Railway operation accident-oriented unsupervised construction method and system for affair atlas

Abstract

The invention relates to an unsupervised construction method and system for a railway operation accident-oriented affair graph, and the method comprises the following steps: obtaining railway operation accident report corpus data, and carrying out the preprocessing to obtain a railway accident report data set; based on the railway accident report data set, classifying all accident report information into a plurality of subsystems according to cause codes of accidents, and extracting keywords of risk events corresponding to cause types as risk event entities; through layer-by-layer decomposition of a subsystem, identifying a causal relationship and a sequential relationship between risk events, and establishing an event relationship to determine a matrix; and determining a matrix based on the risk event entity and the event relationship, calculating the similarity between the to-be-linked railway operation accident report and the risk event entity, selecting and complementing an optimal accident risk chain, and constructing a affair graph model. Compared with the prior art, the method has the advantages of realizing unsupervised construction of the railway operation accident affair atlas and the like.

Description

Technical Field

[0001] The present invention relates to the field of railway safety technology, and in particular to a method and system for unsupervised construction of a causal graph for railway operation accidents. Background Art

[0002] Railway operation involves complex giant systems. External environmental disturbances may spread and amplify railway operation safety risks through equipment, facilities, functional subsystems, etc., and eventually lead to serious railway safety accidents. Traditional accident analysis methods, such as causal chain models such as Fault Tree Analysis (FTA) and Bayesian Network (BN), and systematic analysis framework models such as Human Factor Analysis and Classification System (NFACS) and Complex Network (CN), provide a rich theoretical basis for accident cause analysis by summarizing historical accidents and experiences into models. However, these methods can usually only provide single-dimensional accident cause factors and relationships, and the characteristics and information of individual accident records will be lost in the process of cause induction, which is not conducive to integrating historical accident statistical information to achieve quantitative risk assessment.

[0003] Compared with traditional methods, accident modeling and analysis methods based on knowledge graphs can incorporate heterogeneous information, thus restoring detailed information of a single accident to a greater extent. Although knowledge graphs have been widely used in e-commerce, medical care, finance, government management and other fields, their role in the field of railway safety is still limited. The relevant technology has two significant application bottlenecks in this scenario: First, the commonly used knowledge graph construction method is difficult to capture the complex event evolution logic and dynamic action relationship from external disturbances to the final consequences; second, when extracting knowledge from massive corpora such as railway operation accident reports, a lot of manpower is required for labeling, and considering the professionalism of report terminology, labeling is very difficult. Therefore, in order to apply knowledge graphs to the field of railway safety, related construction and analysis technologies still need to be improved and developed. Summary of the invention

[0004] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and to provide a method and system for unsupervised construction of a causal graph for railway operation accidents, so as to solve or partially solve the problem that it is difficult to capture the complex event evolution logic and dynamic action relationship from external disturbance to final consequence, and that a large amount of manpower is required for labeling when extracting knowledge from massive corpora such as railway operation accident reports.

[0005] The purpose of the present invention can be achieved by the following technical solutions:

[0006] One aspect of the present invention provides an unsupervised construction method of a cause-and-effect graph for railway operation accidents, comprising the following steps:

[0007] Obtain railway operation accident report corpus data, and obtain railway accident report dataset through preprocessing;

[0008] Based on the railway accident report dataset, all accident report information is classified into multiple subsystems according to the cause coding of the accident, and keywords of risk events corresponding to each cause type are extracted as risk event entities;

[0009] Through subsystem decomposition, we can identify the causal and sequential relationships between risk events and establish an event relationship matrix.

[0010] Based on the risk event entity and the event relationship, a matrix is ​​established, the similarity between the railway operation accident report to be linked and the risk event entity is calculated, the optimal accident risk chain is selected and completed, and a causal graph model is constructed.

[0011] As a preferred technical solution, based on the railway operation accident report corpus data, the process of obtaining the railway accident report dataset by preprocessing includes the following steps:

[0012] Obtain railway operation accident report corpus data;

[0013] Extract structured information from railway operation accident report corpus data;

[0014] Extract unstructured information from railway operation accident report corpus data;

[0015] For the unstructured information, numbers, punctuation marks and non-lexical symbols are removed, a stop word list is set, and the unstructured information is converted into n-gram form after word segmentation.

[0016] As a preferred technical solution, the structured information includes one or more of the accident ID, accident time, accident cause code, train type, weather, accident type, number of casualties, and economic losses caused by the accident, and the unstructured information includes a text describing the accident.

[0017] As a preferred technical solution, the multiple subsystems include accident consequences, and one or more of human factors, track and roadbed, mechanical and electrical, signal and communication, and environmental factors.

[0018] As a preferred technical solution, the process of extracting the representation text of the risk event corresponding to each cause type as the risk event entity includes the following steps:

[0019] Based on the unstructured information of railway accident report data under the same subsystem, the pre-trained language model is used to obtain the document embedding vector of the entire accident description;

[0020] Use the language model to obtain the word embedding vector corresponding to each candidate keyword;

[0021] Calculate the cosine similarity between each word embedding vector and the document embedding vector, and select keywords whose similarity meets the requirements;

[0022] The keywords after similarity screening are manually screened and the screened keywords are used as risk event entities.

[0023] As a preferred technical solution, the process of establishing an event relationship adjacency matrix by subsystem layer-by-layer decomposition includes the following steps:

[0024] Obtain a relationship establishment matrix that represents whether there is a mutual influence relationship between subsystems;

[0025] For subsystems with influencing relationships, a risk event relationship establishment matrix is ​​established to characterize whether there is a causal or causal relationship between risk event entities.

[0026] As a preferred technical solution, based on the risk event entity and the event relationship establishment matrix, the similarity between the railway operation accident report to be linked and the risk event entity is calculated, the optimal accident risk chain is selected and completed, and the process of constructing the event graph model includes the following steps:

[0027] Through the depth-first path search with memory mechanism, the full amount of accident risk chain is searched;

[0028] For each railway operation accident record to be linked, initialize a candidate risk event list in dictionary form;

[0029] Based on the structured information of the railway operation accident record to be linked, the risk event entity and the event relationship establishment matrix, the risk event entity is added to the candidate risk event list based on a preset rule to achieve rule-based risk event linking;

[0030] Preprocess the unstructured information in each railway operation accident record to be linked to obtain the unstructured information in the form of n-grams, convert each n-gram and all risk event entities in the event graph into the n-gram word vector to be linked and the risk event entity word vector, calculate the cosine similarity between the word vectors, and assign values ​​to the candidate risk events that meet the similarity requirements;

[0031] For each railway operation accident record to be linked, all accident risk chains in the corresponding candidate risk event list are traversed, and the optimal accident risk chain is obtained by calculating the chain confidence;

[0032] Identify and complete the missing risk event entities in the optimal accident risk chain;

[0033] The structured information of the railway operation accident records to be linked is added to the tail node of the optimal accident risk chain to realize the construction of the event graph model.

[0034] As a preferred technical solution, the process of completing the missing risk event entity includes the following steps:

[0035] If the first risk event entity in the accident risk chain is not identified, then light up the risk event entity;

[0036] If there are unidentified nodes between the first risk event entity on the accident risk chain and the last identified risk event on the accident risk chain, then these unidentified risk event entities are lit up;

[0037] If there are still unidentified risk event entities after the tail node, they will remain unlit.

[0038] As a preferred technical solution, it also includes:

[0039] Based on the constructed event-reasoning graph model, the risk transmission probability between risk events is calculated to evaluate the risk transmission capability;

[0040] Based on the constructed event-diagram model, the cumulative risk transfer probability is calculated to evaluate the ability of exogenous hazards to transfer risks to node events in the railway operation system;

[0041] Based on the constructed event graph model, the weighted out-degree centrality is calculated to evaluate the risk of node event diffusion in the complex railway operation system.

[0042] Based on the constructed causal graph model, the generalized risk of the risk chain is calculated and the overall risk of the accident risk chain is evaluated.

[0043] Another aspect of the present invention provides an unsupervised construction system of a cause-and-effect graph for railway operation accidents, comprising:

[0044] The data processing module is used to obtain the railway operation accident report corpus data and obtain the railway accident report data set through preprocessing;

[0045] A graph construction module is used to divide all accident report information into multiple subsystems based on the railway accident report data set according to the cause coding of the accident, extract keywords of risk events corresponding to each cause type as risk event entities, decompose them layer by layer through the subsystem, identify the causal relationship and sequential relationship between each risk event, and establish an event relationship establishment matrix;

[0046] The analysis and visualization module is used to establish a matrix based on the risk event entity and the event relationship, calculate the similarity between the railway operation accident report to be linked and the risk event entity, select the optimal accident risk chain and complete it, build a causal graph model and perform analysis.

[0047] Compared with the prior art, the present invention has at least one of the following beneficial effects:

[0048] (1) Realize the unsupervised construction of the cause-effect graph of railway operation accidents: First, based on the railway accident report dataset, the present invention divides all accident report information into multiple subsystems according to the cause coding of the accident, extracts the keywords of risk events corresponding to each cause type as risk event entities; then, through the subsystem decomposition layer by layer, identifies the causal relationship and sequential relationship between each risk event, and establishes an event relationship establishment matrix; finally, based on the risk event entity and the event relationship establishment matrix, calculates the similarity between the railway operation accident report to be linked and the risk event entity, selects the optimal accident risk chain and completes it, and constructs a cause-effect graph model.

[0049] (2) High accuracy in keyword screening: The present invention calculates the similarity between document embedding vectors and word embedding vectors and combines manual screening to achieve accurate screening of risk event entities.

[0050] (3) Capable of capturing the complex event evolution logic and dynamic action relationships in the process from external disturbance to final consequences: The present invention constructs a causal graph model by pre-searching the entire accident risk chain, and then performing risk event linking, accident risk chain screening and completion based on rules and chain similarity. This can fully capture the complex event evolution logic and dynamic action relationships in the process from external disturbance to final consequences. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is a flow chart of the unsupervised construction method of the causal graph for railway operation accidents in the embodiment;

[0052] Figure 2 This is an example of structured and unstructured data of a railway operation accident report in the embodiment;

[0053] Figure 3 Schematic diagram of the KeyBERT keyword extraction algorithm flow in the embodiment;

[0054] Figure 4 A schematic diagram of the process of establishing a matrix for the relationship between subsystems and risk events in the embodiment;

[0055] Figure 5 This is an example of a railway operation accident risk event and relationship map in the embodiment;

[0056] Figure 6 Schematic diagrams of three situations in the search process of the path search algorithm with a memory mechanism in the embodiment;

[0057] Figure 7 Schematic diagram of the optimal accident risk chain selection process in the embodiment;

[0058] Figure 8 This is an example of a railway operation event diagram database in the embodiment;

[0059] Fig. 9 Schematic diagram of risk analysis results based on cumulative risk transfer probability in an embodiment;

[0060] Fig.10 Schematic diagram of the risk analysis result based on weighted out-degree centrality in the embodiment;

[0061] Fig.11 Schematic diagram of risk analysis results based on generalized risk of accident chain in the embodiment;

[0062] Fig.12 It is a schematic diagram of the architecture of the cause-and-effect graph construction and analysis system for railway operation accidents in the embodiment. DETAILED DESCRIPTION

[0063] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0064] Example 1

[0065] In view of the problems existing in the above-mentioned prior art, this embodiment provides an unsupervised construction method of a cause-effect graph for railway operation accidents, as shown in the attached Figure 1 As shown in the figure, the unsupervised construction method of the event graph includes the following steps:

[0066] Step S01: Obtain a railway operation accident report corpus, preprocess the corpus, and obtain a railway accident report dataset. Specifically, the accident report corpus and the preprocessing method may include steps S01-1 to S01-4:

[0067] Step S01-1: Download and obtain the railway accident reports from 1996 to 2022 published on the official website of a country's railway operating department. Each accident corresponds to an accident report record. In the case of duplicate reports, they are merged or deleted by accident ID.

[0068] Step S01-2: Unify the structured information format in each accident report. In the embodiment, the structured information of the accident report includes the accident ID, accident time, accident cause code, train type, weather, accident type, accident casualties, and accident economic losses, which are organized into JSON format.

[0069] Step S01-3: Unify the format of the unstructured information in each accident report. In the embodiment, the unstructured information is a description of the accident, which is organized into a TXT format.

[0070] Step S01-4: Preprocessing the unstructured data. The preprocessing step in the embodiment includes removing numbers, punctuation marks and non-lexical symbols from all TXT text content, setting a stop word list, segmenting words and processing the unstructured data into n-grams (n=1, 2, 3).

[0071] The data examples after the above steps are shown in the attached Figure 2 shown.

[0072] Step S02: Integrate the structured and unstructured information in the railway accident report data set, and use rule-based event recognition and keyword extraction technology to extract risk events as entities of the graph. Specifically, the step of extracting risk event entities may include steps S02-1 to S02-3:

[0073] Step S02-1: All accident reports are grouped according to the subsystems to which the causes belong according to the accident cause codes in the structured information. In this embodiment, the subsystems are divided into five subsystems according to the accident cause types: human factors (H), track and roadbed (T), mechanical and electrical (M), signal and communication (S), and environmental factors (E). Accident reports with cause codes belonging to the same category are grouped together. At the same time, an accident consequence (C) subsystem is added to describe and store the final manifestation of the accident.

[0074] Step S02-2: For each group of accident reports with the same cause type, the KeyBERT algorithm is used to extract keywords or terms from the accident description n-grams of similar reports. In this embodiment, the specific steps of the KeyBERT algorithm include: inputting a complete accident description, and obtaining the document embedding vector (document embeddings) of the entire accident description through a pre-trained language model based on the Transformer architecture; then all the n-grams obtained after processing are respectively output through the same model to their corresponding word embedding vectors (word embeddings) of the same dimension; all the word embedding vectors are respectively calculated with the document embedding vector cosine similarity, and finally the n-grams ranked in the top 20% of similarity are taken as the keywords / terms of this type of text. The pre-trained model used in the embodiment is the all-MiniLM-L6-v2 model, and the KeyBERT algorithm process described in the embodiment is as shown in the attached figure. Figure 3 shown.

[0075] Step S02-3: Further, the keywords / terms extracted in step S02-3 are manually screened and processed, including merging terms with the same meaning, standardizing the standard description of terms, deleting terms that do not conform to the risk events of the railway operation subsystem, and listing all the determined risk events and their subsystems as railway operation incident logic map entities. An example of the railway operation incident logic map entity list obtained in this embodiment is shown in the following table.

[0076] Table 1 Example of entity list of railway operation incident logic map

[0077]

[0078] Step S03: Determine the causal relationship and sequential relationship between all risk events from the perspective of subsystem decomposition layer by layer. Specifically, the specific method and steps for determining the functional relationship between subsystems and the causal and sequential relationship between risk events include steps S03-1 to S03-2:

[0079] Step S03-1: First, based on expert experience or historical events, determine whether there is a mutual influence relationship between the above six subsystems: human factors (H), track and roadbed (T), mechanical and electrical (M), signal and communication (S), environmental factors (E), and accident consequences (C), and establish the subsystem relationship establishment matrix RCM S , the expression is as follows:

[0080]

[0081] In this embodiment, the relationship establishment matrix RCM obtained by the expert S As attached Figure 4 shown.

[0082] Step S03-2: For subsystems with an impact relationship, that is, The corresponding two subsystems need to further judge whether there is a causal relationship or a sequential relationship between all entities (risk events) in the two subsystems i and j based on expert experience or historical events, and establish the risk event relationship establishment matrix RCM e , the expression is as follows:

[0083]

[0084] Attached Figure 4 The establishment of a RCM e Example of the process.

[0085] In this embodiment, examples of storage and expression of risk event entities and their adjacency relationships obtained through step S02 and step S03 are shown in the following table.

[0086] Table 2 Examples of storage and expression of risk event entities and their adjacency relationships

[0087]

[0088] Step S04: For all historical accident records in the database or new railway accident records in the future, the accident chain linking method based on similarity is used to link the accident records one by one to the event graph, and finally build a graph database that stores all accident records and the event evolution relationship within the accident. In this embodiment, all weather-related railway operation accidents in the historical data are used as accident records to be linked, and a railway operation event graph database related to meteorological risks is constructed. Specifically, the accident risk chain linking process includes steps S04-1 to S04-8:

[0089] Step S04-1: Based on the risk event entities and event relationship adjacency matrix obtained in step S02 and step S03, a railway operation event logic graph model is constructed. The graph expression of risk event entities and causal / sequential relationships in this embodiment is shown in the attached Figure 5 .

[0090] Step S04-2: In this embodiment, all accident risk chains are searched out through a depth-first path search algorithm with a memory mechanism. The algorithm proposed in the embodiment adds a memory mechanism on the basis of the basic depth-first search (DFS) algorithm, that is, for all searched nodes, a memory variable in the form of a dictionary data structure is used to store the subsequent sub-chains of the node. When the node is searched again, there is no need to search again in a deeper direction. The sub-chain is directly extracted from the memory variable and spliced ​​into the current pre-order chain. The three situations involved in the search process of the above algorithm are as follows. Figure 6Each risk chain is composed of several risk events and event relationships such as causal relationships and sequential relationships. The final accident risk chain (partial examples) is shown in the following table. The set of all accident risk chains is named C.

[0091] Table 3 Accident risk chain set (partial examples)

[0092]

[0093] Step S04-3: Construct a candidate risk event list D in the form of a dictionary for each railway operation accident record to be linked. 1 :α 1 ,…,e i :α i ,…}, where e i represents the i-th candidate risk event, α i is the corresponding confidence, and its value range is [0, 1]. The initial state of each D is an empty dictionary.

[0094] Step S04-4: Perform rule-based risk event linking based on the structured information of the railway operation accident record to be linked. In this embodiment, the corresponding rules between the structured information used and the risk events (entities) involved in the accident process to which it points are shown in the following table. Through this rule, the corresponding risk event is added to D, and its corresponding confidence is assigned a value of 1.

[0095] Table 4 Structured information and risk event correspondence rules used in the embodiment

[0096]

[0097] Step S04-5: Link risk events according to the accident process description in the railway operation accident record d to be linked. Preprocess the accident process description text according to the method described in step S01, including removing special characters, punctuation marks and non-lexical symbols, setting a stop word list, segmenting words and processing unstructured data into n-grams. Each n-gram and all risk event entities in the event graph are converted into n-gram word vectors V to be linked through the all-MiniLM-L6-v2 model. n and risk event entity word vector V e . V n and V e have the same dimensions, V n (i) and V e (i) represents the i-th component of the two vectors, and the cosine similarity is calculated as follows:

[0098]

[0099] In the embodiment, the threshold s is taken c = 0.5, all similarities that meet the condition s>s c Entity u Add to D and its corresponding α u Assign s.

[0100] Step S04-6: For each railway operation accident record to be linked, find the corresponding optimal accident risk chain c d Traverse all accident risk chains in C, calculate the chain confidence, and find the optimal accident risk chain c d For any accident risk chain, traverse the risk event entities on the risk chain one by one. If the risk event entity exists in D, take its corresponding α as the confidence; if the risk event entity is not in D, its confidence is 0. The process of solving the optimal accident risk chain can be expressed as:

[0101]

[0102] Where c represents a certain accident risk chain; is a 0-1 variable indicating the i-th risk event entity on the chain; α i is its corresponding confidence level. Figure 7 The following is an example of selecting the optimal risk chain c for a certain accident record in the embodiment. d The process is shown in Figure 1.

[0103] Step S04-7: Complete the selected optimal risk accident risk chain. For the missing entities in the identified optimal accident risk chain, that is, for some risk event entities that do not exist in D, complete the accident risk chain according to the following rules:

[0104] (1) If the head node (the first risk event entity in the accident risk chain) is not identified, the head node is lit.

[0105] (2) If there are unidentified nodes between the head node and the tail node (the last identified risk event in the accident risk chain), these nodes are lit up.

[0106] (3) If there are unrecognized nodes after the tail node, they remain unlit.

[0107] According to the above rules, each railway accident record related to meteorological risk is linked to a subchain of an accident risk chain in C. The subchain should start from the head node of the accident risk chain, terminate after being transferred to a certain risk event, and its length should be less than or equal to the length of the complete accident risk chain.

[0108] Step S04-8: Add the structured information as attributes to the tail node of the accident risk subchain. In this embodiment, the attributes include accident ID, train type, accident consequences, accident casualties, accident economic losses, and accident occurrence time.

[0109] This embodiment uses the Neo4j database to store the railway operation incident map related to meteorological risks. The complete map after the accident record is linked is shown in the attached figure. Figure 8 As shown, select one of the accident risk chains and select the tail node to display the structured information of the current accident chain.

[0110] Preferably, after constructing the event graph, a railway accident analysis process based on the event graph is also included, specifically including:

[0111] The railway accident analysis method based on the event graph includes the evaluation of risk transfer probability. The risk transfer probability P(v|u) from risk event u to risk event v is used to evaluate the ability of risk events to spread risks. Its calculation expression is as follows:

[0112]

[0113] Among them, chain u,v represents all accident risk chains that include both risk event u and risk event v; Freq C Represents the frequency of a certain risk event occurring in the accident risk chain c, which is obtained based on the statistical information in the graph database. The risk transfer probability is the basis for calculating subsequent indicators. In this embodiment, the risk transfer probability between all risk events with reachable relationships is calculated in advance and stored in a data table for easy access when calculating other indicators later.

[0114] The railway accident analysis method based on the event graph includes the evaluation of the cumulative risk transfer probability. The cumulative risk transfer probability CTP(s,w) is used to evaluate the ability of a certain exogenous hazard s to transfer the risk to the node event w in the complex system of railway operation. Its calculation expression is as follows:

[0115]

[0116] Among them, r is a risk subchain that is transmitted from s to node event w, R s→w is the set of all r. In this embodiment, the ability of all weather factors to transfer risk to a certain risk event is evaluated by accumulating the risk transfer probability. The result can be used to analyze the sensitivity of the railway operation subsystem to different weather factors. An example of the evaluation result is shown in the attached figure. Fig. 9As shown in the evaluation bubble chart, the horizontal axis represents different weather factors, namely exogenous hazards, and the vertical axis represents the intermediate node entities of different accident risk chains. The color of the bubble represents the cumulative risk transfer probability from the exogenous hazards to the node, and the size of the bubble represents the actual accident frequency transmitted to the node.

[0117] The railway accident analysis method based on the event graph includes the evaluation of weighted out-degree centrality. The weighted out-degree centrality WOC(w) is used to evaluate the ability of a node event w in the complex system of railway operation to spread the risk. Its calculation expression is as follows:

[0118]

[0119] Among them, n w Indicates the adjacent node event of w. This embodiment shows a method for analyzing the risk diffusion capacity of risk events using the weighted out-degree centrality index, and visualizes the weighted out-degree centrality of the risk event entity and the actual frequency of the event in historical accidents as the horizontal axis and vertical axis respectively, as shown in the attached figure. Fig.10 The nodes in the first quadrant represent that the risk events occur more frequently, and after the risk events occur, the risks are likely to spread to subsequent nodes, further causing serious accidents.

[0120] The railway accident analysis method based on the event map includes the assessment of the generalized risk of the risk chain, the generalized risk of the risk chain GR (chain s ) is used to evaluate a certain accident risk chain s The overall risk size is calculated as follows:

[0121]

[0122] Cons s represents the set of all risk events identified as tail nodes in the accident risk chain, cons represents one of the risk event elements; s is the head node of the risk chain, Sev cons is the severity of the accident corresponding to the tail node cons. In this embodiment, the severity refers to the accident record structured information in the cons node attributes, namely, casualties and economic losses. Fig.11 The figure shows the evaluation results of all accident risk chains in the embodiment, the left side shows the generalized risk evaluation results of accident chains based on casualties, and the right side shows the generalized risk evaluation results of accident chains based on economic losses.

[0123] Example 2

[0124] Based on Example 1, this example provides an unsupervised construction system for a railway operation accident graph, as shown in the attached Fig.12As shown, based on the aforementioned method for constructing and analyzing a causal graph for railway operation accidents, the present invention also provides a railway operation accident analysis system based on a causal graph, and the system architecture is as follows:

[0125] The data processing module has the following functions or attributes: (1) specifies the supported input data types: txt file (unstructured information of railway accident report), json file (structured information of railway accident report); (2) has data preprocessing function; (3) has the function of importing and exporting external data.

[0126] The graph construction module has the following functions or attributes: (1) built-in core algorithms, including language models and parameters for generating word vectors, similarity calculation methods, path search algorithms, and optimal accident chain search algorithms; (2) it has the function of storing event graph models and specifies the representation methods of graph models, including triples and graph database formats; (3) it has graph database support functions and can select specific databases such as Neo4j and TigerGraph; (4) it has a dynamic update mechanism that can automatically process updated accident records and add them to the event knowledge graph.

[0127] The analysis and visualization module has the following functions or attributes: (1) Quantitative analysis function, which can describe specific accident chain risk analysis indicators, such as cumulative risk transfer probability, weighted out-degree centrality, generalized risk of risk chain, etc.; (2) Visualization function, which can display the analysis results of different risk event nodes or different accident risk chains in charts and generate reports.

[0128] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present invention, and these modifications or replacements should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. An unsupervised construction method for a cause-and-effect graph of railway operation accidents, characterized in that: The steps include: Obtain railway operation accident report corpus data, and obtain railway accident report dataset through preprocessing; Based on the railway accident report dataset, all accident report information is classified into multiple subsystems according to the cause coding of the accident, and keywords of risk events corresponding to each cause type are extracted as risk event entities; Through subsystem decomposition layer by layer, identify the causal relationship and sequential relationship between various risk events, and establish an event relationship matrix; Based on the risk event entity and the event relationship, a matrix is ​​established, the similarity between the railway operation accident report to be linked and the risk event entity is calculated, the optimal accident risk chain is selected and completed, and a causal graph model is constructed.

2. According to claim 1, the unsupervised construction method of the cause-effect graph for railway operation accidents is characterized in that: Based on the railway operation accident report corpus data, the process of obtaining the railway accident report dataset through preprocessing includes the following steps: Obtain railway operation accident report corpus data; Extract structured information from railway operation accident report corpus data; Extract unstructured information from railway operation accident report corpus data; For the unstructured information, numbers, punctuation marks and non-lexical symbols are removed, a stop word list is set, and the unstructured information is converted into n-gram form after word segmentation.

3. The unsupervised construction method of the causal graph for railway operation accidents according to claim 2 is characterized in that: The structured information includes one or more of the accident ID, accident time, accident cause code, train type, weather, accident type, number of casualties, and economic losses caused by the accident, and the unstructured information includes a text describing the accident.

4. According to claim 1, the unsupervised construction method of the cause-effect graph for railway operation accidents is characterized in that: The multiple subsystems include accident consequences, and one or more of human factors, track and roadbed, mechanical and electrical, signal and communication, and environmental factors.

5. According to claim 1, the unsupervised construction method of the cause-effect graph for railway operation accidents is characterized in that: The process of extracting the representation text of the risk event corresponding to each cause type as the risk event entity includes the following steps: Based on the unstructured information of railway accident report data under the same subsystem, the pre-trained language model is used to obtain the document embedding vector of the entire accident description; Use the language model to obtain the word embedding vector corresponding to each candidate keyword; Calculate the cosine similarity between each word embedding vector and the document embedding vector, and select keywords whose similarity meets the requirements; The keywords after similarity screening are manually screened and the screened keywords are used as risk event entities.

6. The unsupervised construction method of the causal graph for railway operation accidents according to claim 1 is characterized in that: The process of building an event relationship adjacency matrix by subsystem layer-by-layer decomposition includes the following steps: Obtain a relationship establishment matrix that represents whether there is a mutual influence relationship between subsystems; For subsystems with influencing relationships, a risk event relationship establishment matrix is ​​established to characterize whether there is a causal or causal relationship between risk event entities.

7. The unsupervised construction method of the causal graph for railway operation accidents according to claim 1 is characterized in that: Based on the risk event entity and the event relationship matrix, the similarity between the railway operation accident report to be linked and the risk event entity is calculated, the optimal accident risk chain is selected and completed, and the process of constructing the event graph model includes the following steps: Through the depth-first path search with memory mechanism, the full amount of accident risk chain is searched; For each railway operation accident record to be linked, initialize a candidate risk event list in dictionary form; Based on the structured information of the railway operation accident record to be linked, the risk event entity and the event relationship establishment matrix, the risk event entity is added to the candidate risk event list based on a preset rule to achieve rule-based risk event linking; Preprocess the unstructured information in each railway operation accident record to be linked to obtain the unstructured information in the form of n-grams, convert each n-gram and all risk event entities in the event graph into the n-gram word vector to be linked and the risk event entity word vector, calculate the cosine similarity between the word vectors, and assign values ​​to the candidate risk events that meet the similarity requirements; For each railway operation accident record to be linked, all accident risk chains in the corresponding candidate risk event list are traversed, and the optimal accident risk chain is obtained by calculating the chain confidence; Identify and complete the missing risk event entities in the optimal accident risk chain; The structured information of the railway operation accident records to be linked is added to the tail node of the optimal accident risk chain to realize the construction of the event graph model.

8. The unsupervised construction method of the causal graph for railway operation accidents according to claim 7 is characterized in that: The process of completing the missing risk event entity includes the following steps: If the first risk event entity in the accident risk chain is not identified, then light up the risk event entity; If there are unidentified nodes between the first risk event entity on the accident risk chain and the last identified risk event on the accident risk chain, then these unidentified risk event entities are lit up; If there are still unidentified risk event entities after the tail node, they will remain unlit.

9. The unsupervised construction method of the causal graph for railway operation accidents according to claim 1 is characterized in that: Also includes: Based on the constructed event-reasoning graph model, the risk transmission probability between risk events is calculated to evaluate the risk transmission capability; Based on the constructed event-diagram model, the cumulative risk transfer probability is calculated to evaluate the ability of exogenous hazards to transfer risks to node events in the railway operation system; Based on the constructed event graph model, the weighted out-degree centrality is calculated to evaluate the risk of node event diffusion in the complex railway operation system. Based on the constructed causal graph model, the generalized risk of the risk chain is calculated and the overall risk of the accident risk chain is evaluated.

10. An unsupervised construction system for the cause-effect graph of railway operation accidents, characterized in that: include: The data processing module is used to obtain the railway operation accident report corpus data and obtain the railway accident report data set through preprocessing; A graph construction module is used to divide all accident report information into multiple subsystems based on the railway accident report data set according to the cause coding of the accident, extract keywords of risk events corresponding to each cause type as risk event entities, decompose them layer by layer through the subsystem, identify the causal relationship and sequential relationship between each risk event, and establish an event relationship establishment matrix; The analysis and visualization module is used to establish a matrix based on the risk event entity and the event relationship, calculate the similarity between the railway operation accident report to be linked and the risk event entity, select the optimal accident risk chain and complete it, build a causal graph model and perform analysis.

Images