Construction method of mining AI assistant

Abstract

The invention discloses a mining AI assistant construction method, and belongs to the field of artificial intelligence and mining engineering cross technology, and the method comprises the steps: building a coal mine multi-source knowledge fusion library, and constructing a mixed index mechanism based on a vector database and a knowledge graph; training a domain adaptive large model through three stages of domain pre-training, instruction fine tuning and human feedback reinforcement learning; six functional modules including intelligent question and answer and decision support, expert experience digitization, safety training and assessment, document processing and generation, response measure and scheme query, and data analysis and report generation are integrated, and four kinds of role services are realized. According to the invention, the accuracy and reliability of the model in a coal mine professional scene are improved, and efficient utilization of multifunctional collaboration and knowledge is realized.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of artificial intelligence and mining engineering, and in particular relates to a method for constructing a mining AI assistant. Background Technology

[0002] In recent years, artificial intelligence and large language model technologies, such as DeepSeek and ChatGPT, have made significant progress in text processing, natural language understanding, and generation, and have been gradually introduced into various industries to assist in daily office work and knowledge management. In the coal mining industry, these general-purpose models can provide employees with basic text processing support, such as organizing technical documents, generating simple documents, and retrieving general information. This reduces the paperwork burden on technical personnel to some extent, improves information processing efficiency, and demonstrates the beneficial effects of general artificial intelligence technology in automating office work and basic interaction.

[0003] However, the aforementioned general-purpose language models suffer from significant limitations in their application within specialized coal mining scenarios due to a lack of in-depth understanding of coal mining expertise, geological conditions, engineering drawings, safety regulations, and industry standards. Specifically, when dealing with critical tasks such as professional design, decision support, safety training, and accident case simulations, the models struggle to accurately interpret domain-specific concepts and complex logical relationships. Their outputs often lack professional precision and scenario adaptability, failing to meet the high reliability requirements of coal mine safety production, technical management, and experience transfer. This hinders the effective integration and in-depth application of artificial intelligence technology in the intelligent transformation of the entire coal mining process. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention proposes a method for constructing a mining AI assistant, thereby resolving the issues present in the prior art.

[0005] To achieve the above objectives, the present invention provides a method for constructing a mining AI assistant, comprising: S1. Establish a multi-source knowledge fusion database for coal mines. The knowledge fusion database is based on vector database and knowledge graph technology. It deeply integrates at least one of the following: professional knowledge of coal mines, geological data, engineering drawings, regulations and systems, equipment archives and accident cases. It also establishes bidirectional links between vector and graph nodes to form a hybrid indexing mechanism. S2, a training domain adaptive large model engine, is trained through three stages: domain pre-training, instruction fine-tuning, and human feedback reinforcement learning. The human feedback reinforcement learning includes evaluation of the model response by experts in the coal mining field, providing preference scores or corrective feedback, and training the reward model using a proximal policy optimization algorithm. S3 integrates multiple functional modules into a unified platform. These modules include an intelligent question-and-answer and decision support module, an expert experience digitization module, a safety training and assessment module, a document processing and generation module, a countermeasures and solution query module, and a data analysis and report generation module. This enables the functions of four roles: senior expert coach, efficient office secretary, safety supervision consultant, and all-employee training mentor.

[0006] Preferably, in S1, establishing a multi-source knowledge fusion library for coal mines includes: integrating multi-source coal mine data and constructing a structured knowledge model; collecting professional data from various business systems through data interfaces; processing text using natural language processing technology, parsing drawings and case images to extract key spatial and entity information; and transforming the extracted information into a structured knowledge model based on a preset coal mine professional domain ontology.

[0007] Preferably, in S1, the implementation of the hybrid indexing mechanism includes: converting standardized knowledge units into high-dimensional vectors through a pre-trained model and storing them in a vector database; constructing a knowledge graph in the form of nodes and edges; and establishing bidirectional links between vectors and graph nodes to support the linkage of semantic search and logical reasoning dual engines.

[0008] Preferably, in S2, the domain pre-training includes: constructing a training dataset based on data from the coal mine multi-source knowledge fusion library, continuously training the basic large model using a self-supervised learning objective, and adjusting the model parameters through domain adaptive optimization.

[0009] Preferably, in S2, the instruction fine-tuning includes: constructing a dataset containing task instructions in the coal mining field, covering a variety of complex scenarios through structured instruction design, and using supervised fine-tuning methods to train the model to improve instruction compliance and task-oriented reasoning capabilities.

[0010] Preferably, in S2, the human feedback reinforcement learning includes: having coal mining experts evaluate the model response mechanism to construct a feedback dataset, training the reward model using a proximal policy optimization algorithm, and adjusting the model output strategy through iterative feedback loops.

[0011] Preferably, in S3, the intelligent question answering and decision support module, based on a multimodal knowledge fusion library and an adaptive large model engine, parses the semantics of user queries, maps them to relevant concepts in the knowledge base, generates explanatory answers through an inference engine, and generates decision suggestions by combining numerical simulations or empirical formulas.

[0012] Preferably, in S3, the expert experience digitization module interacts with experts through a guided dialogue interface, generates progressively challenging questions based on a preset knowledge framework and decision tree logic, analyzes expert answers in real time using natural language processing technology, and fills the narrative into the knowledge framework through entity recognition and relation extraction to form standardized knowledge elements stored in the expert experience knowledge base.

[0013] Preferably, in S3, the safety training and assessment module creates a virtual underground scenario, loads a safety training script through a dynamic story engine, identifies user operations, calculates decision results in real time through a reasoning model, records the assessment process, and generates a structured evaluation report.

[0014] Preferably, in S3, the data analysis and report generation module accesses multiple real-time data sources, stores key parameters in a time-series database, analyzes data and identifies abnormal patterns using mathematical statistics algorithms and LSTM time-series prediction models, and automatically writes analysis conclusions and recommended measures based on a natural language generation engine, outputting a standardized report.

[0015] Compared with the prior art, the present invention has the following advantages and technical effects: This invention combines "establishing a multi-source knowledge fusion library for coal mines, deeply integrating multi-source data based on vector databases and knowledge graph technology, and forming a hybrid indexing mechanism" with "training an adaptive large model through three stages: domain pre-training, instruction fine-tuning, and human feedback reinforcement learning, where human feedback reinforcement learning includes expert evaluation and proximal policy optimization algorithms." This enables the model to deeply understand specific concepts, geological conditions, and safety regulations in the coal mine field, providing accurate and reliable outputs in complex reasoning tasks such as professional design, risk prediction, and accident analysis, effectively overcoming the problem of insufficient professional accuracy caused by the lack of domain knowledge in general models.

[0016] This invention solves the limitations of existing technologies that are functionally isolated and unable to cover the entire coal mine production process by integrating six major functional modules, namely intelligent question answering and decision support, digitization of expert experience, safety training and assessment, document processing and generation, query of countermeasures and solutions, and data analysis and report generation, into a unified platform and realizing four types of role-based functions. It achieves integrated intelligent collaboration of knowledge sharing, decision support and safety management, and significantly improves the system's adaptability and efficiency in complex business scenarios.

[0017] This invention adopts a technical solution of "establishing a hybrid indexing mechanism of vector database and knowledge graph, and realizing the linkage of semantic search and logical reasoning through bidirectional links". It integrates multi-source structured and unstructured data, supports similarity retrieval based on deep semantics and deduction of complex logical relationships, and enables the system to quickly locate target information and mine potential associations in massive coal mine data. It overcomes the shortcomings of single retrieval technology (such as relying solely on vector database or knowledge graph) in terms of knowledge coverage and reasoning ability. Attached Figure Description

[0018] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a diagram of the coal mine multi-source knowledge fusion library architecture according to an embodiment of the present invention; Figure 2 This is an adaptive large model architecture diagram according to an embodiment of the present invention; Figure 3 This is a diagram illustrating the architecture of the six functional modules in an embodiment of the present invention. Detailed Implementation

[0019] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0020] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0021] Example 1 This embodiment provides a method for constructing a mining AI assistant. First, a knowledge fusion library is built based on massive amounts of coal mining expertise, providing the model with precise and professional knowledge sources. Then, through domain pre-training, instruction fine-tuning, and reinforcement learning training, a domain-adaptive large model engine is trained, enabling it to possess professional language understanding and logical reasoning capabilities in the coal mining field. Building upon coal mining expertise, the intelligent question-answering and decision support module not only answers "what" but also "why" and "how." The assistant can systematically extract, record, and structure the experience of senior experts, providing precise guidance to relevant personnel through an expert experience digitization module. Its unique immersive safety training and assessment module simulates real underground scenarios, conducting multiple rounds of interactive question-and-answer sessions and handling process simulations with employees, achieving personalized and practical training and assessment. Through the document processing and generation module, the system can generate diverse documents based on key information provided by users, reducing the paperwork burden on technical personnel. The response measures and solution query module supports precise queries through natural language, instantly locating corresponding processing steps and response solutions. The data analysis and report generation module can access real-time data sources for analysis, automatically generating professional analysis reports to guide production and management work. The methods specifically include: S1. Establish a multi-source knowledge fusion database for coal mines. The knowledge fusion database is based on vector database and knowledge graph technology. It deeply integrates at least one of the following: professional knowledge of coal mines, geological data, engineering drawings, regulations and systems, equipment archives and accident cases. It also establishes bidirectional links between vector and graph nodes to form a hybrid indexing mechanism. Furthermore, in S1, the establishment of a multi-source knowledge fusion library for coal mines includes: integrating multi-source coal mine data and constructing a structured knowledge model; collecting professional data from various business systems through data interfaces; processing text using natural language processing technology, parsing drawings and case images to extract key spatial and entity information; and transforming the extracted information into a structured knowledge model based on a preset coal mine professional domain ontology.

[0022] Furthermore, in S1, the implementation of the hybrid indexing mechanism includes: transforming standardized knowledge units into high-dimensional vectors through a pre-trained model and storing them in a vector database; simultaneously constructing a knowledge graph in the form of nodes and edges; and establishing bidirectional links between vectors and graph nodes to support the linkage of semantic search and logical reasoning dual engines.

[0023] Specifically, S1 provides the model with a precise and traceable source of knowledge. For example... Figure 1 The diagram shows the architecture of a multi-source knowledge fusion database for coal mines. S1 specifically includes the following steps: S11. Integrate multi-source coal mine data to construct a structured knowledge model; This step aims to address the initial challenge of effectively integrating multi-source heterogeneous data in the coal mining sector. First, it deploys configurable data interfaces to comprehensively collect professional data from various business systems. Then, it utilizes natural language processing technology to process the text, and uses computers to analyze drawings and case study images to extract key spatial and entity information. Finally, based on a pre-defined coal mining professional domain ontology, the extracted information is transformed into a structured knowledge model with unified semantic standards. This achieves a standardized transformation from raw, disorganized coal mining knowledge to machine-learnable and understandable information, laying the data foundation for building a multi-source knowledge fusion library.

[0024] S12. Build a hybrid storage and indexing architecture that integrates vector databases and knowledge graphs; This step aims to broaden the scope of retrieval and deepen knowledge mining, overcoming the limitations of a single technical solution. First, the standardized knowledge units obtained from S11 are transformed into high-dimensional vectors through a pre-trained model and stored in a vector database. This enables the system to possess deep semantic-based similarity retrieval capabilities, allowing for rapid retrieval of target information. Then, these knowledge units and their associated relationships are constructed into a knowledge graph in the form of nodes and edges, clearly expressing complex logical relationship chains. A crucial step is establishing bidirectional links between vectors and graph nodes, forming a hybrid indexing mechanism. This enables the dual-engine linkage of "semantic search" and "logical reasoning," allowing the knowledge base to both "find a needle in a haystack" and "follow the clues."

[0025] S2, a training domain adaptive large model engine, is trained through three stages: domain pre-training, instruction fine-tuning, and human feedback reinforcement learning. The human feedback reinforcement learning includes evaluation of the model response by experts in the coal mining field, providing preference scores or corrective feedback, and training the reward model using a proximal policy optimization algorithm. Furthermore, in S2, the domain pre-training includes: constructing a training dataset based on data from the coal mine multi-source knowledge fusion library, continuously training the basic large model using a self-supervised learning objective, and adjusting the model parameters through domain adaptive optimization.

[0026] Furthermore, in S2, the instruction fine-tuning includes: constructing a dataset containing task instructions in the coal mining field, covering a variety of complex scenarios through structured instruction design, and using supervised fine-tuning methods to train the model to improve instruction compliance and task-oriented reasoning.

[0027] Furthermore, in S2, the human feedback reinforcement learning includes: having coal mining experts evaluate the model's response mechanism to construct a feedback dataset, training the reward model using a proximal policy optimization algorithm, and adjusting the model's output policy through iterative feedback loops.

[0028] Specifically, step S2 trains an adaptive large model through three stages: domain pre-training, instruction fine-tuning, and human feedback reinforcement. This transforms underlying knowledge into practical productivity, develops an intelligent service engine for typical coal mine business scenarios, and encapsulates it into standardized application interfaces. This transformation of underlying knowledge into practical productivity solves the "last mile" problem of knowledge-to-business application. Figure 2 The diagram shows the adaptive large model architecture. S2 specifically includes the following steps: S21. Based on data from a coal mine multi-source knowledge fusion library, including technical reports, safety regulations, geological data, equipment manuals, and academic papers, a training dataset is constructed to train an adaptive large model. A self-supervised learning objective is used to continuously train the basic large-scale model. Through iterative optimization, the model absorbs and understands professional knowledge in the coal mining field, improving its generalization performance in coal mining scenarios where specialized data is scarce. Domain-adaptive optimization is used to adjust model parameters, enhancing its understanding of coal mining-specific concepts and transforming it into a large-scale model specifically for mining applications.

[0029] S22. In the instruction fine-tuning stage, based on the pre-trained model, an instruction dataset is constructed. The uncertainty of domain expertise is addressed through structured instruction design, thereby improving the reliability of the model. First, a dataset containing task instructions from the coal mining field is constructed, such as "explaining the relationship between support resistance and roof subsidence" and "analyzing the reasons for excessive gas concentration in the working face." The instruction data is translated into code by professionals to ensure coverage of various complex scenarios. Then, supervised fine-tuning methods are used to train the model, enabling it to respond to instructions accurately and consistently. Subsequent adjustment phases focus on improving the model's instruction compliance and task-oriented reasoning, strengthening the logical chain through multi-turn dialogues and example learning.

[0030] S23. In the human feedback reinforcement learning stage, a human feedback mechanism is introduced to further refine the model output.

[0031] First, experts in the coal mining field evaluate the model's response mechanism, providing preference scores or corrective feedback to construct a feedback dataset. Then, a proximal policy optimization algorithm is used to train the reward model, simulating human preferences. Through iterative feedback loops, the model gradually adjusts its output policy, reducing errors and unsafe content, and improving the accuracy and safety of the model engine in complex reasoning tasks such as risk prediction and accident analysis, enabling efficient and reliable deployment of the model in coal mining environments.

[0032] S3 integrates multiple functional modules into a unified platform. These modules include an intelligent question-and-answer and decision support module, an expert experience digitization module, a safety training and assessment module, a document processing and generation module, a countermeasures and solution query module, and a data analysis and report generation module. This enables the functions of four roles: senior expert coach, efficient office secretary, safety supervision consultant, and all-employee training mentor.

[0033] Furthermore, in S3, the intelligent question answering and decision support module, based on a multimodal knowledge fusion library and an adaptive large model engine, parses the semantics of user queries, maps them to relevant concepts in the knowledge base, generates explanatory answers through an inference engine, and generates decision suggestions by combining numerical simulations or empirical formulas.

[0034] Furthermore, in S3, the expert experience digitization module interacts with experts through a guided dialogue interface, generates progressively challenging questions based on a preset knowledge framework and decision tree logic, analyzes expert answers in real time using natural language processing technology, and fills the narrative into the knowledge framework through entity recognition and relation extraction, forming standardized knowledge elements that are stored in the expert experience knowledge base.

[0035] Furthermore, in S3, the safety training and assessment module creates a virtual underground scenario, loads a safety training script through a dynamic story engine, identifies user operations, calculates decision results in real time through a reasoning model, records the assessment process, and generates a structured evaluation report.

[0036] Furthermore, in S3, the data analysis and report generation module connects to multiple real-time data sources, stores key parameters in a time-series database, analyzes data and identifies abnormal patterns using mathematical statistics algorithms and LSTM time-series prediction models, and automatically writes analysis conclusions and recommended measures based on a natural language generation engine, outputting standardized reports.

[0037] Specifically, through dedicated interaction logic and output templates, six core functional modules are integrated into a unified platform. The models and knowledge base undergo targeted iteration to optimize their service capabilities as the "four main roles," enabling the intelligent collaboration system to leap from a "tool" to an "assistant." This deeply integrates AI technology with mining practices, enhancing its initiative and interactivity, making it an all-around intelligent assistant that empowers employees and improves user experience and work efficiency. For example... Figure 3 The diagram shows the six functional architectures. Step S3 specifically includes: S31, Intelligent Question Answering and Decision Support Module; The intelligent question answering and decision support module, centered on a multimodal knowledge fusion library and an adaptive large model engine, aims to provide not only factual answers but also in-depth analysis of underlying principles and actionable suggestions. First, based on the multimodal knowledge fusion library, when a user asks a question such as "Why does the 5102 working face, which crosses a fault, require reinforced support?", the natural language processing component parses the semantics of the query, identifies keywords such as "5102 working face" and "fault," and maps them to relevant concepts in the knowledge base. Subsequently, the inference engine dynamically derives explanatory answers based on rules and previously trained models. At the decision support level, the system combines numerical simulations or empirical formulas to generate specific support parameter suggestions, along with supporting decision-making evidence, enhancing the credibility of the recommendations. Simultaneously, the system uses natural language generation technology to transform the inference results into coherent paragraph-style outputs, ensuring clear and understandable explanations. The entire module also integrates a continuous learning mechanism, optimizing performance through user feedback and knowledge base updates to ensure accurate and personalized support across various scenarios.

[0038] S32, Expert Experience Digitalization Module; The core of this module is an intelligent "knowledge extraction" system. Through structured interactive design and semantic understanding technology, it transforms the implicit knowledge that senior experts possess—knowledge that is difficult to articulate—into explicit, storable digital assets. First, a knowledge framework is constructed, predefining the information elements to be collected, such as operational experience and troubleshooting methods. Then, through a guided dialogue interface, the system interacts with experts. Based on the pre-set knowledge framework and decision tree logic, the interaction engine generates a series of progressively challenging questions, effectively uncovering the logic behind the expert's experience, judgment, and handling methods from an interview perspective.

[0039] In the stage of structuring expert experience, natural language processing technology is used to analyze experts' free text or voice responses in real time. Entity recognition and relation extraction technologies are then used to automatically populate the corresponding fields of the knowledge framework with the expert's narrative. Ultimately, these structured experiences are stored in the expert experience knowledge base as standardized "knowledge elements," each containing a complete problem context, decision-making logic, and solution. Subsequent users can not only search for this experience using keywords but also discover related knowledge networks, thus transforming individual expert experience into a systematic, searchable, and learnable valuable digital asset.

[0040] S33, Safety Training and Assessment Module; The core of this module lies in creating a highly intelligent and immersive "digital twin" training environment. By integrating 3D simulation, dynamic events, and intelligent evaluation algorithms, it transforms traditional passive listening into active practical simulation.

[0041] First, this module creates a virtual underground scene based on real underground information and actual equipment, covering key areas such as roadways and mining faces. The system loads a safety training script through a dynamic story engine and waits for the trainees to interact. Every action of the trainees will be recognized by the system and input into a database rule-based reasoning model through natural language processing. This model can calculate the decision results in real time and present them dynamically. If an operation error occurs, the corresponding accident scene will be triggered.

[0042] At the assessment level, the system's role goes beyond simply judging right or wrong; it records the entire assessment process of trainees and transforms it into structured data. The system identifies shortcomings in the trainees' procedures, precisely pinpointing operational loopholes and knowledge gaps. For example, "Underground personnel discovered continuous abnormal noises from the roof but did not prioritize 'immediately evacuate from the danger zone'; instead, they performed 'support reinforcement,' revealing the trainee's insufficient understanding of the instantaneous risk of roof collapse." Ultimately, the system generates a detailed individual assessment report for this training, enabling precise personal training and effectively improving employees' emergency response capabilities in real-world hazardous situations.

[0043] S34, Document Processing and Generation Module; The core of the document processing and generation module lies in the tight integration of natural language processing, knowledge base, and rule engine to form a system capable of understanding user intent, retrieving relevant information, and generating standardized language. This module first constructs a multi-layered knowledge system, with a multimodal knowledge fusion database at its core. When the system recognizes key information input by the user, it initiates an intelligent retrieval and assembly process: based on the document type, it retrieves the corresponding document standard framework from the database. Simultaneously, the rule engine triggers corresponding clause settings from the database based on identified key entities, automatically filling the relevant sections of the template.

[0044] At the content generation level, this module employs natural language generation technology to transform structured data and clause-based knowledge into fluent, standardized written language, rather than simply filling in blanks. For more complex documents such as accident analysis reports, the system can access information from the database for integrated diagnosis, generating initial drafts of Word or PDF reports containing "problems," "causes," and "measures." This reduces the burden on technical personnel to write documents from scratch, effectively allowing them to function as efficient and reliable office secretaries familiar with mining regulations.

[0045] S35, Module for querying countermeasures and solutions; The core functionality of this module is to build an intelligent retrieval and push system that can understand natural language and visual information and quickly connect to a multi-source knowledge fusion database. Firstly, based on a deeply integrated coal mine information knowledge base, when a user queries via text input or by uploading an image containing fault codes, the system employs a multi-channel information processing workflow. For images, the system utilizes built-in OCR (Optical Character Recognition) and computer vision models to accurately extract key characters such as equipment nameplates, instrument readings, or fault codes. For natural language, semantic understanding technology is used to interpret the user's intent. Then, the retrieval engine matches the extracted key information with entries in the knowledge base, enabling the system to provide users with detailed response measures and solutions with just a simple "image upload + verbal description" action from the operator.

[0046] S36, Data Analysis and Report Generation Module; The functionality of this module lies in constructing a closed-loop system integrating data acquisition, intelligent analysis, and cognitive generation, transforming massive amounts of real-time monitoring data into reports with practical guiding significance. First, through pre-built data interfaces and APIs, the functional module connects to multiple real-time data sources, such as electro-hydraulic control systems, safety monitoring systems, and equipment health management systems. Key parameters are collected and uniformly stored in a time-series database for subsequent data analysis. Then, using mathematical statistics algorithms and LSTM time-series prediction models, the module learns the inherent patterns in the data, obtaining features that reflect the trend changes and potential influencing factors of each system, and automatically identifying anomaly patterns with potential risks.

[0047] During the report generation phase, the system calls relevant Word or PPT templates based on preset time periods or specific events. The data visualization component automatically generates corresponding charts and embeds them in designated locations within the templates. This module's unique "cognitive generation" capability, implemented by a natural language generation engine, combines analysis results with domain knowledge from a knowledge base based on predefined grammar rules and analytical logic to automatically generate professional and rigorous analytical conclusions and recommended measures. Finally, a standardized report is output and delivered to relevant management personnel, thereby improving the foresight and efficiency of security management.

[0048] The beneficial effects of this embodiment: This embodiment provides a method for constructing a coal mine-specific intelligent assistant. Based on a large language model and coal mine expertise, it proposes an intelligent collaboration system integrating intelligent interaction, decision support, safety training, experience sharing, and report generation. This system can deeply integrate artificial intelligence models into the entire coal mine production and management process. Its core positioning is to serve four key roles for all mine employees: senior expert coach, efficient office secretary, safety supervision consultant, and all-employee training mentor.

[0049] This embodiment can provide real-time and accurate operational guidance to underground engineers and related personnel through natural language dialogue, transforming the experience of top experts into readily accessible digital assets to improve work efficiency. Simultaneously, it can generate personalized training materials and accident case simulations based on the needs of employees with different positions and skill levels, achieving precise safety education tailored to each individual. It can also automatically parse production data, safety logs, and meeting minutes, generating standardized production reports and work summaries with a single click, improving administrative efficiency. Furthermore, it can continuously analyze real-time monitoring data and historical records 24 / 7, proactively identifying potential risks and providing compliance reviews and preventative rectification suggestions to safeguard safe production. Ultimately, it constructs a thinking, early warning, collaborative, and evolving intelligent assistant for coal mines, truly empowering every employee and improving the safety production and operational management level of coal mines.

[0050] This embodiment integrates mine geology, regulations, drawings, and real-time monitoring data to construct a dedicated knowledge base and vector database. Based on an advanced open-source large model, through domain pre-training, instruction fine-tuning, and reinforcement learning, a domain-adaptive AI engine proficient in coal mining expertise is formed. An intelligent question-answering system provides technicians with accurate professional knowledge queries and decision support; an intelligent document module automatically generates technical measures and analysis reports, reducing paperwork burden; and intelligent analysis functions are developed to achieve intelligent analysis of monitoring data and automatic report generation. Ultimately, an intelligent platform integrating expert consultants, office secretaries, and technical specialists is constructed, creating an "AI assistant" for the mine and achieving efficient knowledge utilization and intelligent upgrades to business processes.

[0051] This embodiment aims to create a truly intelligent work partner that understands coal mines, is reliable, and easy to use, comprehensively improving the level of intelligent safety production and office efficiency, and promoting the all-round digital and intelligent transformation of coal mines from production to management.

[0052] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for constructing a mining AI assistant, characterized in that, Includes the following steps: S1. Establish a multi-source knowledge fusion database for coal mines. The knowledge fusion database is based on vector database and knowledge graph technology. It deeply integrates at least one of the following: professional knowledge of coal mines, geological data, engineering drawings, regulations and systems, equipment archives and accident cases. It also establishes bidirectional links between vector and graph nodes to form a hybrid indexing mechanism. S2, a training domain adaptive large model engine, is trained through three stages: domain pre-training, instruction fine-tuning, and human feedback reinforcement learning. The human feedback reinforcement learning includes evaluation of the model response by experts in the coal mining field, providing preference scores or corrective feedback, and training the reward model using a proximal policy optimization algorithm. S3 integrates multiple functional modules into a unified platform. These modules include an intelligent question-and-answer and decision support module, an expert experience digitization module, a safety training and assessment module, a document processing and generation module, a countermeasures and solution query module, and a data analysis and report generation module. This enables the functions of four roles: senior expert coach, efficient office secretary, safety supervision consultant, and all-employee training mentor.

2. The method according to claim 1, characterized in that, In S1, the establishment of a multi-source knowledge fusion library for coal mines includes: integrating multi-source coal mine data and constructing a structured knowledge model; collecting professional data from various business systems through data interfaces; processing text using natural language processing technology, parsing drawings and case images to extract key spatial and entity information; and transforming the extracted information into a structured knowledge model based on a pre-defined coal mine professional domain ontology.

3. The method according to claim 2, characterized in that, In S1, the implementation of the hybrid indexing mechanism includes: transforming standardized knowledge units into high-dimensional vectors through a pre-trained model and storing them in a vector database; constructing a knowledge graph in the form of nodes and edges; and establishing bidirectional links between vectors and graph nodes to support the linkage of semantic search and logical reasoning engines.

4. The method according to claim 1, characterized in that, In S2, the domain pre-training includes: constructing a training dataset based on data from the coal mine multi-source knowledge fusion library, continuously training the basic large model using a self-supervised learning objective, and adjusting the model parameters through domain adaptive optimization.

5. The method according to claim 4, characterized in that, In S2, the instruction fine-tuning includes: constructing a dataset containing task instructions in the coal mining field, covering a variety of complex scenarios through structured instruction design, and using supervised fine-tuning methods to train the model to improve instruction compliance and task-oriented reasoning.

6. The method according to claim 5, characterized in that, In S2, the human feedback reinforcement learning includes: having coal mining experts evaluate the model's response mechanism to construct a feedback dataset, training the reward model using a proximal policy optimization algorithm, and adjusting the model's output policy through iterative feedback loops.

7. The method according to claim 1, characterized in that, In S3, the intelligent question answering and decision support module, based on a multimodal knowledge fusion library and an adaptive large model engine, parses the semantics of user queries, maps them to relevant concepts in the knowledge base, generates explanatory answers through an inference engine, and generates decision suggestions by combining numerical simulations or empirical formulas.

8. The method according to claim 1, characterized in that, In S3, the expert experience digitization module interacts with experts through a guided dialogue interface, generates progressively challenging questions based on a preset knowledge framework and decision tree logic, and uses natural language processing technology to analyze expert answers in real time. Through entity recognition and relation extraction, the narrative is filled into the knowledge framework to form standardized knowledge elements that are stored in the expert experience knowledge base.

9. The method according to claim 1, characterized in that, In S3, the safety training and assessment module creates a virtual underground scenario, loads a safety training script through a dynamic story engine, identifies user operations, calculates decision results in real time through a reasoning model, records the assessment process, and generates a structured evaluation report.

10. The method according to claim 1, characterized in that, In S3, the data analysis and report generation module connects to multiple real-time data sources, stores key parameters in a time series database, analyzes data and identifies abnormal patterns using mathematical statistics algorithms and LSTM time series prediction models, and automatically writes analysis conclusions and recommended measures based on a natural language generation engine, outputting standardized reports.

Images