A shield risk tracing method based on a knowledge graph

Abstract

The application discloses a shield risk tracing method based on a knowledge graph and belongs to the technical field of shield risk analysis. A mode layer of the knowledge graph is constructed by determining an entity set and a relation set; document data is subjected to data preprocessing to obtain processed text, required knowledge triples are extracted, and structured knowledge triples are output in a predetermined format; normalized knowledge triples are obtained through entity alignment to construct the knowledge graph; the membership of a node is calculated based on the knowledge graph, and the posterior probability of the node is obtained in combination with the prior probability of the node; potential risk propagation paths and relation strengths are inferred based on the posterior probability of the node in the knowledge graph, and dynamic analysis and prediction of the knowledge network are realized. The application realizes the structured association of risk factors through the knowledge graph, quantifies uncertainty in combination with fuzzy Bayesian reasoning, and solves the problems that the traditional method is difficult to handle complex causal relationships and information ambiguity.

Description

Technical Field

[0001] This invention belongs to the technical field of shield tunneling risk analysis, and in particular relates to a shield tunneling risk tracing method based on knowledge graphs. Background Technology

[0002] With the rapid development of artificial intelligence technology, knowledge graphs, as an important tool for information management and reasoning, are widely used in scenarios such as risk identification, relationship discovery, and intelligent prediction. Traditional knowledge extraction methods often employ rule templates or phased pipeline processes, requiring a large amount of manually labeled data and struggling to capture complex semantic features.

[0003] The emergence of Large Language Models (LLMs) has brought new ideas to knowledge extraction tasks. By using pre-designed prompts, entities and relationships in text can be extracted in one go, achieving structured knowledge extraction. However, knowledge extraction alone is insufficient to meet the needs of dynamic reasoning. It is necessary to further combine graph databases to build knowledge graphs, thereby realizing relational reasoning and prediction.

[0004] Meanwhile, real-world knowledge networks often contain fuzzy and uncertain information. Traditional deterministic reasoning methods are difficult to accurately reflect the complex dependencies between nodes. Therefore, it is necessary to introduce fuzzy logic and Bayesian theory to quantify knowledge relationships and achieve dynamic analysis that is more in line with reality. Summary of the Invention

[0005] To address the problems of complex risk factor relationships, high information uncertainty, and inaccurate tracing paths in existing shield tunneling risk tracing methods, this invention provides a knowledge graph-based shield tunneling risk tracing method. This method integrates multi-source risk information and quantifies uncertain factors to achieve accurate tracing of risk propagation paths. The method uses a knowledge graph to structurally represent the causal relationships between risk factors, and combines fuzzy Bayesian inference with expert experience and field data to improve the scientific rigor and reliability of risk tracing. The specific implementation steps are as follows:

[0006] This invention adopts the following technical solution: a shield tunneling risk tracing method based on knowledge graphs, comprising the following steps:

[0007] Entity sets are obtained by determining entity types from multiple dimensions. The relationship set is obtained by determining the type of causal relationship between entities based on the progressive nature of risk propagation. Constructing the schema layer of the knowledge graph;

[0008] The document data is preprocessed to obtain processed text. Based on the LIM model, the schema layer is used to extract the required knowledge triples from the processed text and output the structured knowledge triples according to the predetermined format.

[0009] The structured knowledge triples are aligned with entities to obtain normalized knowledge triples, and a knowledge graph based on nodes and directed edges is constructed.

[0010] The membership degree of a node is calculated based on the knowledge graph, and the posterior probability of the node is obtained by combining the prior probability of the node.

[0011] Using tunnel boring machine (TBM) construction risks as the top-level node, the potential risk propagation paths and relationship strengths are inferred based on the posterior probability of the nodes in the knowledge graph, enabling dynamic analysis and prediction of the knowledge network.

[0012] In a further embodiment, the method also includes validating the inference of potential risk propagation paths and continuously optimizing the inference process based on the validation results.

[0013] In a further embodiment, the entity set It includes at least the following entities: entities related to risk events, entities related to equipment factors, entities related to geological factors, entities related to assembly quality factors, entities related to construction phenomena, entities related to parameter performance, entities related to structural design factors, entities related to maintenance and operation factors, and entities related to work wear factors;

[0014] The set of relations It should include at least the following relationships: risk relationships, factor relationships, and parameter relationships.

[0015] In a further embodiment, the mode layer includes: a task description layer, a domain background layer, and an output format layer;

[0016] Correspondingly, the extraction and output process of the mode layer is as follows:

[0017] The LIM model is used to perform semantic analysis on the processed text, and a task description layer is used to extract "" from the processed text. "The triplet, in which," and Representing entities and entity , Representing entity relationship ;

[0018] Constraints are created using the domain context layer to filter out knowledge triples that meet the requirements. These constraints are expressed as follows: ;

[0019] The output format layer formats the knowledge triples to generate a unified structured knowledge triple and outputs it.

[0020] In a further embodiment, the knowledge graph construction process is as follows:

[0021] The structured knowledge triples are normalized to achieve entity alignment, resulting in normalized knowledge triples.

[0022] The normalized knowledge triples are stored, and the entities within the normalized knowledge triples are mapped as nodes, the relations are mapped as directed edges, and the causal relationships are visualized.

[0023] In a further embodiment, the process of obtaining the membership degree distribution is as follows:

[0024] For quantifiable nodes in a knowledge graph, the corresponding membership function is determined based on the measured value of the node and the corresponding threshold range: if the measured value is greater than the maximum threshold of the threshold range, the left half trapezoidal membership function is selected; if the measured value is less than the minimum threshold of the threshold range, the right half trapezoidal membership function is selected; if the measured value is within the threshold range, the trapezoidal membership function is selected.

[0025] For observable nodes, the measured values ​​are substituted into the selected membership function to calculate the membership degree. , Let each node be an unobservable node; for unobservable nodes, their membership degree is temporarily set to 1. .

[0026] In a further embodiment, the posterior probability of the node is calculated as follows:

[0027] In statistical causal relationships, the frequency of contribution of lower-level risk factors to higher-level risk factors is calculated and defined as the prior probability. , For nodes Risk factors;

[0028] The nonnormalized weights are calculated using the following formula. : ;in, Risk factors The prior probability, For nodes Membership degree;

[0029] Calculate the normalized posterior probability : Where E represents the collected risk characteristics, The total number of non-normalized weights. It is the sum of all non-normalized weights.

[0030] In a further embodiment, the method for inferring the potential risk propagation path is as follows:

[0031] Taking the risks of tunnel boring machine construction as the top-level node, and using breadth-first search based on normalized knowledge triples, the causal relationships between entities are traversed layer by layer in the knowledge graph.

[0032] In each risk layer, the node with the highest posterior probability is selected as the key risk factor, and the risk propagation path is generated by tracing back from the top to the bottom according to the risk level.

[0033] Correspondingly, the relationship strength is the posterior probability between adjacent nodes in the risk propagation path.

[0034] In a further embodiment, the mode layer further includes: a role setting layer and a reference instance layer;

[0035] The role setting layer is used to define the roles involved in risk tracing and to clarify the corresponding permissions;

[0036] The reference instance layer is used to determine according to " "A specific example of the structure of a triple is shown."

[0037] The beneficial effects of this invention are as follows: By combining knowledge graphs and fuzzy Bayesian inference techniques, this invention accurately solves the complex causal relationships and information uncertainty problems in the risk tracing of tunnel boring machine (TBM) construction. Through multi-dimensional entity modeling and risk relationship extraction, a clear knowledge graph is constructed, laying a solid foundation for inferring risk propagation paths. Combined with LLM model fuzzy Bayesian inference, dynamic analysis and accurate prediction of potential risks are achieved, especially in handling uncertainties and complex environments, quantifying the membership degree and posterior probability of each node. Through effectiveness verification and optimization mechanisms, the accuracy of path inference is continuously improved, ultimately providing a scientific and reliable decision-making basis for risk prevention and control in TBM construction. Attached Figure Description

[0038] Figure 1 This is the overall framework diagram of the knowledge graph-based shield tunneling construction risk tracing of this invention.

[0039] Figure 2 This is a structural diagram of the Prompt template of the present invention. Detailed Implementation

[0040] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0041] Example 1

[0042] A knowledge graph-based method for tracing risks in tunnel boring machines includes the following steps:

[0043] Entity sets are obtained by determining entity types from multiple dimensions. The relationship set is obtained by determining the type of causal relationship between entities based on the progressive nature of risk propagation. Constructing the schema layer of the knowledge graph;

[0044] The document data is preprocessed to obtain processed text. Based on the LIM model, the schema layer is used to extract the required knowledge triples from the processed text and output the structured knowledge triples according to the predetermined format.

[0045] It should be noted that the documents mentioned in this embodiment, as data sources, can be multi-source information carriers such as construction cases, technical documents, and fault analysis reports. The processed text is obtained by removing redundant information and correcting typos. For example, the phrase "due to high clay content in the formation, the fluidity of the cutterhead center area decreases, ultimately leading to mud cake formation" is segmented into independent short sentences.

[0046] The structured knowledge triples are aligned with entities to obtain normalized knowledge triples, and a knowledge graph based on nodes and directed edges is constructed.

[0047] The membership degree of a node is calculated based on the knowledge graph, and the posterior probability of the node is obtained by combining the prior probability of the node.

[0048] Using tunnel boring machine (TBM) construction risks as the top-level node, the potential risk propagation paths and relationship strengths are inferred based on the posterior probability of the nodes in the knowledge graph, enabling dynamic analysis and prediction of the knowledge network.

[0049] In a further embodiment, the entity set described in this embodiment It includes at least the following entities: entities related to risk events, entities related to equipment factors, entities related to geological factors, entities related to assembly quality factors, entities related to construction phenomena, entities related to parameter performance, entities related to structural design factors, entities related to maintenance and operation factors, and entities related to work wear factors.

[0050] It is worth mentioning that the "entity" described in this embodiment includes not only physical components or mechanisms, but also the states, characteristics or manifestations associated with these components or mechanisms.

[0051] For ease of understanding, combined with Figure 2 The following will illustrate the different types of entities one by one:

[0052] Risk events refer to specific risks that occur during the construction of tunnel boring machines (TBMs). Therefore, the corresponding entities include: cutterhead mud cake formation, sludge removal, main drive seal failure, and shield tail seal failure.

[0053] Equipment factors refer to the type of tunnel boring machine and its performance, and the corresponding physical factors are excessive cutterhead opening ratio, excessively low screw conveyor speed, and abnormal propulsion speed.

[0054] Geological factors refer to the geological conditions described in shield tunneling construction, such as high clay content in the strata, high viscosity in the soil, and argillaceous rock layers.

[0055] Assembly quality factors refer to the assembly quality problems of tunnel boring machine components, such as poor quality of tail brush and large gaps in sealing rings.

[0056] Construction phenomena can be understood as the phenomena that occur during the construction process of the tunnel boring machine, such as reduced fluidity in the central area of ​​the cutterhead, soil adhering to the cutterhead inside the chamber, and difficulty in controlling the attitude of the tunnel boring machine.

[0057] The parameter performance refers to the performance of the tunnel boring machine's tunneling parameters, such as excessively fast tunneling speed.

[0058] Structural design factors can be understood as the structural design factors of the tunnel boring machine's components, such as large gaps in the sealing rings and large intervals in the cutter arrangement.

[0059] Maintenance and operation factors refer to factors caused by improper maintenance and operation of the tunnel boring machine, such as contamination of the main drive gear oil or insufficient grease injection.

[0060] Working wear factors refer to the wear and tear caused by the long-term operation of the tunnel boring machine (such as wear of the main drive seals, damage to the cutting tools, and aging of the tail brush).

[0061] Based on the determination of the above entity types, a result-oriented set of relations is established. It should include at least the following relationships: risk relationships between entities, phenomenon relationships between entities, factor relationships between entities, and parameter relationships between entities.

[0062] The risk relationship between entities can be further understood as a risk event directly caused by a certain phenomenon. For example, poor fluidity in the central area of ​​the cutterhead (phenomenon) leads to the formation of mud cake on the cutterhead (risk event).

[0063] The phenomenological relationship between entities refers to a phenomenon caused by a certain factor. For example, poor soil improvement (factor) leads to poor fluidity in the central area (phenomenon). It should be noted that a factor is the cause or condition that affects the occurrence of a phenomenon, while a phenomenon is the result or observed effect. In this example, "poor soil improvement" is the factor because it is the cause of the phenomenon "poor fluidity in the central area".

[0064] The factor relationship between entities is that one factor triggers another factor. For example, a low foaming rate of the soil conditioner (factor) leads to poor soil improvement effect (factor).

[0065] The parameter relationship between entities can be understood as a factor causing a change in the parameter. For example, a low foaming rate of the modifier (factor) will lead to a low penetration rate (parameter).

[0066] Based on the above entity types and relation sets As explicitly stated, the mode layer described in this embodiment includes: a task description layer, a domain background layer, and an output format layer.

[0067] Based on the layer structure of the pattern layer described above, the process of extracting the output of the pattern layer is as follows:

[0068] The LIM model is used to perform semantic analysis on the processed text, and a task description layer is used to extract "" from the processed text. "The triplet, in which," and Represents numbered entities and entity , For entity relationship ;

[0069] Constraints are created using the domain context layer to filter out knowledge triples that meet the requirements. These constraints are expressed as follows: In other words, the extracted " "The entities and entity relations in the triple should belong to the entity type and causal relation type explicitly stated above."

[0070] The output format layer formats the knowledge triples, generating unified structured knowledge triples for output. This includes converting the knowledge triples to a unified format and specifying entity and causal relationship types, such as JSON, RDF, or other standard formats supporting graph databases, for ease of storage and retrieval.

[0071] Furthermore, the mode layer described in this embodiment also includes: a character setting layer and a reference instance layer;

[0072] The role setting layer is used to define the roles involved in risk tracing and to clarify the corresponding permissions; for example, it clarifies that LLM is an expert in extracting risk knowledge for tunnel boring machine construction, focusing on extracting risk-related knowledge from text.

[0073] The reference instance layer is used to determine according to " "A specific example of the structure of a triple is shown in Table 1."

[0074] Table 1 "Structure of the triplet"

[0075]

[0076] In a further embodiment, the knowledge graph construction process is as follows:

[0077] The structured knowledge triples are normalized to achieve entity alignment, resulting in normalized knowledge triples. Furthermore, entity alignment includes methods such as alias recognition, synonym merging, and semantic similarity calculation to address semantic differences and data redundancy in the extracted data. For example, "insufficient oil injection" and "too low oil injection" are unified as "insufficient oil injection," eliminating node redundancy.

[0078] The normalized knowledge triples are stored, and entities within each triple are mapped to nodes, while relationships are mapped to directed edges, visually representing causal relationships. Different colored nodes represent risk events, geological factors, etc., and relationships are mapped to directed edges (e.g., a red edge indicates "leading to (risk)"). The risk knowledge network is displayed through the Neo4j visualization interface; for example, the association path "geological factor → construction phenomenon → risk event" can be intuitively presented.

[0079] In a further embodiment, the Neo4j Cypher query language development interface supports operations such as "querying all associated factors of 'mud cake'" and "statistically counting the frequency of occurrence of the 'cause (risk)' relationship," providing knowledge retrieval support for subsequent fuzzy Bayesian inference.

[0080] In a further embodiment, the process of obtaining the membership degree distribution is as follows:

[0081] For quantifiable nodes in a knowledge graph, the corresponding membership function is determined based on the node's measured value and the corresponding threshold range: if the measured value is greater than the maximum threshold of the threshold range, a left-half trapezoidal membership function is used; if the measured value is less than the minimum threshold of the threshold range, a right-half trapezoidal membership function is used; if the measured value is within the threshold range, a trapezoidal membership function is used. :

[0082] ;

[0083] For observable nodes, the measured values ​​are substituted into the selected membership function to calculate the membership degree. , For nodes;

[0084] For unobservable nodes, we temporarily assume their membership degree is 1, i.e. .

[0085] For example, for the node "tunneling speed," a maximum and minimum threshold for speed are predefined, and its actual tunneling speed is obtained. If the actual tunneling speed exceeds the maximum threshold, it indicates that the "tunneling speed is too fast," and a left-hand trapezoidal membership function is selected. :

[0086] ;

[0087] In the formula, These are measured values. This marks the starting point for the increase in membership degree. The endpoint is when the membership degree reaches 1.

[0088] Correspondingly, for the node "Spindle Speed", a maximum and minimum threshold for its speed are predefined, and its actual speed value is obtained. If the actual speed value is less than the minimum threshold, it indicates that "Spindle Speed ​​is too low", and a right-hand trapezoidal membership function is selected. :

[0089] ;

[0090] In the formula, This marks the starting point where the membership degree begins to decrease. The endpoint is when the membership degree reaches 0.

[0091] The above method selects the corresponding membership function based on the range between the measured value and the threshold. For observable nodes, the measured value is substituted into the selected membership function to calculate the membership degree. , For nodes.

[0092] However, in reality, some nodes may lack observable values, meaning their state cannot be quantified numerically. For example, a construction management oversight can be temporarily assumed to have a membership degree of 1. This was subsequently verified through manual investigation.

[0093] In a further embodiment, the posterior probability of the node is calculated as follows: the contribution frequency of lower-level risk factors to higher-level risk factors in a statistical causal relationship is calculated and defined as the prior probability. , For nodes Risk factors;

[0094] The nonnormalized weights are calculated using the following formula. : ;in, Risk factors The prior probability, For nodes Membership degree;

[0095] Calculate the normalized posterior probability : Where E represents the collected risk characteristics, and N is the total number of non-normalized weights. It is the sum of all non-normalized weights.

[0096] For example, if "excessive tunneling speed" is a underlying factor of "mud cake formation on the cutterhead," with a prior probability P=0.3, then the non-normalized weights... If there are three other lower-level factors, the total weights are... If the value is 0.6, then the normalized posterior probability is 0.25.

[0097] Based on this, the method for inferring potential risk propagation paths is as follows:

[0098] Taking the risks of tunnel boring machine construction as the top-level node, and using breadth-first search based on normalized knowledge triples, the causal relationships between entities are traversed layer by layer in the knowledge graph.

[0099] In each risk layer, the node with the highest posterior probability is selected as the key risk factor, and the risk propagation path is generated by tracing back from the top to the bottom according to the risk level.

[0100] Correspondingly, the relationship strength is the posterior probability between adjacent nodes in the risk propagation path.

[0101] Using the target risk event (such as "mud cake forming on the cutterhead") as the top-level node, a breadth-first search algorithm is employed to traverse the causal relationships in the knowledge graph layer by layer downwards. At each layer, the node with the highest posterior probability is selected as the key risk factor, and this becomes the starting point for continuing the search for related nodes at the next layer.

[0102] For example, if the posterior probability of "soil is difficult to cut" is the highest in the first layer, then this node is used as the parent node, and the posterior probabilities of its child nodes (such as "high soil viscosity" and "tool damage") are calculated, with the node having the highest value selected. Further, the path is: first layer "soil is difficult to cut (construction phenomenon)" (posterior probability 0.6) → second layer "high soil viscosity (geological factor)" (posterior probability 0.7) → third layer "high soil plasticity index (geological factor)" (posterior probability 0.8), forming the following path: "high soil plasticity index → ​​high soil viscosity → soil is difficult to cut → cutterhead cake formation".

[0103] The risk propagation path can be deduced as follows: high soil plasticity index → ​​high stratum viscosity → poor fluidity in the center area of ​​the cutterhead → soil is difficult to cut → mud cake forms on the cutterhead.

[0104] In another embodiment, the method further includes: validating the inference of potential risk propagation paths, and continuously optimizing the inference process of risk propagation paths based on the validation results.

[0105] For example, taking the mud cake accident case in the WH river-crossing tunnel as an example, the source tracing path of this method is compared with the on-site accident report. If the path matching degree reaches 85% and the accuracy rate reaches 90%, the effectiveness is verified. A comparative experiment is set up. Compared with the method based solely on prior probability, the path matching degree of this method is improved by 20%, proving its advantage in complex risk source tracing.

Claims

1. A method for tracing the source of risks in tunnel boring machines based on knowledge graphs, characterized in that, Includes the following steps: Entity sets are obtained by determining entity types from multiple dimensions. The relationship set is obtained by determining the type of causal relationship between entities based on the progressive nature of risk propagation. Constructing the schema layer of a knowledge graph; The document data is preprocessed to obtain processed text. Based on the LIM model, the schema layer is used to extract the required knowledge triples from the processed text and output the structured knowledge triples according to the predetermined format. The structured knowledge triples are aligned with entities to obtain normalized knowledge triples, and a knowledge graph based on nodes and directed edges is constructed. The membership degree of a node is calculated based on the knowledge graph, and the posterior probability of the node is obtained by combining the prior probability of the node. Using the risks of tunnel boring machine construction as the top-level node, the potential risk propagation paths and relationship strengths are inferred based on the posterior probability of the nodes in the knowledge graph, thereby realizing the dynamic analysis and prediction of the knowledge network. The calculation process for the posterior probability of the node is as follows: In statistical causal relationships, the frequency of contribution of lower-level risk factors to higher-level risk factors is calculated and defined as the prior probability. , For nodes Risk factors; The nonnormalized weights are calculated using the following formula. : ;in, Risk factors The prior probability, For nodes Membership degree; Calculate the normalized posterior probability : Where E represents the collected risk characteristics, The total number of non-normalized weights. It is the sum of all non-normalized weights; The method for inferring the potential risk propagation path is as follows: Taking the risks of tunnel boring machine construction as the top-level node, and using breadth-first search based on normalized knowledge triples, the causal relationships between entities are traversed layer by layer in the knowledge graph. In each risk layer, the node with the highest posterior probability is selected as the key risk factor, and the risk propagation path is generated by tracing back from the top to the bottom according to the risk level. Correspondingly, the relationship strength is the posterior probability between adjacent nodes in the risk propagation path.

2. The shield tunneling risk tracing method based on knowledge graphs according to claim 1, characterized in that, It also includes verifying the effectiveness of inferring potential risk propagation paths and continuously optimizing the inference process based on the verification results.

3. The shield tunneling risk tracing method based on knowledge graphs according to claim 1, characterized in that, The entity set It includes at least the following entities: entities related to risk events, entities related to equipment factors, entities related to geological factors, entities related to assembly quality factors, entities related to construction phenomena, entities related to parameter performance, entities related to structural design factors, entities related to maintenance and operation factors, and entities related to work wear factors; The set of relations It should include at least the following relationships: risk relationships, factor relationships, and parameter relationships.

4. The shield tunneling risk tracing method based on knowledge graphs according to claim 1, characterized in that, The pattern layer includes: a task description layer, a domain background layer, and an output format layer; Correspondingly, the extraction and output process of the mode layer is as follows: The LIM model is used to perform semantic analysis on the processed text, and a task description layer is used to extract semantic information from the processed text. "The triplet, in which," and Representing entities and entity , Representing entity relationship ; Constraints are created using the domain context layer to filter out knowledge triples that meet the requirements. These constraints are expressed as follows: ; The output format layer formats the knowledge triples to generate a unified structured knowledge triple and outputs it.

5. The shield tunneling risk tracing method based on knowledge graphs according to claim 1, characterized in that, The knowledge graph construction process is as follows: The structured knowledge triples are normalized to achieve entity alignment, resulting in normalized knowledge triples. The normalized knowledge triples are stored, and the entities within the normalized knowledge triples are mapped as nodes, the relations are mapped as directed edges, and the causal relationships are visualized.

6. The shield tunneling risk tracing method based on knowledge graphs according to claim 1, characterized in that, The process of obtaining the membership degree distribution is as follows: For quantifiable nodes in a knowledge graph, the corresponding membership function is determined based on the measured value of the node and the corresponding threshold range: if the measured value is greater than the maximum threshold of the threshold range, the left half trapezoidal membership function is selected; if the measured value is less than the minimum threshold of the threshold range, the right half trapezoidal membership function is selected; if the measured value is within the threshold range, the trapezoidal membership function is selected. For observable nodes, the measured values ​​are substituted into the selected membership function to calculate the membership degree. , Let each node be an unobservable node; for unobservable nodes, their membership degree is temporarily set to 1. .

7. The shield tunneling risk tracing method based on knowledge graphs according to claim 4, characterized in that, The mode layer also includes: a character setting layer and a reference instance layer; The role setting layer is used to define the roles involved in risk tracing and to clarify the corresponding permissions; The reference instance layer is used to determine based on " "A specific example of the structure of a triple is shown."

Images