An Automatic Method for Mining Cause Information of TBM Card Machines

By introducing physical mechanism constraints through an improved LDA topic model, the system automatically mines information on the causes of TBM card jamming, solving the problems of data dispersion and strong subjectivity in existing technologies, and achieving accurate identification of card jamming causes and the formation of structured data.

CN123116196APending Publication Date: 2026-11-17YAXIA NATIONAL HYDROPOWER TECHNOLOGY INNOVATION CENTER CO LTD +3
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611370024.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-04
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing TBM (Total Machine Detection) analysis methods rely on manual summarization and general text mining, resulting in scattered data, inconsistent formats, difficulty in accurately identifying the causes of machine malfunctions caused by multiple coupled factors, lack of physical mechanism constraints, and a high degree of subjectivity and misjudgment.

Method used

An improved LDA topic model is adopted, which introduces physical mechanism constraints, including asymmetric priors, order constraints, and word pair co-occurrence potential functions. By combining the mechanism constraint matrix with the keyword statistical correlation matrix, the causal information of TBM card machines is automatically mined.

Benefits of technology

It improves the accuracy and causal consistency of card-related causal subject identification, generates structured case data, and supports risk assessment and decision-making.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123116196A_ABST
    Figure CN123116196A_ABST
Patent Text Reader

Abstract

This invention discloses an automatic method for mining information on the causes of TBM (Tunnel Boring Machine) jams, relating to the field of tunnel excavation construction safety technology. The method includes: preprocessing case text to identify keywords such as geological anomalies, construction conditions, equipment response, abnormal manifestations, and jamming consequences; identifying the context of the jamming causes, and performing staged expression and context enhancement on the keywords; constructing an improved LDA (Local Dependent Aspect) model, introducing physical mechanism constraints, including asymmetric priors, order constraints, and word pair co-occurrence potential functions, to obtain initial topics, keyword weights, and document topic sequences; fusing the mechanism constraint matrix and the keyword statistical correlation matrix to obtain a modified correlation matrix; reconstructing keyword weights to identify core disaster-causing factors; mapping the core factors back to the original text, extracting structured information, and forming case data. This invention can improve the accuracy, engineering rationality, causal consistency, and interpretability of jamming cause topic identification.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of tunnel excavation safety technology, and in particular to an automatic method for mining TBM (Tunnel Boring Machine) causal information. Background Technology

[0002] Tunnel boring machines (TBMs) are widely used in construction scenarios such as railway tunnels, water conservancy tunnels, urban underground engineering, and mine roadways. During TBM advancement under complex geological conditions, they are susceptible to adverse geological conditions such as fault fracture zones, water-rich strata, high ground stress, weak surrounding rock, boulders, and rock bursts. They are also affected by factors such as equipment condition fluctuations, mismatched construction parameters, and differences in on-site handling measures, leading to difficulties in advancement, abnormal cutterhead torque, shield obstruction, and even machine jamming. TBM jamming not only causes construction interruptions, project delays, and economic losses, but can also lead to equipment damage, attitude instability, and secondary safety risks. Therefore, accurate identification and effective analysis of the causes of TBM jamming are of significant engineering importance.

[0003] Existing methods for analyzing TBM jamming events primarily rely on manual review of construction logs, project reports, accident analyses, technical summaries, and handling records to summarize experiences and categorize causes. Some studies have attempted to organize jamming cases using keyword statistics, case classification, or traditional text mining methods. However, the relevant data is usually stored in unstructured text formats, and different projects exhibit significant differences in terminology, recording methods, and descriptive granularity, making direct comparison and unified analysis of case data difficult. Furthermore, the causes of jamming often involve multiple coupled factors, including geological conditions, equipment status, construction conditions, abnormal manifestations, and final consequences. Relying solely on manual summarization, keyword matching, or statistical topic modeling makes it difficult to distinguish between statistical co-occurrence relationships and engineering causal relationships among factors, and also makes it difficult to reveal the causal direction, evolutionary sequence, and primary and secondary mechanisms of different factors.

[0004] Existing general text mining methods typically focus on extracting high-frequency words, feature words, or topic distributions from case texts. The results mainly reflect statistical correlations at the text level, lacking constraints related to the physical mechanisms of TBM (Total Machine Machine) cassettes. When certain anomalous or consequential words frequently appear in case texts, they are easily misidentified as core causative factors, leading to inconsistencies between the identification results and the actual evolutionary mechanisms of cassette machines. Furthermore, relying solely on human experience to judge the strength of factor relationships suffers from high subjectivity, insufficient stability, and difficulty in dynamically updating with new case data.

[0005] In summary, existing technologies have at least the following shortcomings: TBM jamming case texts are scattered, inconsistent in format, and highly unstructured, making it difficult to form case data that can be directly calculated and compared; manual analysis relies on expert experience, resulting in low processing efficiency, strong subjectivity, and difficulty in adapting to the continuous updating of a large number of historical cases; general text mining methods mainly rely on word frequency and co-occurrence relationships, lacking constraints on the physical evolution sequence between geological causes, equipment response, abnormal performance, and jamming consequences, which easily leads to topic mixing, stage mismatch, and misjudgment of causal relationships; standard LDA and other topic models usually adopt symmetric priors and independent topic assignment assumptions, which cannot introduce jamming mechanism knowledge before modeling, and it is also difficult to constrain the topic sequence within the document to conform to the causal chain from cause to consequence of jamming. Summary of the Invention

[0006] The purpose of this invention is to provide an automatic method for mining TBM card cause information, which aims to solve or improve at least one of the above-mentioned technical problems.

[0007] To achieve the above objectives, the present invention provides the following solution: A method for automatically mining TBM card cause information includes: Step 1: Obtain relevant case texts of TBM jamming, preprocess the case texts and identify domain keywords to obtain a set of domain keywords related to geological anomalies, construction conditions, equipment response, abnormal performance and jamming consequences; Step 2: Identify the contextual causes of the card jam in the case text related to the card jam occurrence process, and perform staged expression and context enhancement processing on the keywords in the domain keyword set according to the engineering category and card jam evolution stage of the keywords to obtain the processed keyword set; Step 3: Based on the processed keyword set, construct an improved LDA topic model to perform topic modeling on the TBM card machine case text, and obtain the initial card machine cause topic, topic keyword weights, and document topic sequence; the improved LDA topic model introduces TBM card machine physical mechanism constraints, including one or more of the following: physical guidance asymmetric prior, topic allocation order constraints, and word pair co-occurrence potential function; Step 4: Construct a mechanism constraint matrix based on historical TBM card machine case evidence, and fuse the mechanism constraint matrix with the keyword statistical association matrix constructed from the topic keyword weights to obtain a modified association matrix; Step 5: Reconstruct the topic keyword weights based on the modified correlation matrix to obtain the modified topic keyword weights, and identify the core disaster-causing factors based on the modified topic keyword weights; Step Six: Map the core disaster-causing factors and their related contexts back to the original case text, extract the corresponding geological anomalies, construction conditions, equipment responses, abnormal manifestations, and machine jamming consequences information to form structured TBM jamming case data.

[0008] Optionally, the physically guided asymmetric prior is used to replace the symmetric Dirichlet prior in standard LDA; Constructing a priori matrix of keywords ,in β k,i Keywords i On the topic k The prior weights of the topic k Corresponding to a certain stage of card machine evolution: When keywords i With the topic k When the corresponding card machine evolution stage is matched, β k,i Take the larger value; when the keyword i With the topic k When the corresponding card evolution stage does not match, β k,i Take the smaller value.

[0009] Optionally, the order constraint of topic assignment is used to introduce a hidden Markov unidirectional topic transfer mechanism in the improved LDA topic model: Constructing a topic transition matrix ,in Indicates the topic u Shift to topic v The degree of permissibility, and satisfying: when v = u or v = u When +1, A u,v For transfer to be permitted; when v < u hour, A u,v To prohibit or punish the transfer, when v > u When +1, A u,v This is a jump to a different stage.

[0010] The topic transition matrix A This is used to constrain the topic allocation of subsequent words within the same case text, so that the topic sequence can only remain in the current physical stage or advance to the next physical stage, thereby making the document topic sequence conform to the unidirectional evolution chain of TBM jamming from geological causes to jamming consequences.

[0011] Optionally, the word pair co-occurrence potential function is used to reward or penalize whether the keyword combination conforms to the physical mechanism of the TBM card during the topic update process of Gibbs sampling; Constructing word pair potential functions , indicating word pairs ( w j , w n In the candidate topics k The rationality of the underlying mechanism; among them, This refers to the words within the current word context window. Indicates the current word; When a word pair conforms to the physical mechanism of the TBM card, the potential function takes a positive value; when a word pair violates the physical mechanism, the potential function takes a negative value; when there is no clear mechanism relationship between word pairs, the potential function takes zero or is close to zero. The word pair potential function As a reward or penalty, it is input into the topic update probability formula of Gibbs sampling to affect the topic assignment result of the current word.

[0012] Optionally, the improved LDA topic model uses Gibbs sampling for parameter inference, and the topic update probability formula of the model comprehensively considers text statistical features, the physically guided asymmetric prior, the order constraint of topic assignment, and the word pair co-occurrence potential function: The formula for the topic update probability is: in, Document d The Middle n The theme of the word, This indicates the document after removing the current word. d Assigned to topic k word count, α k Indicates the prior parameters of the document topic. Indicates the topic after removing the current word. k Next word w The number of times it appears, β k,w This represents the asymmetric prior guided by physics. This refers to any term in the thesaurus. This indicates that the current word belongs to the candidate topic. k The probability is directly proportional to the combined effect of all influencing factors. A zd,n-1,k This indicates the order constraint from the previous topic to the current candidate topic. Oh The co-occurrence potential function of word pairs.

[0013] Optionally, the specific processing procedure of step four includes: Based on the logic of rock-machine interaction, surrounding rock instability, and construction disturbance, a TBM jack mechanism constraint diagram G=( V , E ),in V Represents a set of domain keyword nodes. E Represents the set of candidate directed causal relations; For the set of candidate directed causal relationships E Each candidate causal relation keyword in the list i To keywords j Extracting supporting evidence from historical TBM card machine case texts S ij Non-supporting evidence quantity F ij ; Based on the Bayesian update method, according to the amount of supporting evidence S ij Non-supporting evidence quantity F ij Calculate the posterior mechanism correlation strength The mechanism constraint matrix is ​​formed. Among them, when a candidate directed causal relationship exists, ,otherwise .

[0014] Optionally, the keywords in the keyword statistical association matrix i With keywords j The statistical association strength is expressed as: in, Keywords i On the topic k The weights below, K Indicates the number of topics; Any element in the modified correlation matrix is ​​obtained by fusing the statistical correlation matrix and the mechanistic constraint matrix: in, Let λ be an element in the mechanism constraint matrix, λ be the fusion coefficient, and 0 ≤ λ ≤ 1.

[0015] Optionally, the specific process of reconstructing the weights of the topic keywords is as follows: For the topic k Keywords below i Based on the modified correlation matrix, the intermediate modified weights are calculated using the following formula: in, or This is the weighting adjustment factor; Then, the intermediate corrected weights are normalized to obtain the corrected topic keyword weights: in, N Indicates the number of keywords.

[0016] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects: This invention discloses an automatic method for mining information on the causes of TBM (Turbine Machine) jams. The method includes: preprocessing case text; identifying keywords such as geological anomalies, construction conditions, equipment response, abnormal manifestations, and jam consequences; identifying the context of the jam causes, and performing staged expression and context enhancement on the keywords; constructing an improved LDA (Local Least Accuracy Analysis) model, introducing physical mechanism constraints, including asymmetric priors, order constraints, and word pair co-occurrence potential functions, to obtain initial topics, keyword weights, and document topic sequences; fusing the mechanism constraint matrix and the keyword statistical correlation matrix to obtain a modified correlation matrix; reconstructing keyword weights to identify core disaster-causing factors; mapping the core factors back to the original text, extracting structured information, and forming case data. This invention can improve the accuracy, engineering rationality, causal consistency, and interpretability of jam cause topic identification, providing support for jam risk assessment, case management, and early warning decision-making. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of the automatic mining method for TBM card machine cause information of the present invention; Figure 2 This is a schematic diagram of the physical evolution chain of the TBM card machine and the unidirectional sequential constraints of the theme in this embodiment; Figure 3 This is a schematic diagram of Gibbs sampling update in this embodiment, which integrates physical-guided asymmetric priors and word pair co-occurrence potential functions. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] The purpose of this invention is to provide an automatic mining method for TBM card cause information based on physical constraint modeling, which aims to solve or improve at least one of the above-mentioned technical problems.

[0021] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0022] like Figure 1-Figure 3 As shown, this invention provides a method for automatically mining TBM card cause information, including: Step 1: Obtain relevant case texts of TBM jamming, preprocess the case texts and identify domain keywords to obtain a set of domain keywords related to geological anomalies, construction conditions, equipment response, abnormal performance and jamming consequences.

[0023] Step 2: Identify the contextual causes of the card jam in the case text related to the card jam occurrence process, and perform staged expression and context enhancement processing on the keywords in the domain keyword set according to the engineering category and card jam evolution stage of the keywords, to obtain the processed keyword set.

[0024] Step 3: Construct an improved LDA topic model based on the processed keyword set, perform topic modeling on the TBM card machine case text, and obtain the initial card machine causal topic, topic keyword weights, and document topic sequence; the improved LDA topic model introduces TBM card machine physical mechanism constraints, including one or more of the following: physical-guided asymmetric prior, topic assignment order constraints, and word pair co-occurrence potential function.

[0025] Step 4: Construct a mechanism constraint matrix based on historical TBM card machine case evidence, and then merge the mechanism constraint matrix with the keyword statistical association matrix constructed from the topic keyword weights to obtain a modified association matrix.

[0026] Step 5: Reconstruct the topic keyword weights based on the modified correlation matrix to obtain the modified topic keyword weights, and identify the core disaster-causing factors based on the modified topic keyword weights.

[0027] Step Six: Map the core disaster-causing factors and their related contexts back to the original case text, extract the corresponding geological anomalies, construction conditions, equipment responses, abnormal manifestations, and machine jamming consequences information to form structured TBM jamming case data.

[0028] In the above steps, the physically guided asymmetric prior is used to replace the symmetric Dirichlet prior in standard LDA. Different importance weights for keywords are pre-set for different stages of TBM jamming evolution corresponding to different themes, giving the model a stage-specific word preference before data learning. For example, the geological anomaly stage tends to include keywords such as fault, fracture zone, and water-rich strata; the equipment response stage tends to include keywords such as shield obstruction, abnormal cutterhead torque, and sudden thrust increase; and the jamming consequence stage tends to include keywords such as jamming, downtime, and project delay. This makes the theme identification results more consistent with the physical evolution process of TBM jamming from cause to consequence.

[0029] The order constraint of topic assignment is used to change the assumption in standard LDA that word topics are assigned independently. By introducing a hidden Markov unidirectional topic transfer mechanism, the topic of subsequent words within the same case text can only remain in the current physical stage or advance to the next physical stage, but cannot regress. This ensures that the topic sequence within the document conforms to the causal chain of TBM jamming from geological causes, equipment response, abnormal behavior to jamming consequences, avoiding reversed evolutionary order or stage mismatch in topic identification results.

[0030] The word pair co-occurrence potential function is used in the topic update process incorporating Gibbs sampling. By enhancing or suppressing the physical mechanism relationship between keywords, the model tends to generate keyword combinations that conform to the TBM lag mechanism during each topic assignment. For word pairs that conform to the physical mechanism, such as "fault" pointing to "shield obstruction," a positive reward is given; for word pairs that violate the mechanism, such as "fault" and "stuck" directly coexisting in the same topic without intermediate device response or abnormal behavior transition, a penalty is imposed. This reduces erroneous topic aggregation caused by simple text co-occurrence and improves the engineering rationality of topics causing lag.

[0031] Furthermore, based on the interaction between the rock and the machine, the instability of the surrounding rock, and the construction disturbance logic, this invention constructs a TBM jamming mechanism constraint diagram to limit the possible directional causal relationships between keywords. Based on supporting and non-supporting evidence from historical TBM jamming cases, a Bayesian update method is used to calculate the mechanistic association strength of candidate directional causal relationships, forming a mechanism constraint matrix. The mechanism constraint matrix is ​​then fused with the keyword statistical association matrix to correct the topic keyword weights and keyword association relationships, resulting in a corrected jamming causal topic.

[0032] Based on the revised TBM jamming cause topic, the core disaster-causing factors are identified, and the core disaster-causing factors and their related contexts are mapped back to the original case text. The corresponding geological anomalies, construction conditions, equipment responses, abnormal manifestations, and jamming consequences information are extracted to form structured TBM jamming case data.

[0033] Based on the above technical solution, the following embodiments are provided.

[0034] First, relevant case texts of TBM card machines are obtained to construct an original case corpus. The case texts may include one or more of the following: construction logs, engineering reports, accident analysis reports, technical summaries, monitoring records, handling records, and publicly available documents.

[0035] Let the original case corpus be: in, Indicates the first M Each TBM (Tunnel Boring Machine) case text is preprocessed, including text cleaning, Chinese word segmentation, domain terminology recognition, synonym normalization, and stop word filtering. Domain terminology recognition identifies TBM engineering terms such as "fault fracture zone," "water-rich strata," "high ground stress," "shield obstruction," "abnormal cutterhead torque," "sudden thrust increase," "decreased penetration," "jamming," and "shutdown procedures" as complete semantic units, avoiding semantic fragmentation caused by ordinary word segmentation.

[0036] After preprocessing, the following set of domain keywords is obtained: in, w i Indicates the first i Keywords for each field N Indicates the number of keywords.

[0037] After obtaining the domain keyword set, the contextual causes of TBM jamming related to the TBM jamming process are identified in the case text. The contextual causes of jamming include text fragments describing geological anomalies, construction disturbances, equipment responses, abnormal manifestations, jamming occurrences, and consequences of handling.

[0038] As one implementation method, the context causing the machine jam can be identified based on trigger words or trigger expressions. These trigger expressions include, but are not limited to, "machine jam," "stuck," "difficulty in propulsion," "shield obstruction," "abnormal cutterhead torque," "sudden increase in thrust," "decreased penetration," "surrounding rock compression," "collapse," and "shutdown procedures."

[0039] Furthermore, based on the physical evolution process of the TBM card reader, the keywords are divided into different card reader evolution stages. As an example, the card reader evolution stages include: in, S 1 Indicates a stage of geological anomaly. S 2 Indicates the device response phase. S 3 This indicates a stage of abnormal behavior. S 4 This indicates the stage where the machine freezes.

[0040] For example, "fault", "fracture zone", "water-rich strata" and "high ground stress" can be marked as keywords for the geological anomaly stage; "shield obstruction", "abnormal cutterhead torque" and "sudden increase in thrust" can be marked as keywords for the equipment response stage; "difficulty in propulsion", "decreased penetration" and "abnormal attitude" can be marked as keywords for the abnormal manifestation stage; "jamming", "shutdown", "delay in construction period" and "equipment damage" can be marked as keywords for the consequences of jamming stage.

[0041] Keywords appearing in the context of the cause of the card machine can have their weight increased in topic modeling; keywords not appearing in the context of the cause of the card machine can have their original weight maintained or their weight reduced to reduce the interference of irrelevant descriptions on the topic identification results.

[0042] In standard LDA models, topic word distribution typically employs a symmetric Dirichlet prior, meaning different words within the same topic have the same or nearly the same prior weights. The model primarily learns topic structure based on word frequency and co-occurrence relationships in the data. This approach cannot incorporate knowledge of the physical stage of the TBM (Toyota Machine Tool) before modeling, easily leading to a mixture of geological trigger words, equipment response words, abnormal behavior words, and consequence words within the same topic.

[0043] This implementation replaces the symmetric Dirichlet prior in standard LDA with a physically guided asymmetric prior. Specifically, a keyword prior matrix is ​​constructed: in, β k,i Keywords w i On the topic k Prior weights. Topic k This corresponds to a specific stage in the evolution of card machines.

[0044] When keywords w i With the topic k When the corresponding card machine evolution stages are consistent, a larger prior weight is set; when keywords w i With the topic kWhen the corresponding card evolution stages are inconsistent, a smaller prior weight should be set. As one implementation method, it can be set as follows: in, β + > β - >0, S k Indicates the topic k The set of keywords corresponding to the stage.

[0045] For example, when topic k is a geological anomaly topic, terms such as "fault," "breccia zone," and "water-rich strata" correspond to a relatively large number of such terms. The consequences of words like "stuck" and "shutdown" correspond to smaller ones. Therefore, the model can receive physical phase guidance before learning the data, reducing topic phase mismatch. Standard LDA typically assumes that the topic assignments of different words within the same document are independent, without considering the order of words in the document. However, TBM jamming events have a clear physical evolution process, usually manifesting as a geological anomaly or construction disturbance triggering a machine response, which then develops into an abnormal behavior, ultimately resulting in a jam. Therefore, without constraining the topic sequence, the model may produce results that do not conform to engineering logic, such as consequence words preceding triggering words or equipment response words regressing to geological stages.

[0046] This implementation introduces a sequence constraint for topic assignment and employs a Hidden Markov Model unidirectional topic transfer mechanism. Let the document... d The Middle n The theme of each word is z d,n Then, the theme of subsequent words is only allowed to remain at the current stage or move forward to the next stage, and is not allowed to go back.

[0047] Constructing a topic transition matrix: in, A u,v Indicates the topic u Shift to topic v The degree of permissibility. As one implementation method: Here, ϵ is a small non-negative parameter used to represent the weak permissibility of jumping between stages. If strict constraints are required on the card machine evolution process, ϵ can be set to 0, so that the topic sequence can only remain in the current stage or advance to the next stage.

[0048] With the above settings, the thematic sequence in the same case text can form a unidirectional evolution chain from geological anomalies, equipment responses, abnormal manifestations to machine jamming consequences.

[0049] In the topic inference process, this implementation further adds a word pair co-occurrence potential function to the topic update formula of Gibbs sampling, which is used to reward or punish whether the keyword combination conforms to the physical mechanism of TBM card machine.

[0050] Let the current word of the topic to be assigned be... w n The words in its context window are w j Then construct the word pair potential function: in, Oh Indicates word pairs ( w j , w n In candidate topics k The rationality of the underlying mechanism.

[0051] As one implementation method, the potential function takes a positive value when the word pair conforms to the physical mechanism of the TBM card machine; the potential function takes a negative value when the word pair violates the physical mechanism; and the potential function takes zero or close to zero when there is no clear mechanism relationship between the word pairs.

[0052] For example, the word pair "fault" pointing to "shield obstruction" can be set as a positive reward because it conforms to the physical mechanism of fault fracture zones causing surrounding rock compression and shield obstruction. Oh (Fault line, shield blocked) k For combinations where "fault" and "stuck" directly coexist on the same topic and lack a transition in device response or abnormal behavior, a penalty can be set to 0 because it skips the intermediate physical process. Oh (Fault, stuck) k <0; After comprehensively considering asymmetric priors, order constraints, and word pair potential functions, the topic update probability of Gibbs sampling can be expressed as: in, Document d The Middle n The theme of the word, This indicates the document after removing the current word. d Assigned to topic k word count, α k Indicates the prior parameters of the document topic. Indicates the topic after removing the current word. k Next word w The number of times it appears,β k,w This represents the asymmetric prior guided by physics. This refers to any term in the thesaurus. This indicates that the current word belongs to the candidate topic. k The probability is directly proportional to the combined effect of all influencing factors. A zd,n-1,k This indicates the order constraint from the previous topic to the current candidate topic. Oh The co-occurrence potential function of word pairs.

[0053] Therefore, during each topic sampling, the model not only considers the statistical features of the text, but also the physical stage attributes of the keywords, the evolution direction of the topic sequence, and the rationality of the mechanism of word pair combination.

[0054] In this implementation, the physical-guided asymmetric prior, the order constraint of topic assignment, and the word pair co-occurrence potential function can be implemented independently or in combination.

[0055] When only physical-guided asymmetric priors are used, the model is mainly used to improve the keyword discrimination of topics at different card machine evolution stages.

[0056] When only the order constraint of topic assignment is used, the model is mainly used to ensure that the topic sequence within the document conforms to the unidirectional evolution process from cause to effect of the TBM card machine.

[0057] When only the word pair co-occurrence potential function is used, the model is mainly used to enhance keyword combinations that conform to the mechanism and suppress keyword combinations that do not conform to the mechanism.

[0058] When these three elements are combined, the model simultaneously possesses the capabilities of stage prior guidance, topic sequence constraints, and word pair mechanism correction, thereby obtaining cardiogenic topics that are more consistent with the physical laws of TBM engineering.

[0059] After obtaining the initial cause of the TBM jamming, a TBM jamming mechanism constraint diagram is further constructed based on the rock-machine interaction, surrounding rock instability, and construction disturbance logic: G=( V , E )in, V Represents a set of domain keyword nodes. E This represents the set of candidate directed causal relations.

[0060] The candidate directional causal relationships are determined according to the evolution sequence of TBM jamming. For example: geological anomaly to equipment response; equipment response to anomaly manifestation; anomaly manifestation to jamming consequences; construction conditions to equipment response; construction conditions to anomaly manifestation.

[0061] Specifically, when there is a physical mechanism connection between "fault fracture zone" and "shield obstruction", a candidate directed relationship can be constructed: fault fracture zone to shield obstruction. When "shield obstruction" further leads to "difficulty in advancement", a candidate directed relationship can be constructed: shield obstruction to difficulty in advancement.

[0062] For each candidate directed causal relationship i to j We extracted supporting and non-supporting evidence from historical TBM card machine case texts.

[0063] Supporting evidence refers to evidence that demonstrates the keywords i Keywords j Textual expressions that have a causal or contributing effect. For example, "Affected by the fault fracture zone, the shield is squeezed and has difficulty advancing" can serve as supporting evidence for "the fault fracture zone hinders the shield" or "the fault fracture zone leads to difficulty in advancing".

[0064] Non-supporting evidence refers to text content that only shows word co-occurrence but lacks a clear causal direction, or whose expression direction does not conform to the evolutionary logic of TBM card machines.

[0065] Suppose a candidate directed relation i to j The relationship strength parameter is i ij Its prior distribution is: Weighted supporting evidence based on historical case texts S ij and weighted non-supporting evidence quantity F ij Then the mechanistic correlation strength after Bayesian update is: in, S ij and F ij The credibility of the evidence can be weighted according to its credibility. The credibility of the evidence can be determined by a combination of factors, including the co-occurrence distance of keywords, the clarity of causal expression, whether it is located in the context of the cause of the card machine, and whether it conforms to the evolutionary sequence of the card machine.

[0066] Based on the posterior mechanism correlation strength of each candidate directed causal relationship, a mechanism constraint matrix is ​​formed: Among them, when keywords i To keywords j When candidate directed causal relations exist: When keywords iTo keywords j When no candidate directed causal relationship exists: Simultaneously, based on the keyword weights obtained from topic modeling, a keyword statistical association matrix is ​​constructed: As one implementation method, keywords i With keywords j The statistical association strength can be calculated as follows: in, p k,i Keywords i On the topic k The weights below, K Indicates the number of topics.

[0067] Statistical correlation matrix C Mechanism constraint matrix T By merging, the corrected correlation matrix is ​​obtained: Among them, the corrected correlation matrix R any element R ij satisfy: Therefore, the corrected correlation matrix R can be further expanded as follows: λ is the fusion coefficient, and 0 ≤ λ ≤ 1. When λ is large, the model emphasizes textual statistical correlation; when λ is small, the model emphasizes engineering mechanism constraints.

[0068] After obtaining the corrected correlation matrix R Next, the weights of the topic keywords are restructured. For a given topic... k Keywords below i Based on the modified correlation strength between it and other keywords within the same topic, its initial weight is enhanced or suppressed.

[0069] As one implementation method, the intermediate correction weight can be expressed as: in, or This represents the weighting adjustment coefficient; then normalization is performed. in, This represents the corrected topic keyword weights. Based on the corrected topic keyword weights, core disaster-causing factors are identified. These core disaster-causing factors can be determined comprehensively based on keyword weights, mechanistic correlation strength, node centrality, and their frequency of occurrence in the context of the card-causing factors.

[0070] Finally, the core disaster-causing factors and their related keywords are mapped back to the original case text, and the corresponding structured information is extracted to form structured TBM jamming case data. The structured TBM jamming case data includes geological anomaly information, construction condition information, equipment response information, anomaly performance information, jamming consequences information, core disaster-causing factors, causal chains, and the strength of mechanistic correlation. The geological anomaly information includes information related to surrounding rock conditions and adverse geology, such as fault fracture zones, water-rich strata, high ground stress, and weak surrounding rock; the construction condition information includes information related to construction control, such as mismatched propulsion parameters, delayed support, and improper adjustment of tunneling posture; the equipment response information includes information on changes in the TBM equipment's operating status, such as shield obstruction, abnormal cutterhead torque, and sudden increase in thrust; the anomaly manifestation information includes external manifestations during the jamming process, such as difficulty in propulsion, decreased penetration, and abnormal posture; the jamming consequence information includes final results such as jamming, shutdown, equipment damage, and project delay; the core disaster-causing factors include factors that contribute significantly to jamming, such as fault fracture zones, surrounding rock compression, and shield obstruction; the causal chain is used to represent the evolutionary relationship between core disaster-causing factors, for example, fault fracture zones cause shield obstruction, which in turn leads to difficulty in propulsion and further develops into jamming; the mechanism correlation strength is used to represent the posterior strength value corresponding to each candidate directional causal relationship. Through the above implementation methods, the present invention can automatically identify the causal topics and their evolutionary chains of TBM card machine case texts, so that the identification results simultaneously satisfy the statistical regularity of text, the prior of physical stages, the constraint of topic order and the constraint of word pair mechanism, thereby improving the accuracy, engineering rationality and interpretability of TBM card machine causal information mining.

[0071] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for automatically mining TBM jamming cause information, characterized in that, include: Step 1: Obtain relevant case texts of TBM jamming, preprocess the case texts and identify domain keywords to obtain a set of domain keywords related to geological anomalies, construction conditions, equipment response, abnormal performance and jamming consequences; Step 2: Identify the contextual causes of the card jam in the case text related to the card jam occurrence process, and perform staged expression and context enhancement processing on the keywords in the domain keyword set according to the engineering category and card jam evolution stage of the keywords to obtain the processed keyword set; Step 3: Based on the processed keyword set, construct an improved LDA topic model to perform topic modeling on the TBM card machine case text, and obtain the initial card machine cause topic, topic keyword weights, and document topic sequence; the improved LDA topic model introduces TBM card machine physical mechanism constraints, including one or more of the following: physical guidance asymmetric prior, topic allocation order constraints, and word pair co-occurrence potential function; Step 4: Construct a mechanism constraint matrix based on historical TBM card machine case evidence, and fuse the mechanism constraint matrix with the keyword statistical association matrix constructed from the topic keyword weights to obtain a modified association matrix; Step 5: Reconstruct the topic keyword weights based on the modified correlation matrix to obtain the modified topic keyword weights, and identify the core disaster-causing factors based on the modified topic keyword weights; Step Six: Map the core disaster-causing factors and their related contexts back to the original case text, extract the corresponding geological anomalies, construction conditions, equipment responses, abnormal manifestations, and machine jamming consequences information to form structured TBM jamming case data.

2. The TBM jamming cause information automatic mining method of claim 1, wherein, The physically guided asymmetric prior is used to replace the symmetric Dirichlet prior in standard LDA. Constructing a priori matrix of keywords ,in β k,i Keywords i On the topic k The prior weights of the topic k Corresponding to a certain stage of card machine evolution: when the keyword i corresponds to the subject k of the card machine evolution stage, β k,i the larger value is taken; when the keyword i corresponds to the subject k of the card machine evolution stage, β k,i the smaller value is taken. 3.The TBM jamming cause information automatic mining method according to claim 1, characterized in that, The order constraint of topic assignment is used to introduce a hidden Markov unidirectional topic transfer mechanism in the improved LDA topic model: Constructing a topic transition matrix ,in Indicates the topic u Shift to topic v The degree of permissibility, and satisfying: when v = u or v = u When +1, A u,v For transfer to be permitted; when v < u hour, A u,v To prohibit or punish the transfer, when v > u When +1, A u,v This is a jump to a different stage. The subject transition matrix A To constrain the subject assignment of subsequent words within the same case text, so that the subject sequence can only maintain the current physical stage or advance to the next physical stage, thus making the document subject sequence comply with the one-way evolution chain of TBM carding from geological causes to carding consequences. 4.The TBM jamming cause information automatic mining method according to claim 1, wherein, The word pair co-occurrence potential function is used to reward or penalize whether the keyword combination conforms to the physical mechanism of the TBM card during the topic update process of Gibbs sampling. Constructing word pair potential functions , indicating word pairs ( w j , w n In the candidate topics k The rationality of the underlying mechanism; among them, This refers to the words within the current word context window. Indicates the current word; When a word pair conforms to the physical mechanism of the TBM card, the potential function takes a positive value; when a word pair violates the physical mechanism, the potential function takes a negative value; when there is no clear mechanism relationship between word pairs, the potential function takes zero or is close to zero. The word pair potential function As a reward or penalty term, input into the topic update probability formula of Gibbs sampling, to affect the topic assignment results of the current word. 5.The TBM jamming cause information automatic mining method according to claim 1, wherein, The improved LDA topic model uses Gibbs sampling for parameter inference. The topic update probability formula of the model comprehensively considers text statistical features, the physically guided asymmetric prior, the order constraint of topic assignment, and the word pair co-occurrence potential function. The formula for the topic update probability is: in, Document d The Middle n The theme of the word, This indicates the document after removing the current word. d Assigned to topic k word count, α k Indicates the prior parameters of the document topic. Indicates the topic after removing the current word. k Next word w The number of times it appears, β k,w This represents the asymmetric prior guided by physics. This refers to any term in the thesaurus. This indicates that the current word belongs to the candidate topic. k The probability is directly proportional to the combined effect of all influencing factors. A zd,n-1,k This indicates the order constraint from the previous topic to the current candidate topic. Ω The co-occurrence potential function of word pairs. 6.The TBM jamming cause information automatic mining method according to claim 1, wherein, The specific processing steps in step four include: Based on the logic of rock-machine interaction, surrounding rock instability, and construction disturbance, a TBM jack mechanism constraint diagram G=( V , E ),in V Represents a set of domain keyword nodes. E Represents the set of candidate directed causal relations; For the set of candidate directed causal relationships E Each candidate causal relation keyword in the list i To keywords j Extracting supporting evidence from historical TBM card machine case texts S ij Non-supporting evidence quantity F ij ; Based on the Bayesian update method, according to the amount of supporting evidence S ij Non-supporting evidence quantity F ij Calculate the posterior mechanism correlation strength The mechanism constraint matrix is ​​formed. Among them, when a candidate directed causal relationship exists, ,otherwise .

7. The automatic mining method for TBM card machine cause information according to claim 6, characterized in that, The keyword statistical association matrix contains keywords i With keywords j The statistical association strength is expressed as: wherein, denotes the keyword i under the topic k weight, K denotes the number of topics; Any element in the modified correlation matrix is ​​obtained by fusing the statistical correlation matrix and the mechanistic constraint matrix: wherein, is an element in the mechanism constraint matrix, λ is a fusion coefficient, and 0≤λ≤1. 8.The TBM jamming cause information automatic mining method according to claim 7, characterized in that, The specific process of reconstructing the weights of the topic keywords is as follows: For the keywords k under the topic i , the intermediate correction weight is calculated according to the correction correlation matrix, and the calculation formula is: wherein η is a weight correction factor; Then, the intermediate corrected weights are normalized to obtain the corrected topic keyword weights: wherein, N represents the number of keywords.