A knowledge graph construction method and system for shale oil measure evaluation

Abstract

The application discloses a knowledge graph construction method and system for shale oil measure evaluation, which comprises the following steps: building a three-level hierarchical ontology framework for shale oil measure evaluation, and matching a shale oil exclusive term dictionary to define the standard semantic association relationship between entity categories, entity subclasses, inherent attributes and entities; preprocessing shale oil multi-source heterogeneous data, constructing a shale oil annotated data set, and extracting knowledge triples from the shale oil annotated data set; performing entity normalization, relationship logic verification and quality review on the knowledge triples, and importing the verified knowledge triples into a graph database to store a knowledge graph; receiving a query statement for shale oil measure evaluation, performing linkage evaluation reasoning through the knowledge graph and a large language model, and outputting a shale oil measure evaluation result. The method solves the problem that the prior art cannot meet the requirements of dynamic, refined and intelligent measure evaluation under the background of large-scale shale oil development.

Description

Technical Field

[0001] This application relates to the fields of intelligent petroleum engineering and knowledge engineering technology, and in particular to a method and system for constructing a knowledge graph for shale oil measure evaluation. Background Technology

[0002] As shale oil enters the stage of large-scale exploration and development, the frequency and scale of various production enhancement and modification measures are continuously increasing. The effectiveness of these measures has become a core factor determining the productivity of a single well and the economic benefits of block development. Shale oil measure evaluation is a highly complex system engineering project, requiring the integration of multi-source heterogeneous data, including static geological data, dynamic parameters of fracturing operations, long-term production monitoring data, field fault logs, expert tacit experience, and historical effect analysis reports. This data is not only diverse in type but also closely linked in terms of underlying professional business logic. Knowledge graphs, with their structured representation of domain knowledge, semantic association retrieval, and logical reasoning capabilities, have been gradually applied in the oil and gas industry. Currently, four mainstream implementation paths have emerged in the oil and gas field, centered around knowledge graphs, intelligent evaluation, and large language model-assisted solutions: Category 1: General Oil and Gas Knowledge Graph Construction Solution. This solution utilizes general natural language processing tools and basic ontology models to uniformly collect, clean, and extract knowledge from oil and gas industry data. It extracts general entities such as oil wells, formations, equipment, and personnel, along with basic relationships, and stores them using a general graph database. Its core functions focus on basic scenarios such as semantic retrieval and data querying. However, it lacks a targeted design for the business logic of shale oil production enhancement measures. Its ontology system and extraction rules are geared towards a broad oil and gas scenario, lacking specific support for measure evaluation indicators, effect judgment rules, and anomaly attribution logic, thus failing to meet professional constraints.

[0003] The second category is offline oil and gas evaluation knowledge graph solutions. This type of solution involves a simple association between evaluation indicators, basic data, and the graph. It typically employs a one-time offline construction model, where data remains in a fixed state after being stored in the database, relying on periodic manual updates. The graph serves only as a storage medium for the evaluation data. However, this offline graph solution suffers from poor data timeliness and low coupling between the graph and the evaluation model. Essentially, it remains a "manual + database" model, unable to achieve logical tracing of measures, automatic benchmarking against similar measures, or knowledge-based intelligent reasoning.

[0004] The third category: Traditional shale oil treatment evaluation techniques. These primarily rely on manual review of records and historical data, combined with expert experience, to score measures and determine their effectiveness. Some methods simply store structured data in conventional databases for basic information statistics and queries. However, this approach fails to transform unstructured data and expert experience into computable knowledge, resulting in insufficient value extraction from historical data and a lack of transparency in the evaluation process, making it difficult to trace the basis of decision-making.

[0005] The fourth category: Pure Large Language Model (LLM) oil and gas evaluation scheme. This scheme relies on traditional machine learning models, fitting sample data to complete the evaluation task, or directly calling a pure large language model (LLM) to conduct question-and-answer and effect analysis of measures, relying on the general semantic understanding capability of the large language model to output evaluation conclusions. However, traditional machine learning models have limited generalization ability in new blocks, new measures, or complex operating conditions, and the output results lack interpretable evidence; when pure large language models are used for direct analysis, they lack structured domain knowledge constraints, which can easily produce professional errors that do not conform to the laws of technology, geology, or production capacity, resulting in serious deficiencies in the credibility and traceability of the evaluation conclusions.

[0006] In summary, the existing technology system lacks an integrated technical solution that deeply integrates domain ontology, knowledge graph, large language model, incremental updates, and edge-cloud collaborative deployment for shale oil measure evaluation, making it difficult to meet the current needs of large-scale, refined, and dynamic measure evaluation in shale oil development. Summary of the Invention

[0007] This application provides a knowledge graph construction method and system for shale oil measure evaluation, which solves the problem that existing technologies are unable to meet the requirements of dynamic, refined and intelligent measure evaluation in the context of large-scale shale oil development.

[0008] In a first aspect, embodiments of this application provide a knowledge graph construction method for shale oil measure evaluation, including: A three-level hierarchical ontology framework for evaluating shale oil measures was constructed, along with a shale oil-specific terminology dictionary to define the standard semantic relationships between entity categories, entity subcategories, inherent attributes, and entities. Based on the hierarchy and evaluation business dimensions of the three-level hierarchical ontology framework, hierarchical and targeted preprocessing is performed on multi-source heterogeneous shale oil data to construct a shale oil labeled dataset. Knowledge is extracted from the shale oil labeled dataset to obtain knowledge triples containing entities, inherent attributes, and relationships; The knowledge triples are subjected to entity normalization, relational logic verification and quality audit for the shale oil field. The verified knowledge triples are then imported into a graph database to store the knowledge graph. The system receives a query statement for evaluating shale oil measures, extracts primary entities, secondary entities, and evaluation intent from the query statement, and performs semantic similarity matching in the knowledge graph based on the extracted information to retrieve related knowledge subgraphs. It then constructs domain-enhanced prompt text by fusing the query statement and the related knowledge subgraphs, and inputs it into a large language model. The large language model, combined with business knowledge in the shale oil domain, performs reasoning analysis and outputs the evaluation results of shale oil measures.

[0009] In conjunction with the first aspect, one possible implementation involves constructing a three-level hierarchical ontology framework for shale oil measure evaluation, along with a shale oil-specific terminology dictionary, to define entity categories, entity subcategories, inherent attributes, and standard semantic relationships between entities, including: Construct a three-level hierarchical ontology framework comprising a top-level ontology, a middle-level ontology, and a bottom-level ontology; The top-level ontology is used to classify all knowledge objects in the field of shale oil measure evaluation into multiple first-level entities; wherein, the multiple first-level entities include shale oil well entities, production enhancement measure entities, geological environment entities, evaluation index entities, and construction failure and optimization scheme entities; The middle-layer ontology is used to further subdivide each of the first-level entities into multiple second-level entities; The underlying ontology is used to define the inherent attributes of each first-level entity and the corresponding second-level entity, and to standardize various standard semantic associations between each first-level entity and / or second-level entity; Based on a shale oil-specific terminology dictionary, we will unify the industry's synonymous expressions to establish a unified semantic benchmark.

[0010] In conjunction with the first aspect, in one possible implementation, the method of further subdividing each of the first-level entities into multiple second-level entities using the middle-level ontology includes: The aforementioned production enhancement measures are further subdivided into several secondary entities, including: hydraulic fracturing, volumetric fracturing, carbon dioxide huff and puff, profile control and water shut-off, and repeated fracturing. The evaluation index entity is further subdivided into multiple secondary entities, including: construction index, production capacity index, economic index, and risk and failure index; The construction faults in the construction fault and optimization scheme entity are further subdivided into several secondary entities, including: sand blockage, proppant backflow, insufficient formation energy, and tubing puncture leakage. The shale oil well entity and the geological environment entity are provided with an expandable subdivision interface to subdivide secondary entities according to the block geological conditions and well type.

[0011] In conjunction with the first aspect, in one possible implementation, the hierarchical and targeted preprocessing of multi-source heterogeneous shale oil data based on the hierarchy and evaluation business dimensions of the three-level hierarchical ontology framework to construct a shale oil labeled dataset includes: Based on the hierarchy and evaluation business dimensions of the three-level hierarchical ontology framework, the multi-source heterogeneous data of shale oil is divided into the construction measures data layer, the reservoir geological data layer, the production dynamic evaluation data layer, and the expert knowledge text data layer. Outlier removal, field normalization, and dimension unification are performed on the structured data in the construction data layer, reservoir geological data layer, and production dynamic evaluation data layer. A preliminary mapping relationship is established between the processed structured data and the first-level or second-level entities in the three-level hierarchical ontology framework. For the semi-structured data in the construction data layer, reservoir geological data layer and production dynamic evaluation data layer, after parsing their table and form structures, multiple core attribute fields are extracted from them, and each core attribute field is matched to the second-level entity of the three-level hierarchical ontology framework. The unstructured data in the expert knowledge text data layer is processed by text segmentation, noise removal, and invalid fragment filtering to retain the core text content; Entity annotations are performed on the preprocessed data of each layer to construct the shale oil annotation dataset.

[0012] In conjunction with the first aspect, in one possible implementation, the knowledge extraction from the shale oil labeled dataset to obtain knowledge triples containing entities, inherent attributes, and relations includes: Entity extraction is performed on the shale oil annotation dataset using a sequence labeling model to obtain a first knowledge triple containing entity boundary and category label information. The sequence labeling model comprises a pre-trained language encoding layer, a bidirectional long short-term memory network layer, and a conditional random field layer, all cascaded sequentially. The pre-trained language encoding layer performs vector encoding on the text. The bidirectional long short-term memory network layer extracts contextual semantic features based on the vector encoding and predicts entity boundary markers and category labels for each word, as well as first-level or second-level entities. The conditional random field layer globally constrains the entity boundary markers and category labels, outputting an entity label sequence. A large language model is used to jointly extract entities, inherent attributes, and semantic relationships from the shale oil annotation dataset to obtain a second knowledge triple. This triple includes: customizing exclusive prompt words for various entities and embedding domain annotation examples; identifying first-level entities, second-level entities, inherent attributes of various entities, and semantic relationships between different entities within the text; vectorizing the identified entities and comparing their semantic similarity with the shale oil-specific terminology dictionary to filter out suspicious entities and generate verification questions; and using the large language model to perform self-checking on the verification questions, outputting the verified entities, inherent attributes, and semantic relationships. The entity tag sequence of the first knowledge triple, the inherent attributes and semantic relationships in the second knowledge triple, and the unified output format are combined to generate a standardized knowledge triple.

[0013] In conjunction with the first aspect, in one possible implementation, the step of performing entity normalization, relational logic verification, and quality auditing on the knowledge triples for the shale oil domain, and importing the verified knowledge triples into a graph database to store the knowledge graph, includes: Based on the shale oil-specific terminology dictionary, the entities in the knowledge triples are normalized, synonymous entities are merged, ambiguous entities are distinguished, and evaluation indicators and measure names with different expressions are unified. Based on the semantic rules of the three-level hierarchical ontology framework, the association relationships between entities in the knowledge triples are logically verified, and erroneous association relationships that do not conform to the semantic rules are automatically screened and eliminated. The knowledge triples that have undergone entity normalization and relational logic verification are subjected to quality audits. The qualified knowledge triples are then imported into a graph database to construct a knowledge graph for shale oil measure evaluation.

[0014] In conjunction with the first aspect, in one possible implementation, after importing the verified knowledge triples into the graph database to store the knowledge graph, the method further includes: An edge and cloud collaborative layered deployment architecture is established to access new data from shale oil fields in real time to complete incremental knowledge updates. Based on the new data, the method described in the first aspect above is executed to update the knowledge graph.

[0015] Secondly, embodiments of this application provide a knowledge graph construction system for shale oil treatment evaluation, including: The definition module is used to build a three-level hierarchical ontology framework for evaluating shale oil measures, and is equipped with a shale oil-specific terminology dictionary to define the standard semantic relationships between entity categories, entity subcategories, inherent attributes, and entities. The preprocessing module is used to perform hierarchical and targeted preprocessing on multi-source heterogeneous shale oil data based on the hierarchy and evaluation business dimensions of the three-level hierarchical ontology framework, so as to construct a shale oil labeled dataset. The extraction module is used to extract knowledge from the shale oil labeled dataset to obtain knowledge triples containing entities, inherent attributes, and relationships. The review module is used to perform entity normalization, relational logic verification and quality review for the knowledge triples in the shale oil field, and import the qualified knowledge triples into the graph database to store the knowledge graph. The reasoning module receives queries for shale oil measure evaluation, extracts primary and secondary entities and evaluation intent from the queries, and performs semantic similarity matching in the knowledge graph based on the extracted information to retrieve related knowledge subgraphs. It then constructs domain-enhanced prompt text by fusing the query and the related knowledge subgraphs, and inputs it into a large language model. The large language model, combined with shale oil domain business knowledge, performs reasoning analysis and outputs the shale oil measure evaluation results.

[0016] Thirdly, embodiments of this application provide an apparatus comprising: a processor; a memory for storing processor-executable instructions; wherein, when the processor executes the executable instructions, it implements the method as described in the first aspect or any possible implementation of the first aspect.

[0017] Fourthly, embodiments of this application provide a non-volatile computer-readable storage medium, the non-volatile computer-readable storage medium including storage for storing a computer program or instructions that, when executed, cause the method described in the first aspect or any possible implementation of the first aspect to be implemented.

[0018] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages: This application's embodiments construct a three-level hierarchical ontology framework for shale oil measure evaluation. This framework accurately anchors the knowledge boundaries of the shale oil measure evaluation field. A dedicated terminology dictionary unifies industry semantic benchmarks, addressing the issue of insufficient ontology adaptability across the broader oil and gas domain. Hierarchical targeted preprocessing adapts to the characteristics of multi-source heterogeneous shale oil data. A combined extraction method using sequence labeling models and large language models balances the accuracy of entity extraction with the completeness of inherent attributes and relationships, effectively improving the professional accuracy of knowledge extraction. Through the linked reasoning of knowledge graphs and large language models, structured domain knowledge forms professional constraints on the output of the large language model, reducing the probability of professional errors. Furthermore, the semantic understanding capabilities of the large language model enable traceable intelligent evaluation reasoning, solving the problems of insufficient credibility of pure large language model evaluations and insufficient data value mining in traditional manual evaluations. This effectively addresses the difficulty of existing technologies meeting the dynamic, refined, and intelligent requirements of measure evaluation in the context of large-scale shale oil development. It is adaptable to the evaluation needs of large-scale and dynamic shale oil development and provides reliable intelligent technical support for the accurate evaluation of shale oil production enhancement measures. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 A flowchart illustrating a knowledge graph construction method for shale oil measure evaluation provided in this application embodiment; Figure 2 An architecture diagram of a knowledge graph construction system for shale oil measure evaluation provided in this application embodiment; Figure 3 A structural diagram of the three-level hierarchical ontology framework provided in the embodiments of this application; Figure 4 This is a schematic diagram of the entity-relationship topology of the knowledge graph provided in the embodiments of this application. Detailed Implementation

[0021] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0022] The following description of some technologies involved in the embodiments of this application is provided to aid understanding and should be considered merely exemplary. Therefore, those skilled in the art should recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. Similarly, for clarity and brevity, some descriptions of well-known functions and structures are omitted in the following description.

[0023] Figure 1 This is a flowchart of a knowledge graph construction method for shale oil measure evaluation provided in an embodiment of this application, including steps 101 to 105. Figure 1 This is merely one execution order shown in the embodiments of this application and does not represent the only execution order for a knowledge graph construction method for shale oil measure evaluation. Where the final result can be achieved, Figure 1 The steps shown can be performed in parallel or in reverse order.

[0024] Step 101: Construct a three-level hierarchical ontology framework for shale oil measure evaluation, and develop a corresponding shale oil-specific terminology dictionary to define entity categories, entity subcategories, inherent attributes, and standard semantic relationships between entities. In this embodiment, a three-level hierarchical ontology framework is constructed, comprising a top-level ontology, a middle-level ontology, and a bottom-level ontology. The top-level ontology is used to classify all knowledge objects in the shale oil measure evaluation field into various first-level entities. These first-level entities include shale oil well entities, production enhancement measure entities, geological environment entities, evaluation index entities, and construction failure and optimization scheme entities. The middle-level ontology further subdivides each first-level entity into various second-level entities. The bottom-level ontology defines the inherent attributes corresponding to each first-level and second-level entity and standardizes various standard semantic relationships between each first-level and / or second-level entity. A unified industry synonymous expression is established based on the shale oil-specific terminology dictionary to unify semantic benchmarks.

[0025] In this embodiment, the intermediate layer entity further subdivides each primary entity into multiple secondary entities, including: subdividing the production enhancement measures entity into multiple secondary entities, including: hydraulic fracturing, volumetric fracturing, carbon dioxide huff and puff, profile control and water shut-off, and repeated fracturing; subdividing the evaluation index entity into multiple secondary entities, including: construction index, production capacity index, economic index, and risk failure index; subdividing the construction failure part of the construction failure and optimization scheme entity into multiple secondary entities, including: sand plugging, proppant backflow, insufficient formation energy, and tubing leakage; and reserving expandable subdivision interfaces for the shale oil well entity and the geological environment entity to subdivide secondary entities according to the block geological conditions and well type.

[0026] Specifically, such as Figure 3 As shown, a three-level hierarchical ontology framework is constructed, including a top-level ontology, a middle-level ontology, and a bottom-level ontology. At the same time, a shale oil-specific terminology dictionary is provided to unify industry synonyms, eliminate terminological ambiguity, and provide a consistent semantic benchmark for subsequent knowledge extraction, fusion, and reasoning.

[0027] Furthermore, using a top-level ontology, all knowledge objects within the shale oil enhancement measures evaluation domain are divided into five core primary entities: shale oil well entity, production enhancement measures entity, geological environment entity, evaluation index entity, and construction failure and optimization scheme entity, thus establishing the top-level classification boundary of knowledge. Among them, the shale oil well entity and the geological environment entity reserve expandable subdivision interfaces, and can be further refined into secondary entities based on block geological conditions (such as burial depth, brittle mineral content) or well type (such as horizontal well, vertical well) to adapt to the differentiated needs of different work areas.

[0028] The middle-level ontology is used to further classify each primary entity according to business dimensions. The specific subdivision rules are as follows: the production enhancement measure entity is subdivided into various secondary entities such as hydraulic fracturing, volumetric fracturing, CO2 huff and puff, profile control and water shut-off, and repeated fracturing; the evaluation index entity is subdivided into construction indicators (such as construction displacement and proppant addition), production capacity indicators (such as daily oil production), economic indicators (such as input costs), and risk and failure indicators (such as failure rate); the construction failure part of the construction failure and optimization scheme entity is subdivided into specific secondary failure entities such as sand plugging, proppant backflow, insufficient formation energy, and tubing leakage, while the optimization scheme entity is associated with the failure type and is not subdivided separately. The corresponding treatment measures are matched through subsequent relationship definitions.

[0029] The underlying ontology defines the inherent attributes of each entity (including first-level and second-level entities) and standardizes the semantic relationships between entities. For example, the shale oil well entity defines inherent attributes such as well number, well depth, production date, and block to which it belongs; the production enhancement measure entity defines inherent attributes such as construction date, construction scale (e.g., fluid volume, sand volume), and construction parameters; the evaluation index entity defines inherent attributes such as index value, monitoring time, and data source; and the fault entity defines inherent attributes such as the time of fault occurrence and fault severity level. Simultaneously, the standard semantic relationships between entities are standardized, including core logical relationships such as "implementation" (shale oil well—production enhancement measure), "generation" (production enhancement measure—evaluation index), "trigger" (production enhancement measure—construction fault), and "matching" (construction fault—optimization plan), ensuring that the semantic connections between entities are clear and conform to the business rules for shale oil measure evaluation.

[0030] Finally, based on a shale oil-specific terminology dictionary, common synonyms used in the industry (such as unifying "fracturing" as "hydraulic fracturing," "daily production" as "daily oil production under the capacity indicator," and merging "sand plugging" and "sand plugging") are standardized and mapped to eliminate ambiguity caused by differences in terminology, ensuring the consistency of entity recognition and relational reasoning throughout the entire process. This dictionary can be dynamically expanded as business develops, and new terms or synonyms can be added regularly to continuously adapt to actual field usage habits.

[0031] Step 102: Based on the hierarchy and evaluation business dimensions of the three-level hierarchical ontology framework, perform hierarchical and targeted preprocessing on multi-source heterogeneous shale oil data to construct a shale oil labeled dataset. In this embodiment, based on the hierarchy of the three-level hierarchical ontology framework and the evaluation business dimension, the multi-source heterogeneous data of shale oil is divided into a construction data layer, a reservoir geology data layer, a production dynamic evaluation data layer, and an expert knowledge text data layer. Outlier removal, field normalization, and dimensional unification are performed on the structured data in the construction data layer, reservoir geology data layer, and production dynamic evaluation data layer. A preliminary mapping relationship is established between the processed structured data and the first-level or second-level entities in the three-level hierarchical ontology framework. For the semi-structured data in the construction data layer, reservoir geology data layer, and production dynamic evaluation data layer, their table and form structures are parsed, and multiple core attribute fields are extracted. Each core attribute field is matched to a second-level entity in the three-level hierarchical ontology framework. Text segmentation, noise removal, and invalid fragment filtering are performed on the unstructured data in the expert knowledge text data layer, retaining the core text content. Entity annotation is performed on the preprocessed data of each layer to construct a shale oil annotation dataset.

[0032] Specifically, based on the hierarchy of the three-level hierarchical ontology framework and the evaluation business dimensions, the multi-source heterogeneous data of shale oil is divided into four major data levels: the measure construction data layer, the reservoir geological data layer, the production dynamic evaluation data layer, and the expert knowledge text data layer. Each layer has different data sources and structures, and requires differentiated preprocessing strategies.

[0033] Furthermore, the fracturing operation data layer mainly comes from on-site fracturing operation logs and operation reports, including real-time monitoring data such as fracturing displacement, sand addition, fluid volume, and operation pressure, which are mostly structured numerical data; the reservoir geological data layer comes from geological exploration reports and well logging interpretation results, covering geological parameters such as formation pressure, porosity, permeability, and brittle mineral content, which are also mainly structured tables; the production dynamic evaluation data layer comes from daily production ledgers and daily production reports, including dynamic indicators such as daily oil production, water cut, and cumulative oil production, which include both structured data and semi-structured reports; the expert knowledge text data layer comes from fracturing operation summary reports, fault analysis documents, technical solutions, academic literature, etc., which are mainly unstructured text.

[0034] For structured data in the construction data layer, reservoir geological data layer, and production dynamic evaluation data layer, outlier removal is performed first: outliers are identified and removed using statistical thresholding or business experience thresholding (e.g., construction discharge exceeding three times the upper limit of the normal range is considered an anomaly); then, field normalization is performed, standardizing field naming formats (e.g., "date" is standardized to "YYYY-MM-DD", and well names use standardized coding), and missing values ​​are handled using mean imputation or nearest neighbor imputation; finally, unit standardization is implemented, for example, pressure units are standardized to megapascals (MPa), and discharge units are standardized to cubic meters per minute (m³ / min). 3 / min) and sand addition are uniformly expressed in tons (t). After processing, a preliminary mapping association is established between each field and the first-level or second-level entities in the three-level hierarchical ontology framework. For example, "construction discharge" is mapped to the construction index attribute of the production enhancement measures entity, and "daily oil production" is mapped to the second-level entity of production capacity index.

[0035] For semi-structured data in the construction data layer, reservoir geological data layer, and production dynamic evaluation data layer, this mainly refers to documents containing table structures such as construction reports, construction ledgers, and statistical forms. During parsing, a table structure recognition algorithm is used to locate table headers and data rows, accurately extracting multiple core attribute fields, including the start and end times of the measures (e.g., fracturing start and end times), construction scale (e.g., total fluid volume, total sand volume), and input costs (e.g., material costs, engineering costs). After extraction, each core attribute field is matched to its corresponding secondary entity. For example, "measure start and end times" is matched to the inherent attributes of the production enhancement measures entity, and "input costs" is matched to the economic indicators secondary entity.

[0036] For unstructured data in the expert knowledge text data layer, including summary reports of measures, fault logs, technical solutions, expert literature, etc., the preprocessing workflow is as follows: First, the text is segmented into sentences, using periods, question marks, exclamation marks, etc., as delimiters to break long texts into independent sentences; then, noise is removed, including special symbols, garbled characters, irrelevant advertising information, headers and footers, etc.; next, invalid segments are filtered out, such as pure numerical lists, chart descriptions, reference formats, etc., retaining only core text content such as descriptions of measures (e.g., descriptions of fracturing construction processes), effect analyses (e.g., evaluation of increased production effects), and problem diagnoses (e.g., analysis of sand blockage causes). After the above processing, this layer of data enters the annotation stage together with the structured and semi-structured data.

[0037] After preprocessing, the data at each layer were labeled according to preset entity types (i.e., the five primary entities and their subdivided secondary entities defined in the top-level ontology) to construct a shale oil labeled dataset. Labeling employed the BIO labeling method (B for Begin, I for Inside, and O for Outside), performed independently by two labelers. Cohen's Kappa coefficient (κ, a statistic used to measure labeling consistency) was introduced to verify the consistency of the two labelers' work. A κ > 0.85 was required, indicating a high degree of consistency, before the data was included in the final dataset. The dataset sourced from multiple shale oil blocks in China and shale oil wells at different depths (shallow and medium-deep), including full-cycle data such as fracturing logs, well histories, production dynamic reports, and failure analysis reports. The dataset was divided into training and testing sets in an 8:2 ratio to provide high-quality labeled corpus for subsequent knowledge extraction model training.

[0038] Step 103: Extract knowledge from the shale oil annotation dataset to obtain knowledge triples containing entities, inherent attributes, and relationships. In this embodiment, a sequence labeling model is used to extract entities from the shale oil annotation dataset to obtain a first knowledge triple containing entity boundary and category label information. The sequence labeling model includes a pre-trained language encoding layer, a bidirectional long short-term memory network layer, and a conditional random field layer, all cascaded in sequence. The pre-trained language encoding layer is used for vector encoding of the text; the bidirectional long short-term memory network layer is used to extract contextual semantic features based on the vector encoding and predict the entity boundary markers and category labels of first-level or second-level entities corresponding to each word; the conditional random field layer is used to globally constrain the entity boundary markers and category labels, outputting an entity label sequence. A large language model is used to extract entities from the shale oil annotation dataset. The inherent attributes and semantic relationships are jointly extracted to obtain the second knowledge triple, which includes: customizing exclusive prompt words for various entities and embedding domain-annotated examples, and identifying first-level entities, second-level entities, inherent attributes of various entities, and semantic relationships between different entities in the text; vectorizing the identified entities and comparing their semantic similarity with a shale oil-specific terminology dictionary to filter suspicious entities and generate verification questions; performing self-checking on the verification questions through a large language model and outputting the verified entities, inherent attributes, and semantic relationships; and integrating the entity label sequence of the first knowledge triple, the inherent attributes and semantic relationships in the second knowledge triple, and unifying the output format to generate a standardized knowledge triple.

[0039] Specifically, a dual-path extraction architecture combining sequence labeling model and large language model is adopted to obtain the first knowledge triplet and the second knowledge triplet respectively. The two sets of results are then fused to finally output a standardized knowledge triplet. The specific process is as follows.

[0040] I. Entity extraction based on sequence labeling model (first knowledge triples).

[0041] Entity extraction from a shale oil labeled dataset is performed using a sequence labeling model to obtain first knowledge triples containing entity boundaries and category label information. The sequence labeling model comprises a cascaded pre-trained language encoding layer (BERT layer, a Transformers-based bidirectional encoder representation model used to transform input text into dense vector representations), a bidirectional long short-term memory network layer (BiLSTM layer, used to capture bidirectional semantic dependencies in context), and a conditional random field layer (CRF layer, used to globally constrain the label sequence).

[0042] Furthermore, the pre-trained language encoding layer first performs vector encoding on each word in the input text, mapping the words to dense vectors rich in semantic information, which serve as input features for subsequent networks. The bidirectional Long Short-Term Memory (LSTM) network layer receives the vector encoding and extracts contextual semantic features through forward and backward LSTM network units, predicting entity boundary labels (e.g., B - entity type indicating the start of an entity, I - entity type indicating the interior of an entity, O - non-entity) and category labels for first-level or second-level entities (e.g., B - production-increasing measures indicating the start of a production-increasing measure entity). The Conditional Random Field (CRF) layer receives the output from the bidirectional LSTM network layer and comprehensively considers the global dependencies of the entire label sequence, modeling the transition probabilities between adjacent labels to avoid unreasonable label combinations (e.g., two consecutive B labels or an I label lacking a corresponding B label), thus outputting the globally optimal entity label sequence. The first knowledge triple is composed of the entity boundaries and category labels output by this sequence labeling model, recording the position of all identified entities in the text and their corresponding first- or second-level entity types.

[0043] II. Joint extraction based on large language models (second knowledge triples).

[0044] A large language model is used to jointly extract entities, inherent attributes, and semantic relationships from the shale oil annotation dataset, resulting in second knowledge triples. The specific process includes the following three sub-steps: (1) Design of dedicated prompt words and step-by-step reasoning extraction: For five types of primary entities (shale oil well entities, production enhancement measures entities, geological environment entities, evaluation index entities, construction failure and optimization scheme entities) and their subdivided secondary entities, dedicated prompt words (input templates to guide the large language model to complete specific tasks) are customized. A small number of domain-labeled examples are embedded in the prompt words to clarify the identification boundaries, core attribute fields and relationship types of various entities. The thinking chain (a prompting technology that guides the model to decompose complex reasoning tasks into multiple intermediate logic steps) is used to guide the large language model to perform reasoning extraction step by step. That is, first, the core production enhancement measures entities in the text are identified, then the construction anomaly representation entities associated with the measures are located, then the corresponding monitoring and evaluation index entities are extracted, and finally the corresponding optimization and treatment scheme entities are matched to form a complete semantic reasoning chain of "measures → anomalies → indicators → schemes" to ensure that entities, inherent attributes and relationships correspond to each other logically.

[0045] (2) Self-supervised enhanced verification mechanism: The initially extracted entities are vectorized and their semantic similarity is calculated one by one with the vector of each standard term in the shale oil-specific terminology dictionary. Candidate words with high similarity are selected and verification questions are generated accordingly (e.g., "Does this entity belong to the production enhancement measure type?"). Subsequently, the large language model, combined with the domain lexicon, performs a self-check on the initial extraction results to identify erroneous or omitted entities. Uncertain entities are marked as "to be reviewed", and entities with obvious errors (e.g., identifying non-measure words as measure names) are triggered for re-extraction until the consistency requirements are met.

[0046] (3) Inherent attributes and relationship completion: Based on entity recognition, the big language model further extracts the inherent attributes corresponding to various entities (such as the "production date" of shale oil well entities, the "construction displacement" and "sand addition" of production enhancement measures entities, and the "indicator value" of evaluation index entities), as well as the semantic association relationships between different entities (such as "shale oil well - implementation - production enhancement measures", "production enhancement measures - generation - evaluation index", "production enhancement measures - trigger - construction failure", "construction failure - matching - optimization scheme").

[0047] III. Dual-Path Result Fusion and Standardized Output. The entity label sequence in the first knowledge triplet is fused with the inherent attributes and semantic relationships in the second knowledge triplet. Specifically, using the entity boundaries and category labels provided by the first knowledge triplet as entity benchmarks, the inherent attributes confirmed through self-supervised verification in the second knowledge triplet are attached to the corresponding entities, and the confirmed semantic relationships are connected to the corresponding entity pairs, forming a complete (entity, inherent attribute, relationship) knowledge structure. The unified output format is a standardized knowledge triplet, describing the relationships between entities as "(head entity, relationship type, tail entity)," and describing the inherent attributes of entities as "(entity, inherent attribute name, inherent attribute value)," and binding data source tags (such as document name, paragraph number, field source table name) to provide a basis for subsequent knowledge tracing and evaluation attribution. The final generated knowledge triplet covers all core entity types and their associated networks, including production increase measures, construction anomaly representations, evaluation indicators, and optimization schemes.

[0048] Model performance is evaluated using precision (P, the proportion of correctly identified entities out of all correctly identified entities), recall (R, the proportion of correctly identified entities out of all true entities), and F1 score (the harmonic mean of precision and recall, used to comprehensively measure the model's extraction performance). The formulas for these scores are as follows: , , , In the formula, The number of entities correctly identified by the model; The number of non-entities that were incorrectly identified as entities by the model; This represents the number of real entities that were not successfully identified by the model.

[0049] Step 104: Perform entity normalization, relational logic verification, and quality audit on the knowledge triples for the shale oil domain. Import the verified knowledge triples into a graph database to store the knowledge graph. In this embodiment, based on a shale oil-specific terminology dictionary, the entities in the knowledge triples are normalized, synonymous entities are merged, ambiguous entities are distinguished, and evaluation indicators and measure names with different expressions are unified. Based on the semantic rules of a three-level hierarchical ontology framework, the relationships between entities in the knowledge triples are logically verified, and erroneous relationships that do not conform to the semantic rules are automatically screened and eliminated. The knowledge triples that have undergone entity normalization and relational logic verification are quality audited, and the qualified knowledge triples are imported into a graph database to construct a knowledge graph for shale oil measure evaluation.

[0050] Specifically, the standard knowledge triples generated in step 103 undergo pre-database processing, sequentially going through three stages: entity normalization, relational logic verification, and manual quality review. Incorrect, redundant, or non-compliant triples are removed, and finally, the verified triples are stored in the graph database to construct a shale oil measure evaluation knowledge graph, such as... Figure 4 As shown.

[0051] Furthermore, entity normalization is performed based on a shale oil-specific terminology dictionary. Since the same entity may have multiple representations in multi-source data (e.g., "hydraulic fracturing" and "fracturing" refer to the same type of measure, "daily oil production" and "daily output" both point to production capacity indicators, and "sand plugging" and "sand plugging" refer to the same fault phenomenon), the normalization process first scans the head and tail entities in all knowledge triples, performing precise and fuzzy matching of their textual expressions with the standard terms in the terminology dictionary; successfully matched synonymous entities are unified into a unique standard name, and each alias is recorded in the graph as a retrieval entry point. Simultaneously, for cases where the same name represents different semantics in different contexts (e.g., "pressure" may refer to construction pump pressure, formation pressure, or wellhead pressure), a comprehensive judgment is made based on the entity's category (i.e., the ontology level category bound to the entity in step 101) and its associated entity type, correctly classifying ambiguous entities to their corresponding categories to avoid confusion. In addition, all evaluation indicators and measures are named in accordance with the unified naming format of the terminology dictionary (e.g., the unit for "sand addition" is ton, and the unit for "construction discharge" is cubic meters per minute) to ensure that the names of similar entities in the map are unique and unambiguous.

[0052] The relational logic verification is performed automatically based on the semantic rules preset in the three-level hierarchical ontology framework. These semantic rules include, but are not limited to: entity type matching rules (e.g., the head entity of a "trigger" relation must be a production enhancement measure entity, and the tail entity must be a construction failure entity; the head entity of a "generate" relation must be a production enhancement measure entity, and the tail entity must be an evaluation index entity), relational direction constraints (e.g., a "match" relation can only point from a construction failure to an optimization solution, not vice versa), and hierarchical affiliation rules (secondary entities must belong to their corresponding primary entity category; for example, "hydraulic fracturing" cannot be a subclass of a shale oil well entity). The verification engine traverses all relational triples (i.e., head entity—relationship—tail entity form) in all knowledge triples, determining whether each triple conforms to the above rules. Relational triples that do not conform to the semantic rules are automatically removed and anomaly logs are recorded; boundary cases not covered by the rules are retained and transferred to the next stage for expert review and judgment.

[0053] Quality audits involve manual sampling of knowledge triples that have undergone entity normalization and relational logic verification. Shale oil technical personnel (such as oilfield development engineers and measure evaluation experts) conduct focused sampling of core measure entities (e.g., triples related to fracturing parameters), key evaluation indicator entities (e.g., triples corresponding to production capacity and economic indicators), and fault-optimization scheme matching relationships. The sampling rate is no less than 5% of the total number of knowledge triples. Passing knowledge triples are marked as "verified," while failing or questionable triples are returned to step 103 for re-sampling or manual correction. This process ensures the professionalism and accuracy of the knowledge stored in the database.

[0054] Finally, the approved knowledge triples are imported into the graph database. In this embodiment, Neo4j (a high-performance graph database management system with nodes and relationships as its core storage structure, supporting efficient graph traversal and multi-hop queries) is used for storage. During import, each first-level or second-level entity is created as a node in the graph. The node's label corresponds to its ontology category (e.g., "production enhancement measures," "evaluation indicators," "construction faults"). The node's core attribute fields store the entity's normalized inherent attribute key-value pairs (e.g., "construction displacement: 8.5m³"). 3 / min"Saving amount: 120t"). Each relation triple is converted into directed edges between nodes, with the edge type corresponding to the relation type (e.g., "implement", "generate", "trigger", "match"). Edges can have attached attributes (e.g., relation confidence, data source tags). After import, a shale oil measure evaluation knowledge topology graph is constructed based on Neo4j, intuitively displaying the multi-level association network of "shale oil well - production enhancement measures - fault characterization - evaluation indicators - optimization schemes", supporting single-well and single-block knowledge subgraph retrieval and visualization.

[0055] Step 105: Receive a query statement for shale oil measure evaluation; extract primary entities, secondary entities, and evaluation intent from the query statement; based on the extracted information, perform semantic similarity matching in the knowledge graph to retrieve related knowledge subgraphs; construct domain-enhanced prompt text by fusing the query statement and related knowledge subgraphs, and input it into a large language model; the large language model, combined with shale oil domain business knowledge, performs reasoning analysis and outputs the shale oil measure evaluation result. In this embodiment, a query statement for shale oil measure evaluation is received; primary entities, secondary entities, and evaluation intent are extracted from the query statement; based on the extracted information, semantic similarity matching is performed in the knowledge graph to retrieve related knowledge subgraphs; domain-enhanced prompt text is constructed by fusing the query statement and related knowledge subgraphs, and input into a large language model; the large language model, combined with shale oil domain business knowledge, performs reasoning analysis and outputs the shale oil measure evaluation result.

[0056] Specifically, the system receives natural language queries from users regarding shale oil measures evaluation. These queries go through four stages: entity recognition and intent parsing, knowledge graph semantic retrieval, domain-enhanced suggestion construction, and large language model reasoning and response. The final output is an evaluation conclusion that integrates structured knowledge from the knowledge graph with the generation capabilities of the large language model. The specific process is as follows.

[0057] I. Query Entity Identification and Evaluation Intent Analysis.

[0058] The system receives user-input queries (e.g., "Why did the daily oil production of Well XX fail to meet expectations after volumetric fracturing?" or "Is there a risk of sand blockage during hydraulic fracturing in Block YY, and how should it be optimized?"). It uses the sequence labeling model trained in step 103 to perform entity recognition on the queries, extracting primary entities (e.g., the shale oil well entity "Well XX" and the production enhancement measure entity "volumetric fracturing") and secondary entities (e.g., the specific fault type "sand blockage"). Simultaneously, it uses keyword matching and syntactic analysis to determine the evaluation intent type of the query. Evaluation intent types are categorized as follows: construction quality evaluation (focusing on whether construction parameters meet standards and whether the construction process is abnormal), production capacity effect evaluation (focusing on increased production, changes in daily oil production, and decline rate), economic evaluation (focusing on input-output ratio and cost-effectiveness), and fault diagnosis evaluation (focusing on fault type identification, cause location, and optimization scheme matching). The identified entities and intent types are used as input for subsequent graph retrieval and inference.

[0059] II. Knowledge Graph Semantic Similarity Retrieval and Acquisition of Related Knowledge Subgraphs. Based on the extracted entities and evaluation intent, a retrieval is performed in the Neo4j graph database (i.e., the shale oil measure evaluation knowledge graph constructed in step 104). First, semantic similarity matching is performed between the entity names in the query statement and the node names in the Neo4j graph database to overcome the expression differences between user input and graph standard terminology, and to match the most relevant graph nodes. Centered on the matched nodes, the graph is expanded 2 to 3 hops along the relation edges (i.e., layer-by-layer association through relationships such as "implementation," "generation," "trigger," and "matching") to obtain a complete knowledge subgraph containing the well's construction history, the adopted production enhancement measures, the corresponding evaluation index values, the types of failures that have occurred, and the matched optimization schemes. Simultaneously, the indicators are calculated. (That is, the ratio of the number of nodes that match the knowledge nodes in the query statement retrieved from the knowledge graph to the total number of nodes retrieved from the knowledge graph), and its calculation formula is: , In the formula, This represents the number of nodes retrieved from the knowledge graph by the method. The number of knowledge nodes / entities in the query statement; This refers to the number of nodes retrieved from the knowledge graph that match the knowledge nodes in the query statement. The larger this value, the more closely the graph retrieval results match the semantics of the question.

[0060] III. Construct enhanced domain hint text by integrating query statements and related knowledge subgraphs.

[0061] The retrieved knowledge subgraphs are serialized and described in a structured format of entity-inherent attribute-relationship, and then transformed into natural language text paragraphs. The content includes: basic well information, implemented measures and key parameters, corresponding evaluation indicators, historical failure records, and existing optimization schemes. The query statement is concatenated with the serialized text of the knowledge subgraphs, and standard terms from the shale oil-specific terminology dictionary in step 101 are injected as constraints to construct domain-enhanced prompt text. The prompt text explicitly constrains the large language model to answer based on the provided knowledge subgraph information and prohibits the arbitrary fabrication of facts beyond the scope of the graph. It also specifies the output format requirements for the answer according to the evaluation intent type (e.g., diagnosis, comparison, early warning, or optimization recommendation) (e.g., listing cause analysis by item, ranking by impact degree, and optimization suggestions). This enhanced prompt text retains the original semantics of the user query while injecting structured factual evidence from real business knowledge graphs into the large language model.

[0062] IV. Perform reasoning analysis using a large language model and output evaluation results.

[0063] The domain-enhanced prompt text is input into the large language model, which then performs inference analysis based on its business knowledge in the shale oil field. The inference process follows the thought chain reasoning paradigm in step 103, and is completed step by step: First, the data on the effectiveness of measures and failure records related to the well in the knowledge subgraph are summarized; then, the historical average indicators of similar blocks or similar measures are compared (if the knowledge subgraph contains benchmark data); next, the main reasons for anomalies or failure to meet expectations are diagnosed (such as insufficient displacement, low sand addition, formation energy decay, sand blockage, etc.); finally, existing optimization schemes are matched or specific optimization suggestions are given in combination with the rule base. The output includes: evaluation conclusions (such as measures effective / failed to meet expectations / risk exists), attribution analysis (listing the main influencing factors and their contribution ranking), optimization and treatment suggestions (such as adjusting construction parameters, replenishing pre-flush fluid, changing proppant type, etc.), and the corresponding source of evidence (i.e., specific nodes and relationship paths in the knowledge subgraph, which facilitates traceability and verification).

[0064] The evaluation results output by the large language model are also subject to quality quantification, measured using the following three indicators: Indicator F th (That is, the ratio of the number of answers that are faithful to the knowledge graph to the total number of answers), the calculation formula is: , In the formula, 'a' represents the total number of answers generated by the method; 'a' represents the specific answer generated by the method. This is the matching function between the method answer and the knowledge node; it takes the value 1 for a match and 0 otherwise. The larger the value, the higher the accuracy of the domain knowledge contained in the method answer.

[0065] index (That is, the ratio of the number of standard answers matching the knowledge node to the total number of standard answers), the calculation formula is: , In the formula, t represents the total number of segments in the standard answer; t represents a single knowledge segment in the standard answer. This is the matching function between knowledge fragments and knowledge nodes. It takes the value 1 when a match is found and 0 otherwise. The larger the value, the higher the quality of the generated standard answer.

[0066] index (That is, the average semantic similarity between the generated reverse question and the original query statement based on the answer), calculated using the following formula: , In the formula, q represents the number of questions generated based on the method-based answer; q represents the original query statement. To answer the i-th question generated according to the method; To generate a question for query q The semantic similarity between the queries. The higher the value, the more focused the response is on the original query intent.

[0067] Finally, the evaluation results output by the large language model are returned to the user in the form of structured text, while the key nodes and relationship paths in the knowledge graph that support the conclusion are displayed, so as to realize the interpretability and traceability of the evaluation process.

[0068] In this embodiment, after importing the verified knowledge triples into a graph database to store the knowledge graph, the following can be further implemented: An edge-cloud collaborative layered deployment architecture can be built to access new data from shale oil fields in real time to complete incremental knowledge updates. Based on the new data, steps 102 to 104 are executed to update the knowledge graph. Those skilled in the art should understand that step 105 is the reasoning and evaluation step during a query. In each query, step 105 should always be executed based on the latest version of the knowledge graph and does not need to be re-executed with incremental updates.

[0069] Specifically, after completing the initial knowledge graph construction, this application further establishes a layered deployment architecture for edge-cloud collaboration (i.e., a layered deployment mode where the edge and cloud work together), which is adapted to the low-computing-power industrial control computer hardware environment of well sites and the high-performance computing resources of oilfield data centers. This enables real-time access and incremental knowledge updates of newly added construction, production, and fault data, allowing for dynamic iteration without a full reconstruction of the graph, forming a continuous closed-loop update mechanism. The architecture and update process are as follows.

[0070] I. Layered deployment architecture for edge-cloud collaboration.

[0071] The system is divided into two layers based on computing resources and business timeliness requirements: edge and cloud. The edge layer (i.e., the well site industrial control computer) is deployed at the shale oil well site and equipped with a low-power industrial control computer (computing power is limited, typically CPU (Central Processing Unit) level or low-end GPU (Graphics Processing Unit)). The edge layer deploys a lightweight large language model (i.e., LLM, a simplified version processed by model pruning, quantization, and other compression techniques, with significantly fewer parameters than the cloud model), a local knowledge sub-graph (containing only relevant knowledge nodes and relationships of the well and adjacent well groups, stored in a Neo4j embedded version or a lightweight copy of a graph database), and a data preprocessing and rapid entity recognition module. The edge terminal is responsible for receiving construction parameters (displacement, pressure, sand addition, etc.) and daily production data collected by well site sensors in real time, performing preliminary preprocessing (outlier removal, field normalization, and unit unification), and using a lightweight model to complete rapid entity recognition and preliminary evaluation of the effects of simple measures. It outputs simple optimization suggestions (such as "current construction pressure is too high, it is recommended to reduce displacement") on site with low latency for on-site engineers to refer to immediately.

[0072] Cloud-based (Oilfield Decision Center): Deployed in the oilfield data center or cloud service platform, equipped with high-performance GPU clusters (such as multi-GPU A100 / H100) and large-capacity storage. The cloud deploys a 72B (billion-parameter scale) large language model (i.e., a full LLM with stronger reasoning and semantic understanding capabilities), a full global knowledge graph (storing complete knowledge triples and relational networks for all blocks and wells), and a deep reasoning and analysis engine. The cloud receives preliminary analysis results and incremental raw data uploaded from the edge, performs complex multi-hop reasoning (such as cross-well and cross-block measure effect benchmarking), long-term development trend analysis (such as block production capacity decline patterns), and global measure scheme optimization (such as recommending optimal fracturing parameters based on multi-well data), generating in-depth evaluation reports for decision-making.

[0073] II. Incremental dynamic update mechanism.

[0074] The continuously updated construction data, production capacity data, fault records, etc. are connected to the edge terminal in real time, triggering the incremental update process, which is divided into the following sub-steps.

[0075] (1) Incremental Data Acquisition and Preprocessing: The edge device receives incremental data packets (including parameters of new construction phases, daily oil production, and newly occurring fault alarms) from well site sensors and the production management system in real time through the OPC unified architecture (a communication protocol for industrial automation data exchange) or MQTT (Message Queuing Telemetry Transmission Protocol, a lightweight Internet of Things communication protocol). The edge device then performs hierarchical targeted preprocessing of the incremental data in step 102: outlier removal, field normalization, and unit unification are performed on structured data; core attribute fields (start and end time of measures, construction scale, and input cost) are extracted from the tables of semi-structured reports (such as newly added measure ledgers); and sentence segmentation, noise reduction, and core content retention are performed on unstructured text (such as newly written fault analysis reports). After preprocessing, only the effective part of the incremental data (not all historical data) is pushed to the subsequent stages.

[0076] (2) Incremental Knowledge Extraction and Fusion: The edge device utilizes a lightweight LLM or calls the cloud-based large language model interface to perform the knowledge extraction process (including entity recognition, inherent attribute extraction, and relation extraction) on the incremental preprocessing results, generating incremental knowledge triples. Subsequently, the edge device performs entity normalization and relation logic verification in step 104, only performing conflict detection between the newly added triples and existing knowledge in the local subgraph (e.g., whether the newly extracted construction displacement value contradicts historical records, or whether the new fault is a duplicate of a known fault). The incremental triples that pass the verification first update the local knowledge subgraph of the edge device.

[0077] (3) Cloud synchronization and full graph update: The edge end will periodically (e.g., every 6 hours or after each construction phase) upload incremental triples and preliminary evaluation results to the cloud via an encrypted channel. After receiving the data, the cloud will first perform global conflict resolution (because the cloud has full data, it can discover cross-well and cross-block related conflicts, such as ambiguity of the same well name in different blocks), and then re-perform entity normalization and relationship verification based on the three-level hierarchical ontology framework to ensure global consistency. After the verification is qualified, the incremental triples will be inserted into the full knowledge graph in the cloud (Neo4j supports incremental insertion of nodes and relationships without rebuilding the full graph). At the same time, the cloud will periodically feed back new cases (such as new types of faults and their successful handling solutions) to the shale oil-specific terminology dictionary, expand the synonyms, new terms (such as the newly emerging measure type "supercritical CO2 fracturing") and evaluation index names in the dictionary, and simultaneously distribute the updated terminology dictionary to the edge end to continuously optimize the self-supervised verification capability of the edge end (the semantic similarity comparison benchmark in step 103 will be updated accordingly).

[0078] (4) Closed-loop iteration trigger: After each incremental update, the system automatically feeds back the new data and related context to the preprocessing stage in step 102 (i.e., the incremental data itself is used as new input to go through the complete link again), and at the same time updates the version of the knowledge graph used for evaluation and reasoning in step 105 to ensure that subsequent queries can be answered based on the latest knowledge. The entire process does not involve the repeated processing of all historical data, and only steps 102 to 104 are executed for the incremental part, thereby significantly reducing the computation and storage overhead and adapting to the normalized demand for high-frequency data input under continuous shale oil operations.

[0079] In this embodiment, a first knowledge triplet is extracted using a sequence labeling model, and a second knowledge triplet is jointly extracted using a large language model. The two extraction results complement each other, and after entity normalization and relation verification, a high-quality knowledge graph for shale oil measure evaluation is obtained. This solves the problems of extraction omissions, extraction errors, and poor domain adaptability in traditional knowledge extraction methods. By retrieving and matching related knowledge subgraphs based on the knowledge graph, domain-enhanced prompt text is constructed to drive the large language model's reasoning. This not only utilizes the reasoning ability of the large language model but also constrains its answers to strictly follow business facts, solving the problems of distorted answers and lack of traceability in general large language models. Through an edge-cloud collaborative layered architecture and incremental dynamic update mechanism, the requirements for low-latency response on-site and deep analysis in the cloud are taken into account, while reducing the computational overhead of full reconstruction. This can adapt to the business scenario of continuous access to dynamic data in shale oil development and provide stable knowledge support for rapid and accurate evaluation of the effectiveness of shale oil measures.

[0080] While this application provides the method operation steps as described in the embodiments or flowcharts, more or fewer operation steps may be included based on conventional or non-inventive labor. The order of steps listed in this embodiment is merely one possible execution order among many and does not represent the only execution order. In actual device or client product execution, the methods shown in this embodiment or the accompanying drawings can be executed sequentially or in parallel (e.g., in a parallel processor or multi-threaded processing environment).

[0081] like Figure 2 As shown in the figure, this application embodiment also provides a knowledge graph construction system 200 for shale oil measure evaluation. The system includes: a definition module 201, a preprocessing module 202, an extraction module 203, an auditing module 204, and a reasoning module 205, as detailed below.

[0082] The definition module 201 is used to build a three-level hierarchical ontology framework for evaluating shale oil measures, and is equipped with a shale oil-specific terminology dictionary to define the entity categories, entity subcategories, inherent attributes and standard semantic relationships between entities.

[0083] The preprocessing module 202 is used to perform hierarchical and targeted preprocessing on multi-source heterogeneous shale oil data based on the hierarchy and evaluation business dimensions of the three-level hierarchical ontology framework, so as to construct a shale oil labeled dataset.

[0084] The extraction module 203 is used to extract knowledge from the shale oil labeled dataset to obtain knowledge triples containing entities, inherent attributes and relationships.

[0085] The audit module 204 is used to perform entity normalization, relational logic verification and quality audit on knowledge triples for the shale oil field, and import the qualified knowledge triples into the graph database to store the knowledge graph.

[0086] The reasoning module 205 receives the query statement for shale oil measure evaluation, extracts the primary entities, secondary entities, and evaluation intent from the query statement, and performs semantic similarity matching in the knowledge graph based on the extracted information to retrieve the associated knowledge subgraph; it integrates the query statement and the associated knowledge subgraph to construct domain-enhanced prompt text, and inputs it into the large language model; the large language model combines business knowledge in the shale oil field to carry out reasoning analysis and output the shale oil measure evaluation result.

[0087] Some modules in the apparatus described in this application can be described in the general context of computer-executable instructions that are executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, classes, etc., that perform a specific task or implement a specific abstract data type. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0088] The apparatus or module described in the above embodiments can be implemented by a computer chip or physical entity, or by a product with a certain function. For ease of description, the above apparatus is described by dividing it into various modules according to their functions. When implementing the embodiments of this application, the functions of each module can be implemented in one or more software and / or hardware. Of course, a module that implements a certain function can also be implemented by combining multiple sub-modules or sub-units.

[0089] The methods, apparatus, or modules described in this application can be implemented in a computer-readable program code manner. The controller can be implemented in any suitable manner, for example, as a microprocessor or processor and a computer-readable medium storing computer-readable program code (e.g., software or firmware) executable by the (micro)processor, logic gates, switches, application-specific integrated circuits (ASICs), programmable logic controllers, and embedded microcontrollers. Examples of controllers include, but are not limited to, the following microcontrollers: ARC 625D, Atmel AT91SAM, and Microchip PIC18F26K20. Memory controllers can also be implemented as part of the control logic of memory. Those skilled in the art will also recognize that, in addition to implementing the controller in purely computer-readable program code manner, the same functions can be achieved by logically programming the method steps to make the controller take the form of logic gates, switches, ASICs, programmable logic controllers, and embedded microcontrollers. Therefore, such a controller can be considered a hardware component, and the means included within it for implementing various functions can also be considered as structures within the hardware component. Alternatively, the means for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0090] This application also provides an apparatus, the apparatus comprising: a processor; a memory for storing processor-executable instructions; wherein, when the processor executes the executable instructions, it implements the method described in this application.

[0091] This application also provides a non-volatile computer-readable storage medium storing a computer program or instructions thereon, which, when executed, enables the method described in this application embodiment to be implemented.

[0092] Furthermore, in the various embodiments of this application, each functional module can be integrated into one processing module, or each module can exist independently, or two or more modules can be integrated into one module.

[0093] The aforementioned storage media include, but are not limited to, Random Access Memory (RAM), Read-Only Memory (ROM), Cache, Hard Disk Drive (HDD), or Memory Card. The memory can be used to store computer program instructions.

[0094] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary hardware. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product, or it can be embodied in the process of data migration. The computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, mobile terminal, server, or network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments of this application.

[0095] The various embodiments described in this specification are presented in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on its differences from other embodiments. All or part of this application can be used in numerous general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, mobile communication terminals, multiprocessor systems, microprocessor-based systems, programmable electronic devices, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices, etc.

[0096] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit this application. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of this application.

Claims

1. A method for constructing a knowledge graph for shale oil treatment evaluation, characterized in that, include: A three-level hierarchical ontology framework for evaluating shale oil measures was constructed, along with a shale oil-specific terminology dictionary to define the standard semantic relationships between entity categories, entity subcategories, inherent attributes, and entities. Based on the hierarchy and evaluation business dimensions of the three-level hierarchical ontology framework, hierarchical and targeted preprocessing is performed on multi-source heterogeneous shale oil data to construct a shale oil labeled dataset. Knowledge is extracted from the shale oil labeled dataset to obtain knowledge triples containing entities, inherent attributes, and relationships; The knowledge triples are subjected to entity normalization, relational logic verification and quality audit for the shale oil field. The verified knowledge triples are then imported into a graph database to store the knowledge graph. The system receives a query statement for evaluating shale oil measures, extracts primary entities, secondary entities, and evaluation intent from the query statement, and performs semantic similarity matching in the knowledge graph based on the extracted information to retrieve related knowledge subgraphs. The query statement and the associated knowledge subgraph are combined to construct domain-enhanced prompt text, which is then input into a large language model. The large language model combines business knowledge in the shale oil field to conduct reasoning analysis and output the evaluation results of shale oil measures.

2. The method according to claim 1, characterized in that, The aforementioned three-level hierarchical ontology framework for evaluating shale oil measures is constructed, along with a shale oil-specific terminology dictionary to define entity categories, entity subcategories, inherent attributes, and standard semantic relationships between entities, including: Construct a three-level hierarchical ontology framework comprising a top-level ontology, a middle-level ontology, and a bottom-level ontology; The top-level ontology is used to classify all knowledge objects in the field of shale oil measure evaluation into multiple first-level entities; wherein, the multiple first-level entities include shale oil well entities, production enhancement measure entities, geological environment entities, evaluation index entities, and construction failure and optimization scheme entities; The middle-layer ontology is used to further subdivide each of the first-level entities into multiple second-level entities; The underlying ontology is used to define the inherent attributes of each first-level entity and the corresponding second-level entity, and to standardize various standard semantic associations between each first-level entity and / or second-level entity; Based on a shale oil-specific terminology dictionary, we will unify the industry's synonymous expressions to establish a unified semantic benchmark.

3. The method according to claim 2, characterized in that, The method of further subdividing each primary entity into multiple secondary entities using the intermediate ontology includes: The aforementioned production enhancement measures are further subdivided into several secondary entities, including: hydraulic fracturing, volumetric fracturing, carbon dioxide huff and puff, profile control and water shut-off, and repeated fracturing. The evaluation index entity is further subdivided into multiple secondary entities, including: construction index, production capacity index, economic index, and risk and failure index; The construction faults in the aforementioned construction faults and optimization scheme entities are further subdivided into several secondary entities, including: sand blockage, proppant backflow, insufficient formation energy, and tubing puncture leakage. The shale oil well entity and the geological environment entity are provided with an expandable subdivision interface to subdivide secondary entities according to the block geological conditions and well type.

4. The method according to claim 2, characterized in that, Based on the hierarchy and evaluation business dimensions of the aforementioned three-level hierarchical ontology framework, hierarchical and targeted preprocessing is performed on multi-source heterogeneous shale oil data to construct a shale oil labeled dataset, including: Based on the hierarchy and evaluation business dimensions of the three-level hierarchical ontology framework, the multi-source heterogeneous data of shale oil is divided into the construction measures data layer, the reservoir geological data layer, the production dynamic evaluation data layer, and the expert knowledge text data layer. Outlier removal, field normalization, and dimension unification are performed on the structured data in the construction data layer, reservoir geological data layer, and production dynamic evaluation data layer. A preliminary mapping relationship is established between the processed structured data and the first-level or second-level entities in the three-level hierarchical ontology framework. For the semi-structured data in the construction data layer, reservoir geological data layer and production dynamic evaluation data layer, after parsing their table and form structures, multiple core attribute fields are extracted from them, and each core attribute field is matched to the second-level entity of the three-level hierarchical ontology framework. The unstructured data in the expert knowledge text data layer is processed by text segmentation, noise removal, and invalid fragment filtering to retain the core text content; Entity annotations are performed on the preprocessed data of each layer to construct the shale oil annotation dataset.

5. The method according to claim 1, characterized in that, The knowledge extraction from the shale oil labeled dataset yields knowledge triples containing entities, inherent attributes, and relationships, including: Entity extraction is performed on the shale oil annotation dataset using a sequence labeling model to obtain a first knowledge triple containing entity boundary and category label information. The sequence labeling model comprises a pre-trained language encoding layer, a bidirectional long short-term memory network layer, and a conditional random field layer, all cascaded sequentially. The pre-trained language encoding layer performs vector encoding on the text. The bidirectional long short-term memory network layer extracts contextual semantic features based on the vector encoding and predicts entity boundary markers and category labels for each word, as well as first-level or second-level entities. The conditional random field layer globally constrains the entity boundary markers and category labels, outputting an entity label sequence. A large language model is used to jointly extract entities, inherent attributes, and semantic relationships from the shale oil annotation dataset to obtain a second knowledge triple. This triple includes: customizing exclusive prompt words for various entities and embedding domain annotation examples; identifying first-level entities, second-level entities, inherent attributes of various entities, and semantic relationships between different entities within the text; vectorizing the identified entities and comparing their semantic similarity with the shale oil-specific terminology dictionary to filter out suspicious entities and generate verification questions; and using the large language model to perform self-checking on the verification questions, outputting the verified entities, inherent attributes, and semantic relationships. The entity tag sequence of the first knowledge triple, the inherent attributes and semantic relationships in the second knowledge triple, and the unified output format are combined to generate a standardized knowledge triple.

6. The method according to claim 1, characterized in that, The process of performing entity normalization, relational logic verification, and quality auditing on the knowledge triples for the shale oil domain, and importing the verified knowledge triples into a graph database for storing the knowledge graph, includes: Based on the shale oil-specific terminology dictionary, the entities in the knowledge triples are normalized, synonymous entities are merged, ambiguous entities are distinguished, and evaluation indicators and measure names with different expressions are unified. Based on the semantic rules of the three-level hierarchical ontology framework, the relationships between entities in the knowledge triples are logically verified, and erroneous relationships that do not conform to the semantic rules are automatically screened and eliminated. The knowledge triples that have undergone entity normalization and relational logic verification are subjected to quality audits. The qualified knowledge triples are then imported into a graph database to construct a knowledge graph for shale oil measure evaluation.

7. The method according to claim 1, characterized in that, After importing the verified knowledge triples into the graph database to store the knowledge graph, the process also includes: An edge and cloud collaborative layered deployment architecture is established to access new data from shale oil fields in real time to complete incremental knowledge updates. Based on the new data, the method described in claim 1 is executed to update the knowledge graph.

8. A knowledge graph construction system for implementing the method described in any one of claims 1-7, oriented towards shale oil measure evaluation, characterized in that, include: The definition module is used to build a three-level hierarchical ontology framework for evaluating shale oil measures, and is equipped with a shale oil-specific terminology dictionary to define the standard semantic relationships between entity categories, entity subcategories, inherent attributes, and entities. The preprocessing module is used to perform hierarchical and targeted preprocessing on multi-source heterogeneous shale oil data based on the hierarchy and evaluation business dimensions of the three-level hierarchical ontology framework, so as to construct a shale oil labeled dataset. The extraction module is used to extract knowledge from the shale oil labeled dataset to obtain knowledge triples containing entities, inherent attributes, and relationships. The review module is used to perform entity normalization, relational logic verification and quality review for the knowledge triples in the shale oil field, and import the qualified knowledge triples into the graph database to store the knowledge graph. The reasoning module is used to receive query statements for shale oil measure evaluation, extract primary entities, secondary entities and evaluation intent from the query statements, and perform semantic similarity matching in the knowledge graph based on the extracted information to retrieve related knowledge subgraphs. The query statement and the associated knowledge subgraph are combined to construct domain-enhanced prompt text, which is then input into a large language model. The large language model combines business knowledge in the shale oil field to conduct reasoning analysis and output the evaluation results of shale oil measures.

9. An apparatus for constructing a knowledge graph for evaluating shale oil measures, characterized in that, include: processor; Memory used to store processor-executable instructions; When the processor executes the executable instructions, it implements the method as described in any one of claims 1 to 7.

10. A non-volatile computer-readable storage medium, characterized in that, Includes storage of computer programs or instructions that, when executed, cause the method as described in any one of claims 1 to 7 to be implemented.

Images