Intelligent question-answering system and method for molecular oil refining

Abstract

The invention discloses an intelligent question-answering system and method for molecular oil refining, and relates to the technical field of oil refining chemical engineering and computer software, and the system comprises a molecular database which is used for storing various types of crude oil data vectors; the intention recognition module is used for extracting intentions in user query based on a pre-trained intention recognition model; the vector data processing module is used for retrieving a plurality of crude oil data vectors from the molecular database according to an accurate matching retrieval mode and / or a similarity retrieval mode, wherein the matching value of the crude oil data vectors and the user intention exceeds a preset value; and the response generation module is used for generating an expression form matched with the intention based on the plurality of crude oil data vectors with the matching values exceeding the preset value, and pushing the expression form to the user. Through natural language processing and machine learning technologies, user demands are responded and understood in real time, and automatic analysis and processing of crude oil data are realized.

Description

Technical Field

[0001] This invention relates to the fields of oil refining and chemical engineering, and computer software technology, and in particular to a molecular oil refining intelligent question-and-answer system and method. Background Technology

[0002] This section is intended to provide background or context for the embodiments of the invention set forth in the claims. The description herein is not an admission that it is prior art simply because it is included in this section.

[0003] With the development of refining technology, the demand for accurate analysis and processing of crude oil data is increasing. Traditional data processing methods rely on manual operation and experience-based judgment, which are inefficient and prone to errors. Processing methods based on preset rule matching lack flexibility and are difficult to cope with complex and changing operating conditions, abnormal scenarios, or new types of data. At the same time, the problem of data silos is serious, requiring manual collection by users, resulting in low levels of intelligence and difficulty in supporting production decisions. Summary of the Invention

[0004] This invention provides a molecular refining intelligent question-and-answer system to integrate crude oil data and respond to and understand user needs in real time through natural language processing and machine learning technologies, thereby achieving automated analysis and processing of crude oil data. The system includes: a molecular database, an intent recognition module, a vector data processing module, and a response generation module.

[0005] The molecular database is used to store various types of crude oil data vectors. These crude oil data vector types include molecular structure codes, process parameter vectors, and crude oil property vectors. Different types of crude oil data vectors are linked together through association indexes.

[0006] The intent recognition module is used to: extract the intent from user queries based on a pre-trained intent recognition model; the intent recognition model is obtained by fine-tuning a pre-trained language model using technical data from the oil refining field;

[0007] The vector data processing module is used to: retrieve multiple crude oil data vectors from the molecular database that match the user's intent with a value exceeding a preset value, based on the exact match retrieval method and / or the similarity retrieval method; the exact match retrieval method uses an inverted index to retrieve crude oil data vectors that perfectly match the user's intent;

[0008] The response generation module is used to: generate an expression form for intent adaptation based on multiple crude oil data vectors whose matching values ​​exceed a preset value, and push it to the user.

[0009] This invention also provides a molecular refining intelligent question-answering method to integrate crude oil data, and through natural language processing and machine learning technologies, to respond to and understand user needs in real time, thereby achieving automated analysis and processing of crude oil data. The method includes:

[0010] Receive user query requests;

[0011] The intent in a user's query request is extracted based on a pre-trained intent recognition model; the intent recognition model is obtained by fine-tuning a pre-trained language model using technical data from the oil refining field.

[0012] Based on the exact match retrieval method and / or the similarity retrieval method, multiple crude oil data vectors with a matching value exceeding a preset value with the user's intent are retrieved from the molecular database; the exact match retrieval method uses an inverted index to retrieve crude oil data vectors that perfectly match the user's intent; the molecular database is used to store various types of crude oil data vectors; the crude oil data vector types include molecular structure codes, process parameter vectors, and crude oil property vectors, and different types of crude oil data vectors are linked to each other through association indexes;

[0013] Based on multiple crude oil data vectors whose matching values ​​exceed preset values, an expression for intent adaptation is generated and pushed to the user.

[0014] This invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the above-described molecular refining intelligent question-and-answer method.

[0015] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described molecular refining intelligent question-and-answer method.

[0016] This invention also provides a computer program product, which includes a computer program that, when executed by a processor, implements the above-described molecular refining intelligent question-and-answer method.

[0017] This invention constructs a molecular database to store various types of crude oil data vectors, including molecular structure codes, process parameter vectors, and crude oil property vectors. Different types of crude oil data vectors are linked together through association indexes, achieving the integration of crude oil data. This allows for the full exploration of the correlations between crude oil data and the effective use of crude oil data for decision-making. Furthermore, this invention utilizes natural language processing and machine learning technologies to identify user intent, respond to and understand user needs in real time, and assist users in processing molecular refining data. This enables automated analysis and processing of crude oil data, significantly improving the efficiency and accuracy of refining data processing, helping users make better refining decisions, and enhancing the user experience. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art 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. In the drawings:

[0019] Figure 1 This is a schematic diagram of the molecular refining intelligent question-and-answer system in an embodiment of the present invention;

[0020] Figure 2 This is a specific example diagram of the molecular refining intelligent question-and-answer system in an embodiment of the present invention;

[0021] Figure 3 This is a schematic diagram of the precise matching retrieval process in an embodiment of the present invention;

[0022] Figure 4 This is a schematic diagram of the similarity retrieval process in an embodiment of the present invention;

[0023] Figure 5 This is a flowchart illustrating the intelligent question-and-answer method for molecular refining in an embodiment of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. Here, the illustrative embodiments of the present invention and their descriptions are used to explain the present invention, but are not intended to limit the present invention.

[0025] The acquisition, storage, use, and processing of data in this application all comply with the relevant provisions of national laws and regulations.

[0026] In the integration of computer software technology with the oil refining and chemical industry, existing technologies typically rely on manual operation, experience-based judgment, and pre-defined rule matching to achieve accurate analysis and processing of crude oil data. While these traditional methods meet basic needs to a certain extent, the increasing demands on data processing as refining technology develops have revealed their shortcomings. Furthermore, most existing refining data processing systems focus on data collection and storage, lacking the ability for in-depth data analysis and application.

[0027] Therefore, this invention overcomes the limitations of existing technologies and proposes a molecular refining intelligent question-and-answer system capable of understanding and responding to user queries, thereby achieving automated analysis and processing of crude oil data. Based on user queries, this system can provide accurate crude oil data and refining suggestions, improving the efficiency and accuracy of the refining process.

[0028] Figure 1 This is a schematic diagram of the molecular refining intelligent question-and-answer system in an embodiment of the present invention, such as... Figure 1 As shown, the system 100 includes: a molecular database 101, an intent recognition module 102, a vector data processing module 103, and a response generation module 104;

[0029] Molecular database 101 is used to store various types of crude oil data vectors; the types of crude oil data vectors include molecular structure codes, process parameter vectors, and crude oil property vectors, and different types of crude oil data vectors are linked to each other through association indexes;

[0030] The intent recognition module 102 is used to: extract the intent from the user query based on a pre-trained intent recognition model; the intent recognition model is obtained by fine-tuning the pre-trained language model using technical data from the oil refining field;

[0031] The vector data processing module 103 is used to: retrieve multiple crude oil data vectors from the molecular database 101 with a matching value exceeding a preset value for the user's intent based on the exact match retrieval method and / or the similarity retrieval method; the exact match retrieval method uses an inverted index to retrieve crude oil data vectors that perfectly match the user's intent;

[0032] The response generation module 104 is used to: generate an expression form for intent adaptation based on multiple crude oil data vectors whose matching values ​​exceed preset values, and push it to the user.

[0033] The following describes in detail the molecular refining intelligent question-and-answer system in the embodiments of the present invention.

[0034] This invention provides an embodiment of a molecular database for storing crude oil data vectors. This molecular database is designed to efficiently store and retrieve large amounts of crude oil data vectors. The database design considers the storage requirements of high-dimensional data and optimizes the data index structure to support fast data querying and similarity calculation. Each crude oil data vector stored in the database contains key information such as basic information, molecular structure, and properties of the crude oil. This information is encoded in vector form to facilitate efficient calculation and comparison.

[0035] Figure 2 This is a specific example diagram of the molecular refining intelligent question-and-answer system in an embodiment of the present invention, as shown below. Figure 2 As shown, the molecular database in this system can receive crude oil molecular structure data, process parameter data, physical property index data, etc., and perform molecular structure encoding, process parameter vectorization, etc., to form different types of crude oil data vectors, which are finally stored in the molecular database. Different types of crude oil data vectors in the molecular database are linked to each other through association indexes.

[0036] In one embodiment, the system may further include a data preprocessing module for: receiving crude oil data and preprocessing and vectorizing the crude oil data; the preprocessing includes data cleaning and / or standardization. The preprocessing flow is as follows:

[0037] Molecular data cleaning: Clean the crude oil molecular composition data, remove abnormal molecular structures, standardize molecular naming rules, and ensure the consistency of molecular structures.

[0038] Process data standardization: Standardize the refining process parameters, including unifying the temperature unit (degrees Celsius), pressure unit (MPa), and flow rate unit (t / h).

[0039] Quality data verification: Detect and correct outliers for crude oil quality indicators, and use statistical methods to identify and correct outlier data points.

[0040] In one embodiment, the molecular structure encoding is obtained by encoding the hydrocarbon molecular structure in crude oil using principal component analysis (PCA); the process parameter vector is obtained by converting multiple refining process parameters into standardized vectors, and the refining process parameters are associated with each other using process knowledge graph technology; the crude oil property vector is obtained by mapping the physicochemical properties of crude oil into a high-dimensional vector space and establishing the correspondence between properties and molecular structure.

[0041] In this embodiment, the molecular database design considers the storage and retrieval needs of high-dimensional data. The molecular database employs a distributed storage system to support the storage and parallel processing of large-scale data. The index structure of the molecular database is optimized to support fast vector retrieval and similarity calculation. Different types of data are converted into vector data based on different encoding methods:

[0042] Molecular structure encoding: The molecular structures of hydrocarbons (such as alkanes, alkenes, aromatics, cycloalkanes, etc.) in crude oil are converted into 24-dimensional molecular structure vectors, and principal component analysis is used to encode the molecular structures.

[0043] Process parameter vectorization: Refining process parameters (such as temperature, pressure, catalyst type, reaction time, etc.) are converted into standardized vectors, and process knowledge graph technology is used to establish the relationship between parameters.

[0044] In this system, a process knowledge graph is a large semantic network used for the formal representation and storage of process knowledge in the oil refining field. The process knowledge graph consists of "nodes" and "edges":

[0045] 1) "Node" represents a core concept in the oil refining field, mainly including:

[0046] Refining unit operation: The refining unit in this system refers to a specific process or device in a refinery that uses hydrocarbon molecules as the processing object and aims to achieve their chemical structure transformation and physical property separation. Examples of refining unit operations include catalytic cracking, hydrotreating, catalytic reforming, and atmospheric and vacuum distillation.

[0047] Process parameters include reaction temperature, reaction pressure, weight hourly space velocity, hydrogen-to-oil ratio, and catalyst type.

[0048] Material or molecular type: such as light naphtha, heavy gas oil, n-alkanes, isoalkanes, aromatics, etc.

[0049] Product properties: such as research octane number, cetane number, sulfur content, density, etc.

[0050] 2) "Edges" describe the rich semantic relationships and industry knowledge between these entities, for example:

[0051] Available parameters: Catalytic cracking unit - Available parameters -> Reaction temperature.

[0052] Catalyst used: Hydrotreating unit - catalyst used -> Cobalt-molybdenum catalyst.

[0053] Raw material relationship: Heavy gas oil - is the raw material for... -> Catalytic cracking unit.

[0054] Production relationship: Catalytic cracking unit - production -> high octane gasoline.

[0055] The optimization effect achieved: Increase reaction temperature - for the purpose of optimization -> olefin yield.

[0056] Incompatible relationship: High-calcium crude oil - incompatible with... -> Residue oil hydrotreating catalyst.

[0057] Crude oil property mapping: The physicochemical properties of crude oil (such as density, viscosity, sulfur content, nitrogen content, metal content, etc.) are mapped into a high-dimensional vector space to establish the correspondence between properties and molecular structure.

[0058] In one embodiment, the association index includes: a cross-modal bidirectional index of molecular structure encoding and crude oil property vectors, and association rules between process parameters and molecular structure crude oil properties; the association rules are used to match molecular structure and crude oil properties using process parameters.

[0059] For example, molecular structure encoding can be represented as "V_mol", process parameter vector can be represented as "V_proc", and crude oil property vector can be represented as "V_prop". These vectors are not stored in isolation, but are constructed into a multimodal knowledge graph through associated indexes and metadata to achieve deep-level association.

[0060] First, molecular structure - crude oil properties ("V_mol" - "V_prop").

[0061] A cross-modal bidirectional indexing method is used to achieve an association index between molecular structure encoding and crude oil property vectors.

[0062] In this embodiment, the data preprocessing module can also be used to: supplement the crude oil property vector with a pre-trained molecular structure property prediction model; the molecular structure property prediction model describes the nonlinear mapping relationship between the molecular structure encoding and the crude oil property vector, and the molecular structure property prediction model adopts a regression model based on a graph neural network.

[0063] For example, based on known physicochemical principles and a large amount of experimental and production data, a molecular structure-property prediction model has been established, such as a regression model based on graph neural networks (GNNs). This model learns the nonlinear mapping relationship between the molecular structure encoding vector "V_mol" and the key property vector "V_prop". During the construction phase of the vector molecular database, for newly entered crude oil samples, their "V_prop" not only comes from direct measurement data, but can also be generated or supplemented by the molecular structure property prediction model based on their "V_mol", forming a more complete property description. At the same time, a cross-modal index of "V_mol" and "V_prop" has been established in the molecular database, supporting the reverse retrieval of molecular structures with specific properties through property.

[0064] Second, process parameters - molecular structure / properties ("V_proc" - "V_mol" / "V_prop").

[0065] By utilizing the association rules between process parameters and crude oil properties based on molecular structure, we can achieve the association index between process parameter vectors and molecular structure codes, as well as the association index between process parameter vectors.

[0066] The process parameter vector "V_proc" is associated with specific refining unit operations (such as catalytic cracking and hydrotreating) through a process knowledge graph. Each refining unit operation has its typical molecular transformation pathway and target product property range. For example, under a specific "V_proc", the catalytic cracking unit tends to crack large molecule alkanes to generate small molecule olefins and isoalkanes, resulting in an increased octane number of the product. The molecular database stores these "process-molecule-property" association rules and association indexes. When querying processes involving optimization or feedstock recommendations, this system can utilize these rules, association indexes, process knowledge graphs, etc., combined with the "V_proc" conditions, to screen out crude oil / fractions with molecular structures ("V_mol") that match the expected reaction pathway or that meet the target properties ("V_prop").

[0067] In one embodiment, the molecular database is a graph database.

[0068] In this embodiment, graph-based storage and retrieval can be employed. Besides vector storage, the core "molecule-process-property" association index, association rules, and process knowledge graph are stored and managed using a graph database, such as JanusGraph. Graph databases excel at handling complex relationship networks, supporting efficient path queries, such as "finding process routes that can convert high-sulfur heavy crude oil into low-sulfur gasoline," and neighbor queries, such as "finding the range of feedstock properties compatible with catalytic cracking catalyst P." They can also be linked and filtered with vector retrieval results.

[0069] In one embodiment, the following index is established based on the characteristics of molecular refining metadata:

[0070] (1) Molecular structure index: Establish an index system based on molecular structure to support rapid retrieval based on molecular type, carbon number range, functional group and other conditions.

[0071] (2) Process parameter index: Establish a multi-dimensional index of process parameters to support efficient querying based on process condition combinations.

[0072] (3) Property association index: Establish an association index between crude oil properties and molecular structure to support property-driven molecular retrieval.

[0073] Data storage in the fractional database: Preprocessed crude oil data vectors are stored in the database. Each vector contains detailed information about the crude oil, such as density, boiling point, and sulfur content. This information is encoded as numerical vectors for calculation and comparison.

[0074] The embodiment includes an intent recognition module for: extracting the intent from user queries based on a pre-trained intent recognition model; the intent recognition model is obtained by fine-tuning a pre-trained language model using technical data from the oil refining field. Through the intent recognition model, this system can identify and extract key information and intent from user queries. After training, the intent recognition model can understand complex language structures and semantics, thereby accurately identifying the user's query intent.

[0075] refer to Figure 2 In the intent recognition module, the BERT model can be used to perform technical terminology feature engineering, and intent classification training and recognition can be carried out.

[0076] In practice, a supervised learning method is adopted, using technical data from the oil refining field as the training dataset to fine-tune and train the BERT model.

[0077] To further adapt the pre-trained language model to the data processing of the molecular refining intelligent question-answering system, in one embodiment, the intent recognition model is designed by adding a classification layer to the pre-trained language model. The intent recognition model is trained using the cross-entropy loss function and the AdamW optimizer.

[0078] Adding a classification layer can more accurately output user intent. The cross-entropy loss function measures the difference between the model's predicted probability distribution and the true label. For user intent classification problems, the Softmax activation function can be used to transform the network output into a probability distribution, and the loss can be calculated by combining it with negative log-likelihood. AdamW is an improvement on the Adam optimizer, which can update model parameters more efficiently and stably, accelerate convergence, and reduce the risk of overfitting.

[0079] Specifically, the model training method can be as follows: processing steps such as dataset source, feature engineering, hyperparameter setting, transfer learning and multi-stage fine-tuning, and optimization strategies.

[0080] (1) Data source.

[0081] Refining process manuals (collecting refinery process manuals, operating procedures, technical standards, and other documents to build a refining professional corpus); expert knowledge base (integrating Q&A records, technical reports, and failure cases from experts in the refining field to form an expert knowledge dataset); molecular structure descriptions (collecting professional descriptions of crude oil molecular structures, chemical reaction equations, and process parameter descriptions to build a molecular refining professional terminology database); online technical documents (collecting online resources such as refining technology websites, technical forums, and academic papers to expand the training dataset); intermediate domain datasets, etc.

[0082] (2) Data preprocessing.

[0083] Annotation: Extract text data from refining process manuals, operating procedures, and expert knowledge bases to construct a high-quality annotated dataset.

[0084] Feature engineering is employed to establish a refining-specific dictionary, including molecular names (such as n-alkanes, isoalkanes, cycloalkanes, etc.), process terms (such as catalytic cracking, hydrocracking, reforming, etc.), and equipment names (such as reactors, separation towers, heat exchangers, etc.). Molecular structure descriptions are converted into 24-dimensional molecular structure vectors, including molecular descriptors such as molecular weight, carbon number, hydrogen-to-carbon ratio, and degree of unsaturation. Process parameters are converted into numerical features, including process indices such as temperature, pressure, space velocity, and conversion rate. Analytical rules for chemical reaction formulas are established to identify key information such as reactants, products, catalysts, and reaction conditions.

[0085] (3) Model fine-tuning.

[0086] The model architecture can use BERT-base as the base model, with fine-tuning of the architecture and parameters for the oil refining field.

[0087] In terms of architecture, a classification layer is added to the backend of the BERT-base model to output user intent classification.

[0088] Hyperparameter settings: Initial learning rate set to 2e-5, employing a learning rate decay strategy, decreasing by 10% every 1000 steps. Batch size set to 16, considering the complexity of refining-related texts and GPU memory limitations. Training epochs set to 5 to avoid overfitting and ensure model generalization ability in the refining field. Sequence length set to 2048, adapting to the average length of refining-related texts.

[0089] After fine-tuning the BERT-based model, the intent recognition model's ability to recognize intents in the oil refining field was significantly improved, enabling it to accurately understand technical terms and domain-specific expressions. Entity recognition performance was also significantly improved, laying a solid foundation for subsequent tasks.

[0090] (4) Transfer learning and multi-stage fine-tuning.

[0091] In one embodiment, to further improve the adaptability of the pre-trained language model to the oil refining field and enhance the intent recognition model's ability to understand the user's intent regarding oil refining issues, the intent recognition model is trained as follows:

[0092] Technical data from the oil refining field are collected to form an oil refining dataset, and technical data from intermediate fields are collected to form an intermediate dataset; the intermediate fields include any combination of chemistry, chemical engineering, and materials science.

[0093] A progressive transfer learning strategy is adopted. First, the pre-trained language model is fine-tuned for intermediate domain adaptation using an intermediate domain dataset. Then, the pre-trained language model is fine-tuned for the final specified task using an oil refining domain dataset.

[0094] This example employs a progressive transfer learning strategy. First, a language model is pre-trained on a large, general-purpose corpus. Then, intermediate domain adaptation is performed on texts (academic papers, patents, technical manuals) from a broader range of chemistry, chemical engineering, and materials science fields. Finally, final task-specific fine-tuning is conducted on a carefully constructed specialized dataset from the oil refining field.

[0095] The training corpus was constructed using academic literature and professional documents in the fields of chemistry, chemical engineering, and materials science. A masked language model pre-training task was used, focusing on adaptive training for scientific and technical terminology. A moderate learning rate was set, and multiple rounds of iterative training were conducted. The masked language model is one of the core tasks used in the pre-training phase of the BERT model. By randomly masking parts of the text, the model predicts the masked content based on the context. This intermediate domain-adaptive training enabled the model to establish a solid foundation of scientific knowledge and substantially improve its semantic understanding of technical terms. While maintaining general language capabilities, the model initially acquired domain-specific expertise.

[0096] In the final stage, task-specific fine-tuning was performed on a specialized refining domain dataset. Using a meticulously labeled refining domain question-answering dataset, a multi-task learning framework was employed to simultaneously train intent classification and named entity recognition. A refined learning rate scheduling strategy was configured, weight decay regularization was added, and an early stopping mechanism was implemented to avoid overfitting. This final fine-tuning stage endowed the model with deep domain specialization capabilities, resulting in excellent performance on intent recognition and entity extraction tasks. It also maintained good generalization performance for complex queries and unknown use cases, significantly improving the stability and reliability of the model output.

[0097] (5) Optimization strategy.

[0098] Domain-adaptive techniques are employed to adapt a general language model to the specific domain of oil refining. Simultaneously, the model is trained for multiple tasks, including intent recognition, entity recognition, and relation extraction, thereby enhancing its overall capabilities.

[0099] The trained BERT-base is used for intent recognition, and user intents are categorized as follows:

[0100] a. Molecular query intent: Identify the user's intent to query molecular structure, molecular properties, molecular reactions, etc.

[0101] b. Process query intent: Identify the user's intent to query process parameters, process conditions, process optimization, etc.

[0102] c. Device query intent: Identify the user's intent to query device information, device maintenance, device malfunction, etc.

[0103] d. Quality query intent: Identify the user's intent to query product quality, quality indicators, quality optimization, etc.

[0104] In this embodiment, user intent classification can be combined to improve the retrieval of crude oil data vectors that match the user intent in the molecular database. An expression form adapted to the user intent classification can be generated and pushed to the user.

[0105] For example, by dividing crude oil data vectors in a molecular database into categories that match user intent classifications, once the user intent classification is determined, the category matching the user intent classification can be directly found in the molecular database, narrowing the search scope and improving search efficiency.

[0106] For example, different user intent categories can be set to different forms of expression for push notifications to users. The expression forms corresponding to molecular query intents include 2D chemical structure diagrams, text metadata, etc., the expression forms corresponding to process query intents include structured tables, dashboard highlights, etc., the expression forms corresponding to equipment query intents include information cards, icons / images, category tags, etc., and the expression forms corresponding to quality query intents include test report-style tables, color / symbol labels, etc.

[0107] In this embodiment of the invention, the intelligent question-and-answer system for molecular refining includes a vector data processing module for retrieving, based on the user's query intent, the most relevant data vector from the vector database.

[0108] refer to Figure 2 An inverted index retrieval system is established. Using exact match and / or similarity-based retrieval methods, multiple crude oil data vectors with matching values ​​exceeding a preset threshold are retrieved from the molecular database. For example, when retrieving multiple crude oil data vectors matching the user's intent using exact match and / or similarity-based retrieval methods, the degree of matching between the retrieved crude oil data vectors and the user's intent is quantified. For instance, the exact match retrieval method might calculate a matching degree of 100%, while the similarity-based retrieval method uses a suitable preset value based on the similarity retrieval algorithm employed. For example, a preset value of 90% or 95% is set to filter out multiple crude oil data vectors with matching degrees exceeding the preset value.

[0109] In this embodiment, for vector data such as molecular structure encoding (24-dimensional vector), process parameter vectors, and physical property data, an inverted index is established to leverage the sparsity of the data. The key is a feature ID with a non-zero dimension, and the value is the inverted list. Each inverted list contains all vector IDs (DocIDs) with non-zero values ​​in that dimension, along with their corresponding non-zero values ​​(Values). The vector ID is a unique identifier for the entire data vector; each complete vector stored in the molecular database has a unique ID. The feature ID is an identifier for a specific dimension (field) within the vector, such as the name of a specific feature or field, indicating the type of information stored at that location. The complete vector is no longer stored in the database; only a list of non-zero elements (index and value) for each vector is stored.

[0110] Assume the vector data is as follows:

[0111] Molecular structure encoding: for example (a1,a2,...,a24), where a1 represents the mole fraction of n-alkanes, a2 represents the mole fraction of isoalkanes, and so on.

[0112] Process parameter vector: for example (b1,b2,...,b18), where b1 represents temperature ( b2 represents pressure (MPa), etc.

[0113] Crude oil property vector: for example (c1,c2,...,c16), where c1 represents density ( C2 represents sulfur content (ppmwt), etc.

[0114] To achieve efficient retrieval, an inverted index is created for each feature dimension (i.e., each a_i, b_i, c_i). a_i is any non-zero dimension data of molecular structure encoding with non-zero values, b_i is any non-zero dimension data of process parameter vector with non-zero values, and c_i is any non-zero dimension data of crude oil property vector with non-zero values.

[0115] Key: Feature ID. For example, F_a1 (representing the feature 'molar fraction of n-alkane').

[0116] Value: An inverted list containing all vector IDs that have a non-zero value for this feature, along with their corresponding values.

[0117] For example:

[0118] Suppose there are 3 crude oil vectors in the database:

[0119] 1) Crude Oil A: V001

[0120] Molecular structure: a1=0.25 (25% n-alkanes),

[0121] Physical property data: c2=15000 (sulfur content 15000ppm).

[0122] 2) Crude Oil B:V002

[0123] Molecular structure: a1=0.18

[0124] Process parameters: b1=485 (the temperature of a certain process is...) ).

[0125] 3) Crude oil C:V003

[0126] Molecular structure: a1=0.30

[0127] Physical property data: c2=5000.

[0128] Therefore, the inverted list constructed for feature F_a1 (mole fraction of n-alkane) is as follows:

[0129] Key: F_a1 (feature ID),

[0130] Value (Inverted List):

[0131] (V001,0.25),

[0132] (V002,0.18),

[0133] (V003,0.30),

[0134] Similarly, the inverted list created for feature F_c2 (sulfur content) is as follows:

[0135] Key: F_c2 (feature ID),

[0136] Value (Inverted List):

[0137] (V001,15000),

[0138] (V003,5000),

[0139] The inverted index created for feature F_b1 (temperature) is as follows:

[0140] Key: F_b1 (Feature ID),

[0141] Value (Inverted List):

[0142] (V002,485).

[0143] Search process: When a user queries "crude oil with n-alkane content higher than 20%", the system converts this condition into a constraint on feature F_a1 (value > 0.2). The system finds the inverted list corresponding to F_a1, quickly scans the list to find all entries with a value > 0.2, namely (V001, 0.25) and (V003, 0.30), and immediately obtains the results: V001 and V003.

[0144] In one embodiment, the exact match retrieval method can be processed as follows:

[0145] List all non-zero dimension values ​​in the vector representation of the user intent;

[0146] An inverted index is pre-built on the crude oil data vectors in the molecular database to obtain multiple inverted lists. Each inverted list includes the vector ID and the non-zero values ​​corresponding to multiple dimensions of the vector ID.

[0147] For each non-zero dimension value in the vector representation of the user's intent, inverted lists with the same non-zero value in the same dimension are matched sequentially from multiple inverted lists. The intersection of the matched inverted lists is then taken to obtain the vector ID for exact match retrieval.

[0148] Figure 3 This is a schematic diagram of the precise matching retrieval process in an embodiment of the present invention, for reference only. Figure 3 :

[0149] Query: For example, user intent exists in the form of a query vector. Given a query vector (which is also sparse), list all its non-zero dimensions (Query Features), each non-zero dimension corresponding to a feature ID.

[0150] deal with:

[0151] Intersection (AND): To find a perfect match, every non-zero dimension of the query vector must exist in the target vector and have the same value.

[0152] Speeding up using inverted indexes:

[0153] For the first non-zero dimension of the query vector, retrieve the corresponding inverted list L1 (containing all vector IDs with the same feature ID and matching values) from the inverted index based on the feature ID.

[0154] For the second non-zero dimension of the query vector, retrieve its inverted list L2.

[0155] calculate (Find the intersection, and only keep the vector IDs that appear in both lists).

[0156] Similarly, iterate through the remaining non-zero dimensions, finding the intersection with the inverted list of each subsequent dimension, until there are no more unprocessed non-zero dimensions.

[0157] Final result: The intersection result is the IDs of all exact matching query vectors.

[0158] This method achieves precise and efficient retrieval through exact matching. Precision: It guarantees that the returned results strictly satisfy all dimensional conditions specified in the query vector (numerical equality, string matching, etc.). This is crucial for scenarios requiring the search for specific, explicit numerical values ​​or categories. Efficiency (for queries with few specified dimensions): When the number of non-zero dimensions specified in the query vector is small (i.e., few constraints), the retrieval efficiency is very high because only a small number of inverted lists need to be searched and intersection operations performed. Inverted index lookups and set intersection calculations (especially using efficient set operation libraries) are typically much faster than nearest neighbor searches in high-dimensional spaces.

[0159] Similarity retrieval methods can be optimized using cosine similarity calculation; see reference for more information. Figure 4 , Figure 4 This is a schematic diagram of the similarity retrieval process in an embodiment of the present invention:

[0160] Query: Given a query vector Q (sparse), find the K vectors with the highest similarity to it (K is a preset positive integer). First, list all non-zero dimensions of the vectors.

[0161] Processing (Top-K similarity calculation):

[0162] Initialization: Create an empty result priority queue, i.e., a max heap.

[0163] Traverse the non-zero dimensions f of the query: For each non-zero dimension f of the query vector Q and its value q_f:

[0164] Find the inverted list L_f: Find the inverted list L_f corresponding to dimension f from the inverted index.

[0165] Traverse the inverted list: For each entry (vector D, value d_f) in L_f:

[0166] Partial score calculation: The contribution of this item to the similarity between D and Q is q_f × d_f (the numerator of the similarity).

[0167] Accumulated score: The sum of q_f × d_f is added to the current total score score_D of vector D.

[0168] Normalization:

[0169] Query vector magnitude : It can be calculated before the query (because the query vector is known).

[0170] Target vector magnitude The magnitude of each vector D needs to be pre-calculated during the index construction phase. And store it.

[0171] Final cosine similarity:

[0172] ;

[0173] Maintain the vector D with the highest Top-K scores: During or after the score accumulation process, maintain a priority queue (min-heap) to store the current Top-K highest scores.

[0174] Finally, the Top-K vector IDs are output.

[0175] In this embodiment of the invention, the exact match retrieval method and the similarity retrieval method are two parallel and complementary vector retrieval mechanisms provided by the system. They are two core retrieval modes designed within the vector data processing module to meet different query semantic requirements. They can be used independently or in combination (exact match filtering first, similarity sorting later) to provide the system with flexible, efficient and accurate data retrieval capabilities.

[0176] In this embodiment of the invention, the system further includes a response generation module, which, based on the results calculated by the vector data processing module, filters out the vector data that best meets the user's needs. The logic includes threshold settings, sorting algorithms, and priority rules to ensure that the system can select the most suitable response based on the user's query intent and the relevance of the data.

[0177] refer to Figure 2 The response generation module filters the results based on the calculations of the vector data processing module. It can use natural language generation technology to generate the filtered data and an expression for intent adaptation, and then push it to the user.

[0178] Example 1 is given below: Crude oil data query and analysis.

[0179] (1) User input: This system provides a user input interface. Users can input query requests through the interface, such as "query data on light crude oil with high sulfur content".

[0180] (2) Intent recognition: The system uses natural language processing technology to analyze user input and identify key information such as "high sulfur content" and "light crude oil".

[0181] (3) Vector retrieval: The system retrieves data vectors related to "high sulfur content" and "light crude oil" from the molecular database.

[0182] (4) Vector data processing: The system calculates the similarity between the search results and the user query, and selects the most relevant crude oil data.

[0183] (5) Results presentation: The system will present the filtered crude oil data to the user in a structured form, including information such as the molecular structure, sulfur content, and API degree of crude oil.

[0184] Example 2 is given below: Refining suggestion generation.

[0185] (1) User input: The user inputs specific refining process requirements, such as "a type of crude oil suitable for catalytic cracking is needed".

[0186] (2) Intent recognition: The system recognizes that the user’s need is to find crude oil suitable for catalytic cracking.

[0187] (3) Vector data processing: The system retrieves crude oil data vectors related to catalytic cracking from the database and calculates the similarity.

[0188] (4) Refining suggestion generation: The system generates refining suggestions based on the search results, such as recommending specific types of crude oil and providing corresponding refining parameters.

[0189] (5) Response output: The system outputs refining suggestions to the user in the form of natural language to help the user make decisions.

[0190] The molecular refining intelligent question-and-answer system of this invention can achieve the following effects:

[0191] 1) Improve data processing efficiency: Through automated data processing and analysis, the processing efficiency of oil refining data has been greatly improved.

[0192] 2) Improve data accuracy: By utilizing advanced natural language processing and machine learning technologies, the accuracy and reliability of data processing have been improved.

[0193] 3) Enhance user experience: By using natural language generation technology, data is presented to users in an easy-to-understand form, thus improving the user experience.

[0194] 4) Support for decision making: The refining suggestions provided by the system can help users make better refining decisions and improve the efficiency and quality of the refining process.

[0195] This invention also provides a molecular refining intelligent question-and-answer method, as described in the following embodiments. Since the problem-solving principle of this device is similar to that of the molecular refining intelligent question-and-answer system, the implementation of this device can refer to the implementation of the molecular refining intelligent question-and-answer system; repeated details will not be elaborated further.

[0196] Figure 5 This is a flowchart illustrating the intelligent question-and-answer method for molecular refining in an embodiment of the present invention, as shown below. Figure 5 As shown, the method includes:

[0197] Step 501: Receive user query request;

[0198] Step 502: Extract the intent from the user's query request based on the pre-trained intent recognition model; the intent recognition model is obtained by fine-tuning the pre-trained language model using technical data from the oil refining field;

[0199] Step 503: Based on the exact match retrieval method and / or the similarity retrieval method, retrieve multiple crude oil data vectors from the molecular database whose matching value with the user intent exceeds a preset value; the exact match retrieval method uses an inverted index to retrieve crude oil data vectors that perfectly match the user intent; the molecular database is used to: store various types of crude oil data vectors; the crude oil data vector types include molecular structure codes, process parameter vectors, and crude oil property vectors, and different types of crude oil data vectors are linked to each other through association indexes;

[0200] Step 504: Generate an expression for intent adaptation based on multiple crude oil data vectors whose matching values ​​exceed a preset value, and push it to the user.

[0201] In one embodiment, the molecular structure encoding is obtained by encoding the hydrocarbon molecular structure in crude oil using principal component analysis; the process parameter vector is obtained by converting multiple refining process parameters into standardized vectors, and the refining process parameters are associated with each other using process knowledge graph technology; the crude oil property vector is obtained by mapping the physicochemical properties of crude oil into a high-dimensional vector space and establishing the correspondence between properties and molecular structure.

[0202] In one embodiment, the method further includes constructing a molecular database. Before constructing the molecular database, crude oil data is received, and the crude oil data is preprocessed and vectorized. The preprocessing includes data cleaning and / or standardization.

[0203] In one embodiment, the preprocessing and vectorization of crude oil data in this method may include:

[0204] The crude oil property vector is supplemented by a pre-trained molecular structure property prediction model; the molecular structure property prediction model describes the nonlinear mapping relationship between molecular structure encoding and crude oil property vector, and the molecular structure property prediction model adopts a regression model based on graph neural network.

[0205] In one embodiment, the association index includes: a cross-modal bidirectional index of molecular structure encoding and crude oil property vectors, and association rules between process parameters and molecular structure crude oil properties; the association rules are used to match molecular structure and crude oil properties using process parameters.

[0206] In one embodiment, the molecular database is a graph database.

[0207] In one embodiment, the intent recognition model is constructed by adding a classification layer to a pre-trained language model and training it using a cross-entropy loss function and an AdamW optimizer.

[0208] In one embodiment, the intent recognition model is trained as follows:

[0209] Technical data from the oil refining field are collected to form an oil refining dataset, and technical data from intermediate fields are collected to form an intermediate dataset; the intermediate fields include any combination of chemistry, chemical engineering, and materials science.

[0210] A progressive transfer learning strategy is adopted. First, the pre-trained language model is fine-tuned for intermediate domain adaptation using an intermediate domain dataset. Then, the pre-trained language model is fine-tuned for final task-specific fine-tuning using an oil refining domain dataset.

[0211] In one embodiment, the exact match retrieval method is processed as follows:

[0212] List all non-zero dimension values ​​in the vector representation of the user intent;

[0213] An inverted index is pre-built on the crude oil data vectors in the molecular database to obtain multiple inverted lists. Each inverted list includes the vector ID and the non-zero values ​​corresponding to multiple dimensions of the vector ID.

[0214] For each non-zero dimension value in the vector representation of the user's intent, inverted lists with the same non-zero value in the same dimension are matched sequentially from multiple inverted lists. The intersection of the matched inverted lists is then taken to obtain the vector ID for exact match retrieval.

[0215] This invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the above-described molecular refining intelligent question-and-answer method.

[0216] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described molecular refining intelligent question-and-answer method.

[0217] This invention also provides a computer program product, which includes a computer program that, when executed by a processor, implements the above-described molecular refining intelligent question-and-answer method.

[0218] This invention provides a molecular refining intelligent question-and-answer system that utilizes advanced natural language processing and machine learning technologies to achieve efficient processing and intelligent analysis of crude oil data. Specifically, this invention constructs an efficient molecular database to store and manage various types of crude oil data vectors, including molecular structure codes, process parameter vectors, and crude oil property vectors. An optimized index structure supports rapid data retrieval and similarity calculation. Furthermore, the system accurately parses user query intent through an intent recognition module, and combined with a vector data processing module and a response generation module, can quickly and accurately provide users with the refining data and refining suggestions they require. Compared with existing technologies, this invention not only significantly improves data processing efficiency and accuracy but also enhances user experience through natural language generation technology, supporting users in making better refining decisions and improving the efficiency and quality of the refining process.

[0219] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0220] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0221] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0222] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0223] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A molecular refining intelligent question-and-answer system, characterized in that, It includes: a molecular database, an intent recognition module, a vector data processing module, and a response generation module; The molecular database is used to store various types of crude oil data vectors. These crude oil data vector types include molecular structure codes, process parameter vectors, and crude oil property vectors. Different types of crude oil data vectors are linked together through association indexes. The intent recognition module is used to: extract the intent from user queries based on a pre-trained intent recognition model; the intent recognition model is obtained by fine-tuning a pre-trained language model using technical data from the oil refining field; The vector data processing module is used to: retrieve multiple crude oil data vectors from the molecular database that match the user's intent with a value exceeding a preset value, based on the exact match retrieval method and / or the similarity retrieval method; the exact match retrieval method uses an inverted index to retrieve crude oil data vectors that perfectly match the user's intent; The response generation module is used to: generate an expression form for intent adaptation based on multiple crude oil data vectors whose matching values ​​exceed a preset value, and push it to the user.

2. The molecular refining intelligent question-and-answer system as described in claim 1, characterized in that, The molecular structure encoding is obtained by encoding the hydrocarbon molecular structure in crude oil using principal component analysis; the process parameter vector is obtained by converting multiple refining process parameters into standardized vectors, and the refining process parameters are associated with each other using process knowledge graph technology; the crude oil property vector is obtained by mapping the physicochemical properties of crude oil into a high-dimensional vector space and establishing the correspondence between properties and molecular structure.

3. The molecular refining intelligent question-and-answer system as described in claim 2, characterized in that, It also includes a data preprocessing module, used for: Receive crude oil data, preprocess and vectorize the crude oil data; preprocessing includes data cleaning and / or standardization.

4. The molecular refining intelligent question-and-answer system as described in claim 3, characterized in that, The data preprocessing module is also used for: The crude oil property vector is supplemented by a pre-trained molecular structure property prediction model; the molecular structure property prediction model describes the nonlinear mapping relationship between molecular structure encoding and crude oil property vector, and the molecular structure property prediction model adopts a regression model based on graph neural network.

5. The molecular refining intelligent question-and-answer system as described in claim 2, characterized in that, The association index includes: a cross-modal bidirectional index of molecular structure encoding and crude oil property vectors, and association rules between process parameters and molecular structure crude oil properties; the association rules are used to match molecular structure and crude oil properties using process parameters.

6. The molecular refining intelligent question-and-answer system as described in claim 1, characterized in that, The molecular database is a graph database.

7. The molecular refining intelligent question-and-answer system as described in claim 1, characterized in that, The intent recognition model is constructed by adding a classification layer to a pre-trained language model. The intent recognition model is trained using the cross-entropy loss function and the AdamW optimizer.

8. The molecular refining intelligent question-and-answer system as described in claim 1, characterized in that, The intent recognition model is trained as follows: Technical data from the oil refining field are collected to form an oil refining dataset, and technical data from intermediate fields are collected to form an intermediate dataset; the intermediate fields include any combination of chemistry, chemical engineering, and materials science. A progressive transfer learning strategy is adopted. First, the pre-trained language model is fine-tuned for intermediate domain adaptation using an intermediate domain dataset. Then, the pre-trained language model is fine-tuned for the final specified task using an oil refining domain dataset.

9. The molecular refining intelligent question-and-answer system as described in claim 1, characterized in that, The exact match retrieval method is processed as follows: List all non-zero dimension values ​​in the vector representation of the user intent; An inverted index is pre-built on the crude oil data vectors in the molecular database to obtain multiple inverted lists. Each inverted list includes the vector ID and the non-zero values ​​corresponding to multiple dimensions of the vector ID. For each non-zero dimension value in the vector representation of the user's intent, inverted lists with the same non-zero value in the same dimension are matched sequentially from multiple inverted lists. The intersection of the matched inverted lists is then taken to obtain the vector ID for exact match retrieval.

10. A molecular refining intelligent question-and-answer method, characterized in that, include: Receive user query requests; Extract the intent from the user's query request based on a pre-trained intent recognition model; The intent recognition model was obtained by fine-tuning a pre-trained language model using technical data from the oil refining field. Based on the exact match retrieval method and / or the similarity retrieval method, multiple crude oil data vectors with a matching value exceeding a preset value with the user's intent are retrieved from the molecular database; the exact match retrieval method uses an inverted index to retrieve crude oil data vectors that perfectly match the user's intent; the molecular database is used to store various types of crude oil data vectors; the crude oil data vector types include molecular structure codes, process parameter vectors, and crude oil property vectors, and different types of crude oil data vectors are linked to each other through association indexes; Based on multiple crude oil data vectors whose matching values ​​exceed preset values, an expression for intent adaptation is generated and pushed to the user.

11. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method of claim 10.

12. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method of claim 10.

13. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the method of claim 10.

Images