Construction method of oil and gas production large model platform

Abstract

The invention discloses a construction method of an oil and gas production large model platform, and relates to the technical field of oil and gas production intellectualization, and the method comprises the following steps: pre-constructing an oil and gas field knowledge graph, employing a BERT-BiLSTM-CRF model to carry out the automatic extraction of entities and attributes of massive unstructured engineering documents in oil and gas production, the extraction result is used for filling and perfecting the knowledge graph; a WellGPT model focusing on professional knowledge reasoning and a WellGPT-VL model focusing on view fusion understanding are constructed respectively. According to the method, on the basis of Qwen2.5 double bases and the dynamic weight hybrid LoRA technology based on intention recognition, efficient fine adjustment of a universal base model is achieved through precise classification of an intention recognition scheduling center and special training of an independent LoRA adapter, the generalization ability of the universal model is reserved, professional knowledge of the oil and gas industry is deeply injected, and the method has the advantages of being high in universality, high in practicability and high in practicability. And professional terms, physical laws and multi-modal data characteristics in the field can be accurately understood, and the professional adaptability and output reliability of the model in an oil and gas production scene are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent oil and gas production technology, and more specifically, to a method for constructing a large-scale oil and gas production model platform. Background Technology

[0002] Currently, the global energy industry is at a critical juncture of digital transformation and leapfrogging towards intelligent manufacturing. Oil and gas production, as a typical process industry, faces numerous challenges, including complex geological conditions, changing extraction environments, and heterogeneous multi-source data, placing extremely high demands on the accuracy and timeliness of production decisions. With the deepening development of the Industrial Internet, oil and gas fields have accumulated massive amounts of production data and technical documents. How to activate these data assets and transform them into explicit knowledge to guide production has become a core issue for the oil and gas industry to achieve cost reduction, efficiency improvement, and high-quality development.

[0003] In recent years, large language model (LLM) technology based on the Transformer architecture has made groundbreaking progress, providing a new technological paradigm for the development of industrial intelligence. However, directly applying general-purpose large models to the oil and gas production field still faces significant challenges: First, general-purpose models lack in-depth knowledge accumulation in vertical domains, making them prone to factual errors or "machine illusions" when faced with oil and gas-specific corpora; second, oil and gas production scenario data has highly multimodal characteristics, making it difficult for existing models to achieve semantic alignment and fusion analysis of cross-modal data; third, industrial scenarios have extremely high requirements for the interpretability and security of decision-making, and purely data-driven models cannot meet the stringent standards of physical mechanism constraints in engineering sites. Meanwhile, in the vertical oil and gas sector, although some companies have built large model ecosystems, existing large models mostly focus on the generalization capabilities of the entire industry chain, lacking deep adaptation to the coupling relationships of multiple factors in the "geology, engineering, and production" stages of production, and facing difficulties in lightweight deployment and engineering-based closed-loop implementation, making it difficult to support precise production decision-making.

[0004] Therefore, there is an urgent need for a dedicated large-scale model construction method that adapts to the core needs of oil and gas production scenarios, and can specifically address the aforementioned technical pain points. Summary of the Invention

[0005] In response to the problems in related technologies, this invention proposes a method for constructing a large-scale oil and gas production model platform to overcome the aforementioned technical problems existing in the existing related technologies.

[0006] The technical solution of this invention is implemented as follows:

[0007] A method for constructing a large-scale oil and gas production model platform includes the following steps:

[0008] A knowledge graph in the oil and gas field is pre-built, and the BERT-BiLSTM-CRF model is used to automatically extract entities and attributes from massive unstructured engineering documents in oil and gas production. The extraction results are used to fill and improve the knowledge graph.

[0009] Based on the Qwen2.5-7B and Qwen2.5-VL-7B dual bases, the dynamic weight hybrid LoRA technology based on intent recognition is used to efficiently fine-tune the parameters, and the WellGPT model focusing on professional knowledge reasoning and the WellGPT-VL model focusing on visual text fusion understanding are constructed respectively.

[0010] A five-layer technical architecture is constructed, comprising a data layer, a platform layer, a model layer, a tool layer, and a workflow layer. These layers work collaboratively to achieve end-to-end operation from data support to model collaboration and scenario application. Specifically, the data layer uses a knowledge graph as its core to build a data foundation, enabling multi-source data fusion; the platform layer relies on open-source tools, building a basic technical framework based on the Neo4j database, LLaMA-Factory, and Python; the model layer performs scenario-based fine-tuning of the Qwen series models, using the fine-tuned large model as a central hub for summary and scheduling; the tool layer builds various domain-specific small models using FastAPI as a tool reserve, including numerical simulation models, dynamic production capacity prediction models, and equipment fault early warning models; and the workflow layer integrates the capabilities of the first four layers, providing customized workflow support for various oil and gas production scenarios.

[0011] By orchestrating collaborative workflows between large models and specialized small models through the Dify platform, customized application solutions can be formed to adapt to different oil and gas production business scenarios.

[0012] The construction of the knowledge graph in the oil and gas field adopts a hybrid domain ontology construction method, which includes the following steps:

[0013] The model layer is constructed from top to bottom, taking the business logic of the entire process of oil and gas field development as the framework, and referring to industry technical standards, professional textbooks and domain theoretical systems to build a top-level ontology, including five major categories: reservoir, drilling, completion, oil production and support entities, which are then subdivided into second-level sub-domains and third-level core entities.

[0014] The entity layer is built from the bottom up, extracting four levels of specific instances and semantic relationships from actual data sources such as engineering manuals and failure cases to fill in the details of the schema layer framework.

[0015] The entity data standard, combined with the characteristics of the oil and gas field development industry, optimizes entity granularity, improves cross-domain relationships, and corrects deviations through domain expert review.

[0016] The BERT-BiLSTM-CRF model includes: a BERT module, a Bi-LSTM module, and a CRF module, wherein;

[0017] The BERT module is used to transform the input text sequence into a word vector sequence that integrates multi-dimensional information. The word vector sequence is the sum of elements of word embedding, position embedding and sentence information embedding.

[0018] The Bi-LSTM module is used to further model the vector sequence output by BERT, capture the bidirectional contextual dependencies of the sequence, and output a feature representation containing global semantics.

[0019] The CRF module is used to learn the transition probabilities between labels based on the output of Bi-LSTM, and output the optimal entity label sequence by constraining the legality of the label sequence.

[0020] The dynamic weighted hybrid LoRA technology based on intent recognition for efficient parameter fine-tuning includes the following steps:

[0021] An intent recognition scheduling center is constructed, and an intent recognition module is built based on a lightweight TextCNN network. The TextCNN network transforms discrete instruction text into a vector matrix, and extracts key semantic features from the instructions through supervised training to determine the business affiliation of user instructions.

[0022] Independent LoRA adapter training was conducted based on the Qwen2.5 base model. The LLaMA-Factory framework was used to fine-tune LoRA for three sub-corpora: working condition diagnosis, oil and gas reservoir geology and engineering technology, to train three independent LoRA adapter weights.

[0023] By performing dynamic weight combination inference, TextCNN maps user input commands to intent vectors corresponding to business categories. The model output weights are a linear combination of basic weights and weighted LoRA increments, dynamically adjusting the contribution of knowledge from different domains to inference.

[0024] The training data for the TextCNN network is a command fine-tuning dataset, which is constructed from approximately 1,500 typical Prompt commands randomly selected from three sub-corpora: working condition diagnosis, oil and gas reservoir geology, and oil production engineering, forming standardized training sample pairs.

[0025] The specialized mini-models include: numerical simulation mini-model, dynamic capacity prediction mini-model, operating condition diagnosis mini-model, and equipment fault early warning mini-model. The collaborative workflow enables the tool-based collaboration between the large model and the specialized mini-models.

[0026] The beneficial effects of this invention are:

[0027] This invention deeply integrates a hybrid ontology construction method with the BERT-BiLSTM-CRF model to build a specialized knowledge graph covering the entire oil and gas production process. This enables precise and automated extraction of entities and attributes from massive amounts of unstructured engineering documents, laying a solid and high-quality domain data foundation for model operation. It fundamentally solves the core pain point of general-purpose large models lacking professional knowledge in the oil and gas field. Relying on the Qwen2.5 dual-base architecture and intention-recognition-based dynamic weight hybrid LoRA technology, through precise classification by the intention recognition scheduling center and specialized training of independent LoRA adapters, it achieves efficient fine-tuning of the general-purpose base model. This retains the generalization ability of the general model while deeply injecting professional knowledge of the oil and gas industry. It can accurately understand domain-specific terminology, physical laws, and multimodal data characteristics, significantly reducing the risk of machine illusion and significantly improving the model's professional adaptability and output reliability in oil and gas production scenarios. The invention also innovatively designs… The five-layer technical architecture of the system, through the collaborative linkage of the data layer, platform layer, model layer, tool layer, and workflow layer, achieves a closed-loop operation across the entire chain, including data support, technology construction, model scheduling, professional tool reserves, and scenario adaptation. Combined with the workflow orchestration capabilities of the Dify platform, it effectively opens up the collaborative channel between large models and specialized small models, meeting the stringent requirements for lightweight deployment and engineering implementation in oil and gas production sites. It can quickly respond to customized application demands from different business scenarios. The application of this large model platform in core oil and gas production scenarios has successfully promoted the transformation of production decision-making from experience-driven to data-driven intelligent decision-making. It not only assists engineering technicians in improving the accuracy and scientific nature of decision-making, but also optimizes business processes, simplifies work links, and reduces excessive reliance on human experience. It provides a complete and feasible technical path for the digital and intelligent transformation of the oil and gas industry, and plays an important supporting and promoting role in the industry's cost reduction, efficiency improvement, and high-quality development. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0029] Figure 1 This is a flowchart of the knowledge graph construction method for a large-scale oil and gas production model platform according to an embodiment of the present invention.

[0030] Figure 2 This is a BERT-BiLSTM-CRF model diagram of a method for constructing a large-scale oil and gas production model platform according to an embodiment of the present invention;

[0031] Figure 3 This is a WellGPT architecture diagram of an oil and gas production large-scale model platform construction method according to an embodiment of the present invention.

[0032] Figure 4 This is a LoRA fine-tuning diagram of a method for constructing a large-scale oil and gas production model platform according to an embodiment of the present invention;

[0033] Figure 5 This is a scenario application example of a method for constructing a large oil and gas production model platform according to an embodiment of the present invention. In the figure, (a) is a multimodal pumping unit condition diagnosis diagram, (b) is an oil and gas reservoir development scheme recommendation diagram, and (c) is an oil production engineering scheme preparation diagram.

[0034] Figure 6 This is a confusion matrix diagram of a method for constructing a large-scale oil and gas production model platform according to an embodiment of the present invention;

[0035] Figure 7 This is an expert scoring radar chart of a method for constructing a large-scale oil and gas production model platform according to an embodiment of the present invention;

[0036] Figure 8 This is a bar chart showing the retrieval recall rate of a method for constructing a large-scale oil and gas production model platform according to an embodiment of the present invention.

[0037] Figure 9 This is a comparison chart of the operational efficiency of a method for constructing a large-scale oil and gas production model platform according to an embodiment of the present invention. Detailed Implementation

[0038] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention are within the scope of protection of the present invention.

[0039] According to an embodiment of the present invention, a method for constructing a large-scale oil and gas production model platform is provided.

[0040] like Figure 1 As shown, the method for constructing a large-scale oil and gas production model platform according to an embodiment of the present invention includes the following steps:

[0041] Step S1: Pre-construct a knowledge graph in the oil and gas field, and use the BERT-BiLSTM-CRF model to automatically extract entities and attributes from massive unstructured engineering documents in oil and gas production. Use the extraction results to fill and improve the knowledge graph.

[0042] Among them, such as Figure 1 As shown, a knowledge graph for the oil and gas field is constructed using a hybrid domain ontology construction method. First, based on the business logic of the entire oil and gas field development process—reservoir evaluation, drilling, completion, and production—a top-level ontology is built, referencing industry technical standards, professional textbooks, and domain theoretical systems. This ontology encompasses five major categories: reservoir, drilling, completion, production, and support entities. Then, second-level subdomains are subdivided; for example, the reservoir domain is subdivided into reservoir type, parameters, and fluids. Third-level core entities, such as drill bits and drill pipes under drilling tools, complete the model layer construction. Next, fourth-level specific instances, such as PDC drill bits and CO2 flooding, along with semantic relationships (e.g., oil well cement used in cementing processes), are extracted from actual data sources such as engineering manuals and failure cases to fill the entity layer. Finally, the entity granularity is optimized based on industry characteristics; for example, casing is subdivided into surface, technical, and production casing. Cross-domain relationships are improved, and deviations are corrected through domain expert review, forming the final oil and gas domain knowledge graph.

[0043] This involves extracting entities and attributes from unstructured documents and using the BERT-BiLSTM-CRF model to process massive amounts of unstructured engineering documents such as geological reports, drilling records, and production data reports.

[0044] This technical solution, such as Figure 2 As shown, the BERT-BiLSTM-CRF model consists of three parts: the underlying BERT module transforms the input text sequence into a word vector sequence that incorporates multi-dimensional information. The intermediate Bi-LSTM module further models the vector sequence output by BERT, capturing the bidirectional contextual dependencies of the sequence and outputting feature representations containing global semantics. The top-level CRF module learns the transition probabilities between labels based on the output of Bi-LSTM. By constraining the legality of the label sequence (e.g., B-Title can be followed by I-Title but not B-Hazard), it outputs the optimal entity label sequence, achieving accurate structured extraction of entities and their attributes from drilling and completion engineering design documents. The BERT module output... The calculation formula is the sum of the elements of word embedding, position embedding, and sentence information embedding, expressed as:

[0045]

[0046] in, The word embedding encoding result for word vectors. The location information encoding result, The three components—word vectors, lexical semantics, positional context, and sentence role—are integrated to encode the information of the sentence.

[0047] Step S2, based on the Qwen2.5-7B and Qwen2.5-VL-7B dual-base, uses the dynamic weighted hybrid LoRA technology based on intent recognition for efficient parameter fine-tuning, and constructs the WellGPT model focused on professional knowledge reasoning and the WellGPT-VL model focused on visual-text fusion understanding, respectively; specifically including the following steps:

[0048] An intent recognition and scheduling hub is constructed based on a lightweight TextCNN network, responsible for determining the business attribute of user commands. TextCNN first transforms the discrete command text into a vector matrix, represented as:

[0049] ;

[0050] Where X is the input matrix of TextCNN, For the words after word segmentation, Let n be a d-dimensional vector, and n be the sentence length.

[0051] Its training data directly originates from the instruction fine-tuning dataset constructed in this invention: approximately 1500 typical Prompt instructions are randomly extracted from three sub-corpora—operating condition diagnosis, oil and gas reservoir geology, and oil production engineering—to construct standardized training sample pairs (Instruction, Label). The Label corresponds to the three major business categories mentioned above. Through supervised training, TextCNN can efficiently extract key semantic features from instructions, such as "dynamometer card diagnosis," "remaining oil distribution," and "fracturing scheme." A 2-4 gram window filter is used, and a single convolution is calculated as follows:

[0052] ;

[0053] By concatenating all window outputs, we obtain the feature map:

[0054] ;

[0055] Where h is the window size. These are the filter weights. Core features are selected using max pooling:

[0056] ;

[0057] splicing pooling results The fully connected layer maps the result to intent probabilities.

[0058] ;

[0059] Where F is the total number of filters, y is the probability distribution of the three types of intents, and the term with the highest probability is the instruction intent, thus enabling TextCNN to have high-precision intent classification capabilities and providing a reliable basis for subsequent dynamic routing.

[0060] After completing the routing module construction, in order to create an "expert model" focused on specific tasks, a decoupling strategy is adopted during the large model training phase, such as... Figure 4 The LoRAAdapters section on the right. Based on the Qwen2.5 base model, the LLaMA-Factory framework is used to independently fine-tune LoRA for three sub-corpora: operational condition diagnosis, reservoir geology, and engineering technology, training three independent LoRA adapter weights respectively. , and The yellow hourglass-shaped structure in the image illustrates the low-rank decomposition principle of LoRA, which simulates parameter updates through the product of two low-rank matrices A and B, significantly reducing GPU memory usage. The underlying Qwen2.5-7B base parameters are shown during model training. Keep the system frozen and only update the parameters of the upper-level LoRA adapter.

[0061] Finally, in the inference phase, this fine-tuning strategy achieves a closed-loop chain from intent perception to parameter combination. When the user inputs a question, such as... Figure 4 On the left, TextCNN first analyzes the instruction text and maps it into three types of intent vectors: operational condition diagnosis, reservoir development, or engineering plan. The model then outputs weights. This can be expressed as a linear combination of the basic weights and the weighted LoRA increments:

[0062] ;

[0063] in, These are the frozen pre-trained weights; For the first A domain-specific low-rank update matrix, consisting of two low-rank matrices. and The product constitutes ; These are the gating coefficients or attention weights output by the intent recognition module, used to dynamically adjust the contribution of different domain knowledge to the final inference. For example, when a dynamometer card diagnostic request is input, only the relevant domain knowledge is activated. Adapter. This not only ensures the professional accuracy of the answers, but also enables efficient reuse of parameters and logical consistency across modal scenarios.

[0064] Step S3: Build a five-layer technical architecture that includes a data layer, platform layer, model layer, tool layer, and workflow layer. Each layer works together to achieve full-link operation from data support to model collaboration and then to scenario application.

[0065] Specifically, such as Figure 3As shown, the data layer uses the knowledge graph constructed in step S1 as its core to achieve the fusion and storage of multi-source data such as geology, engineering, and technology; the platform layer relies on open-source tools such as Neo4j database, LLaMA-Factory, and Python to build a basic technical framework and provide basic support for data storage and model training; the model layer uses the WellGPT and WellGPT-VL models constructed in step S2 as the summary and scheduling center, responsible for knowledge reasoning and task scheduling; the tool layer uses FastAPI to build small models in various fields such as numerical simulation, dynamic production capacity prediction, and equipment fault early warning, forming a professional tool reserve; the workflow layer integrates the capabilities of the first four layers and realizes the full-link automated operation of data processing, model calling, and tool collaboration through process orchestration.

[0066] Step S4 involves orchestrating a collaborative workflow between large models and specialized small models through the Dify platform to adapt to different oil and gas production business scenarios and form customized application solutions.

[0067] This technical solution orchestrates a collaborative workflow between large models and specialized small models through the Dify platform, adapting to core scenarios such as multimodal pumping unit condition diagnosis, oil and gas reservoir development scheme recommendation, and oil production engineering scheme preparation, as detailed below:

[0068] In the multimodal pumping unit condition diagnosis scenario, WellGPT leverages its multimodal fusion capabilities to perform multimodal fusion analysis on dynamometer cards and reservoir fluid property data. Combined with prompts, it diagnoses the pumping unit's operating conditions, achieving a comprehensive assessment of its status, such as normal production, wax deposition, insufficient fluid supply, decreased pump efficiency, and sand production in the well. (See [link to relevant documentation]). Figure 5 (a) Once an anomaly is identified, the fault type is automatically located, and it is accurately determined whether the leakage is caused by pump barrel wear or the dynamometer card distortion caused by rod breakage. The degree of impact of the fault and the solution are output to help maintenance personnel respond quickly and upgrade the handling of pumping well faults from "experience-driven, passive maintenance" to "data-driven, intelligent diagnosis" to ensure continuous and stable oil and gas production.

[0069] The recommended scenario for oil and gas reservoir development schemes uses WellGPT as the core collaborative hub, relying on standardized data and professional toolchains to achieve standardized operation of the entire process through three stages: "data input and preprocessing—similar oil and gas reservoir retrieval—development scheme generation and analysis." First, it inputs data such as geological reports and well logging curves, and uses industry feature extraction scripts to form a structured feature set including reservoir structure and production capacity. Then, it connects to a knowledge graph containing over 500 cases, and uses vector retrieval tools adapted to the oil and gas industry to filter the Top 5 similar oil and gas reservoirs and output analogy results with sources. Finally, it integrates professional tools such as geological model analysis and production capacity prediction, combines development thresholds to determine the direction, simulates production dynamics, and generates feasible schemes, such as... Figure 5 (b);

[0070] In the oil production engineering scheme development scenario, WellGPT is used as the central hub. Relying on a knowledge graph of "oil and gas reservoir-process-equipment" and specialized computing tools, a fully automated system is constructed, encompassing "data integration-process matching-parameter optimization-scheme generation-dynamic adaptation." This system achieves a transformation from "experience-driven" to "data-driven intelligent generation" through standardized integration of multi-source data, a dual mechanism of "rule-based reasoning + case analogy" for process matching, parameter optimization using linked small models, standardized scheme text generation, and dynamic iterative adaptation. The output schemes conform to industry standards, such as... Figure 5 (c).

[0071] To comprehensively verify the actual effectiveness of WellGPT in complex oil and gas production scenarios, this paper constructs a test set based on real oilfield production data, and conducts a systematic quantitative evaluation from four dimensions: fault diagnosis accuracy, knowledge retrieval capability, expert comprehensive evaluation, and operational efficiency. The experimental results are as follows: Figure 6-9 As shown.

[0072] Among them, the multimodal operating condition diagnostic accuracy analysis: For the operating condition diagnostic scenario of the pumping unit, six typical operating condition samples were selected for testing, including normal, insufficient fluid supply, gas influence, pump leakage, valve leakage, and rod breakage. For example... Figure 6 As shown in the confusion matrix, WellGPT-VL exhibits a significant diagonal advantage across all categories, demonstrating extremely high overall classification accuracy. Particularly in the easily confused categories of "Pump Leakage" and "Valve Leakage," the model, leveraging its multimodal visual-semantic alignment capabilities, accurately identified 290 and 260 samples respectively, with an extremely low false positive rate, proving its robustness in feature extraction under complex operating conditions.

[0073] The evaluation included a multi-dimensional expert assessment: to assess the professionalism of the generated content, five senior engineers in the oil and gas field were invited to conduct blind reviews and scoring of the oil production engineering plans generated by the model. The evaluation dimensions covered safety, accuracy, format compliance, feasibility, and completeness. Figure 7 The radar chart shows that WellGPT (blue area) encompasses and surpasses the general-purpose, untuned large model (red dashed line) in all dimensions. Its advantages are particularly evident in the "safety" and "accuracy" metrics, indicating that the vertically tuned WellGPT effectively complies with industrial safety standards, avoids the "illusion" risks common in general-purpose models, and better meets the requirements of field engineering.

[0074] Among them, knowledge retrieval and question answering performance evaluation: In the scenario of recommending oil and gas reservoir development plans, accurate knowledge recall is a prerequisite for generating high-quality plans. Experiments compared the hit rates (HitRate) of Keyword Search, GeneralRAG, and the proposed WellGPT method under different truncation thresholds. Figure 8 As shown, thanks to the enhanced knowledge graph in the oil and gas field, WellGPT outperforms all other metrics in HitRate@1, @5, and @10. Particularly noteworthy is its HitRate@1 (first hit rate), where WellGPT achieves 52.0%, significantly outperforming the general RAG's 45.2%. This demonstrates that the model can more accurately understand the deep semantics of geological engineering terms, greatly reducing noise in information retrieval.

[0075] Among them, the overall process efficiency analysis: In terms of actual operational efficiency, the time consumption of the traditional manual mode and the WellGPT-assisted mode in the entire process was compared. For example... Figure 9 As shown, the traditional model takes an average of 165 minutes to complete a full fault diagnosis and report preparation, while the WellGPT model only takes 7.5 minutes, achieving a 22.0-fold efficiency improvement. Among them, in the most time-consuming "Report Writing" stage, the time cost is drastically reduced from 90 minutes to 5.0 minutes, greatly freeing up the productivity of technical personnel and enabling them to focus on higher-value decision-making work.

[0076] In summary, by utilizing the technical solutions described above in this invention, through knowledge graph construction, precise entity extraction, efficient parameter fine-tuning, and multi-layer architecture design, the adaptability problem of general large models in the oil and gas production field is effectively solved, realizing intelligent and efficient production decision-making and providing strong support for the high-quality development of the oil and gas industry.

[0077] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Those skilled in the art, upon considering the disclosure in the specification and embodiments, will readily conceive of other embodiments of this disclosure. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the claims.

[0078] It should be understood that this disclosure is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this disclosure is limited only by the appended claims.

Claims

1. A method for constructing a large-scale oil and gas production model platform, characterized in that, Includes the following steps: A knowledge graph in the oil and gas field is pre-built, and the BERT-BiLSTM-CRF model is used to automatically extract entities and attributes from massive unstructured engineering documents in oil and gas production. The extraction results are used to fill and improve the knowledge graph. Based on the Qwen2.5-7B and Qwen2.5-VL-7B dual bases, the dynamic weight hybrid LoRA technology based on intent recognition is used to efficiently fine-tune the parameters, and the WellGPT model focusing on professional knowledge reasoning and the WellGPT-VL model focusing on visual text fusion understanding are constructed respectively. A five-layer technical architecture is built, comprising a data layer, a platform layer, a model layer, a tool layer, and a workflow layer. Each layer works collaboratively to achieve end-to-end operation from data support to model collaboration and scenario application. Among them, the data layer builds a data foundation with knowledge graphs as the core to achieve multi-source data fusion; the platform layer is supported by open-source tools and relies on Neo4j database, LLaMA-Factory, and Python to build the basic technical framework. The model layer performs scenario-based fine-tuning on the Qwen series models, and uses the fine-tuned large model as the summary and scheduling center; the tool layer builds various domain-specific small models in the form of FastAPI as a tool reserve, including numerical simulation small models, dynamic production capacity prediction small models, and equipment fault early warning small models; the workflow layer integrates the capabilities of the first four layers to provide customized workflow support for various oil and gas production scenarios. By orchestrating collaborative workflows between large models and specialized small models through the Dify platform, customized application solutions can be formed to adapt to different oil and gas production business scenarios.

2. The method for constructing a large-scale oil and gas production model platform according to claim 1, characterized in that, The construction of the knowledge graph in the oil and gas field adopts a hybrid domain ontology construction method, including the following steps: The model layer is constructed from top to bottom, taking the business logic of the entire process of oil and gas field development as the framework, and referring to industry technical standards, professional textbooks and domain theoretical systems to build a top-level ontology, including five major categories: reservoir, drilling, completion, oil production and support entities, which are then subdivided into second-level sub-domains and third-level core entities. The entity layer is built from the bottom up, extracting four levels of specific instances and semantic relationships from engineering manuals, failure cases, and actual data sources to fill in the details of the schema layer framework. The entity data standard, combined with the characteristics of the oil and gas field development industry, optimizes entity granularity, improves cross-domain relationships, and corrects deviations through domain expert review.

3. The method for constructing a large-scale oil and gas production model platform according to claim 1, characterized in that, The BERT-BiLSTM-CRF model includes: a BERT module, a Bi-LSTM module, and a CRF module, wherein; The BERT module is used to transform the input text sequence into a word vector sequence that integrates multi-dimensional information. The word vector sequence is the sum of elements of word embedding, position embedding and sentence information embedding. The Bi-LSTM module is used to further model the vector sequence output by BERT, capture the bidirectional contextual dependencies of the sequence, and output a feature representation containing global semantics. The CRF module is used to learn the transition probabilities between labels based on the output of Bi-LSTM, and output the optimal entity label sequence by constraining the legality of the label sequence.

4. The method for constructing a large-scale oil and gas production model platform according to claim 1, characterized in that, The intent-based dynamic weighted hybrid LoRA technique for efficient parameter fine-tuning includes the following steps: An intent recognition scheduling center is constructed, and an intent recognition module is built based on a lightweight TextCNN network. The TextCNN network transforms discrete instruction text into a vector matrix, and extracts key semantic features from the instructions through supervised training to determine the business affiliation of user instructions. Independent LoRA adapter training was conducted based on the Qwen2.5 base model. The LLaMA-Factory framework was used to fine-tune LoRA for three sub-corpora: working condition diagnosis, oil and gas reservoir geology and engineering technology, to train three independent LoRA adapter weights. By performing dynamic weight combination inference, TextCNN maps user input commands to intent vectors corresponding to business categories. The model output weights are a linear combination of basic weights and weighted LoRA increments, dynamically adjusting the contribution of knowledge from different domains to inference.

5. The method for constructing a large-scale oil and gas production model platform according to claim 4, characterized in that, The training data for the TextCNN network is an instruction fine-tuning dataset, which is constructed from approximately 1,500 typical Prompt instructions randomly selected from three sub-corpora: working condition diagnosis, oil and gas reservoir geology, and oil production engineering, forming standardized training sample pairs.

6. The method for constructing a large-scale oil and gas production model platform according to claim 1, characterized in that, The specialized mini-models include: numerical simulation mini-model, dynamic capacity prediction mini-model, operating condition diagnosis mini-model, and equipment fault early warning mini-model. The collaborative workflow enables the tool-based collaboration between the large model and the specialized mini-models.

Images