A bridge underwater detection general knowledge graph construction method, device and medium

Abstract

The application discloses a bridge underwater detection general knowledge graph construction method and device and medium, including the following steps: acquiring standard, paper, report bridge underwater detection field knowledge source, extracting key information, obtaining bridge underwater detection knowledge base and standard consistent bridge detection report text information; based on the bridge underwater detection knowledge base, building entity, relationship and attribute directory; according to the entity directory, performing report bridge underwater detection field knowledge source entity annotation, obtaining a data set; identifying entities in the data set, identifying relationships in the data set, using a rule matching method to identify attributes in the bridge underwater detection data set, and summarizing entity, relationship and attribute sets to obtain a "entity-relation-entity" and "entity-attribute-attribute value" triple set; performing entity alignment and ablation, and based on a visualization platform and an instruction set, realizing bridge underwater detection general graph construction. The application improves the accuracy and timeliness of monitoring.

Description

Technical Field

[0001] This invention pertains to knowledge graphs in the field of bridge engineering, and particularly relates to a method, equipment, and medium for constructing a general knowledge graph for underwater bridge inspection. Background Technology

[0002] Defects in underwater foundations can easily cause bridge accidents. Most bridges have not undergone underwater foundation inspections during their operation, resulting in underwater foundations exhibiting scouring, tilting, subsidence, pile necking, cracking, and steel reinforcement corrosion. Heavy traffic, environmental conditions, and hidden dangers left over from the bridge construction phase all contribute to and threaten bridge safety.

[0003] With the introduction of relevant standards and specifications for underwater bridge structure inspection, and the increasing maturity of new inspection technologies such as 3D sonar point clouds and underwater robots, underwater bridge structure inspection is developing towards standardization and quantification. Underwater bridge structure inspection reports contain a wealth of information on apparent defects, scour conditions, and riverbed topography, necessitating an efficient, convenient, and visualized storage method for subsequent retrieval and decision support.

[0004] A knowledge graph is a structured semantic knowledge base used to describe concepts and their relationships in the real world. It can contain massive amounts of knowledge from specialized fields and can simultaneously perform multiple data tasks such as storage, querying, processing, and visualization. Currently, research on knowledge graphs for underwater bridge structure inspection is still in its early stages and is limited by the specialized nature of bridges and the unique characteristics of underwater structures. The creation of knowledge graphs in this field requires the participation of senior industry experts. Furthermore, the textual information in underwater bridge inspections is complex, and the writing styles, organizational structures, and problem descriptions of inspection reports issued by different inspection units vary. Therefore, a significant amount of manual work is required for entity and relationship extraction, ablation, and alignment, resulting in high construction costs and low automation in knowledge graph creation. Summary of the Invention

[0005] Purpose of the invention: The purpose of this invention is to provide a general knowledge graph construction method for underwater bridge inspection, which improves the accuracy and timeliness of monitoring and provides strong support for the safe operation, inspection, maintenance and intelligent operation and maintenance of bridges.

[0006] The second objective of this invention is to provide a point cloud-image registration system for detecting the spatial morphology of bridge defects.

[0007] A third objective of this invention is to provide an electronic device.

[0008] A fourth objective of this invention is to provide a computer-readable storage medium.

[0009] Technical Solution: To achieve the above objectives, this invention discloses a method for constructing a general knowledge graph for underwater bridge inspection, comprising the following steps:

[0010] S1. Define the general knowledge graph information for underwater bridge inspection. The general knowledge graph information for underwater bridge inspection includes basic bridge information, basic information of underwater components, inspection methods and technologies, types and characteristics of underwater defects, and maintenance measures.

[0011] S2. Obtain knowledge sources in the field of underwater bridge inspection, including standards, papers, and reports; extract key information from the knowledge sources; perform structured processing; and obtain bridge inspection report text information that is consistent with the underwater bridge inspection knowledge base and standards.

[0012] S3. Construct a general ontology catalog for underwater bridge inspection: Based on the constructed underwater bridge inspection knowledge base, construct an entity, relation, and attribute catalog applicable to underwater bridge inspection; according to the entity catalog, use the Label Studio tool library to annotate knowledge source entities in the field of underwater bridge inspection for report-type inspection, and obtain a dataset for the field of underwater bridge inspection.

[0013] S4. Knowledge Source Identification and Calculation in the Bridge Inspection Domain: The BERT-BiLSTM-CRF algorithm is used to identify entities in the underwater bridge inspection dataset, the GNN graph neural network is used to identify relationships in the underwater bridge inspection dataset, and the rule matching method is used to identify attributes in the underwater bridge inspection dataset. The entity, relationship, and attribute sets are summarized to obtain the triple sets of "entity-relationship-entity" and "entity-attribute-attribute value".

[0014] S5. Construction of a general underwater inspection map for bridges: Based on the triplet set obtained in step S4, entity alignment and ablation are performed. Based on the Neo4j knowledge graph data visualization platform and Cypher instruction set, a general underwater inspection map for bridges is constructed.

[0015] Optionally, the basic information of the bridge in S1 includes the bridge name, bridge location, bridge construction year, bridge structural form and total bridge length; the basic information of the underwater components includes component category, component number and location, component type, construction technology, component size and component material; the detection methods and technologies include three-dimensional multibeam radar, three-dimensional sonar point cloud, artificial diving exploration and underwater robot photography; the types and characteristics of underwater defects include defect category, defect location and defect size; and the maintenance measures include riprap protection, pier grooving, sacrificial pile protection and protective ring protection.

[0016] Optionally, step S2 specifically includes the following steps:

[0017] Knowledge sources in the field of underwater bridge inspection are acquired, processed, and imported into a knowledge base for underwater bridge inspection. This involves downloading electronic documents of standards and regulations from the National Standards Information Platform and official websites of relevant international departments and industry associations. These standards and regulations are then categorized, with current national and industry standards representing classic knowledge sources, and current local and group standards representing non-classical knowledge sources. The following weighting coefficients are assigned to each category:

[0018]

[0019] in, The weight of a certain standard is denoted by level, which is the standard level. The values ​​are: domestic national / industry standards level=5, foreign national / industry standards level=4, domestic local standards level=3, foreign local standards level=2, and domestic and foreign group standards level=1.

[0020] According to the organizational structure of the standard document, review it sentence by sentence, retain the content related to the field of underwater structure inspection, management and maintenance, delete irrelevant content, number each sentence of the retained content, and indicate the chapter and clause or clause explanation and appendix of the standard document in which the sentence is located;

[0021] We sorted out the core points of the standards regarding underwater bridge inspection procedures, technical requirements, acceptance standards, and maintenance measures, extracted the core points from the standard clauses, and converted them into structured information.

[0022] Optionally, step S2 specifically includes the following steps:

[0023] Knowledge sources in the field of underwater bridge inspection, including both academic papers and reports, were acquired, processed, and imported into a knowledge base for the field. Specifically, keyword searches were performed on the Web of Science (WOS) paper platform to extract and export relevant literature information, including article titles, authors, abstracts, journal impact factors, and citation counts. Original texts were obtained from publisher platforms and categorized based on citation counts and publication dates. For journal articles in the field of underwater bridge structure inspection and maintenance, classic knowledge sources were defined as: papers published within the last 3 years with a journal impact factor ≥ 5 and WOS citation count ≥ 10; and papers published between 3 and 10 years ago with a journal impact factor ≥ 5 and WOS citation count ≥ 20. All other papers were classified as non-classical knowledge sources. The following weighting coefficients were assigned to each of these papers:

[0024]

[0025] in, The weight of a journal article is given by IF, which is the impact factor of the journal in which the article was published, and cite is the number of citations of the article on the WOS platform.

[0026] The topics are divided into three categories: disease treatment, detection technology, and disease mechanism. The key information related to the applicable conditions, usage steps, and final effects of the methods described in the papers are listed one by one.

[0027] The obtained information is structured and imported into a knowledge base for underwater bridge inspection in plain text format.

[0028] Optionally, step S2 specifically includes the following steps:

[0029] The process involves acquiring knowledge sources in the field of underwater bridge inspection reports, processing them, and importing them into the underwater bridge inspection knowledge base. Specifically, the original underwater bridge structure inspection reports undergo text extraction and cleaning. This includes extracting relevant information such as component type, component number, defect type, defect location, and proposed treatment measures. Redundant information irrelevant to underwater structure inspection, management, and maintenance is removed. The reports are broken down into smaller, granular descriptions, from complete paragraphs to short sentences and phrases. The results are then stored in TXT format documents, sentence by sentence.

[0030] The text is standardized by converting English units into Chinese units and English measurement terms into their corresponding Chinese expressions.

[0031] Correct and align the text, converting typos and grammatical errors in the report into correct spelling, and pre-unify different expressions and structures of the same semantic meaning in different bridge inspection reports;

[0032] The text is encoded to be consistent, removing special characters such as periods, question marks and exclamation marks at the end of the text, converting all punctuation marks in the sentence to full-width characters, and converting the text information into UTF-8 encoding format that is easy for computers to read and process;

[0033] Text segmentation is performed, and combined with the commonly used vocabulary of underwater bridge structures in the constructed knowledge base for underwater bridge inspection, the SnowNLP Chinese word segmentation tool is used to segment the text in the cleaned underwater bridge structure inspection report, finally obtaining standardized and consistent bridge inspection report text information.

[0034] Optionally, step S3 specifically includes the following steps:

[0035] S301. Based on the constructed knowledge base for underwater bridge inspection, clarify the level L of underwater bridge inspection:

[0036] L=∑ i L i

[0037] Among them, L i The detection levels are defined as follows: i = {1, 2, ..., n}, n = 4; the specific definitions of each detection level are: L1 - bridge level, L2 - structural level, L3 - component level, L4 - component level.

[0038] S302. Based on the knowledge base for underwater bridge inspection constructed in step S2 and the analysis results in step S301, construct the underwater bridge inspection ontology directory, which includes an entity directory, a relationship directory and an attribute directory.

[0039] Step S302 specifically includes the following steps:

[0040] S302-1. Based on the knowledge base for underwater bridge structure inspection constructed in step S2 and the results of S301, define the named entities in the bridge inspection field as 10 categories, map the underwater bridge inspection level L in S301, and add entities for describing defects, attributes and numerical entities for describing component and defect features, location entities for representing positions, negative modifier entities for assisting in text extraction, and measure entities for describing maintenance measures. Establish the entity directory E for the underwater bridge structure inspection field:

[0041]

[0042] Among them, E i To define entities, i = {1, 2, ..., n}, n = 10; specific entity directory sub-items are defined as: E1-Bridge entity, E2-Structural entity, E3-Component entity, E4-Component entity, E5-Disease entity, E6-Attribute entity, E7-Numerical entity, E8-Location entity, E9-Negative modifier entity, E... 10 - Entities involved in the measures;

[0043] S302-2. Based on the knowledge base for underwater bridge inspection constructed in step S2 and the results of S302-1, remove E6-attribute entities, E7-numerical entities used to construct attribute-key-value pairs, and E9-negation modifier entities used for auxiliary judgment. Analyze the remaining E1-bridge entities, E2-structural entities, E3-component entities, E4-component entities, E5-disease entities, E8-location entities, and E... 10 - The measures involve relationships between 7 types of entities, establishing a relationship directory for the field of underwater bridge structure inspection (R):

[0044]

[0045] Where, r iFor the field of underwater structure inspection, i = {1, 2, ..., n}, n = 7; the specific relationship is defined as: r1 - structure subordinate, r2 - component subordinate, r3 - component subordinate, r4 - existing defects, r5 - existing treatment measures, r6 - bridge structure location, r7 - defect location;

[0046] S302-3. Based on the knowledge base for underwater bridge inspection constructed in step S2 and the E6-attribute entity in the results of S302-1, classify and construct the attributes in the underwater bridge structure inspection text, and build a basic attribute vocabulary P for the underwater bridge structure inspection field:

[0047]

[0048] Where, p i For underwater structure detection attribute categories, i = {1, 2, ..., n}, n = 5; specifically defined as: p1 - entity common attribute, p2 - bridge body attribute, p3 - pier body attribute, p4 - foundation body attribute, p5 - defect attribute;

[0049] S303. Based on the entity catalog constructed in step S302, create a dataset of the standardized bridge inspection report text information obtained in step S203.

[0050] Step S303 specifically includes the following steps:

[0051] S303-1. Annotate the text information. Based on the entity directory E of the bridge maintenance field constructed in step S302-1, use the same field automatic annotation function of Label Studio tool library to annotate the entity start part B, the middle part I and the non-entity part O in the text information of the bridge maintenance field according to the three-dimensional sequence annotation method, namely BIO method. Then, perform manual correction to correct any errors that may exist in the annotation process, and form a bridge underwater structure detection text dataset.

[0052] S303-2. Dataset partitioning: Using uniform random sampling, the dataset is divided into a training set, a validation set, and a test set in a ratio of 8:1:1. The labels of the three datasets are similar.

[0053] Optionally, step S4 specifically includes the following steps:

[0054] S401. Based on the BERT-BiLSTM-CRF algorithm, identify the entities in the underwater bridge detection dataset labeled in S303, and obtain the extracted entity set;

[0055] Step S401 specifically includes the following steps:

[0056] S401-1, Construction and Training of BERT-BiLSTM-CRF Named Entity Recognition Algorithm for Underwater Bridge Detection: This paper integrates the BERT model, BiLSTM model, and CRF model into a BERT-BiLSTM-CRF model to achieve the function of named entity recognition; The training set, validation set, and test set of the S302-2 dataset are input into the BERT-BiLSTM-CRF model for training.

[0057] S401-2, Text Tokenization and Masking: Input the text to be extracted into the BERT model's built-in encoder Tokenizer. Assign text embedding tokens according to a predefined dictionary, perform full word masking on some words, add a sequence beginning token [CLS] to the beginning of the sequence, and separate sentences with clause tokens [SEP]. Output three embedding layers: character embedding Q, text embedding K, and position embedding V.

[0058] S401-3, BERT Large Language Model Feature Extraction: The character embedding Q, text embedding K, and position embedding V output from step S401-2 are imported into the BERT model. The language features of the bridge underwater structure detection text are extracted through the Transformer encoder, multi-head attention mechanism, and fully connected network, and the language feature sequence is output.

[0059] S401-4, BiLSTM context semantic dependency capture: The sequence output from step S401-3 is imported into the BiLSTM model. Useless information is filtered out through forward and backward propagation and related mechanisms such as input gate, output gate and forget gate, and effective information is passed. The bidirectional semantic dependency of the context information of the underwater bridge detection text is captured and the sequence is output.

[0060] S401-5, CRF sequence label prediction: The output of step S401-4 is imported into a conditional random field. Based on the dependency relationship between adjacent entity labels, the predicted sequence is scored and its probability is evaluated. The highest-scoring predicted sequence is obtained by decoding the maximum likelihood function. The entity with the highest score is selected and the results are summarized to obtain the bridge underwater detection entity set E×E.

[0061] S402. Use GNN (Graph Neural Network) to identify the relationships between entities in the underwater bridge detection dataset extracted in step S401.

[0062] Step S402 specifically includes the following steps:

[0063] S402-1. Convert the underwater bridge detection text data into a graph structure. The nodes in the graph correspond to the entities in the entity set E×E extracted in S401, and the edges represent the relationships between entities. Construct a dependency graph based on the syntactic analysis results, where entities are nodes and dependency relationships are edges.

[0064] S402-2. Use the Word2Vec word embedding method to obtain word vector representations, thereby capturing contextual semantic information. Assign an embedding vector to each entity type, integrating entity type information into node features. Assuming the entity type set is T, for an entity of type t∈T, its type embedding is e. t The initial feature vector x of the node i It is obtained by concatenating word embeddings and entity type embeddings, and its expression is x. i =[h i ;e t ];

[0065] S402-3, GNN layer propagation, aggregates information from neighboring nodes through a message passing mechanism to update the features of the current node; wherein, the feature update formula for the node in the l-th layer is:

[0066]

[0067] in, It is the feature vector of node i in the (l+1)th layer. W is the set of neighboring nodes of node i. l This is the weight matrix of the l-th layer, and σ is the ReLU activation function. It is the feature vector of node j in the l-th layer;

[0068] S402-4. For adjacent entities e1 and e2 in the entity set E×E, match the potential relation triples (e1, r, e2) according to the relation directory obtained in step S302-2, and aggregate the corresponding node features using the concatenation method. Concatenate the final GNN feature vectors of e1 and e2 to obtain:

[0069]

[0070] in, and It is the node feature vector after passing through an L-layer GNN;

[0071] S402-5. Input the aggregated feature vector into the MLP (Multilayer Perceptron) classifier to predict the relationship types between entities, and calculate the probability of each relationship type using the Softmax function:

[0072] P(r|e1,e2)=Softmax(W s z+b s )

[0073] Among them, W s and b s These are the weights and biases of the classifier; the relationship is identified by taking the result with the highest probability.

[0074] S403. Identify the attributes in the underwater bridge inspection dataset using a rule matching method;

[0075] The specific steps of step S403 are as follows:

[0076] S403-1. Expand the attribute word list in the field of underwater bridge inspection, summarize the attribute keys obtained through text specification research and entity recognition in the early stage, filter the stop words and low-frequency words in the attribute set, remove the common stop words such as "de", "shi", "zai" that have no practical function and the words with low occurrence frequency, perform word sense disambiguation, perform word sense disambiguation on polysemous words, determine the actual meaning of each attribute, classify them according to their subordinate categories, and expand the attribute word list;

[0077] S403-2. Entity attribute extraction. Based on the attribute word list, use the string numerical brute-force matching algorithm and the orthogonal expression matching algorithm to match the attribute entities extracted from step S401, and extract the attributes of the entities in the text;

[0078] S404. Summarize the entities extracted in S401, the relationships extracted in S402, and the attribute sets extracted in S403 to obtain a triple set of "(entity)-(relationship)-(entity)" and "(entity)-(attribute)-(attribute value)";

[0079] The specific steps of step S404 are as follows:

[0080] S404-1. Construction of the (entity)-(relationship)-(entity) triple. Uniformly establish a directed relationship between the identified entities (e i , e j ) and the relationship r i obtained in step S402 to form an (entity)-(relationship)-(entity) triple (e i , r i , e j );

[0081] S404-2. Construction of the (entity)-(attribute)-(attribute value) triple. In the bridge inspection text, the attributes and key values appear in pairs. Extract the attribute values, filter out the characters with a separating function such as "wei", "shi", ":" in the text, and perform manual verification to form a triple of "(entity)-(attribute)-(attribute value)" (e i , p i , v i ).

[0082] Optionally, the specific steps of step S5 are as follows:

[0083] S501, Bridge Underwater Inspection Knowledge Graph Alignment and Ablation: Based on the cosine similarity principle, this method aligns and ablates multiple entities in all (entity)-(relationship)-(entity) triples.

[0084]

[0085] Among them, e i and e k They are (e) i ,r i ,e i+1 ) and (e k ,r k ,e k+1 ) Entities in two different triples, if entity e i and e k If the cosine value of the distance between the two entities is ≥0.9, it indicates that the two entities are similar. i ,e k Two entities are aligned to become one entity e i Make two triples (e i ,r i ,e i+1 ) and (e k ,r k ,e k+1 ) merge to form a new map structure (e i+1 ,r i ,e i ,r k ,e k+1 The process involves iterative loops to align and merge all triples, thereby constructing a unified and complete knowledge network.

[0086] S502. Knowledge Graph Construction and Visualization: Based on the constructed triple relationships, complex triples are converted into a Cypher language instruction set that the database can read, and then imported into the newly built Neo4j knowledge graph platform to complete the construction of the knowledge graph; by selecting and displaying all nodes in the database, the knowledge graph visualization of the underwater bridge structure detection can be realized.

[0087] Based on the same inventive concept, the present invention provides an electronic device including a processor and a storage medium;

[0088] The storage medium is used to store instructions;

[0089] The processor is configured to operate according to the instructions to perform the steps of the method described above.

[0090] Based on the same inventive concept, the computer-readable storage medium of the present invention stores a computer program thereon, characterized in that the program, when executed by a processor, implements the steps of the method described above.

[0091] Beneficial Effects: Compared with existing technologies, this invention has the following significant advantages: This invention achieves automated and rapid construction and application of a knowledge graph for underwater bridge inspection, greatly saving time and manpower costs in its construction, and solving the current problems of high cost, low automation, and low intelligence in underwater bridge inspection knowledge graph construction; This invention integrates three types of knowledge sources—standards, papers, and reports—in the knowledge graph construction, taking into account standardization, academic rigor, and engineering practice, avoiding the limitations of a single knowledge source; This invention addresses the differences in expression style and format of inspection reports by preprocessing and combining it with a domain thesaurus, achieving unified knowledge extraction, which is superior to general graph construction methods that cannot handle text heterogeneity; The knowledge graph constructed by this invention supports dynamic updates and intuitive visualization, facilitating the subsequent integration of new inspection technologies, defect types, and other knowledge, providing continuous support for decision-making assistance and maintenance of underwater bridge inspections. Attached Figure Description

[0092] Figure 1 This is a schematic diagram of the framework in this invention;

[0093] Figure 2 This is a schematic diagram of the underwater bridge detection entities and their relationships in this invention;

[0094] Figure 3 This is a schematic diagram of the underwater bridge inspection entity annotation in this invention;

[0095] Figure 4 This is a diagram illustrating the underwater bridge detection entity recognition method of the present invention;

[0096] Figure 5 This is a diagram of the relationship recognition model in the field of underwater bridge inspection in this invention;

[0097] Figure 6 This is a diagram of the attribute recognition model for underwater bridge inspection in this invention.

[0098] Figure 7 This is a model diagram for the automatic construction of a knowledge graph for underwater bridge detection in this invention;

[0099] Figure 8 This is a visualization example of the knowledge graph for underwater bridge detection in this invention. Detailed Implementation

[0100] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0101] like Figure 1 As shown, this invention discloses a method for constructing a general knowledge graph for underwater bridge detection, comprising the following steps:

[0102] S1. The general knowledge graph information for underwater bridge inspection includes basic bridge information, basic information on underwater components, inspection methods and technologies, types and characteristics of underwater defects, and maintenance measures. This general knowledge graph information is represented and stored in the graph database in the form of triples of "entity-relationship-entity" and "entity-attribute-attribute value," forming the general knowledge graph for underwater bridge inspection. Basic bridge information includes bridge name, bridge location, year of construction, bridge structural form, and total bridge length. Basic information on underwater components includes component category, component number and location, component type, construction technology, component size, and component material. Inspection methods and technologies include 3D multibeam radar, 3D sonar point cloud, artificial diving exploration, and underwater robot photography. Types and characteristics of underwater defects include defect type, defect location, and defect size. Maintenance measures include riprap protection, pier grooving, sacrificial pile protection, and protective ring protection.

[0103] S2. Obtain knowledge sources in the field of underwater bridge inspection, including standards, papers, and reports. Extract key information from these knowledge sources, perform structured processing, and obtain bridge inspection report text information that is consistent with the standards and specifications.

[0104] Step S2 specifically includes the following steps:

[0105] S201. Obtain knowledge sources in the field of underwater bridge inspection in accordance with standards, process them, and import them into the knowledge base of underwater bridge inspection.

[0106] Step S201 specifically includes the following steps:

[0107] S201-1. Download electronic documents of standards and specifications from the National Standards Information Platform and the official websites of relevant foreign departments and industry associations. Classify the obtained standards and specifications, with current national and industry standards and specifications being considered classic knowledge sources and current local and group standards and specifications being considered non-classical knowledge sources. Assign the following weighting coefficients to each:

[0108]

[0109] in, The level represents the weight of a certain standard or regulation. The level values ​​are: domestic national / industry standards level=5, foreign national / industry standards level=4, domestic local standards level=3, foreign local standards level=2, and domestic and foreign group standards level=1.

[0110] S201-2. According to the organizational structure of the standard document, review each sentence, retain the content related to the underwater structure inspection, management and maintenance, delete irrelevant content, number each sentence of the retained content, and indicate the chapter and clause or clause explanation and appendix in the standard document in which the sentence is located, such as: "

[001] (5.2.6-2) During the flood season, monitor the passage of floating objects under the bridge, and if necessary, use hooks or other means to guide them to pass smoothly through the bridge opening. Floating objects blocking the bridge should be removed or retrieved in time." In this sentence, "

[001] " is the sentence number, and "(5.2.6-2)" is the chapter number and clause number;

[0111] S201-3. Sorting out the core points of the specifications regarding the inspection process, technical requirements, acceptance standards, and maintenance measures for underwater bridge inspection, extracting the core points from the specifications, and converting them into structured information, such as: "

[001] (5.2.6-2) Inspection process: Detect and monitor the passage of floating objects. Handling measures: Use hooks or other means to guide floating objects to pass smoothly through the bridge opening, and promptly remove or retrieve floating objects blocking the bridge.", importing it into the knowledge base for the field of underwater bridge inspection in plain text form.

[0112] S202. Obtain knowledge sources in the field of underwater bridge inspection from academic papers, process them, and import them into the knowledge base of underwater bridge inspection.

[0113] Step S202 specifically includes the following steps:

[0114] S202-1. Conduct keyword searches on the Web of Science (WOS) paper platform, extracting and exporting relevant literature information including article title, author, abstract, journal impact factor, and citation count. Obtain the original text from the paper publisher's platform and categorize the papers according to citation count and publication date. In the field of bridge underwater structure inspection and maintenance journal articles, the definition of classic knowledge sources is: papers published within the last 3 years with a journal impact factor ≥ 5 and WOS citation count ≥ 10, and papers published within the last 3-10 years with a journal impact factor ≥ 5 and WOS citation count ≥ 20; the remaining papers are classified as non-classical knowledge sources. The following weighting coefficients are assigned to the above papers respectively:

[0115]

[0116] in, The weight of a journal article is given by IF, which is the impact factor of the journal in which the article was published, and cite is the number of citations of the article on the WOS platform.

[0117] S202-2, categorized by topic into three types: disease treatment, detection technology, and disease mechanism, listing the applicable conditions, usage steps, and key information related to the final effect of the methods described in the paper, such as: "《Riprap Protection at Bridge Piers》-Disease Treatment-Applicable Conditions: Moving bed water flow environment, various high-speed and low-speed flow conditions-Usage Steps: 1. Determine the pier diameter D and the median particle size d of the bed sand." 50 1. Calculate parameters such as water depth y0 and design flow velocity U; 2. Calculate the Fraud number and critical velocity; 3. Calculate the dimensions of the riprap; 4. Install the riprap, ensuring the dimensions are close to the calculated values, and the riprap layer thickness is 2d. 50 Covering diameter of 4D - Final effect: Changes the main failure mode from bed-induced instability to sediment scouring near the piers, reducing the scour depth by 62.4%;

[0118] S202-3. The information obtained in step S202-2 is processed in a structured manner and imported into the knowledge base for underwater bridge inspection in plain text form.

[0119] S203. Obtain knowledge sources in the field of underwater bridge inspection reports, process them, and obtain standardized bridge inspection report text information.

[0120] Step S203 specifically includes the following steps:

[0121] S203-1. Extract and clean the original underwater bridge structure inspection report. Extract the effective information related to component type, component number, defect type, defect location, and proposed treatment measures from the inspection report. Delete redundant information that is not related to underwater structure inspection, management, and maintenance. Decompose the whole paragraph description in the small-granular text of the report into short sentences and phrases. Store the results in a TXT format document in units of sentences.

[0122] S203-2. Standardize the text by converting English units into Chinese units and English measurement terms into their corresponding Chinese expressions, such as converting “D” into “diameter”, “H” into “depth”, and “m” into “meter”.

[0123] S203-3. Correct and align the text, convert typos and grammatical errors in the report into correct spellings, and unify different expressions and structures of the same meaning in different bridge inspection reports in advance, such as correcting "group foundation pile foundation" to "group pile foundation".

[0124] S203-4. Perform encoding standardization processing on the text, remove special characters such as periods, question marks and exclamation marks at the end of the text, convert all punctuation marks in the sentence to full-width characters, and convert the text information into UTF-8 encoding format that is easy for computers to read and process.

[0125] S203-5. Perform text segmentation. Combining the commonly used terminology for underwater bridge structures in the knowledge base built in S201 and S202, use the SnowNLP Chinese word segmentation tool to segment the text in the cleaned underwater bridge structure inspection report, and finally obtain standardized and consistent bridge inspection report text information.

[0126] S3. Construct a general ontology catalog for underwater bridge inspection: Based on the underwater bridge inspection knowledge base built in steps S201 and S202, construct an entity, relation, and attribute catalog suitable for underwater bridge inspection; according to the entity catalog, use the LabelStudio tool library to annotate the knowledge source entities in the field of underwater bridge inspection for report-type inspection, and obtain the underwater bridge inspection domain dataset.

[0127] Step S3 specifically includes the following steps:

[0128] S301. Based on the knowledge base for underwater bridge inspection constructed in steps S201 and S202, clarify the level L of underwater bridge inspection:

[0129] L=∑ i L i

[0130] Among them, L i The detection levels are defined as follows: i = {1, 2, ..., n}, n = 4; the specific definitions of each detection level are: L1 - bridge level, L2 - structural level, L3 - component level, L4 - component level.

[0131] S302. Based on the knowledge base for underwater bridge inspection constructed in step S2 and the analysis results in step S301, construct the underwater bridge inspection ontology directory, which includes an entity directory, a relationship directory and an attribute directory.

[0132] Step S302 specifically includes the following steps:

[0133] S302-1. Based on the knowledge base for underwater bridge structure inspection constructed in step S2 and the results of S301, define the named entities in the bridge inspection field as 10 categories, map the underwater bridge inspection level L in S301, and add entities for describing defects, attributes and numerical entities for describing component and defect features, location entities for representing positions, negative modifier entities for assisting in text extraction, and measure entities for describing maintenance measures. Establish the entity directory E for the underwater bridge structure inspection field:

[0134]

[0135] Among them, E iTo define entities, i = {1, 2, ..., n}, n = 10; specific entity directory sub-items are defined as: E1-Bridge entity, E2-Structural entity, E3-Component entity, E4-Component entity, E5-Disease entity, E6-Attribute entity, E7-Numerical entity, E8-Location entity, E9-Negative modifier entity, E... 10 - Entities involved in the measures;

[0136] S302-2. Based on the knowledge base for underwater bridge inspection constructed in step S2 and the results of S302-1, remove E6-attribute entities, E7-numerical entities used to construct attribute-key-value pairs, and E9-negation modifier entities used for auxiliary judgment. Analyze the remaining E1-bridge entities, E2-structural entities, E3-component entities, E4-component entities, E5-disease entities, E8-location entities, and E... 10 - The measures involve relationships between 7 types of entities, establishing a relationship directory for the field of underwater bridge structure inspection (R):

[0137]

[0138] Where, r i For the field of underwater structure inspection, i = {1, 2, ..., n}, n = 7; the specific relationship is defined as: r1 - structure subordinate, r2 - component subordinate, r3 - component subordinate, r4 - existing defects, r5 - existing treatment measures, r6 - bridge structure location, r7 - defect location;

[0139] Please see Figure 2 , Figure 2 It explains the six types of relationships that exist between the seven types of entities.

[0140] S302-3. Based on the knowledge base for underwater bridge inspection constructed in step S2 and the E6-attribute entity in the results of S302-1, classify and construct the attributes in the underwater bridge structure inspection text, and build a basic attribute vocabulary P for the underwater bridge structure inspection field:

[0141]

[0142] Where, p i The underwater structure detection attribute categories are i = {1, 2, ..., n}, n = 5; specifically defined as: p1 - entity common attribute, p2 - bridge body attribute, p3 - pier body attribute, p4 - foundation body attribute, p5 - defect attribute.

[0143] S303. Based on the entity catalog constructed in step S302, create a dataset of the standardized bridge inspection report text information obtained in step S203.

[0144] Step S303 specifically includes the following steps:

[0145] S303-1. Annotate the text information. Based on the entity directory E of the bridge maintenance field constructed in step S302-1, use the same-field automatic annotation function of the Label Studio tool library to annotate the entity start part B, the middle part I, and the non-entity part O in the bridge maintenance field text information according to the three-dimensional sequence annotation method, i.e., the BIO method. Then, perform manual correction to correct any errors that may exist in the annotation process, forming a bridge underwater structure detection text dataset. For example, the sentence "The No. 8 pier of the Guxi River Bridge has scraped corners, and no obvious defects are found in other locations" in the report is annotated and corrected to form the sentence "Guxi[B-BRI]Xi[I-BRI]He[I-BRI]Da[I-BRI]Qiao[I-BRI]8[B-MEM]#[I-M [EM] Bridge pier [I-MEM] exists [O] scraped [B-DIS] rubbed [I-DIS] corner [I-DIS], [O] other [O] positions [O] no [B-NEG] visible [B-NEG] obvious [O] damage [O]"; In the bridge entity (BRI) of "Guxi River Bridge", "Gu" is the starting character of the entity, so it is marked as [B-BRI], and the other 4 characters such as "Xi" are the middle part of the entity, so they are marked as [I-BRI]; while "exists" and "other positions" are non-entity parts, so they are marked as [O]. Please refer to Figure 3 This illustrates the text annotation process.

[0146] S303-2. Dataset partitioning: Using uniform random sampling, the dataset is divided into a training set, a validation set, and a test set in a ratio of 8:1:1. The labels of the three datasets are similar.

[0147] S4. Knowledge Source Identification and Calculation in the Bridge Inspection Domain: The BERT-BiLSTM-CRF algorithm is used to identify entities in the underwater bridge inspection dataset, the GNN graph neural network is used to identify relationships in the underwater bridge inspection dataset, and the rule matching method is used to identify attributes in the underwater bridge inspection dataset. The entity, relationship, and attribute sets are summarized to obtain the triple sets of "entity-relationship-entity" and "entity-attribute-attribute value".

[0148] like Figure 4 As shown, step S4 specifically includes the following steps:

[0149] S401. Based on the BERT-BiLSTM-CRF algorithm, identify the entities in the underwater bridge detection dataset labeled in S303, and obtain the extracted entity set;

[0150] Step S401 specifically includes the following steps:

[0151] The S401-1 algorithm for underwater bridge detection and named entity recognition is constructed and trained by integrating the BERT model (a large language model based on the Transformer architecture), the BiLSTM model (bidirectional long short-term memory network), and the CRF model (conditional random field) into a BERT-BiLSTM-CRF model to achieve named entity recognition. The training, validation, and test sets in the S302-2 dataset are input into the BERT-BiLSTM-CRF model to train the model, enabling the algorithm to accurately capture text features in the underwater bridge structure domain and complete the entity extraction task.

[0152] S401-2, Tokenization and Masking: Input the text to be extracted into the BERT model's built-in encoder Tokenizer. Assign text embedding tokens according to a predefined dictionary, perform full word masking on some words, add a sequence beginning token [CLS] to the beginning of the sequence, and separate sentences with clause tokens [SEP]. Output three embedding layers: character embedding Q, text embedding K, and position embedding V.

[0153] S401-3, BERT Large Language Model Feature Extraction: The character embedding Q, text embedding K, and position embedding V output from step S401-2 are imported into the BERT model. The language features of the bridge underwater structure detection text are extracted through the Transformer encoder, multi-head attention mechanism, and fully connected network, and the language feature sequence is output.

[0154] S401-4, BiLSTM context semantic dependency capture: The sequence output from step S401-3 is imported into the BiLSTM model. Useless information is filtered out through forward and backward propagation and related mechanisms such as input gate, output gate and forget gate, and effective information is passed. The bidirectional semantic dependency of the context information of the underwater bridge detection text is captured and the sequence is output.

[0155] S401-5, CRF sequence label prediction: The output of step S401-4 is imported into a conditional random field. Based on the dependency relationship between adjacent entity labels, the predicted sequence is scored and its probability is evaluated. The highest-scoring predicted sequence is obtained by decoding the maximum likelihood function. The entity with the highest score is selected and summarized to obtain the bridge underwater detection entity set E×E.

[0156] S402. Use GNN (Graph Neural Network) to identify the relationships between entities in the underwater bridge detection dataset extracted in step S401.

[0157] like Figure 5As shown, step S402 specifically includes the following steps:

[0158] S402-1. Convert the underwater bridge detection text data into a graph structure. The nodes in the graph correspond to the entities in the entity set E×E extracted in S401, and the edges represent the relationships between entities. Construct a dependency graph based on the syntactic analysis results, where entities are nodes and dependency relationships are edges.

[0159] S402-2. Use the Word2Vec word embedding method to obtain word vector representations, thereby capturing contextual semantic information. Assign an embedding vector to each entity type, integrating entity type information into node features. Assuming the entity type set is T, for an entity of type t∈T, its type embedding is e. t The initial feature vector x of the node i It is obtained by concatenating word embeddings and entity type embeddings, and its expression is x. i =[h i ;e t ];

[0160] S402-3, GNN layer propagation, aggregates information from neighboring nodes through a message passing mechanism to update the features of the current node; wherein, the feature update formula for the node in the l-th layer is:

[0161]

[0162] in, It is the feature vector of node i in the (l+1)th layer. W is the set of neighboring nodes of node i. l This is the weight matrix of the l-th layer, and σ is the ReLU activation function. It is the feature vector of node j in the l-th layer.

[0163] S402-4. For adjacent entities e1 and e2 in entity set E×E, match potential relation triples (e1, r, e2) according to the relation directory obtained in step S302-2, and aggregate the corresponding node features using the concatenation method. Concatenate the final GNN feature vectors of e1 and e2 to obtain the following:

[0164]

[0165] in, and It is the node feature vector after passing through an L-layer GNN;

[0166] S402-5. Input the aggregated feature vector into the MLP (Multilayer Perceptron) classifier to predict the relationship types between entities, and calculate the probability of each relationship type using the Softmax function:

[0167] P(r|e1,e2) = Softmax(W s z + b s )

[0168] where W s and b s are the weights and biases of the classifier, and the result with the highest probability is taken as the recognized relationship.

[0169] As shown in Figure 6 , in S403, the rule matching method is used to identify the attributes in the underwater bridge detection dataset;

[0170] Step S403 specifically includes the following steps:

[0171] S403-1. Expand the attribute word list for the underwater bridge detection field.汇总 the attribute keys obtained through text specification research and entity recognition in the early stage, filter the stop words and low-frequency words in the attribute set, remove common stop words such as "的", "是", "在" that have no practical effect and words with low occurrence frequencies, perform word sense disambiguation, perform word sense disambiguation on polysemous words, determine the actual meanings of each attribute, and classify them according to their subordinate categories to expand the attribute word list;

[0172] S403-2. Entity attribute extraction. Based on the attribute word list, use the brute-force string matching algorithm and the orthogonal expression matching algorithm to match the attribute entities extracted from step S401, and extract the attributes of the entities in the text.

[0173] S404.汇总 the entities extracted in S401, the relationships extracted in S402, and the attribute set extracted in S403 to obtain the "entity-relationship-entity" and "entity-attribute-attribute value" triple sets;

[0174] The specific steps of step S404 include the following steps:

[0175] S404-1. Construction of the (entity)-(relationship)-(entity) triple.统一 establish a directed relationship between the recognized entities (e i , e j ) and the relationship r i obtained in step S402 to form the (entity)-(relationship)-(entity) triple (e i , r i , e j );

[0176] S404-2. Construction of the (entity)-(attribute)-(attribute value) triple. In the bridge detection text, the attributes and key values appear in pairs. Extract the attribute values, filter out the characters with delimiter functions such as "为", "是", ":" in the text, and perform manual verification to form the "entity-attribute-attribute value" triple (e It should be noted that there are some inaccuracies and unclear expressions in the original Chinese text, and the translation is adjusted as accurately as possible based on the existing content.i ,p i ,v i ).

[0177] S5. Construction of a general underwater inspection map for bridges: Based on the triplet set obtained in step S4, entity alignment and ablation are performed. Based on the Neo4j knowledge graph data visualization platform and Cypher instruction set, a general underwater inspection map for bridges is constructed.

[0178] like Figure 7 As shown, step S5 specifically includes the following steps:

[0179] S501, Bridge Underwater Inspection Knowledge Graph Alignment and Ablation: Based on the cosine similarity principle, this method aligns and ablates multiple entities in all (entity)-(relationship)-(entity) triples.

[0180]

[0181] Among them, e i and e k They are (e) i ,r i ,e i+1 ) and (e k ,r k ,e k+1 ) Entities in two different triples, if entity e i and e k If the cosine value of the distance between the two entities is ≥0.9, it indicates that the two entities are similar. i ,e k Two entities can be aligned to form a single entity e. i Make two triples (e i ,r i ,e i+1 ) and (e k ,r k ,e k+1 ) merge to form a new map structure (e i+1 ,r i ,e i ,r k ,e k+1 The process involves iterative loops to align and merge all triples, thereby constructing a unified and complete knowledge network.

[0182] S502, Knowledge Graph Construction and Visualization: Based on the constructed triple relationships, complex triples are converted into a Cypher language command set that the database can read, and imported into the newly built Neo4j knowledge graph platform to complete the construction of the knowledge graph; selecting and displaying all nodes in the database enables visualization of the bridge underwater structure detection knowledge graph, such as... Figure 8As shown, this is the knowledge graph constructed after S5 is completed.

[0183] Example 2: An electronic device according to the present invention includes a processor and a storage medium;

[0184] Storage media are used to store instructions;

[0185] The processor is configured to operate according to the instructions to perform the steps of the method described above.

[0186] Example 3: The computer-readable storage medium of the present invention stores a computer program thereon, which, when executed by a processor, implements the steps of the method described above.

Claims

1. A method for constructing a general knowledge graph for underwater bridge inspection, characterized in that, Includes the following steps: S1. Define the general knowledge graph information for underwater bridge inspection. The general knowledge graph information for underwater bridge inspection includes basic bridge information, basic information of underwater components, inspection methods and technologies, types and characteristics of underwater defects, and maintenance measures. S2. Obtain knowledge sources in the field of underwater bridge inspection, including standards, papers, and reports; extract key information from the knowledge sources; perform structured processing; and obtain bridge inspection report text information that is consistent with the underwater bridge inspection knowledge base and standards. S3. Construct a general ontology catalog for underwater bridge inspection: Based on the constructed underwater bridge inspection knowledge base, construct an entity, relation, and attribute catalog applicable to underwater bridge inspection; according to the entity catalog, use the Label Studio tool library to annotate knowledge source entities in the field of underwater bridge inspection for report-type inspection, and obtain a dataset for the field of underwater bridge inspection. Step S3 specifically includes the following steps: S301. Based on the constructed knowledge base for underwater bridge inspection, clarify the level L of underwater bridge inspection: ， Among them, L i The detection levels are defined as follows: i = {1, 2, ..., n}, n = 4; the specific definitions of each detection level are: L1 - bridge level, L2 - structural level, L3 - component level, L4 - component level. S302. Based on the knowledge base for underwater bridge inspection constructed in step S2 and the analysis results in step S301, construct the underwater bridge inspection ontology directory, which includes an entity directory, a relationship directory and an attribute directory. Step S302 specifically includes the following steps: S302-1. Based on the knowledge base for underwater bridge structure inspection constructed in step S2 and the results of S301, define the named entities in the bridge inspection field as 10 categories, map the underwater bridge inspection level L in S301, and add entities for describing defects, attributes and numerical entities for describing component and defect features, location entities for representing positions, negative modifier entities for assisting in text extraction, and measure entities for describing maintenance measures. Establish the entity directory E for the underwater bridge structure inspection field: ， Among them, E i To define an entity, n=10; the specific entity directory sub-items are defined as follows: E1-Bridge entity, E2-Structural entity, E3-Component entity, E4-Component entity, E5-Disease entity, E6-Attribute entity, E7-Numerical entity, E8-Location entity, E9-Negative modifier entity, E 10 - Entities involved in the measures; S302-2. Based on the knowledge base for underwater bridge inspection constructed in step S2 and the results of S302-1, remove E6-attribute entities, E7-numerical entities used to construct attribute-key-value pairs, and E9-negation modifier entities used for auxiliary judgment. Analyze the remaining E1-bridge entities, E2-structural entities, E3-component entities, E4-component entities, E5-disease entities, E8-location entities, and E... 10 - The measures involve relationships between 7 types of entities, establishing a relationship directory for the field of underwater bridge structure inspection (R): ， in, This relates to the field of underwater structure inspection. , n=7; the specific relationships are defined as follows: r1-structure subordinate, r2-component subordinate, r3-part subordinate, r4-existing defects, r5-existing treatment measures, r6-bridge structure located, r7-defect located; S302-3. Based on the knowledge base for underwater bridge inspection constructed in step S2 and the E6-attribute entity in the results of S302-1, classify and construct the attributes in the underwater bridge structure inspection text, and build a basic attribute vocabulary P for the underwater bridge structure inspection field: ， in, For underwater structure detection attribute categories, n=5; specifically defined as: p1-entity common attributes, p2-bridge body attributes, p3-pier body attributes, p4-foundation body attributes, p5-disease attributes; S303. Based on the entity catalog constructed in step S302, create a dataset of the standardized bridge inspection report text information obtained in step S203. Step S303 specifically includes the following steps: S303-1. Annotate the text information. Based on the entity directory E of the bridge maintenance field constructed in step S302-1, use the same field automatic annotation function of Label Studio tool library to annotate the entity start part B, the middle part I and the non-entity part O in the text information of the bridge maintenance field according to the three-dimensional sequence annotation method, namely BIO method. Then, perform manual correction to correct any errors that may exist in the annotation process, and form a bridge underwater structure detection text dataset. S303-2. Dataset partitioning: Using uniform random sampling, the dataset is divided into a training set, a validation set, and a test set. S4. Knowledge Source Identification and Calculation in the Bridge Inspection Domain: The BERT-BiLSTM-CRF algorithm is used to identify entities in the underwater bridge inspection dataset, the GNN graph neural network is used to identify relationships in the underwater bridge inspection dataset, and the rule matching method is used to identify attributes in the underwater bridge inspection dataset. The entity, relationship, and attribute sets are summarized to obtain the triple sets of "entity-relationship-entity" and "entity-attribute-attribute value". S5. Construction of a general underwater inspection map for bridges: Based on the triplet set obtained in step S4, entity alignment and ablation are performed. Based on the Neo4j knowledge graph data visualization platform and Cypher instruction set, a general underwater inspection map for bridges is constructed.

2. The method for constructing a general knowledge graph for underwater bridge inspection according to claim 1, characterized in that: The basic information of the bridge in S1 includes the bridge name, bridge location, bridge construction year, bridge structural form and total length. The basic information of the underwater components includes component category, component number and location, component type, construction technology, component size and component material. The detection methods and technologies include three-dimensional multibeam radar, three-dimensional sonar point cloud, artificial diving exploration and underwater robot photography. The types and characteristics of underwater defects include defect category, defect location and defect size. The maintenance measures include rock dumping protection, pier grooving, sacrificial pile protection and protective ring protection.

3. The method for constructing a general knowledge graph for underwater bridge inspection according to claim 1, characterized in that: Step S2 specifically includes the following steps: Knowledge sources in the field of underwater bridge inspection are acquired, processed, and imported into a knowledge base for underwater bridge inspection. This involves downloading electronic documents of standards and regulations from the National Standards Information Platform and official websites of relevant international departments and industry associations. These standards and regulations are then categorized, with current national and industry standards representing classic knowledge sources, and current local and group standards representing non-classical knowledge sources. The following weighting coefficients are assigned to each category: ， in, As the weight of a certain standard, To standardize the hierarchy, the values ​​are: domestic national / industry standards level=5, foreign national / industry standards level=4, domestic local standards level=3, foreign local standards level=2, and domestic and foreign group standards level=1. According to the organizational structure of the standard document, review it sentence by sentence, retain the content related to the field of underwater structure inspection, management and maintenance, delete irrelevant content, number each sentence of the retained content, and indicate the chapter and clause or clause explanation and appendix of the standard document in which the sentence is located; We sorted out the core points of the standards regarding underwater bridge inspection procedures, technical requirements, acceptance standards, and maintenance measures, extracted the core points from the standard clauses, and converted them into structured information.

4. The method for constructing a general knowledge graph for underwater bridge inspection according to claim 1, characterized in that: Step S2 specifically includes the following steps: Knowledge sources in the field of underwater bridge inspection are acquired from academic papers, processed, and imported into a knowledge base for the field. This involves keyword searching on the Web of Science platform to extract and export relevant literature information, including article title, author, abstract, journal impact factor, and citation count. The original texts are also retrieved from publisher platforms, and papers are categorized based on citation count and publication date. For journal articles in the field of underwater bridge structure inspection and maintenance, classic knowledge sources are defined as: papers published within the last 3 years with a journal impact factor ≥ 5 and WOS citation count ≥ 10; and papers published within the last 3-10 years with a journal impact factor ≥ 5 and WOS citation count ≥ 20. All other papers are classified as non-classical knowledge sources. The following weighting coefficients are assigned to each of these papers: ， in, The weight of a journal article The impact factor of the journal in which the paper was published. This represents the number of citations to this paper on the WOS platform; The topics are divided into three categories: disease treatment, detection technology, and disease mechanism. The key information related to the applicable conditions, usage steps, and final effects of the methods described in the papers are listed one by one. The obtained information is structured and imported into a knowledge base for underwater bridge inspection in plain text format.

5. The method for constructing a general knowledge graph for underwater bridge inspection according to claim 1, characterized in that: Step S2 specifically includes the following steps: The process involves acquiring knowledge sources in the field of underwater bridge inspection reports, processing them, and importing them into the underwater bridge inspection knowledge base. Specifically, the original underwater bridge structure inspection reports undergo text extraction and cleaning. This includes extracting relevant information such as component type, component number, defect type, defect location, and proposed treatment measures. Redundant information irrelevant to underwater structure inspection, management, and maintenance is removed. The reports are broken down into smaller, granular descriptions, from complete paragraphs to short sentences and phrases. The results are then stored in TXT format documents, sentence by sentence. The text is standardized by converting English units into Chinese units and English measurement terms into their corresponding Chinese expressions. Correct and align the text, converting typos and grammatical errors in the report into correct spelling, and pre-unify different expressions and structures of the same semantic meaning in different bridge inspection reports; The text is encoded to be consistent, removing special characters such as periods, question marks and exclamation marks at the end of the text, converting all punctuation marks in the sentence to full-width characters, and converting the text information into UTF-8 encoding format that is easy for computers to read and process; Text segmentation is performed, and combined with the commonly used vocabulary of underwater bridge structures in the constructed knowledge base for underwater bridge inspection, the SnowNLP Chinese word segmentation tool is used to segment the text in the cleaned underwater bridge structure inspection report, finally obtaining standardized and consistent bridge inspection report text information.

6. The method for constructing a general knowledge graph for underwater bridge inspection according to claim 5, characterized in that: Step S4 specifically includes the following steps: S401. Based on the BERT-BiLSTM-CRF algorithm, identify the entities in the underwater bridge detection dataset labeled in S303, and obtain the extracted entity set; Step S401 specifically includes the following steps: S401-1, Construction and Training of BERT-BiLSTM-CRF Named Entity Recognition Algorithm for Underwater Bridge Detection: This paper integrates the BERT model, BiLSTM model, and CRF model into a BERT-BiLSTM-CRF model to achieve the function of named entity recognition; The training set, validation set, and test set of the S302-2 dataset are input into the BERT-BiLSTM-CRF model for training. S401-2, Text Tokenization and Masking: Input the text to be extracted into the BERT model's built-in encoder Tokenizer. Assign text embedding tokens according to a predefined dictionary, perform full word masking on some words, add a sequence beginning token [CLS] to the beginning of the sequence, and separate sentences with clause tokens [SEP]. Output three embedding layers: character embedding Q, text embedding K, and position embedding V. S401-3. Feature extraction of the BERT large language model. The output character embedding Q, text embedding K, and position embedding V from step S401-2 are imported into the BERT model as inputs. The language features of the underwater bridge structure detection text are extracted through the Transformer encoder, multi-head attention mechanism, and fully connected network, and a language feature sequence is output. S401-4. Capturing context semantic dependencies of BiLSTM. The sequence output from step S401-3 is imported into the BiLSTM model. Useless information is filtered out through forward and backward propagation and mechanisms related to input gates, output gates, and forget gates, and effective information is passed to capture the bidirectional semantic dependencies of the context information of the underwater bridge detection text for sequence output. S401-5. CRF sequence label prediction. The output from step S401-4 is imported into the conditional random field. According to the dependencies between adjacent entity labels, the predicted sequence is scored and probabilistically evaluated. The predicted sequence with the highest score is obtained by decoding the maximum likelihood function, and the entity with the highest score is taken and aggregated to obtain the underwater bridge detection entity set E×E. S402. Use the GNN graph neural network to identify the relationships between the entities extracted from the underwater bridge detection domain dataset in step S401. The specific steps of step S402 are as follows: S402-1. Convert the underwater bridge detection text data into a graph structure. The nodes in the graph correspond to the entities in the entity set E×E extracted in S401, and the edges represent the connections between the entities. A dependency graph is constructed based on the syntactic analysis results, where the entities are used as nodes and the dependency relationships are used as edges. S402-2. Use the Word2Vec word embedding method to obtain word vector representations, and then capture the context semantic information. An embedding vector is assigned to each entity type, and the entity type information is incorporated into the node features. Assume the entity type set is For type The entity whose type is embedded as ; Initial feature vector of the node It is obtained by concatenating word embeddings and entity type embeddings, and its expression is: ; S402-3. GNN layer propagation. The information of neighboring nodes is aggregated through the message passing mechanism to update the features of the current node. The node feature update formula for the l-th layer is: ， in, It is the first The feature vector of layer node i It is the set of neighboring nodes of node i. It is the first The layer's weight matrix, where σ is the ReLU activation function. Let be the feature vector of node j in the l-th layer; S402-4, For adjacent entities in entity set E×E and Match potential relation triples based on the relation directory obtained in step S302-2. The corresponding node features are aggregated using a concatenation method. , The final GNN feature vectors are concatenated to obtain ;in, and It is the node feature vector after passing through an L-layer GNN; S402-5. Input the aggregated feature vectors into the MLP multi-layer perceptron classifier to predict the relationship types between entities, and calculate the probability of each relationship type through the Softmax function: ， in, and These are the weights and biases of the classifier; the relationship is identified by taking the result with the highest probability. S403. Use the rule matching method to identify the attributes in the underwater bridge detection dataset. The specific steps of step S403 are as follows: S403-1. Expand the attribute vocabulary of the underwater bridge detection domain. Summarize the attribute keys obtained through previous text specification research and entity recognition, filter the stop words and low-frequency words in the attribute set, remove common stop words such as "de", "shi", "zai" that have no practical effect and words with low occurrence frequencies, perform word sense disambiguation on polysemous words, determine the actual meanings of each attribute, and classify them according to their subordinate categories to expand the attribute vocabulary. S403-2. Entity attribute extraction. Based on the attribute vocabulary, use the string numerical brute-force matching algorithm and orthogonal expression matching algorithm to match the attribute entities extracted from step S401, and extract the attributes of the entities in the text. S404. Summarize the entities extracted from S401, the relations extracted from S402, and the attribute sets extracted from S403 to obtain the triple sets of "entity-relationship-entity" and "entity-attribute-attribute value". Step S404 specifically includes the following steps: S404-1, (entity)-(relationship) triple construction, uniformly identifying entities ( , The relationship between the two steps is as follows: Establish directed relations to form (entity)-(relation)-(entity) triples. , , ); S404-2. Construction of (entity)-(attribute)-(attribute value) triples. In the bridge inspection text, attributes and key values appear in pairs. Extract the attribute values, filter out the characters with a separating function such as "is", "are", and ":", and conduct manual verification to form "(entity-attribute-attribute value)" triples , , ).

7. The method for constructing a general knowledge graph for underwater bridge inspection according to claim 6, characterized in that: Step S5 specifically includes the following steps: S501, Bridge Underwater Inspection Knowledge Graph Alignment and Ablation: Based on the cosine similarity principle, this method aligns and ablates multiple entities in all (entity)-(relationship)-(entity) triples. ， in, and They are ( , , )and( , , The entities in two different triples, if the entity and Cosine value between This indicates that the two entities are similar. , Two entities are aligned to become one entity. Make two triples ( , , )and( , , ) merge to form a new map structure ( , , , , The process involves iterative loops to align and merge all triples, thereby constructing a unified and complete knowledge network. S502. Knowledge Graph Construction and Visualization: Based on the constructed triple relationships, complex triples are converted into a Cypher language instruction set that the database can read, and then imported into the newly built Neo4j knowledge graph platform to complete the construction of the knowledge graph; by selecting and displaying all nodes in the database, the knowledge graph visualization of the underwater bridge structure detection can be realized.

8. An electronic device, characterized in that, Including processor and storage media; The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps of the method according to any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the method according to any one of claims 1 to 7.

Images