C1, the Jaccard similarity between the specific location and cause of the incident and the target case is calculated, and then ranked according to... The formula calculates the overall similarity.

[0095] The third layer is a hierarchical expansion: if the number of cases in C1 is less than K, then historical cases that match the target case in at least one of the categories of incident location or incident cause are retrieved to form a candidate set C2; if the sum of C1 and C2 is still less than K, then a global search is performed on the accident case knowledge graph G_c.

[0096] Where C2 = {V_i | P_cat_i = P_cat_new or C_cat_i = C_cat_new};

[0097] In the experiment, approximately 95% of accidents yielded more than 30 candidate cases through the first-level hard screening, while about 5% of accidents were mainly passenger-related, requiring the triggering of the second-level expanded search. Global search was not triggered in the actual test. Search results were sorted according to the principle of "hierarchical priority, descending similarity within a hierarchy," i.e., C1 takes precedence over C2, and C2 takes precedence over global search results; within the same hierarchy, results were sorted according to overall similarity. Arranged from highest to lowest. K is a configurable parameter; during the verification phase, K=3 can be used, and more similar cases can be returned as needed by business requirements.

[0098] After the Top-K cases are returned, Neo4j is used to query the complete action chain corresponding to each similar case, including the accident case, specific measures, standard measures, contingency plan basis, and basis level. This chain ensures that subsequent decision-making recommendations are supported by both historical cases and emergency plans.

[0099] In this embodiment, the Top-K cases returned by structured retrieval exist in the form of knowledge graph nodes and edges, including information such as accident characteristics, specific measures, standard measures, contingency plan basis, and basis level. This format ensures the accuracy of the data structure, but it is still relatively fragmented when directly presented to the user. Therefore, this invention introduces a large language model to semantically reorganize and enhance the interpretability of the retrieval results.

[0100] The input to the large language model consists of three parts: the first part is the original description text of the new accident and its standardized retrieval vector V_new; the second part is the accident features, similarity scores, and historical handling measures of the Top-K similar cases; the third part is the measure chain corresponding to each similar case, namely, specific measures, standard measures, contingency plan basis, and basis level. The above information is encapsulated into structured JSON and integrated prompt words are constructed.

[0101] The integrated prompts consist of four parts: role setting, input data, task constraints, and output format. The role is set as "Urban Rail Transit Emergency Decision Assistant" or "Metro Emergency Incident Analysis Expert"; the task constraints require the model to summarize only based on the input cases and contingency plans, and not to generate measures without supporting evidence; the output format is limited to a four-part structure: "Case Analysis - Similar Cases - Suggested Measures - Standard Basis".

[0102] The Top-K similarity cases are fused with a large language model to generate emergency decision-making solutions for urban rail transit accidents, specifically including:

[0103] Read Top-K similar cases: Based on the search results, obtain the case number, accident feature fields, similarity score, and historical handling measures for each similar case.

[0104] Query the measure mapping chain: Starting from each similar case node, query the corresponding specific measures, standard measures, emergency plan basis and basis level in Neo4j along the relationship path of taking, mapping to, based on and belonging to.

[0105] Constructing structured input: New accident descriptions, standardized retrieval vectors, Top-K similar case information, and measure mapping chains are uniformly encapsulated into JSON format and used as input data for the large language model.

[0106] Constructing integrated prompts: Prompts include role settings, input data, task constraints, and output format. Among them, task constraints require the model to generate suggestions only based on input cases and contingency plans, and not to fabricate measures without a source.

[0107] Semantic restructuring is performed: the large language model compares and analyzes similar cases, merges repeated or similar specific measures, and associates them with standard measures and contingency plans.

[0108] Generate structured decision-making solutions: The model outputs natural language text in the format of "case analysis - similar cases - recommended measures - standard basis" (as shown in Table 2). This transforms the search results from discrete triples into readable, interpretable, and traceable emergency decision-making suggestions.

[0109] The large language model performs three types of transformations on the input information: First, it merges similar case information according to case dimensions, so that each case forms an independent structured expression unit; second, it summarizes and removes duplicates of the disposal measures in each case, forming a unified set of measures; third, it associates and organizes the measures with corresponding standard measures, contingency plan clauses and basis levels, forming a traceable emergency decision-making text.

[0110] Table 2

[0111]

[0112] This invention also relates to a matching and decision-making system, which can be understood as a system for implementing the intelligent matching and decision-making method for urban rail transit accidents described in the above embodiments. This matching and decision-making system has a three-layer architecture: a data layer, a service layer, and a presentation layer. The data layer stores accident cases, emergency plans, and semantic mapping relationships based on the Neo4j graph database; the service layer encapsulates data management, graph query, intelligent retrieval, and large model invocation logic; and the presentation layer provides an interface for natural language accident description input, similar case display, interactive knowledge graph browsing, and decision suggestion output.

[0113] The matching and decision-making system is divided into four functional modules based on the core business of CBR intelligent question answering: a case library management module, a knowledge graph visualization module, an intelligent retrieval module, and a decision suggestion generation module. The intelligent retrieval module uses two similarity features: the specific location of the incident and the specific cause of the incident, with weight parameters set to 0.4 and 0.6 respectively; the categories of incident location and cause of the incident serve as hard filtering conditions for hierarchical retrieval. The descriptions of the functional modules are shown in Table 3.

[0114] Table 3

[0115]

[0116] The matching and decision-making system uses Neo4j to store several urban rail transit operation accident cases and their emergency response chains. After a user inputs a new accident description, the system first uses a large language model to extract four standardized search fields, forming V_new; if the accident location is also extracted, it is saved as an auxiliary attribute. Subsequently, a first-level hard screening is performed based on the accident location category and the accident cause category; within the candidate set, the weighted Jaccard similarity of the specific accident location and the specific accident cause is calculated; finally, the Top-K similar cases are returned.

[0117] Taking three types of verification cases—signal equipment failure, power supply equipment failure, and personnel entering the track—as examples, the system can retrieve historical cases that are highly consistent with the target accident in terms of the location, specific cause, and cause category. It further generates suggested measures including train speed limits, extended departure intervals, passenger information guidance, operational adjustments, and necessary emergency responses. The output of the large language model is organized according to "case analysis - similar cases - suggested measures - standard basis," transforming the search results from phrase-level measures into readable, explainable, and traceable emergency decision-making text.

[0118] The recommended measures are derived from common handling measures extracted from a Top-K similar case set, and their coverage relationship is compared with the actual handling measures after semantic reorganization of the large model. The core measures in the three verification cases of signal failure, power supply equipment failure, and personnel entering the line can all be covered by the recommended results, indicating that the system can support the task of matching emergency response plans for rail transit.

[0119] This invention uses a large language model to semantically standardize accident descriptions, identifying the same accident cause pointed to by different expressions such as "the car door cannot be opened" and "the car door cannot be closed," reducing missed detections in keyword searches. Hard filtering by accident location and cause categories ensures consistency in the macro-type of candidate cases; weighted similarity ranking by specific accident location and cause improves fine-grained matching. Through a three-layer mapping of "source measures, standard measures, and contingency plan basis," each recommended measure can be traced back to historical accident cases and emergency plan clauses. The knowledge graph provides structured facts and evidence chains, and the large language model reassembles discrete triples into coherent, point-based, and readable emergency decision text. After adding new accident cases or updating contingency plan text, nodes, relationships, and mapping chains can be incrementally added in Neo4j without rebuilding the entire knowledge graph.

[0120] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A knowledge graph-based intelligent matching and decision-making method for urban rail transit accidents, characterized in that: include: Obtain historical accident data and emergency response plan text data of urban rail transit, and construct accident case knowledge graph and emergency response plan knowledge graph respectively; Semantic analysis and standardization are performed on historical accident handling measures in the accident case knowledge graph to form standard measure nodes. Semantic mapping relationships are established between standard measure nodes and emergency plan basis nodes in the emergency plan knowledge graph, forming a link between accident handling measures and emergency plan basis. Obtain the description text of the urban rail transit accident to be processed, and perform semantic feature extraction and standardization on the accident description text to obtain a new accident feature vector corresponding to the feature space of historical accident cases; Based on the new accident feature vector, hierarchical case retrieval is performed in the accident case knowledge graph. Candidate cases are screened through macro-level accident category features, and the similarity between candidate cases and target accidents is calculated based on fine-grained accident features to obtain multiple similar historical accident cases. Based on similar historical accident cases, corresponding accident handling measures, standard measures, and emergency response plan basis are obtained from the accident case knowledge graph and the emergency plan knowledge graph; The target accident information, similar historical accident case information, and corresponding measures are input into the big language model, which then performs semantic recombination of the information to generate an emergency decision-making plan for urban rail transit accidents.

2. The knowledge graph-based intelligent matching and decision-making method for urban rail transit accidents according to claim 1, characterized in that: Construct an accident case knowledge graph using the following method: Data cleaning and field extraction were performed on historical urban rail transit accident cases to obtain accident name, location, type, cause, result, response measures, and loss information. A large language model is used to perform multi-round semantic parsing on accident report texts to extract basic accident information, accident classification information, accident causal information, and response measure information. The extracted results are normalized according to the preset data format, and accident components or fault types with different expressions but the same semantics are mapped to a unified entity, resulting in an accident case knowledge graph including accident case nodes, accident attribute nodes, and handling measure nodes.

3. The knowledge graph-based intelligent matching and decision-making method for urban rail transit accidents according to claim 1, characterized in that: Establishing a link between accident response measures and the basis of contingency plans, specifically including: The source measures text corresponding to historical accidents are extracted from the accident case knowledge graph. The semantic coding model is used to convert each source measure text into a semantic vector. Based on the semantic similarity, the source measure text is clustered and a standard measure node with a unified semantic description is generated according to the clustering results. Based on the implementing entity, handling stage, action content, and plan chapter information corresponding to the standard measures, the standard measure nodes are mapped to the plan clause nodes, work nodes, or responsibility nodes in the emergency plan knowledge graph, forming a cross-knowledge graph mapping chain that includes source measure nodes, standard measure nodes, plan basis nodes, and basis level nodes.

4. The knowledge graph-based intelligent matching and decision-making method for urban rail transit accidents according to claim 3, characterized in that: Cluster analysis was performed on the source measures texts, specifically including: The BERT model is used to encode the text of each source measure to obtain the corresponding high-dimensional semantic vector; Agglomerative hierarchical clustering is performed on the source measure text using the cosine similarity between semantic vectors as the clustering distance. Clustering thresholds are determined based on the silhouette coefficient, Davies-Bouldin index, Calinski-Harabasz index, and the accuracy of standard measure mapping to obtain a set of standard measures that meet the semantic consistency requirements.

5. The knowledge graph-based intelligent matching and decision-making method for urban rail transit accidents according to claim 1, characterized in that: Semantic feature extraction and standardization of the accident description text are performed, specifically including: Input the text describing the accident to be processed into the large language model, constrain the output structured fields of the large language model through the preset prompt template, and extract four types of retrieval fields: the category of the accident location, the specific location of the accident, the category of the accident cause, and the specific cause of the accident. Normalization is performed on the same accident entity corresponding to different natural language expressions to generate a new accident retrieval vector with the same field structure as the feature matrix of historical accident cases.

6. The knowledge graph-based intelligent matching and decision-making method for urban rail transit accidents according to claim 1, characterized in that: The hierarchical case retrieval specifically includes: The first layer of retrieval involves a hard screening of historical accident cases in the accident case knowledge graph based on the type of location and cause of the target accident, to obtain a set of first candidate cases. The second layer of retrieval calculates the comprehensive similarity based on the matching relationship between the specific location and cause of the accident in the historical accident cases in the first candidate case set and the corresponding fields of the target accident. The third-level retrieval, when the number of cases in the first candidate case set is less than the preset return number, further expands the retrieval to at least one historical accident case that matches a macro category. Output the Top-K similar historical accident cases according to the retrieval hierarchy priority and comprehensive similarity ranking rules.

7. The knowledge graph-based intelligent matching and decision-making method for urban rail transit accidents according to claim 6, characterized in that: The overall similarity is calculated using the following method: The weights of different fine-grained accident characteristics are determined using the analytic hierarchy process (AHP). The specific location of the accident corresponds to the first weight, the specific cause of the accident corresponds to the second weight, and the second weight is greater than the first weight. Calculate the Jaccard similarity between the target accident and historical accident cases in terms of both the specific location of the accident and the specific cause of the accident. The similarity scores of each dimension are weighted and summed based on the first and second weights to obtain a comprehensive matching score for historical accident cases.

8. The knowledge graph-based intelligent matching and decision-making method for urban rail transit accidents according to claim 1, characterized in that: The accident case knowledge graph and emergency response plan knowledge graph are stored in a graph database; The graph database is used to store accident case nodes, accident attribute nodes, measure nodes, standard measure nodes, contingency plan basis nodes, and basis level nodes, and establishes the association path between nodes through occurrence relationship, action relationship, mapping relationship, basis relationship, and attribution relationship; By querying the graph database, we can obtain historical cases and complete action chains corresponding to the target accident.

9. The knowledge graph-based intelligent matching and decision-making method for urban rail transit accidents according to claim 1, characterized in that: Generate emergency decision-making plans for urban rail transit accidents, specifically including: The target accident description text, new accident retrieval vector, Top-K similar historical accident cases, and corresponding measures reference chains are encapsulated into structured input data; Construct integrated prompts that include character settings, input data, task constraints, and output format; Using a large language model, based on fusion prompts, the handling measures in multiple similar historical accident cases are merged, analyzed, and correlated. Output a structured emergency decision-making text organized according to case analysis, similar cases, recommended measures, and standard criteria.

10. The knowledge graph-based intelligent matching and decision-making method for urban rail transit accidents according to claim 9, characterized in that: When generating emergency decision-making text, only suggested measures are allowed to be generated based on the input historical accident case information, standard measure information, and emergency plan basis nodes; each suggested measure is associated with the corresponding historical accident source and emergency plan basis; and accident feature information, similar case information, knowledge graph association paths, and emergency decision-making suggestions are output through the presentation layer.

Images