A comprehensive identification method for risk points in urban rail transit operations based on accident text

Abstract

The present invention provides a comprehensive identification method for urban rail transit operation risk points based on accident texts. It includes: obtaining original accident text data; establishing a professional vocabulary for urban rail transit operation safety and an accident normalization template; performing normalization preprocessing of accident text data; establishing a framework word set based on the results of semantic co-occurrence analysis of the accident text; characterizing the importance of the framework word set based on the TF-IDF algorithm, and weighting it using the coefficient of variation weighting method; extracting key words and probabilities based on the BTM topic model; calculating the importance of words oriented to text semantics; proposing a comprehensive identification method for urban rail transit operation risk points based on accident texts, and executing to obtain identification results. The method of the present invention can effectively identify and extract operational risk points and risk values ​​for urban rail system operation accident texts, and can provide reliable support for the organization and implementation of urban rail operation safety management.

Description

Technical Field

[0001] The present invention relates to the technical field of urban rail transit, and in particular to a comprehensive identification method for urban rail transit operation risk points based on accident texts. Background Art

[0002] Urban rail transit systems are usually large in scale, with complex structural functions and operating conditions, strong coupling relationships between various parts of the system, high density of personnel and equipment, and are generally in a closed environment, with the characteristics of "small failures, big impacts". The operation of urban rail transit systems not only covers human factors, but is also closely related to equipment facilities and mobile equipment, environment and management. If the potential dangerous factors in all aspects are triggered, it is likely to cause local failures or even operational accidents, causing huge personal and property losses. In addition, the rail transit systems of many large and medium-sized cities today adopt a large-scale, networked operation model, with extremely long line mileage and extremely high service frequency, which also puts higher requirements on the safe operation of today's urban rail transit. In a large-scale and complex road network, once an operational safety accident occurs, the depth and breadth of its impact will far exceed the consequences of a single operational accident itself.

[0003] The operation process of the urban rail transit system is closely related to the numerous component nodes in the system, which involve multiple types of people, machines, environment and management. A considerable number of component nodes may be affected by unsafe factors and are "risky" or may spread "risks". Risk points are component nodes in the system that may generate or spread risks. At present, there is no effective method to detect and control risk points that are prone to failure or problems during the operation of urban rail transit systems. Summary of the invention

[0004] The embodiment of the present invention provides a comprehensive identification method for urban rail transit operation risk points based on accident texts, so as to effectively identify the operation risk points and risk values ​​of the urban rail transit system.

[0005] In order to achieve the above object, the present invention adopts the following technical scheme.

[0006] A comprehensive identification method for urban rail transit operation risk points based on accident texts, including:

[0007] Obtain text data of urban rail transit accident cases, correspond each accident case to an accident text data, and divide the accident category text set;

[0008] Establish a professional safety vocabulary database and standardized accident reporting template for urban rail transit operations;

[0009] Performing normalization preprocessing on the accident text data based on the accident report normalization template;

[0010] Based on the safety professional vocabulary, semantic co-occurrence analysis is performed on the accident text data after normalization preprocessing, and a framework word set is established according to the analysis results;

[0011] The importance of the framework word set is characterized based on the term frequency-inverse text frequency index TF-IDF algorithm, and the framework word set is weighted using the coefficient of variation weighting method;

[0012] Extract keywords from accident text data based on the BTM topic model, and calculate the probability of keywords based on the weights of the framework word set;

[0013] Calculating the importance of the words in the frame word set in the accident text;

[0014] The product of the probability of the key words and the importance of the words is used as the risk judgment index of the words, and the comprehensive identification of the operational risk points of the urban rail transit system is carried out based on the professional safety vocabulary of urban rail transit operation and the risk judgment index of the words.

[0015] Preferably, the establishment of a professional safety vocabulary and a standardized template for accident reporting for urban rail transit operations includes:

[0016] The safety professional vocabulary of the urban rail transit operation system includes four categories: personnel, physics, management and environment. The physics category includes multiple equipment and facility subsystems such as vehicle system, power supply system, electromechanical system, civil construction facility system, line system and signal system.

[0017] The format of the standardized template for urban rail transit operation accident report is shown in Table 4:

[0018] Table 4 Standardized template for accident reporting in urban rail transit operations

[0019]

[0020] Preferably, the normalization preprocessing of the accident text data based on the accident report normalization template includes:

[0021] Based on the standardized template of accident reports for urban rail transit operations, the format of accident text data is consistent, some incomplete and missing parts of the original accident text data are completed, all accident text data are checked, purely descriptive sentences are simplified, and informal terms are standardized.

[0022] Preferably, the semantic co-occurrence analysis is performed on the accident text data after normalization preprocessing based on the safety professional vocabulary, and a framework word set is established according to the analysis result, including:

[0023] The Jaccard similarity coefficient is used as an indicator to characterize the degree of word co-occurrence in semantics. Given two sets A and B, the Jaccard coefficient is defined as the ratio of the size of the intersection of A and B to the size of the union of A and B, expressed as follows:

[0024]

[0025] Where J(A,B) is the Jaccard coefficient value, J(A,B)∈[0,1]. When both sets A and B are empty, J(A,B) is defined as 1.

[0026] The co-occurring word pairs are determined based on the calculation results of word frequency and similarity. It is stipulated that when the number of word occurrences is greater than 30 and must appear in the top 1% of the total number of cases, and the Jaccard coefficient of the inter-word connection is greater than 0.2, the word is included in the framework word set. KH coder software is used for coding and visualization, and a semantic co-occurrence analysis is conducted on the safety accident cases in the operation of the urban rail transit system. Based on the obtained word analysis results, five framework word sets are established, namely, personnel, equipment and environment, physical process and participant activity, technical measures and management methods, and basic information description, which are expressed by formulas (2), (3), (4), (5), and (6), respectively.

[0027] V H = {v H1 ,v H2 ,v H3 ,...,v Hn} (2)

[0028] V E = {v E1 ,v E2 ,v E3 ,...,v Em} (3)

[0029]

[0030] V T = {v T1 ,v T2 ,v T3 ,...,v Tj} (5)

[0031] V B = {v B1 ,v B2 ,v B3 ,...,v Bq} (6)

[0032] V H ——A set of framework words for personnel;

[0033] V E ——A collection of framework terms for equipment and environment;

[0034] V P - A set of framework words for physical processes and participant activities;

[0035] V T ——A set of framework terms for technical measures and management methods;

[0036] V B ——A set of basic information description framework words;

[0037] n——the total number of words in the personnel-related framework word set;

[0038] m——the total number of words in the equipment and environment framework word set;

[0039] i——the total number of words in the frame word set of physical process and participant activity;

[0040] j——the total number of words in the technical measures and management methods framework word set;

[0041] q——Words in the basic information description framework word set.

[0042] Preferably, the TF-IDF algorithm is used to characterize the importance of the framework word set, and the coefficient of variation weighting method is used to weight the framework word set, including:

[0043] The coefficient of variation weighting method is used to weight each frame word set, and the weight of each frame word set is used as the feature value of the full text of the accident from a semantic perspective. The feature value includes word frequency TF, inverse document frequency IDF, word frequency-inverse document frequency TF-IDF and coefficient of variation and weight;

[0044] The word frequency TF is used to measure the frequency of the representative words in the entire accident text. The normalization of the word frequency TF is completed using the following formula (7):

[0045]

[0046] In the formula ——Word v x The frequency of words;

[0047] ——Word v x In the accident text d w The number of times it appears in

[0048] ——Accident textd w The total number of all participles in .

[0049] The inverse document frequency (IDF) is a characteristic parameter used to measure the general importance of keywords, and the denominator is added with 1.

[0050]

[0051] In the formula ——Word v x The full-text inverse document frequency of

[0052] |D|——The total number of all accident texts;

[0053] 1+|{k:v x ∈d k}|——Contains the word v x The total number of accident texts.

[0054] Term frequency-inverse document frequency TF-IDF is used to calculate the comprehensive frequency of a word in the full text, reflecting the global importance of this keyword in the accident text set. The higher the weight value, the higher the importance of the word in the accident text. The calculation method is as follows (9).

[0055]

[0056] In the formula ——Keyword v x The full-text word frequency of the accident text corpus;

[0057] ——Keyword v x The full-text inverse document frequency of

[0058] TF-IDF(v x )——Keyword v x Term frequency-inverse document frequency value in the accident text set.

[0059] Each indicator is weighted according to the degree of variation between the current value and the specified value of each evaluation indicator. When the difference between the current value of each indicator and the target value is large, it means that it is difficult for the indicator to achieve the target value and a larger weight should be given. Otherwise, a smaller weight should be given. Taking the personnel-based frame word set as an example, its coefficient of variation is calculated as formula (10), and the corresponding weight formula of the personnel sub-frame is formula (11);

[0060]

[0061] Where C(V H )——the coefficient of variation of the set of personnel-related frame words;

[0062] σ(V H)——Standard deviation of TF-IDF value of each word in the personnel-related framework word set;

[0063] ——The average TF-IDF value of each word in the personnel-related framework word set;

[0064] W(V H )——the weight value of the staff subframe of the semantic network framework;

[0065] According to the above method, for the full text semantics, the weight values ​​of the frame word sets of the accident text data corresponding to each semantic network are set to 0.213, 0.206, 0.167, 0.214, and 0.200 respectively. The detailed calculation results are shown in Table 5:

[0066] Table 5. Calculation results of subframe weights of semantic network framework

[0067]

[0068] Preferably, the extracting of keywords from the accident text data based on the BTM topic model and calculating the probability of keywords based on the weights of the framework word set include:

[0069] Textrank is used to extract the segmented summaries of the long text in the accident text data to obtain the short text. The Jieba word segmentation tool is used to segment the extracted short text. The BTM topic model is used to extract keywords from the short text after word segmentation. The extracted keywords are determined as some words under each topic distribution in the BTM model operation results. These words together constitute the initial word set, and the full-text word topic probability of the words under each topic is calculated. The full-text word topic probability is defined as the probability of word w in topic z (w│z) and the probability of topic z θ Z The product of .

[0070] Preferably, the step of calculating the importance of the words in the framework word set in the accident text includes:

[0071] Each extracted word is regarded as a morpheme, and the importance of a word in the full text of the accident is defined as the word importance, which is expressed by formula (18):

[0072]

[0073] In the formula, score(v x ,d)——morpheme (word)v x Importance in the full text d;

[0074] p(s i |d)——a certain category of text set s iThe similarity with the full text d, s i ∈d;

[0075] n i ×if(v x ins i )——Text Sets i Contains morphemes (words) x If a certain category of text set s i Contains morphemes (words) v x ;

[0076] m——the total number of category text sets, 0 <i≦m。

[0077] According to the accident type, the full text is divided into multiple accident category text sets A i (i=1,2,…10), use the vector space model method to obtain the similarity between the category text set and the full text, and convert the similarity comparison between texts into the similarity comparison between vectors. Assume that two texts d a , d b The eigenvectors of And the angle between the vectors is ω, then use d a , d b The cosine coefficient value of the angle between represents the similarity between texts, and the cosine value is positively correlated with the similarity between texts;

[0078]

[0079] Based on the previous results, construct a semantic feature vector of the full text of the accident For text category s i And feature t, the mutual information calculation method is as follows:

[0080]

[0081] Where MI(t,s i )——text category s i and the mutual information of feature t;

[0082] P(t)——The probability of feature t appearing in the corpus, which refers to the ratio of the total number of texts containing word t to the total number of texts in the full text;

[0083] P(s i )——Category is s i The probability of a text appearing in all text sets is s i The ratio of the total number of texts in a category to the total number of texts in all categories;

[0084] P(t,s i )——The text contains both the feature word t and belongs to the category s i The probability that it contains both the feature word t and belongs to the category s i The ratio of the total number of texts in the whole text to the total number of texts in the whole text;

[0085] Based on the above method, after removing outliers, the words in the framework word set and the text set A of each accident category are calculated. i (i=1,2,…10), the logarithm of the mutual information is based on 2, and the unit of the mutual information is bit. Then, for each frame set V H 、V E 、V P 、V T 、V B , find the text set A of each accident category i (i=1,2,…10) and the mean of the mutual information MI(V H ,A i )、MI(V E ,A i )、MI(V P ,A i )、MI(V T ,A i )、MI(V B ,A i ), then, for each accident category text set A i (i=1,2,…10), sort out the above-obtained means and normalize them respectively. The normalization formula is shown in formula (21):

[0086]

[0087] In the formula, x is a certain mean value;

[0088] x′——normalized value corresponding to the mean;

[0089] X is the set of all means of the accident category text set corresponding to the mean.

[0090] Get the feature vector of the semantic-based accident category text set A1 As shown in formula (22):

[0091]

[0092] Combined with the calculation method, the similarity between each accident category text set and the full text is obtained from the semantic perspective. The value of a word v x The importance score (vx ,d).

[0093] Preferably, the product of the probability of the key words and the importance of the words is used as the risk judgment index of the words, and the comprehensive identification of the operational risk points of the urban rail transit system is performed based on the professional safety vocabulary of urban rail transit operation and the risk judgment index of the words, including:

[0094] The product of the two comprehensive indicators, the full-text word topic probability and the word importance, is used as the word risk judgment indicator. Based on the safety professional word library of urban rail transit operation, a comprehensive identification method for the operation risk points of the urban rail transit system is proposed. The process is shown in Steps 1 to 8, where the α value of the BTM model is set to α = 50 / T, T is the number of topics, the β value is taken as 0.01 according to general experience, and the number of iterations is set to 100 by default;

[0095] Step 1: Import the text segmentation results, enter the number of topics T = 3, and start the BTM model test iteration, calculate the model results and output the perplexity (D test )3;

[0096] Step 2: Repeat Step 1 and output the perplexity (D) of the model under different test conditions with different number of topics T (T>3 and is an integer). test ) T ;

[0097] Step 3: Draw the perplexity curve under different number of topics T, and determine the number of topics T=K corresponding to the minimum value within the reasonable fitting range;

[0098] Step 4: Output the BTM model running results under the condition of T = K, and calculate the full-text word topic probability value of the words under each topic. Sort the full-text word topic probability values ​​under each topic respectively, and take the top M1 words under each topic to form the initial word set λ1;

[0099] Step 5: Traverse the professional vocabulary of urban rail transit operation safety, filter out all words in λ1 that are synonymous or strongly related to a part of the professional vocabulary, merge the synonyms, and output the related word set λ2;

[0100] Step 6: Calculate the word importance and risk judgment index value of each word in λ2 in turn, sort the risk judgment index values ​​of all words in λ2 and take the top M2 words to form the candidate word set λ3;

[0101] Step 7: Compare with the professional vocabulary, and inquire with multiple industry experts based on the output of λ3, integrate various descriptive information about the same risk point and eliminate redundancy, so as to form risk points that correspond to the professional vocabulary and meet the actual situation;

[0102] Step 8: The risk value of a risk point is calculated as the sum of the risk judgment index values ​​of all words associated with it in λ3. All risk points are sorted by risk value, and the process is terminated and output.

[0103] It can be seen from the technical solutions provided by the above-mentioned embodiments of the present invention that the embodiments of the present invention help researchers to effectively analyze the risk knowledge and occurrence mechanisms behind operational accidents, provide support for safety managers to take proactive safety decisions, and propose a comprehensive method for effectively identifying and extracting operational risk points and risk values ​​of urban rail transit systems based on the text of urban rail transit system operations accidents.

[0104] Additional aspects and advantages of the present invention will be given in part in the following description, which will become obvious from the following description, or may be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0105] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0106] Figure 1 A processing flow chart of a comprehensive identification method for urban rail transit operation risk points based on accident text provided by an embodiment of the present invention;

[0107] Figure 2 A schematic diagram of an urban rail transit operation active safety knowledge graph ontology model provided by an embodiment of the present invention;

[0108] Figure 3 A schematic diagram of a graphical model representation of a BTM provided in an embodiment of the present invention;

[0109] Figure 4 A schematic diagram of a perplexity curve provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0110] The embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be interpreted as limiting the present invention.

[0111] It will be understood by those skilled in the art that, unless expressly stated, the singular forms "one", "said", and "the" used herein may also include plural forms. It should be further understood that the term "comprising" used in the specification of the present invention refers to the presence of the features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof. It should be understood that when we refer to an element as being "connected" or "coupled" to another element, it may be directly connected or coupled to the other element, or there may be intermediate elements. In addition, the "connection" or "coupling" used herein may include wireless connection or coupling. The term "and / or" used herein includes any unit and all combinations of one or more associated listed items.

[0112] It will be understood by those skilled in the art that, unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as those generally understood by those skilled in the art in the art to which the present invention belongs. It should also be understood that terms such as those defined in common dictionaries should be understood to have meanings consistent with the meanings in the context of the prior art, and will not be interpreted with idealized or overly formal meanings unless defined as herein.

[0113] To facilitate understanding of the embodiments of the present invention, several specific embodiments will be further explained below with reference to the accompanying drawings, and each embodiment does not constitute a limitation on the embodiments of the present invention.

[0114] The processing flow of a comprehensive identification method for urban rail transit operation risk points based on accident text provided by an embodiment of the present invention is as follows: Figure 1 As shown, the following processing steps are included:

[0115] Step S1: Acquire text data of urban rail transit accident cases, each accident case corresponds to one piece of accident text data, and divide the accident category text set.

[0116] Step S2: Establish a professional safety vocabulary and a standardized template for accident reporting for urban rail transit operations.

[0117] Step S3, performing normalization preprocessing on the accident text data based on the accident report normalization template;

[0118] Step S4: perform semantic co-occurrence analysis on the accident text data based on the safety professional vocabulary, and establish a framework word set according to the analysis results.

[0119] Step S5: characterize the importance of the framework word set based on the TF-IDF (Term Frequency-Inverse Document Frequency) algorithm, and weight the framework word set using the coefficient of variation weighting method.

[0120] Step S6: extract keywords from the accident text based on the BTM topic model (BiTerm Topic Model), and calculate the probability of the keywords based on the weights of the framework word set.

[0121] Step S7: define the importance of the words in the framework word set in the accident text - word importance and its calculation method, and calculate the word importance oriented to the text semantics.

[0122] Step S8: The product of the probability of the key words and the importance of the words is used as the risk judgment index of the words, and based on the constructed professional vocabulary of urban rail transit operation safety, a comprehensive identification method for the operation risk points of the urban rail transit system is proposed.

[0123] The steps of a comprehensive identification method for urban rail transit operation risk points based on accident texts are proposed, and the identification results are obtained through execution.

[0124] Furthermore, the step 1 is specifically as follows:

[0125] Conduct online and offline data research and summary, collect relatively complete urban rail transit system operation accident cases, stipulate accident classification principles, each accident case corresponds to a text data, divide the accident category, and the text of each category constitutes an accident category text set.

[0126] In order to understand the current status of operational safety management and data analysis in the urban rail transit industry, we conducted field visits and surveys in typical subway operating companies and operating stations in Beijing, Qingdao, Shanghai, Nanjing, Chengdu, Urumqi and other places. We also conducted online data surveys and summaries on the websites of many foreign urban rail transit operating companies and relevant administrative departments and transportation departments in many provinces and cities at home and abroad. We collected a total of 910 typical and relatively complete cases of urban rail transit system operation accidents.

[0127] According to statistics from scholars, there are 13 types of accidents (disaster events) that have occurred or may occur in subways in various countries around the world. my country's "Regulations on Information Reporting and Analysis Management of Risky Events in Urban Rail Transit Operations" (Jiaoyungui

[2019] No. 10) also stipulates 16 types of risky events in urban rail transit operations. Figure 2A schematic diagram of an active safety knowledge graph ontology model for urban rail transit operation provided by an embodiment of the present invention. Combined with the above information, each accident case corresponds to a text data, and all texts are divided into 10 categories according to the actual content of the data: fire (including explosion), train derailment, train collision, facility equipment failure, terrorist attack, flood, suspension of operation, large passenger flow (including trampling), passenger fall from the track and others. The text of each category constitutes an accident category text set A. i (i=1,2,…10). Among them, facility and equipment failures mainly include elevator accidents, signal system failures, communication equipment failures and other major equipment failures, deformation and collapse of infrastructure, and people and objects being trapped. Outages refer to accidents that cause a long-term interruption of the operation of one or several lines. The collection of all accident text data is called the full text, represented by d. The statistics by type and time period are shown in Table 1, among which equipment and facility failures are the most important form of accidents.

[0128] Table 1 Statistics of accident occurrence time periods

[0129]

[0130] Furthermore, the step 2 is specifically as follows:

[0131] Based on the research work, a professional safety vocabulary database for urban rail transit operations was established; after analyzing and summarizing a large amount of accident data, a standardized template format for urban rail transit operation accident reports was formed.

[0132] In order to complete the recognition of professional vocabulary in the field of urban rail transit during the text analysis and extraction process, and avoid the occurrence of vocabulary loss and ambiguity, the present invention has established a professional safety vocabulary for urban rail transit operations based on a large amount of research work, and summarized the core components, elements and various environments involved in the operation of the urban rail transit operation system. The professional safety vocabulary has four levels, and the words are divided into four categories: personnel, physics, management and environment. The physical category includes multiple equipment and facility subsystems such as vehicle system, power supply system, electromechanical system, civil construction facility system, line system and signal system.

[0133] Some results of the safety professional vocabulary for urban rail transit operation constructed are presented. Tables 2 and 3 are the physical category - vehicle system vocabulary (part) and the personnel, environment and management vocabulary respectively.

[0134] Table 2 Physics - Vehicle System Vocabulary (Part)

[0135]

[0136]

[0137] Table 3 Vocabulary of personnel, environment and management

[0138]

[0139]

[0140] After analyzing and summarizing a large amount of acquired accident data, the present invention forms a standardized template for urban rail transit operation accident reports, the format of which is shown in Table 4.

[0141] Table 4 Standardized template for accident reporting in urban rail transit operations

[0142]

[0143] Submitted by: Submitted on:

[0144] Furthermore, the step 3 is specifically as follows:

[0145] All accident texts were unified into Chinese. Based on expert consultation, the initial accident text data was standardized and pre-processed based on the accident report standardization template for urban rail transit operations. The key points of the processing are as follows:

[0146] (1) Standardize the format of accident text data

[0147] Based on the standardized template of accident report for urban rail transit operation, the data format of accident data (accident report, event record, etc.) is unified, the format of accident text data is consistent, and it is uniformly stored in Word and Excel for subsequent research.

[0148] (2) Repair missing information in the original data

[0149] After in-depth data review and field visits, some of the incomplete and missing parts in the original accident text data were completed, such as some omitted common sense information, missing process information, location information, specific information about people and objects, etc., to supplement the text structure of the data.

[0150] (3) Standardized description text information

[0151] Based on the standardized template for accident reporting in urban rail transit operations, professional safety managers were further consulted to check all accident text data in turn, simplify purely descriptive sentences, and standardize informal terms. At the same time, a professional vocabulary was established to standardize the description of synonymous or near-synonymous words in the text for subsequent analysis. For example, "door mechanism, door panel, door" and so on are standardized as "door", and "roadbed, sleeper, rail, rail, track" and so on are standardized as "track".

[0152] Furthermore, the step 4 is specifically as follows:

[0153] The Jaccard similarity coefficient is used as an indicator to characterize the degree of word co-occurrence in semantics. It is often used to compare the similarities and differences between finite sample sets. The larger the Jaccard coefficient value, the higher the sample similarity. Given two sets A and B, the Jaccard coefficient is defined as the ratio of the size of the intersection of A and B to the size of the union of A and B, expressed as follows:

[0154]

[0155] Where J(A,B) is the Jaccard coefficient, J(A,B)∈[0,1]. When both sets A and B are empty, J(A,B) is defined as 1.

[0156] According to the calculation results of word frequency and similarity, the most common co-occurring word pairs are determined. Since words with high correlation can represent specific topics covered by the text data, while words with low correlation scores are general and cannot represent any specific topic, in order to ensure the scientific nature of the semantic network, to be included in the final semantic network, it is stipulated that the number of occurrences of the word must be greater than 30 and must appear in the top 1% of the total number of cases, and the Jaccard coefficient of the connection between words must be greater than 0.2. Based on this method, a semantic co-occurrence analysis was conducted on all 910 urban rail transit system operation safety accident cases. The process used KHcoder software for coding and visualization, which enabled a rapid review of the literature and generated quantitative mining results. According to the obtained word analysis results, five framework word sets were established, namely, personnel, equipment and environment, physical process and participant activity, technical measures and management methods, and basic information description, which are expressed by formulas (2), (3), (4), (5), and (6), respectively.

[0157] V H = {v H1 ,v H2 ,v H3 ,...,v Hn} (2)

[0158] V E = {v E1 ,v E2 ,v E3 ,...,v Em} (3)

[0159]

[0160] V T = {v T1 ,v T2 ,v T3 ,...,vTj} (5)

[0161] V B = {v B1 ,v B2 ,v B3 ,...,v Bq} (6)

[0162] Where V H ——A set of framework words for personnel;

[0163] V E ——A collection of framework terms for equipment and environment;

[0164] V P - A set of framework words for physical processes and participant activities;

[0165] V T ——A set of framework terms for technical measures and management methods;

[0166] V B ——A set of basic information description framework words;

[0167] n——the total number of words in the personnel-related framework word set;

[0168] m——the total number of words in the equipment and environment framework word set;

[0169] i——the total number of words in the frame word set of physical process and participant activity;

[0170] j——the total number of words in the technical measures and management methods framework word set;

[0171] q——The total number of words in the basic information description framework word set.

[0172] Furthermore, the step 5 is specifically as follows:

[0173] The TF-IDF algorithm, which can balance the popularity and specificity of words, is introduced to characterize the importance of the frame word set. The coefficient of variation weighting method is used to weight each frame word set, and the weight of each frame word set is used as the feature value of the full text of the accident from a semantic perspective. The relevant concepts are as follows:

[0174] (1) Term frequency (TF)

[0175] Word frequency is used to measure the frequency of occurrence of representative words in all accident texts. In order to eliminate the influence of different accident text lengths on the calculation of this measurement parameter, the following formula (7) is usually used to complete the normalization process.

[0176]

[0177] In the formula ——Word v x The frequency of words;

[0178] ——Word v x In the accident text d w The number of times it appears in

[0179] ——Accident textd w The total number of all participles in .

[0180] (2) Inverse Document Frequency (IDF)

[0181] Inverse document frequency is a characteristic parameter used to measure the general importance of keywords. In order to ensure that the formula makes sense and avoid the denominator of the calculation formula being 0, the denominator is often added with 1.

[0182]

[0183] In the formula ——Word v x The full-text inverse document frequency of

[0184] |D|——The total number of all accident texts;

[0185] 1+|{k:v x ∈d k}|——contains the word v x The total number of accident texts.

[0186] (3) Term Frequency-Inverse Document Frequency (TF-IDF)

[0187] In order to ensure the accuracy of the extraction results, the above two representation parameters are often combined to calculate the comprehensive frequency of a word in the whole text, reflecting the global importance of this keyword in the accident text set. The higher the weight value, the higher the importance of the word in the accident text. The calculation method is as follows (9).

[0188]

[0189] In the formula ——Keyword v x The full-text word frequency of the accident text corpus;

[0190] ——Keyword v x The full-text inverse document frequency of

[0191] TF-IDF(v x )——Keyword v xTerm frequency-inverse document frequency value in the accident text set.

[0192] (4) Coefficient of variation and weight

[0193] The coefficient of variation method is a relatively objective weighting method. This method weights each indicator according to the degree of variation between the current value of each evaluation indicator and the target value. When the current value of each indicator is far from the target value, it means that it is difficult for the indicator to achieve the target value and a larger weight should be given. Otherwise, a smaller weight should be given. Taking the personnel-related framework word set as an example, its coefficient of variation is calculated as formula (10), and the corresponding weight formula of the personnel subframe is formula (11). The calculation method of the coefficient of variation and weight of other frameworks is similar.

[0194]

[0195] Where C(V H )——the coefficient of variation of the set of personnel-related frame words;

[0196] σ(V H )——Standard deviation of TF-IDF value of each word in the personnel-related framework word set;

[0197] ——The average TF-IDF value of each word in the personnel-related framework word set;

[0198] W(V H )——The weight value of the staff subframe of the semantic network framework.

[0199] According to the above method, for the full-text semantics, the characteristic values ​​(weights) of the accident text data corresponding to each semantic network sub-frame are set to 0.213, 0.206, 0.167, 0.214, and 0.200 respectively. The detailed calculation results are shown in Table 5.

[0200] Table 5. Calculation results of subframe weights of semantic network framework

[0201]

[0202] Furthermore, the step 6 is specifically as follows:

[0203] Based on the extraction and segmentation of long text summaries, the perplexity is used to measure the effectiveness of the topic model in extracting text topics. The BTM topic model is combined to complete keyword extraction and obtain the full-text word topic probability.

[0204] The BTM topic model (Biterm Topic Model) considers the entire corpus to be a mixture of various topics, and performs topic learning on short texts from the perspective of the entire corpus to solve the sparsity problem of the corpus. Biterm refers to an unordered word pair that appears simultaneously in the context. Each biterm comes from a specific topic, and the author extracts any two different words from a short text to form a biterm. The graph model of BTM is represented as follows: Figure 3 As shown in Figure 2, α and β are the hyperparameters of the Dirichle prior distribution; θ is the topic probability distribution in the BTM text corpus; Ф k is the probability distribution of topic-word pairs; z is the topic number of the word pair biterm, T is the number of topics; w i 、w j are the two words in the biterm; B is the number of word pairs in the BTM corpus.

[0205] (1) BTM modeling process and reasoning method

[0206] For the word pairs of the entire corpus, the general modeling process of the BTM topic model is as follows:

[0207] 1) For each topic z, generate the topic word distribution z~Dirichlet(β) under z;

[0208] 2) Generate a global topic distribution θ~Dirichlet(ɑ) for the entire text corpus;

[0209] 3) For each biterm b in the bigram pair set B, i, w j );

[0210] ① Randomly extract a z from θ in the entire text corpus;

[0211] ② Extract two words from the extracted topic z, that is, a word pair.

[0212] Through the above process, the joint probability formula of word pair b is:

[0213] P(b)=∑ z P(z)P(w i │z)P(w j │z)=∑ z θ z i|z j|z (12)

[0214] The formula for calculating the probability of biterm for the entire text corpus is:

[0215] P(B)=∏ (i,j)∑ z θ z i|z j|z (13)

[0216] BTM cannot directly obtain the topic probability distribution of the document during the topic learning process, but can obtain it through inference calculation. First, the initial state of the Markov chain is randomly selected, and the conditional probability P(z│z ―b ,B,α,β), as shown in formula (14).

[0217]

[0218] Where z ―b ——topic distribution of all word pairs except word pair b;

[0219] B——global word pair set;

[0220] n z ——the number of times the word pair is assigned topic z;

[0221] n (w│z) is the number of times word w is assigned topic z;

[0222] M – the number of distinct words in the corpus.

[0223] Based on the co-occurrence of topic word pairs, we calculate the topic-word distribution and the topic distribution θ of the corpus. The formulas are shown in (15) and (16) respectively.

[0224]

[0225] Where (w│z) – the probability of word w in topic z;

[0226] |B|——the total number of word pairs;

[0227] θ Z ——the probability of topic z.

[0228] (2) Determination of the optimal number of topics

[0229] Figure 4A schematic diagram of a perplexity curve provided for an embodiment of the present invention. Perplexity is used to measure the effectiveness of a topic model in extracting text topics. The lower the perplexity, the stronger the topic model capability, the higher the degree of generalization, and the better the model, as defined in formula (17). For models trained under different numbers of topics (Topic), the number of topics (and other parameters) corresponding to the minimum perplexity is the optimal one. However, the more topics the better, when the number of topics is large, the generated model will often overfit. An effective method is to observe the curve change of the Perplexity index with the number of Topic, find the minimum value within a reasonable range of curve fitting, and thus select a suitable topic value.

[0230]

[0231] Where D test ——The text collection of the test corpus;

[0232] ——In the document words;

[0233] For Documentation The number of words in

[0234] M – the size of the test corpus text collection (the number of documents);

[0235]

[0236] (3) Word extraction results and full-text word topic probability calculation

[0237] The extracted keywords are determined as some words under each topic distribution in the BTM model operation results, and these words together constitute the initial word set. The indicator for evaluating and selecting words under each topic - the full-text word topic probability Probability (w, d) is defined as the probability of word w in topic z (w│z) and the probability of topic z θ Z The product of .

[0238] (4) Short text acquisition and text segmentation

[0239] In order to obtain the short texts with strong summarization ability required in the BTM model and facilitate the subsequent text mining work, text summary extraction and text segmentation (including part-of-speech filtering) were carried out. Using the cloud service platform launched by the natural language processing tool HanLP in 2020, by calling the HanLP-API interface, Textrank was used to complete the segmented summary extraction of a large amount of long text in the full text of the accident, and urban rail transit safety professionals were consulted to improve the summary content. Afterwards, the Jieba word segmentation tool was used to complete the word segmentation of all short texts before and after the extraction, and filtered with reference to the Chinese part-of-speech comparison table [Peking University standard].

[0240] Furthermore, the step 7 is specifically as follows:

[0241] Based on the results of semantic analysis and the concept of importance in text search, each extracted word is regarded as a morpheme, and the importance of a word in the full text of the accident is defined as the word importance. Text similarity and mutual information are introduced, and a method for calculating word importance is proposed.

[0242] Importance refers to the importance of each morpheme in the entire text search string, and is an important component of core word extraction. Combining the theoretical algorithm of importance in text search, each extracted word is regarded as a morpheme, and the importance of a word in the full text of the accident is defined as word importance, which can be expressed by formula (18).

[0243]

[0244] In the formula, score(v x ,d)——morpheme (word)v x Importance in the full text d;

[0245] p(s i |d)——a certain category of text set s i The similarity with the full text d, s i ∈d;

[0246] n i ×if(v x ins i )——Text Sets i Contains morphemes (words) v x If a certain category of text set s i Contains morphemes (words) v x ;

[0247] m——the total number of category text sets, 0 <i≦m。

[0248] According to the above formula (18), the problem is transformed into dividing the full text into several categories of text sets and obtaining the similarity between each category text set and the full text. The full text has been divided into multiple accident category text sets A according to the accident type. i (i=1,2,…10). Therefore, the next step of the study needs to find the similarity between the category text set and the full text.

[0249] To calculate the text similarity, we use a typical statistical method, Vector Space Model (VSM), whose basic idea is to convert the similarity comparison between texts into the similarity comparison between vectors. a , d b The eigenvectors of And the angle between the vectors is ω, then use d a , d b The cosine coefficient value of the angle between represents the similarity between texts, and the cosine value is positively correlated with the similarity between texts.

[0250]

[0251] Based on the previous results, construct a semantic feature vector of the full text of the accident To calculate the text set A for each accident category i (i=1,2,…10) feature vectors, this paper introduces the concept of mutual information (MI). Mutual information is a measure of the correlation between feature items and text categories. The core idea is that if a feature item appears frequently in a certain category and less frequently in other categories, it can be considered that the mutual information between the feature item and the former is relatively large. For text category s i And feature t, the mutual information calculation method is as follows:

[0252]

[0253] Where MI(t,s i )——text category s i and the mutual information of feature t;

[0254] P(t)——The probability of feature t appearing in the corpus, which refers to the ratio of the total number of texts containing word t to the total number of texts in the full text;

[0255] P(s i )——Category is s i The probability of a text appearing in all text sets is s i The ratio of the total number of texts in a category to the total number of texts in all categories;

[0256] P(t,s i )——The text contains both the feature word t and belongs to the category s i The probability that it contains both the feature word t and belongs to the category s i The ratio of the total number of texts in the full text to the total number of texts in the full text.

[0257] Based on the above method, after removing outliers, the words in the frame word set and the text set A of each accident category can be calculated. i (i=1,2,…10), the logarithm of the mutual information is based on 2, and the unit of the mutual information is bit. Then, for each frame set V H 、V E 、V P 、V T 、V B , find the text set A of each accident category i (i=1,2,…10) and the mean of the mutual information MI(V H ,A i )、MI(V E ,A i )、MI(V P ,A i )、MI(V T ,A i )、MI(V B ,A i ). Then, for each accident category text set A i (i=1,2,…10), sort out the means obtained above and normalize them respectively. The normalization formula is shown in formula (21).

[0258]

[0259] In the formula, x is a certain mean value;

[0260] x′——normalized value corresponding to the mean;

[0261] X is the set of all means of the accident category text set corresponding to the mean.

[0262] Finally, the feature vector of the semantic-based accident category text set A1 can be obtained: As shown in formula (22). similar.

[0263]

[0264] Combined with the calculation method, the similarity between each accident category text set and the full text can be obtained from a semantic perspective. The value of a word vx The importance of the word in the full text of the accident - word importance score (v x ,d).

[0265] Furthermore, the step 8 is specifically as follows:

[0266] The product of the two comprehensive indicators, the full-text word topic probability and the word importance, is used as the word risk judgment indicator. Based on the safety professional word library of urban rail transit operation, a comprehensive identification method for the operation risk points of the urban rail transit system is proposed. The process is shown in Steps 1 to 8. The α value of the BTM model is set to α = 50 / T (T is the number of topics), the β value is 0.01 according to general experience, and the number of iterations is set to 100 by default.

[0267] Step 1: Import the text segmentation results, enter the number of topics T = 3, and start the BTM model test iteration, calculate the model results and output the perplexity (D test )3;

[0268] Step 2: Repeat Step 1 and output the perplexity (D) of the model under different test conditions with different number of topics T (T>3 and is an integer). test ) T ;

[0269] Step 3: Draw the perplexity curve under different number of topics T, and determine the number of topics T=K corresponding to the minimum value within the reasonable fitting range;

[0270] Step 4: Output the BTM model running results under the condition of T = K, and calculate the full-text word topic probability value of the words under each topic. Sort the full-text word topic probability values ​​under each topic respectively, and take the top M1 words under each topic to form the initial word set λ1;

[0271] Step 5: Traverse the professional vocabulary of urban rail transit operation safety, filter out all words in λ1 that are synonymous or strongly related to a part of the professional vocabulary, merge the synonyms, and output the related word set λ2;

[0272] Step 6: Calculate the word importance and risk judgment index value of each word in λ2 in turn, sort the risk judgment index values ​​of all words in λ2 and take the top M2 words to form the candidate word set λ3;

[0273] Step 7: Compare with the professional vocabulary, and inquire with multiple industry experts based on the output of λ3, integrate various descriptive information about the same risk point and eliminate redundancy, so as to form risk points that correspond to the professional vocabulary and meet the actual situation;

[0274] Step 8: The risk value of a risk point is calculated as the sum of the risk judgment index values ​​of all words associated with it in λ3. All risk points are sorted by risk value, and the process is terminated and output.

[0275] Through the method, some results of the process are shown as follows Figure 3 As shown in Table 6. Figure 3 It can be seen that when T>19, the model fitting effect gradually tends to be stable. When T=32, the perplexity value reaches a minimum value. After T>32, the perplexity value shows an upward trend, and the model generation effect becomes worse.

[0276] Table 6 Similarity between accident category text set and full text

[0277]

[0278]

[0279] Under the condition of T=32, M1=25, M2=125 are taken according to the data situation, and Steps 4 to 8 are carried out. Finally, the operational risk points of the urban rail transit system are extracted and identified and divided into four categories: personnel, physical, environmental, and management. They are sorted by risk value as shown in Tables 7 to 10.

[0280] Table 7 Risk point identification results - personnel

[0281]

[0282] Table 8 Risk point identification results - physics

[0283]

[0284]

[0285] Table 9 Risk point identification results - environmental category

[0286]

[0287] Table 10 Risk point identification results - management

[0288]

[0289] To sum up, the embodiment of the present invention provides a comprehensive identification method for urban rail transit operation risk points based on accident texts, which can effectively identify and extract urban rail transit system operation risk points and risk values ​​for urban rail transit system operation accident texts, and can provide support for better organization and implementation of urban rail transit operation safety management and proactive safety decision-making.

[0290] The embodiment of the present invention is very necessary to achieve the purpose of actively preventing operational accidents. From the perspective of possibility and importance, it comprehensively considers the full-text word topic probability and word importance in the accident text, and proposes a comprehensive identification method for identifying risk points in urban rail transit operations based on accident text data analysis.

[0291] Those skilled in the art can understand that the accompanying drawings are only schematic diagrams of an embodiment, and the modules or processes in the accompanying drawings are not necessarily required to implement the present invention.

[0292] It can be known from the description of the above implementation methods that those skilled in the art can clearly understand that the present invention can be implemented by means of software plus a necessary general hardware platform. Based on such an understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which can be stored in a storage medium such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in the various embodiments of the present invention or certain parts of the embodiments.

[0293] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device or system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment. The device and system embodiments described above are merely schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0294] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A comprehensive identification method for urban rail transit operation risk points based on accident text, characterized in that: include: Obtain text data of urban rail transit accident cases, correspond each accident case to an accident text data, and divide the accident category text set; Establish a professional safety vocabulary database and standardized accident reporting template for urban rail transit operations; Performing normalization preprocessing on the accident text data based on the accident report normalization template; Based on the safety professional vocabulary, semantic co-occurrence analysis is performed on the accident text data after normalization preprocessing, and a framework word set is established according to the analysis results; The importance of the framework word set is characterized based on the term frequency-inverse text frequency index TF-IDF algorithm, and the framework word set is weighted using the coefficient of variation weighting method; Extract keywords from accident text data based on the BTM topic model, and calculate the probability of keywords based on the weights of the framework word set; Calculating the importance of the words in the frame word set in the accident text; The product of the probability of the key words and the importance of the words in the accident text is used as the risk judgment index of the words, and the comprehensive identification of the operational risk points of the urban rail transit system is carried out based on the professional safety vocabulary of the urban rail transit operation and the risk judgment index of the words.

2. The method according to claim 1, characterized in that The establishment of a professional safety vocabulary and accident report standardization template for urban rail transit operations includes: The safety professional vocabulary of the urban rail transit operation system includes four categories: personnel, physics, management and environment. The physics category includes multiple equipment and facility subsystems such as vehicle system, power supply system, electromechanical system, civil construction facility system, line system and signal system. The format of the standardized template for urban rail transit operation accident report is shown in Table 4: Table 4 Standardized template for accident reporting in urban rail transit operations 3. The method according to claim 2, characterized in that The method of performing normalization preprocessing on the accident text data based on the accident report normalization template includes: Based on the standardized template of accident reports for urban rail transit operations, the format of accident text data is consistent, some incomplete and missing parts of the original accident text data are completed, all accident text data are checked, purely descriptive sentences are simplified, and informal terms are standardized.

4. The method according to claim 3, characterized in that The said semantic co-occurrence analysis is performed on the accident text data after normalization preprocessing based on the said safety professional vocabulary, and a framework word set is established according to the analysis results, including: The Jaccard similarity coefficient is used as an indicator to characterize the degree of word co-occurrence in semantics. Given two sets A and B, the Jaccard coefficient is defined as the ratio of the size of the intersection of A and B to the size of the union of A and B, expressed as follows: Where J(A,B) is the Jaccard coefficient value, J(A,B)∈[0,1]. When both sets A and B are empty, J(A,B) is defined as 1. The co-occurring word pairs are determined based on the calculation results of word frequency and similarity. It is stipulated that when the number of word occurrences is greater than 30 and must appear in the top 1% of the total number of cases, and the Jaccard coefficient of the word connection is greater than 0.2, the word is included in the framework word set. KH coder software is used for coding and visualization, and semantic co-occurrence analysis is performed on safety accident cases in urban rail transit system operations. Based on the obtained word analysis results, five framework word sets are established, namely, personnel, equipment and environment, physical process and participant activity, technical measures and management methods, and basic information description, which are expressed by formulas (2), (3), (4), (5), and (6) respectively. V H ={v H1 ,v H2 ,v H3 ,...,v Hn } (2) V E ={v E1 ,v E2 ,v E3 ,...,v Em } (3) V T ={v T1 ,v T2 ,v T3 ,...,v Tj } (5) V B ={v B1 ,v B2 ,v B3 ,...,v Bq } (6) V H ——A set of framework words for personnel; V E ——A collection of framework terms for equipment and environment; V P - A set of framework words for physical processes and participant activities; V T ——A set of framework terms for technical measures and management methods; V B ——A set of basic information description framework words; n——the total number of words in the personnel-related framework word set; m——the total number of words in the equipment and environment framework word set; i——the total number of words in the frame word set of physical process and participant activity; j——the total number of words in the technical measures and management methods framework word set; q——The total number of words in the basic information description framework word set.

5. The method according to claim 4, characterized in that The TF-IDF algorithm based on term frequency-inverse text frequency index characterizes the importance of the framework word set, and the coefficient of variation weighting method is used to weight the framework word set, including: The coefficient of variation weighting method is used to weight each frame word set, and the weight of each frame word set is used as the feature value of the full text of the accident from a semantic perspective. The feature value includes word frequency TF, inverse document frequency IDF, word frequency-inverse document frequency TF-IDF and coefficient of variation and weight; The word frequency TF is used to measure the frequency of the representative words in the entire accident text. The normalization of the word frequency TF is completed using the following formula (7): In the formula ——Word v x The frequency of words; ——Word v x In the accident text d w The number of times it appears in ——Accident textd w The total number of all participles in ; The inverse document frequency (IDF) is a characteristic parameter used to measure the general importance of keywords, and the denominator is added with 1. In the formula ——Word v x The full-text inverse document frequency of |D|——The total number of all accident texts; 1+|{k:v x ∈d k }|——contains the word v x The total number of accident texts; TF-IDF is used to calculate the comprehensive frequency of a word in the full text, reflecting the global importance of this keyword in the accident text set. The higher the weight value, the higher the importance of the word in the accident text. The calculation method is as follows (9): In the formula ——Keyword v x The full-text word frequency of the accident text corpus; ——Keyword v x The full-text inverse document frequency of TF-IDF(v x )——Keyword v x The term frequency-inverse document frequency value in the accident text set; Each indicator is weighted according to the degree of variation between the current value and the specified value of each evaluation indicator. When the difference between the current value of each indicator and the target value is large, it means that it is difficult for the indicator to achieve the target value and a larger weight should be given. Otherwise, a smaller weight should be given. Taking the personnel-based frame word set as an example, the coefficient of variation calculation formula is formula (10), and the corresponding personnel sub-frame weight calculation formula is formula (11); Where C(V H )——the coefficient of variation of the set of personnel-related frame words; σ(V H )——Standard deviation of TF-IDF value of each word in the personnel-related framework word set; ——The average TF-IDF value of each word in the personnel-related framework word set; W(V H )——the weight value of the staff subframe of the semantic network framework; According to the above method, for the full text semantics, the weight values ​​of the frame word sets of the accident text data corresponding to each semantic network are set to 0.213, 0.206, 0.167, 0.214, and 0.200 respectively. The detailed calculation results are shown in Table 5: Table 5 Calculation results of subframe weights of semantic network framework 6. The method according to claim 5, characterized in that The method of extracting keywords from accident text data based on the BTM topic model and calculating the probability of keywords based on the weights of the framework word set includes: Textrank is used to extract the segmented summaries of the long text in the accident text data to obtain the short text. The Jieba word segmentation tool is used to segment the extracted short text. The BTM topic model is used to extract keywords from the short text after word segmentation. The extracted keywords are determined as some words under each topic distribution in the BTM model operation results. These words together constitute the initial word set, and the full-text word topic probability of the words under each topic is calculated. The full-text word topic probability is defined as the probability w│z of word w in topic z and the probability θ of topic z Z The product of .

7. The method according to claim 6, characterized in that The step of calculating the importance of the words in the framework word set in the accident text includes: Each extracted word is regarded as a morpheme, and the importance of a word in the full text of the accident is defined as the word importance, which is expressed by formula (18): In the formula, score(v x ,d)——morpheme v x The importance of the accident in the full text; p(s i |d)——a certain category of text set s i The similarity with the full text of the accident, s i ∈d; n i ×if(v x ins i )——Text Sets i Contains morphemes v x If a certain category of text set s i Contains morphemes v x ; m——the total number of category text sets, 0 <i≤m; According to the accident type, the full text of the accident is divided into multiple accident category text sets A i , i = 1, 2, ... 10, use the vector space model method to obtain the similarity between the accident category text set and the accident full text text, convert the similarity comparison between texts into the similarity comparison between vectors, assuming that the two texts d a , d b The eigenvectors of are And the angle between the vectors is ω, then use d a , d b The cosine coefficient value of the angle between represents the similarity between texts, and the cosine value is positively correlated with the similarity between texts; Based on the previous results, construct a semantic feature vector of the full text of the accident 0.214, 0.200), for text category s i And feature t, the mutual information calculation method is as follows: Where MI(t,s i )——text category s i and the mutual information of feature t; P(t)——the probability of feature t appearing in the corpus, which refers to the ratio of the total number of texts containing word t to the total number of texts in the full text of the accident; P(s i )——Category is s i The probability of a text appearing in all text sets is s i The ratio of the total number of texts in a category to the total number of texts in all categories; P(t,s i )——The text contains both the feature word t and belongs to the category s i The probability that it contains both the feature word t and belongs to the category s i The ratio of the total number of texts in the accident to the total number of texts in the full text of the accident; Based on the above method, after removing outliers, the words in the framework word set and the text set A of each accident category are calculated. i The mutual information of each frame set V is then H 、V E 、V P 、V T 、V B , find the text set A of each accident category i The mean of the mutual information between the words MI(V H , A i )、MI(V E , A i )、MI(V P , A i )、MI(V T , A i )、MI(V B , A i ), then, for each accident category text set A i , i = 1, 2, ... 10, sort out the above-obtained means and normalize them respectively. The normalization formula is shown in formula (21): In the formula, x is a certain mean value; x′——normalized value corresponding to the mean; X——the set of all means of the accident category text set corresponding to the mean; Get the feature vector of the semantic-based accident category text set A1 As shown in formula (22): Combined with the calculation method, the similarity between each accident category text set and the full text of the accident is obtained from the semantic perspective. The value of a word v x The importance score (v x , d).

8. The method according to claim 7, characterized in that The method of using the product of the probability of the key words and the importance of the words in the accident text as the risk judgment index of the words, and performing comprehensive identification of the operational risk points of the urban rail transit system based on the safety professional vocabulary of urban rail transit operation and the risk judgment index of the words, includes: The product of the full-text word topic probability and word importance is used as the word risk judgment indicator, and based on the constructed safety professional word library of urban rail transit operation, a comprehensive identification method for the operational risk points of the urban rail transit system is proposed. The process is shown in Steps 1 to 8, where the α value of the BTM model is set to d = 50 / T, T is the number of topics, the β value is taken as 0.01 according to general experience, and the number of iterations is set to 100 by default; Step 1: Import the text segmentation results, enter the number of topics T = 3, and start the BTM model test iteration. Calculate the model results and output the perplexity (D test )3; Step 2: Repeat Step 1 and output the perplexity (D) of the model under different test conditions of the number of topics T. test ) T , T>3 and is an integer; Step 3: Draw the perplexity curve under different number of topics T, and determine the number of topics T=K corresponding to the minimum value within the reasonable fitting range; Step 4: Output the BTM model running results under the condition of T = K, calculate the full-text word topic probability value of the words under each topic, sort the full-text word topic probability values ​​under each topic respectively, take the top M1 words under each topic, and together form the initial word set λ1; Step 5: Traverse the professional vocabulary of urban rail transit operation safety, filter out all words in λ1 that are synonymous or strongly related to a part of the professional vocabulary, merge the synonyms, and output the related word set λ2; Step 6: Calculate the word importance and risk judgment index value of each word in λ2 in turn, sort the risk judgment index values ​​of all words in λ2 and take the top M2 words to form the candidate word set λ3; Step 7: Compare with the professional vocabulary, and inquire with multiple industry experts based on the output of λ3, integrate various descriptive information about the same risk point and eliminate redundancy, so as to form risk points that correspond to the professional vocabulary and meet the actual situation; Step 8: The risk value of a risk point is calculated as the sum of the risk judgment index values ​​of all words associated with it in λ3. All risk points are sorted by risk value, and the process is terminated and output.

Images