Railway engineering unfavorable geology knowledge base construction method and system

Abstract

The invention discloses a railway engineering unfavorable geology knowledge base construction method and system, and the method comprises the steps: collecting and preprocessing multi-source railway engineering text data, and obtaining a multi-source text data set; performing knowledge extraction, knowledge fusion and knowledge processing based on the multi-source file data set; constructing a railway engineering unfavorable geology knowledge base; and carrying out user interaction design based on a railway unfavorable geology knowledge base. According to the method, unfavorable geological knowledge of railway engineering can be automatically extracted from a large amount of text data, a structured knowledge base is constructed, the knowledge base not only improves the knowledge extraction efficiency and accuracy, but also provides scientific and accurate information support for safety management and decision-making of railway engineering in the form of the knowledge graph, and the method is suitable for popularization and application. Therefore, key fault information can be better mastered for railway operation enterprises, and accurate and effective information data and technical support are provided for decision making.

Description

Technical Field

[0001] The present invention relates to the technical field of railway engineering, and in particular to a method and system for constructing a knowledge base of adverse geology for railway engineering. Background Art

[0002] As an important national infrastructure, railway engineering plays an irreplaceable role in promoting regional economic development and improving people's lives. However, during the construction and operation of railway engineering, it often faces the challenges of adverse geological conditions, such as landslides and mud-rock flows. These adverse geological phenomena not only increase the difficulty of engineering construction, but also pose a serious threat to the safety of railway operations. At the same time, with the continuous expansion of the railway network, how to effectively manage and deal with these adverse geological problems has become a key issue that needs to be urgently solved in the field of railway engineering.

[0003] With the rapid development of information technology in recent years, text mining technology has shown great potential in information extraction and knowledge discovery. Especially in the field of railway engineering, a large number of text materials such as engineering documents and geological research reports contain rich knowledge about railway adverse geology. Through text mining technology, key information of concern can be extracted from these unstructured text data to provide support for railway engineering decision-making in adverse geology. However, due to the complexity and professionalism of railway engineering text data, traditional text mining methods face many challenges in processing these data.

[0004] As a structured knowledge representation method, knowledge graph provides a new perspective for knowledge storage, retrieval and reasoning by organizing entities, attributes and relationships in the form of graphs. In the field of railway engineering, building a knowledge base of adverse geology based on text mining and knowledge graphs can not only systematically organize and display knowledge related to adverse geology, but also reveal the intrinsic connection between different geological phenomena through the association analysis function of the graph, providing more accurate and effective information for risk assessment and decision making. Knowledge graphs have achieved remarkable application results in other fields. However, in the field of railway engineering, especially the construction of knowledge bases for adverse geological problems, it is still in the exploratory stage. The following major problems still exist:

[0005] (1) Knowledge extraction and integration of unstructured data from poor geological information data: How to effectively extract key knowledge entities, attributes and relationships from unstructured text data in the railway engineering field and integrate them into a structured knowledge base. This includes determining appropriate text mining techniques and methods to handle the recognition of professional terms and complex sentences, and how to ensure the accuracy and comprehensiveness of the knowledge extraction process.

[0006] (2) A knowledge base of adverse geological information constructed and effectively applied based on knowledge graphs: How to effectively connect the constructed knowledge graphs with the actual application of railway engineering, especially in the application of adverse geological risk assessment and fault diagnosis. This involves the construction method of knowledge graphs, the depth and breadth of knowledge representation, and how to use knowledge graphs for complex decision support and risk management. Summary of the invention

[0007] The present application provides a method and system for constructing a knowledge base of adverse geology for railway engineering, so as to solve the problems of accurate and efficient extraction of adverse geological information knowledge and effective connection between the constructed knowledge graph and the practical application of railway engineering.

[0008] According to the first aspect, an embodiment provides a method for constructing a railway engineering adverse geological knowledge base, the method comprising:

[0009] Collect and preprocess multi-source railway engineering text data to obtain a multi-source text dataset;

[0010] Performing knowledge extraction, knowledge fusion and knowledge processing based on multi-source document data sets, wherein the knowledge extraction includes entity recognition and extraction, relationship extraction and attribute extraction, the knowledge fusion includes entity unified alignment and entity disambiguation, and the knowledge processing includes ontology construction, knowledge reasoning and quality assessment;

[0011] Constructing a knowledge base of adverse geology for railway engineering, including constructing a knowledge graph based on the extracted entities and relationships and visualizing the knowledge graph;

[0012] User interaction design based on the railway adverse geology knowledge base includes the development of relevant applications based on the knowledge graph and the provision of a user interface for interactive query and analysis.

[0013] Furthermore, multi-source railway engineering text data is collected and preprocessed to obtain a multi-source text dataset, which specifically includes:

[0014] Railway engineering text data is collected from multiple channels and sources, including engineering reports, geological survey records, and accident analysis reports, and the collected data is used as the raw material for knowledge extraction; the collected data is cleaned, irrelevant content is removed, and text standardization is performed in preparation for knowledge extraction.

[0015] Furthermore, entity recognition and extraction specifically include:

[0016] For entity recognition, a deep learning Bi-LSTM model is used. The model uses a single-layer Bi-LSTM. The single-layer Bi-LSTM includes two LSTMs, one LSTM is used to process the sequence forward, and the other LSTM is used to process the sequence backward. The forward LSTM obtains a result vector after n time steps, and the reverse LSTM also obtains another result vector after n time steps. The two result vectors are then concatenated as the final output of the BiLSTM.

[0017] Furthermore, relation extraction specifically includes:

[0018] Convolutional deep neural network (CDNN) is used to extract features at the vocabulary and sentence levels. The CDNN network includes the following components: word vector representation, feature extraction and output. The sentence-level feature vector construction process is that each word vector is composed of word features and position features. In order to extract vocabulary-level and sentence-level features, the word vector is placed in the convolution layer to extract sentence-level features. At the same time, the extracted sentence-level features are directly connected to form the final sentence feature vector. Finally, the final classification result is obtained through the hidden layer and the Softmax layer, and then the final relationship extraction result is obtained.

[0019] Furthermore, attribute extraction includes:

[0020] The joint framework of bidirectional recurrent neural network JRNN is used for attribute extraction. Combined with the characteristics of the processed data, the selected JRNN model uses GRU units to construct RNN; the model combines Bi-RNN with discrete features, and learns the representation of sentences from forward and reverse directions through two RNNs respectively, automatically generates features, and uses memory vectors and memory matrices to store trigger words, entities, and the dependencies between entities and trigger words.

[0021] Further,

[0022] Entity uniform alignment, including:

[0023] Entity unification is achieved using a twin network-based attribute attention entity alignment model. The twin network is used to share the network weight parameters of entity pairs during encoding. In the twin network, two entity encoding methods are used to learn the integrity of the entity and the importance of the attribute to the entity respectively.

[0024] Entity disambiguation, including:

[0025] The end-to-end entity linking (JEEEL) model based on joint encoding is used for entity disambiguation. The model input is the original mention text, and directly outputs the results of multiple mention entities consisting of triplets of a single mention entity, that is, end-to-end output.

[0026] Furthermore, ontology construction specifically includes:

[0027] Based on the characteristics of railway adverse geological data, the inductive reasoning method of TACT is used to effectively handle the topological perception association between relations in an independent entity manner; TACT uses an entity-independent way to deal with the given triple (u, r t ,v) for scoring, where r t It is the target relationship between entity u and entity v; TACT consists of two modules: relationship association module and graph structure module.

[0028] Furthermore, the relationship association module is specifically used for:

[0029] Divide all relations in the knowledge graph into multiple topological patterns, and based on the definition of different association patterns, transform the original knowledge graph into a relational association graph, where nodes represent relations and edges represent the correlation patterns between any two relations in the original knowledge graph;

[0030] The relationship correlation network (RCG) is used to simulate the importance of different correlation patterns on perceptual link prediction. RCG includes a correlation pattern part and a correlation coefficient part. The correlation pattern part considers the influence of different topological structures between any two relations, while the correlation coefficient part is to understand different correlation levels.

[0031] In order to distinguish the relationship t The different correlation levels of the relationship are learned by using the attention network. t to obtain its domain embedding in the local graph.

[0032] Furthermore, the graph structure module is specifically used for:

[0033] For a triple (u,r t ,v),The local graph surrounding it contains information about how triples are connected to its domain. A GralL-based graph structure network embeds the local graph into the vector to make full use of the local graph information.

[0034] According to the second aspect, an embodiment provides a system for constructing a railway engineering adverse geological knowledge base, the system comprising:

[0035] The data acquisition and processing module is used to collect and preprocess multi-source railway engineering text data to obtain a multi-source text data set;

[0036] A knowledge extraction and fusion module is used to perform knowledge extraction, knowledge fusion and knowledge processing based on multi-source document data sets, wherein the knowledge extraction includes entity recognition and extraction, relationship extraction and attribute extraction, the knowledge fusion includes entity unified alignment and entity disambiguation, and the knowledge processing includes ontology construction, knowledge reasoning and quality assessment;

[0037] The knowledge base construction module is used to construct a knowledge base of unfavorable geology for railway engineering, including constructing a knowledge graph based on the extracted entities and relationships and realizing the visualization of the knowledge graph;

[0038] The application module is used for user interaction design based on the railway adverse geology knowledge base, including the development of relevant applications based on the knowledge graph and providing a user interface for interactive query and analysis.

[0039] The present application provides a method and system for constructing a bad geological knowledge base for railway engineering, which collects multi-source railway engineering text data and performs preprocessing to obtain a multi-source text data set; performs knowledge extraction, knowledge fusion and knowledge processing based on the multi-source file data set, wherein the knowledge extraction includes entity recognition and extraction, relationship extraction and attribute extraction, the knowledge fusion includes entity unified alignment and entity disambiguation, and the knowledge processing includes ontology construction, knowledge reasoning and quality assessment; constructs a bad geological knowledge base for railway engineering, including constructing a knowledge graph based on the extracted entities and relationships and realizing the visualization of the knowledge graph; user interaction design based on the bad geological knowledge base for railway engineering includes developing related applications based on the knowledge graph and providing a user interface for interactive query and analysis. The present invention can automatically extract bad geological knowledge for railway engineering from a large amount of text data and construct a structured knowledge base, which not only improves the efficiency and accuracy of knowledge extraction, but also provides scientific and accurate information support for the safety management and decision-making of railway engineering in the form of knowledge graph, so as to better grasp the key information of faults for railway operating enterprises and provide accurate and effective information data and technical support for decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 A flowchart of a method for constructing a knowledge base of adverse geology for railway engineering provided by one embodiment of the present invention;

[0041] Figure 2 A process framework diagram of a method for constructing a knowledge base of adverse geology for railway engineering provided by an embodiment of the present invention;

[0042] Figure 3 A knowledge extraction process across data sources for a method for constructing a knowledge base of adverse geology for railway engineering provided by an embodiment of the present invention;

[0043] Figure 4A schematic diagram of event extraction of a method for constructing a knowledge base of adverse geology for railway engineering provided by one embodiment of the present invention;

[0044] Figure 5 A technical flow chart related to knowledge processing of a method for constructing a knowledge base of adverse geology for railway engineering provided by one embodiment of the present invention;

[0045] Figure 6 A TACT framework diagram of a method for constructing a knowledge base of adverse geology for railway engineering provided by one embodiment of the present invention;

[0046] Figure 7 An illustration of the topological structure and relationship correlation diagram of a method for constructing a knowledge base of adverse geology for railway engineering provided by one embodiment of the present invention;

[0047] Figure 8 A schematic diagram of the logical structure of a system for constructing a knowledge base of adverse geology for railway engineering provided by one embodiment of the present invention. DETAILED DESCRIPTION

[0048] The present invention is further described in detail below by specific embodiments in conjunction with the accompanying drawings. Wherein similar elements in different embodiments adopt associated similar element numbers. In the following embodiments, many detailed descriptions are for making the present application better understood. However, those skilled in the art can easily recognize that some features can be omitted in different situations, or can be replaced by other elements, materials, methods. In some cases, some operations related to the present application are not shown or described in the specification, this is to avoid the core part of the present application being overwhelmed by too much description, and for those skilled in the art, it is not necessary to describe these related operations in detail, and they can fully understand the related operations according to the description in the specification and the general technical knowledge in the art.

[0049] In addition, the features, operations or characteristics described in the specification can be combined in any appropriate manner to form various implementations. At the same time, the steps or actions in the method description can also be interchanged or adjusted in a manner that is obvious to those skilled in the art. Therefore, the various sequences in the specification and the drawings are only for the purpose of clearly describing a certain embodiment and are not meant to be a required sequence, unless otherwise specified that a certain sequence must be followed.

[0050] The technical solution of this invention aims to build a knowledge base of unfavorable geology for railway engineering based on text mining and knowledge graph, so as to automatically extract and integrate typical unfavorable geology information for railway engineering from multiple sources, and help related enterprises better grasp the key information of unfavorable geology, thereby providing strong technical support for the safety management and decision-making of railway engineering. The overall research framework of this task is as follows Figure 2 This embodiment will study the hierarchical and graph-based railway engineering bad geology knowledge representation model for the real background of multi-source heterogeneous data such as texts and industrial elements involved in the field of railway engineering bad geology, and extract the knowledge graph framework of railway engineering bad geology from multi-source heterogeneous data.

[0051] The first embodiment of the present invention provides a method for constructing a railway engineering adverse geological knowledge base. Figure 1 and Figure 2 Provide detailed explanation.

[0052] like Figure 1 As shown, in step S100, multi-source railway engineering text data is collected and pre-processed to obtain a multi-source text data set.

[0053] Specifically, railway engineering text data are collected from multiple channels and sources, including engineering reports, geological survey records, accident analysis reports, etc., and these data are used as raw materials for knowledge extraction. The collected data are cleaned, irrelevant content is removed, and text standardization is performed to obtain a multi-source text dataset as the multi-source text dataset to be processed in preparation for knowledge extraction.

[0054] like Figure 1 As shown, in step S200, knowledge extraction, knowledge fusion and knowledge processing are performed based on a multi-source document data set, wherein the knowledge extraction includes entity recognition and extraction, relationship extraction and attribute extraction, the knowledge fusion includes entity unified alignment and entity disambiguation, and the knowledge processing includes ontology construction, knowledge reasoning and quality assessment.

[0055] Entity recognition is mainly based on the filtered text data set obtained in step S100, using text mining technology (such as Chinese language processing system (Language Technology Platform, LTP)) to identify key entities from the text, such as landslides, mudslides, karst and other adverse geological phenomena, and identify attributes related to the entity, such as the scale of the landslide, geological structure, hydrogeological conditions, etc.

[0056] Relationship extraction is to determine the relationship between entities based on the identified entities, such as the relationship between "landslide" and "geological structure", the relationship between "mudslide" and "hydrogeological conditions", and then identify the mutual influence and connection between various geological phenomena mentioned in multi-source data sets.

[0057] Knowledge representation: The acquired relational data is used to construct a domain knowledge graph using the Neo4j graph database. The knowledge graph represents the identified entities and relationships in the form of a graph. Each node represents an entity, and the edge represents the relationship between entities and defines the types of entities and relationships, as well as their attributes.

[0058] The specific contents are as follows:

[0059] (1) Knowledge extraction:

[0060] Knowledge extraction can be mainly divided into three tasks: entity recognition and extraction, relationship extraction, and attribute extraction.

[0061] 11) Entity relations include two aspects, namely, relation type and relation parameters. The entity relation extraction problem can be regarded as a classification problem, which is mainly carried out through supervision, semi-supervision or unsupervised.

[0062] For entity extraction, entity extraction can clearly identify an object in a group of objects with similar attributes. It can be understood as an entity with a text identifier, and the entity does not exist without being attached to other things. Entity extraction is to extract entity information elements from text.

[0063] There are mainly rule-based and dictionary-based methods, supervised learning-based methods, and deep learning-based methods for entity extraction. In this embodiment, the main deep learning method used for entity recognition is the Bi-LSTM research method. This method uses a single-layer BiLSTM. A single-layer BiLSTM is actually two LSTMs, one for forward processing of the sequence and one for reverse processing of the sequence. The forward LSTM obtains a result vector after n time steps, and the reverse LSTM also obtains another vector result after n time steps. These two result vectors are then concatenated as the final output of the BiLSTM. The main path of entity recognition is as follows: Figure 3 .

[0064] 12) Relation extraction is one of the important tasks of knowledge extraction. For unstructured text data, relationship extraction is to extract the semantic relationship between two or more entities from the text.

[0065] Relation extraction is closely related to entity extraction. Generally, after identifying entities in a text, possible relationships between entities are extracted. Traditional relationship classification models require a lot of manpower to design features, and many implicit features are difficult to define. Therefore, traditional methods do not work well in large-scale relationship extraction tasks.

[0066] Relation extraction based on deep learning can automatically learn effective features. Supervised relationship extraction is a main method in deep learning methods, which has a good effect in solving problems such as manual feature selection and feature extraction error propagation. Another main method of relationship extraction based on deep learning is the remote supervision method, which uses known knowledge base information to reduce manual processing.

[0067] According to the characteristics of data processing in this embodiment, this embodiment chooses to use Convolutional Deep Neural Network (CDNN) to extract vocabulary and sentence-level features, taking all word tags as input without complex preprocessing, solving the problem that features extracted from the preprocessing system may cause error propagation and hinder system performance. The CDNN network extracts multiple levels of feature vectors for the input sentence, which mainly includes the following three components: word vector representation, feature extraction and output, in which the sentence-level feature vector construction process is that each word vector is composed of word features and position features. In order to extract vocabulary and sentence-level features, the word vector is placed in the convolution layer to extract sentence-level features, and the extracted sentence-level features are directly connected to form the final sentence feature vector. Finally, the final classification result is obtained through the hidden layer and the Softmax layer, and then the final relationship extraction result is obtained.

[0068] 13) Attribute extraction refers to extracting event information that users are interested in from natural language text and presenting it in a structured form. The main tasks of event extraction include finding trigger words from text and identifying the role played by elements from text.

[0069] In event extraction, pattern matching-based methods extract events through pattern matching algorithms. However, pattern matching methods can achieve good performance in specific fields, but have poor portability, especially when performing event extraction across fields, which requires reconstruction.

[0070] This embodiment adopts a joint framework of bidirectional recurrent neural networks (JRNN). In combination with the characteristics of data processing in this embodiment, the JRNN model selected in this embodiment uses GRU (Gated Recurrent Unit) units to construct RNN. The GRU unit controls the flow of information through a gating mechanism to alleviate the gradient vanishing problem. At the same time, a memory vector is used to store the information of trigger words (triggers), and two memory matrices are used to store entities (arguments) and their dependencies with the trigger words. This model combines Bi-RNN with discrete features, and automatically generates features by using two RNNs to learn the representation of sentences from the forward and reverse directions respectively. At the same time, memory vectors and memory matrices are used to store trigger words, entities, and the dependencies between them. The model avoids the problem of error accumulation and propagation leading to a decrease in model performance. All event information can be extracted at the same time, and the global features learned from the overall structure are used to improve the prediction ability of local information. The main recognition results of attribute recognition are as follows: Figure 4 .

[0071] (2) Knowledge Fusion

[0072] The process of knowledge fusion can usually be decomposed into the following two sub-steps, as follows:

[0073] 21) Entity Unification

[0074] In the process of integrating multi-source heterogeneous data, it is usually the case that one real-world entity corresponds to multiple representations. This is because the same entity may have different meanings in different contexts, and ambiguity can be eliminated through dictionaries and supervised methods. Knowledge fusion, the knowledge graph is a semantic network that describes the relationship between entities in a formalized way. In the process of constructing the knowledge graph, the knowledge graph formed by a single data source is prone to form information islands and insufficient knowledge coverage. Knowledge fusion technology can integrate multi-source data to improve the richness and accuracy of the knowledge graph. Therefore, in the process of constructing a knowledge graph based on the evaluation of the level of adverse geological development of modern railway engineering, there are many problems in the fusion of multi-source heterogeneous data based on multiple industries, multiple modes, and multiple levels. There are great problems between the unified alignment of various entities and entity detection and disambiguation. Therefore, this embodiment uses an attribute attention entity alignment model based on a twin network for the problem of entity unification in the first part of knowledge fusion.

[0075] The entity pairs in the entity unified alignment task are all entities in the knowledge base and have similar structures. Therefore, the twin network is used to share the network weight parameters of the entity pairs in the encoding process. In the twin network, two entity encoding methods are used to learn the integrity of the entity and the importance of the attribute to the entity respectively, thereby realizing the unified alignment of entities in the process of integrating typical industry knowledge of adverse geology in modern railway engineering.

[0076] 22) Entity Disambiguation

[0077] Currently, most entity linking tasks combine entity detection and disambiguation modules, which leads to error propagation problems. They also usually encode mentions and candidates separately and then calculate the similarity of their vectors, which makes it impossible to learn the association between entities from the encoding process.

[0078] Based on the entity disambiguation problem, this embodiment proposes an end-to-end entity linking (Joint Encoding End-to-end Entity Linking, JEEEL) based on joint encoding. The model input is the original mention text, and directly outputs the results of multiple mention entities consisting of a single mention entity triple (mention entity, starting position, knowledge base ID), that is, end-to-end output, which reduces the error transmission between tasks and strengthens the connection between entity detection and entity disambiguation, so that the model can fully learn the correlation between the two during the training process. At the same time, the mention entity and the candidate entity are jointly encoded to learn the connection between entities from the bottom-level encoding.

[0079] The model first uses BERT to encode the original text, uses CRF to decode to obtain entity mentions, and then inputs the description text of the mention and the candidate entity candidate corresponding to the mention into the shared BERT for joint encoding. The output of BERT is then input into the bidirectional LSTM to extract the upper-level features, and then decomposed into the mention vector and the mean vector and spliced ​​together and input into the fully connected layer to determine whether the input mention and the candidate are linkable, and finally output the triple result of all the mentioned entities in the text. Among them, (m, o, k) represents the triple result of a single mentioned entity, that is, (mentioned entity, starting position, knowledge base ID). Therefore, based on this model, the entity detection and disambiguation modules are performed simultaneously in combination with the entity linking task, which optimizes the error propagation problem and completes the entity disambiguation unification.

[0080] The above two parts complete the knowledge integration part in the process of establishing the knowledge map of adverse geology for modern railway engineering.

[0081] (3) Knowledge Processing

[0082] Knowledge processing refers to the knowledge reasoning of the "facts" obtained after information extraction / knowledge fusion to expand existing knowledge and obtain new knowledge, such as Figure 5 As shown, it mainly includes three steps: ontology construction, knowledge reasoning and quality assessment.

[0083] Ontology construction and knowledge graph implementation: Based on the characteristics of knowledge in this field, a top-down construction technique is used, starting with ontology construction, defining the conceptual hierarchy and classification system of the knowledge graph, and then filling in specific entities and attributes. Define the ontology model, including concepts, attributes, and relationship types, to achieve a structured representation of the knowledge graph.

[0084] Ontology construction refers to the standardized description of concepts and the relationships between concepts in a specified field at the model layer, which mainly includes three stages: calculation of similarity of parallel relationships between entities, extraction of hierarchical and subordinate relationships between entities, and generation of ontology.

[0085] Specifically, based on the characteristics of railway poor geological data, this embodiment uses a new inductive reasoning method called TACT (Topology-Aware Correlation), which is a method that effectively handles topology-aware correlations between relationships in an independent entity manner. The TACT method can effectively model the semantic correlations between relationships. This method has been applied to many fields and has a good effect on semantic correlation problems. Figure 6 As shown in Figure 2, TACT aims to classify a given triple (u, r t ,v) for scoring, where r t is the target relationship between entity u and entity v. It consists of two modules: the relationship association module and the graph structure module. The relationship association module is based on the observation that the semantic correlation between any two relations is highly correlated with their topological structure. In addition, this method designs a graph structure module based on the idea of ​​GraalL to make full use of the graph structure information. In order to model the semantic correlation between relations, the correlation is considered from two aspects, namely the correlation mode and the correlation coefficient. The correlation between any two relations in the knowledge graph is highly correlated with their topological structure. The correlation coefficient is used to represent the degree of semantic correlation between any two relations.

[0086] 31) Relationship Association Module

[0087] In order to model the association pattern between any two relations, this method divides all the relation pairs in the knowledge graph into 7 topological patterns. Figure 7 As shown in the figure, the topological modes are "head to tail", "tail to tail", "head to head", "tail to head", "parallel", "loop" and "unconnected", and the corresponding related modes are defined as "HT", "TT", "HH", "TH", "PARA", "LOOP" and "NC". For example (r 1 ,HT,r 2 ) represents r 1 and r 2 The association between them is the "HT" mode.

[0088] Based on the definition of different association patterns, this method converts the original knowledge graph into a relational association graph, where nodes represent relations and edges represent the correlation patterns between any two relations in the original knowledge graph. For a topological pattern where two relations are not connected, the corresponding RCG consists of two isolated nodes.

[0089] 32) Relationship Network

[0090] Based on RCG, this method proposes a relational correlation network to simulate the importance of different correlation patterns for perceptual link prediction. RCG consists of two parts: the correlation pattern part and the correlation coefficient part. The correlation pattern part considers the influence of different topological structures between any two relations. The correlation coefficient part is to understand different degrees of correlation.

[0091] For the relationship r t The edges of the RCG can be divided into six groups according to the topological patterns "HT", "TT", "HH", "TH", "PARA" and "LOOP". Note that the topological pattern "NC" is not considered because it means that the edges (relations) in the knowledge graph are not connected. For these six groups, this embodiment uses six linear transformations to learn different semantic relevances corresponding to the topological patterns. In order to distinguish the relationship r t The attention network is used to learn the correlation coefficients of all relations. t All relevant relations of , to obtain its domain embedded in the local graph, using r t N express.

[0092]

[0093] in, is the weight parameter, represents the embedding of all relations. Assume r i The embedding is Yes [i,:] =r i , R [i,:] represents the i-th slice along the first dimension. is a characteristic vector, that is, when r i and r t When the relationship is the i-th mode, the value is 1; otherwise, the value is 0. is the weight parameter, indicating that in the pth mode, the relationship r t different degrees of relevance. Satisfy [Λ t p ] i ≥0 and

[0094] Furthermore, the patent will t and r t N Connect them together to get the final embedding r t F .

[0095]

[0096] 33) Graph structure modeling

[0097] For a triple (u,r t ,v), the local graph around it contains information about how the triples are connected to their domain. In order to make full use of the local graph information, this embodiment embeds the local graph into the vector based on a GraalL graph structure network. The steps are as follows:

[0098] (a) Subgraph extraction:

[0099] For a triple (u,r t ,v), first extract the closed subgraph around the target nodes u and v. First, calculate the neighborhood set N of the two nodes k (u) and N k (v), where k represents the maximum distance of the domain set; secondly, the intersection of different triples is calculated; finally, N nodes whose distance to u and v is greater than k are pruned. k (u)∩N k (v) to calculate the closed subgraph G(u,r t ,v).

[0100] (b) Node labeling: Label the nodes in the closed subgraph extracted in the previous step. Label each node i around nodes u and v in the subgraph with a tuple (d(i,u),d(i,u)), where d(i,u) represents the shortest distance between nodes i and u, without counting any paths passing through v. This captures the topological position of each node relative to the target node. The two target nodes u and v are uniquely labeled as (0,1) and (1,0). The node feature is defined as [one-hot(d(i,u))⊕one-hot(d(i,v))], where one-hot(p) represents a one-hot vector with only the p-th item being 1, and d represents the dimension of the embedding.

[0101] (c) Graph Embedding

[0102] After the nodes of the closed subgraph are labeled, the nodes in the subgraph have initial embedding values. Use R-GCN to learn the extracted closed subgraph G(u,r t ,v) on embedding.

[0103]

[0104] in Represents entity e at level k i Embeddings in R-GCN. Represents the domain index set of node i under the relationship r∈R. is a normalization constant; is the weight parameter.

[0105] Assuming that R-GCN has L layers, the subgraph G(u,r t ,v)Embedding:

[0106]

[0107] in, Represents a subgraph G(u,r t ,v) of the node set. Combining the target node and subgraph embedding, use the vector Represents structural information.

[0108]

[0109] Knowledge reasoning: Use the constructed knowledge graph to perform reasoning, discover the potential connections and rules between entities, and apply reasoning rules, such as the relationship between the scale of the landslide and the geological structure, to predict the development trend of adverse geological phenomena.

[0110] Quality assessment: The core purpose of quality assessment is to quantify the credibility of railway bad address data and ensure the accuracy and reliability of the knowledge base. The first step is to quantify the credibility of the above knowledge, such as determining the standards for evaluating the quality of the knowledge base, which may include accuracy, consistency, completeness, reliability, and timeliness. By identifying and excluding data points with insufficient confidence, the quality of the knowledge base can be significantly improved. Implementing quality assessment is an ongoing process that requires continuous iteration and improvement. Through quality assessment, the quality of the constructed knowledge base can be effectively ensured to provide users with reliable and accurate information.

[0111] like Figure 1 As shown, in step S300, a knowledge base of unfavorable geology for railway engineering is constructed, including constructing a knowledge graph based on the extracted entities and relationships and realizing visualization of the knowledge graph.

[0112] Knowledge base construction: store the organized knowledge and rules in the knowledge base, establish query and update interfaces, design the organizational structure of the knowledge base, and ensure that the knowledge is easy to retrieve and update. Specifically, based on the obtained triples, use Cypher language to query the knowledge graph, perform data analysis and visualization, and use Neo4j's import tool or Cypher query language to batch import railway adverse geological information into the Neo4j database, and use Neo4j browser or Neo4j Bloom for data visualization.

[0113] like Figure 1 As shown, in step S400, user interaction design based on the railway adverse geology knowledge base includes developing relevant applications based on the knowledge graph and providing a user interface for interactive query and analysis.

[0114] User interaction design: Given the industry knowledge base that has been constructed, design a user interface so that users can easily query the knowledge base, submit feedback and suggestions, and provide visualization tools to help users understand complex geological relationships. Specifically, based on the Neo4j knowledge graph, develop related applications and provide a user interface for interactive query and analysis to query specific paths, patterns, or statistical information. For example, provide a search box that allows users to enter the name of a railway section or the type of adverse geology; provide filters that allow users to filter results based on geographic location or problem severity; provide result display to display query results in the form of charts or lists. Based on the complexity of the query and the amount of data, tune Neo4j for performance, including adjusting hardware resources, optimizing indexes and queries, etc.

[0115] The present invention collects railway engineering geological data through a data acquisition unit, and then the data processing unit cleans, extracts and keyword-classifies the data. Key information is effectively integrated into the knowledge graph unit based on the Neo4j graph database to form a tree graph, and a railway engineering adverse geological knowledge base is constructed to intuitively display geological information and its relationship. The present invention not only improves the data retrieval speed, but also enhances the analysis and judgment capabilities of railway engineering adverse geological information, providing strong support for the safety and efficiency of railway engineering.

[0116] Corresponding to the above-disclosed method for constructing a knowledge base of unfavorable geology for railway engineering, the embodiment of the present invention further discloses a system for constructing a knowledge base of unfavorable geology for railway engineering, such as Figure 5 As shown, it specifically includes:

[0117] The data acquisition and processing module is used to collect and preprocess multi-source railway engineering text data to obtain a multi-source text data set;

[0118] A knowledge extraction and fusion module is used to perform knowledge extraction, knowledge fusion and knowledge processing based on multi-source document data sets, wherein the knowledge extraction includes entity recognition and extraction, relationship extraction and attribute extraction, the knowledge fusion includes entity unified alignment and entity disambiguation, and the knowledge processing includes ontology construction, knowledge reasoning and quality assessment;

[0119] The knowledge base construction module is used to construct a knowledge base of unfavorable geology for railway engineering, including constructing a knowledge graph based on the extracted entities and relationships and realizing the visualization of the knowledge graph;

[0120] The application module is used for user interaction design based on the railway adverse geology knowledge base, including the development of relevant applications based on the knowledge graph and providing a user interface for interactive query and analysis.

[0121] It should be noted that for the detailed description of a system for constructing a railway engineering adverse geological knowledge base provided in an embodiment of the present invention, reference can be made to the relevant description of a method for constructing a railway engineering adverse geological knowledge base provided in an embodiment of the present application, which will not be repeated here.

[0122] In addition, an embodiment of the present invention also provides an electronic device, which includes: a processor and a memory; the memory is used to store one or more program instructions; the processor is used to run one or more program instructions to execute the steps of a method for constructing a railway engineering adverse geological knowledge base as described in any of the above items.

[0123] It should be noted that the detailed description of an electronic device provided in an embodiment of the present invention can refer to the relevant description of a method for constructing a railway engineering adverse geological knowledge base provided in an embodiment of the present application, which will not be repeated here.

[0124] In addition, an embodiment of the present invention further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of a method for constructing a knowledge base of adverse geology for railway engineering as described in any of the above items are implemented.

[0125] It should be noted that, for the detailed description of a computer-readable storage medium provided in an embodiment of the present invention, reference can be made to the relevant description of a method for constructing a railway engineering adverse geological knowledge base provided in an embodiment of the present application, which will not be repeated here.

[0126] Those skilled in the art will appreciate that all or part of the functions of the various methods in the above-mentioned embodiments can be implemented by hardware or by computer programs. When all or part of the functions in the above-mentioned embodiments are implemented by computer programs, the program can be stored in a computer-readable storage medium, and the storage medium can include: read-only memory, random access memory, disk, optical disk, hard disk, etc., and the program is executed by a computer to implement the above-mentioned functions. For example, the program is stored in the memory of the device, and when the program in the memory is executed by the processor, all or part of the above-mentioned functions can be implemented. In addition, when all or part of the functions in the above-mentioned embodiments are implemented by computer programs, the program can also be stored in a storage medium such as a server, another computer, disk, optical disk, flash disk or mobile hard disk, and can be downloaded or copied and saved in the memory of the local device, or the system of the local device is updated, and when the program in the memory is executed by the processor, all or part of the functions in the above-mentioned embodiments can be implemented.

[0127] The above specific examples are used to illustrate the present invention, which is only used to help understand the present invention and is not intended to limit the present invention. For those skilled in the art, according to the concept of the present invention, some simple deductions, modifications or substitutions can be made.

Claims

1. A method for constructing a knowledge base of unfavorable geology for railway engineering, characterized in that: The method comprises: Collect and preprocess multi-source railway engineering text data to obtain a multi-source text dataset; Performing knowledge extraction, knowledge fusion and knowledge processing based on multi-source document data sets, wherein the knowledge extraction includes entity recognition and extraction, relationship extraction and attribute extraction, the knowledge fusion includes entity unified alignment and entity disambiguation, and the knowledge processing includes ontology construction, knowledge reasoning and quality assessment; Constructing a knowledge base of adverse geology for railway engineering, including constructing a knowledge graph based on the extracted entities and relationships and visualizing the knowledge graph; User interaction design based on the railway adverse geology knowledge base includes the development of relevant applications based on the knowledge graph and the provision of a user interface for interactive query and analysis.

2. A method for constructing a railway engineering adverse geological knowledge base as claimed in claim 1, characterized in that: Collect and preprocess multi-source railway engineering text data to obtain a multi-source text dataset, including: Railway engineering text data is collected from multiple channels and sources, including engineering reports, geological survey records, and accident analysis reports, and the collected data is used as the raw material for knowledge extraction; the collected data is cleaned, irrelevant content is removed, and text standardization is performed in preparation for knowledge extraction.

3. A method for constructing a railway engineering adverse geological knowledge base as claimed in claim 1, characterized in that: Entity recognition and extraction specifically include: For entity recognition, a deep learning Bi-LSTM model is used. The model uses a single-layer Bi-LSTM. The single-layer Bi-LSTM includes two LSTMs, one LSTM is used to process the sequence forward, and the other LSTM is used to process the sequence backward. The forward LSTM obtains a result vector after n time steps, and the reverse LSTM also obtains another result vector after n time steps. The two result vectors are then concatenated as the final output of the BiLSTM.

4. A method for constructing a railway engineering adverse geological knowledge base as claimed in claim 3, characterized in that: Relation extraction specifically includes: Convolutional deep neural network (CDNN) is used to extract features at the vocabulary and sentence levels. The CDNN network includes the following components: word vector representation, feature extraction and output. The sentence-level feature vector construction process is that each word vector is composed of word features and position features. In order to extract vocabulary-level and sentence-level features, the word vector is placed in the convolution layer to extract sentence-level features. At the same time, the extracted sentence-level features are directly connected to form the final sentence feature vector. Finally, the final classification result is obtained through the hidden layer and the Softmax layer, and then the final relationship extraction result is obtained.

5. A method for constructing a railway engineering adverse geological knowledge base as claimed in claim 4, characterized in that: Attribute extraction, including: The joint framework of bidirectional recurrent neural network JRNN is used for attribute extraction. Combined with the characteristics of the processed data, the selected JRNN model uses GRU units to construct RNN; the model combines Bi-RNN with discrete features, and learns the representation of sentences from forward and reverse directions through two RNNs respectively, automatically generates features, and uses memory vectors and memory matrices to store trigger words, entities, and the dependencies between entities and trigger words.

6. A method for constructing a railway engineering adverse geological knowledge base as claimed in claim 1, characterized in that: Entity uniform alignment, including: Entity unification is achieved using a twin network-based attribute attention entity alignment model. The twin network is used to share the network weight parameters of entity pairs during encoding. In the twin network, two entity encoding methods are used to learn the integrity of the entity and the importance of the attribute to the entity respectively. Entity disambiguation, including: The end-to-end entity linking (JEEEL) model based on joint encoding is used for entity disambiguation. The model input is the original mention text, and directly outputs the results of multiple mention entities consisting of triplets of a single mention entity, that is, end-to-end output.

7. A method for constructing a railway engineering adverse geological knowledge base as claimed in claim 1, characterized in that: The ontology construction specifically includes: Based on the characteristics of railway adverse geological data, the inductive reasoning method of TACT is used to effectively handle the topological perception association between relations in an independent entity manner; TACT uses an entity-independent way to deal with the given triple (u, r t ,v) for scoring, where r t It is the target relationship between entity u and entity v; TACT consists of two modules: relationship association module and graph structure module.

8. A method for constructing a railway engineering adverse geological knowledge base as claimed in claim 7, characterized in that: The relationship association module is specifically used for: Divide all relations in the knowledge graph into multiple topological patterns, and based on the definition of different association patterns, transform the original knowledge graph into a relational association graph, where nodes represent relations and edges represent the correlation patterns between any two relations in the original knowledge graph; The relationship correlation network (RCG) is used to simulate the importance of different correlation patterns on perceptual link prediction. RCG includes a correlation pattern part and a correlation coefficient part. The correlation pattern part considers the influence of different topological structures between any two relations, while the correlation coefficient part is to understand different correlation levels. In order to distinguish the relationship t The different correlation levels of the relationship are learned by using the attention network. t to obtain its domain embedding in the local graph.

9. A method for constructing a railway engineering adverse geological knowledge base as claimed in claim 7, characterized in that: The graph structure module is specifically used for: For a triple (u,r t ,v),The local graph surrounding it contains information about how triples are connected to its domain. A GralL-based graph structure network embeds the local graph into the vector to make full use of the local graph information.

10. A railway engineering adverse geological knowledge base construction system, characterized in that: The system comprises: The data acquisition and processing module is used to collect and preprocess multi-source railway engineering text data to obtain a multi-source text data set; A knowledge extraction and fusion module is used to perform knowledge extraction, knowledge fusion and knowledge processing based on multi-source document data sets, wherein the knowledge extraction includes entity recognition and extraction, relationship extraction and attribute extraction, the knowledge fusion includes entity unified alignment and entity disambiguation, and the knowledge processing includes ontology construction, knowledge reasoning and quality assessment; The knowledge base construction module is used to construct a knowledge base of unfavorable geology for railway engineering, including constructing a knowledge graph based on the extracted entities and relationships and realizing the visualization of the knowledge graph; The application module is used for user interaction design based on the railway adverse geology knowledge base, including the development of relevant applications based on the knowledge graph and providing a user interface for interactive query and analysis.

Images