Rock burst intelligent early warning and decision making system and method based on large language model driving

Abstract

The invention discloses a rock burst intelligent early warning and decision making system and method based on large language model driving. The system is characterized in that a data acquisition module, a data preprocessing module, a domain self-adaption large model module and a multi-agent early warning and decision making module are connected in sequence; the multi-agent early-warning and decision-making module comprises five agents including a seismic source mechanism intelligent analysis early-warning agent, an impact danger early-warning analysis daily report, a rock burst knowledge question and answer agent, a rock burst risk special event analysis agent and a rock burst early-warning and prevention and control auxiliary decision-making agent, and a knowledge database; and the user input module is connected with the rock burst knowledge question-answer agent. The method comprises the following steps: collecting multi-source monitoring data; extracting feature data; carrying out deep semantic understanding and professional reasoning by utilizing a domain self-adaptive large model module; and based on the input data, the multi-agent early warning and decision module outputs result data through a multi-agent communication and negotiation mechanism. According to the invention, accurate early warning and intelligent decision-making of the rock burst risk can be realized.

Description

Technical Field

[0001] This invention belongs to the fields of mine safety engineering and artificial intelligence technology, specifically a rockburst intelligent early warning and decision-making system and method driven by a large language model. Background Technology

[0002] In recent years, large language models (LLMs) such as DeepSeek, GPT-5, and Wenxin Yiyan have made significant progress in the field of natural language processing. Through pre-training on massive corpora, they have demonstrated powerful general knowledge representation and multi-task generalization capabilities, and have been applied in various professional fields such as medicine and law. However, when these general-purpose models are transferred to high-risk and highly specialized industrial scenarios such as rockburst monitoring, early warning, and prevention decision-making in mining engineering, their actual effectiveness faces severe challenges.

[0003] Rockbursts are a major dynamic disaster in deep coal mining, characterized by complex mechanisms, multiple causes, suddenness, and high prediction difficulty. These disasters not only cause severe damage to mine shafts and equipment, threatening personnel lives, but can also trigger secondary disasters such as coal and gas outbursts and gas-dust explosions, posing a systemic threat to mine safety. Currently, general-purpose large-scale language models have three key shortcomings in dealing with this scenario: First, the professional knowledge system in the field of rockbursts is highly specialized, involving a large number of scientific and technological terms, technical procedures, and engineering cases, and heavily relies on the comprehensive analysis of real-time monitoring data (such as microseismic activity, stress, and ground sounds). Because the proportion of relevant domain content in the training corpus of general-purpose models is extremely low, there are knowledge gaps in understanding professional concepts such as "stress concentration factor" and "optimization of stress relief borehole parameters," making it difficult to achieve accurate semantic understanding and deep reasoning. Second, rockburst prevention and control is a typical dynamic decision-making process, requiring a comprehensive judgment combining static technical specifications and real-time monitoring information. Although Retrieval Augmentation (RAG) technology has been introduced to supplement static knowledge, general-purpose models still fall short in terms of result filtering, weight allocation, and multi-step logical reasoning. They struggle to effectively integrate multi-source information, resulting in generated content lacking contextual coherence and causal consistency. Third, rockburst safety decision-making demands extremely high response speed and accuracy, with very low tolerance for error. If the model generates incorrect or ambiguous content, it may delay prevention and control efforts, or even mislead on-site decisions, leading to serious safety hazards. Therefore, there is an urgent need to develop a dedicated large-scale language model for rockburst early warning and prevention, possessing high reliability and strong domain adaptability, to support intelligent mine construction and ensure safe production.

[0004] To address the shortcomings of existing technologies, there is an urgent need to provide an integrated solution that combines knowledge distillation, efficient parameter fine-tuning, and enhanced retrieval generation. This solution would enable the construction of a robust, responsive, and highly reliable intelligent question-and-answer and decision support system for rockburst, providing crucial technical support for mine safety. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention provides a rockburst intelligent early warning and decision-making system and method driven by a large language model. This system adopts a modular design and utilizes an efficient communication mechanism to enable real-time collaborative work among various intelligent agents, ensuring seamless data transmission between modules during rockburst monitoring and early warning. Furthermore, by integrating multiple artificial intelligence agents, it achieves intelligent management of the entire process of rockburst risk assessment, from perception and analysis to early warning and decision support. This method, through multi-source data fusion, domain-adaptive models, multi-agent collaboration, and RAG technology, achieves accurate early warning and intelligent decision-making for rockburst risks, significantly improving the automation level and decision quality of disaster prevention and control.

[0006] To achieve the above objectives, the present invention provides an intelligent early warning and decision-making system for rockbursts based on a large language model, comprising a data acquisition module, a data preprocessing module, a domain adaptive large model module, a multi-agent early warning and decision-making module, and a user input module; The data acquisition module consists of multiple high-precision sensors, which are respectively arranged on multiple monitoring nodes in the area to be monitored to collect multi-source monitoring data in real time. The data preprocessing module is connected to the data acquisition module and is used to preprocess the multi-source monitoring data and extract features from the processed multi-source monitoring data to obtain feature data. The domain-adaptive large model module is connected to the data preprocessing module and is used to perform deep semantic understanding and professional reasoning based on feature data, and output multi-source monitoring data, professional semantic parsing results, and risk reasoning results. The multi-agent early warning and decision-making module is connected to the domain adaptive large model module. The multi-agent early warning and decision-making module includes an intelligent agent for source mechanism analysis and early warning, an intelligent agent for daily reports on shock hazard early warning analysis, an intelligent agent for answering questions about rockburst knowledge, an intelligent agent for analyzing special events related to rockburst risk, an intelligent agent for assisting in early warning and prevention of rockburst, and a knowledge database. These intelligent agents are interconnected, and the knowledge database is connected to each of them. The intelligent source mechanism analysis and early warning system is used to process and analyze earthquake waveform data, complete source location and moment tensor inversion, identify rupture types, and generate and output earthquake risk early warning maps and reports. The intelligent daily report system for shock hazard early warning analysis is used to process and analyze multi-source monitoring data, perform trend comparison and anomaly identification, and generate and output a daily comprehensive safety situation analysis report integrating text and graphics. The intelligent question-and-answer system for rockburst knowledge includes a retrieval enhancement generation module, used to parse regulations and standards, answer professional questions, and provide case-based operational suggestions. The intelligent system for analyzing special events of rockburst risk is used to automatically identify sudden increases in microseismic frequency and precursor signals of high-energy events, and comprehensively analyze geological and mining conditions to obtain an abnormal risk index and infer the causes of rockburst. The intelligent system for rockburst early warning and prevention and control auxiliary decision-making is used to integrate multi-source monitoring data and geological and mining environment data to comprehensively assess the risk of rockburst, and based on the assessment results, intelligently match and optimize prevention and control plans from the knowledge database, and generate and output pressure relief plans or prevention and control measures. The user input module is connected to the rockburst knowledge question-and-answer intelligent agent and is used to receive user input data.

[0007] In this invention, high-precision sensors can collect multi-source monitoring data such as seismic waveforms, stress, and energy in real time. Combined with preprocessing and feature extraction techniques from the data preprocessing module, noise interference can be effectively eliminated, highlighting key risk signals. Through the setting of a domain-adaptive large model module, professional reasoning can be performed on the feature data to output risk analysis results. Semantic understanding integrates domain knowledge, improving the professionalism and accuracy of risk identification, avoiding the generality bias of traditional models, and ensuring accuracy in subsequent early warning and decision-making processes. For the multi-agent early warning and decision-making module, each agent is designed for different tasks related to rockburst early warning and prevention. By simultaneously encapsulating multiple agents performing specific functions, and enabling these agents to negotiate and share information, tasks such as diagnosis, early warning, and scheduling are dynamically allocated. Resource optimization and strategy collaboration are achieved during execution, allowing the multi-agent early warning and decision-making module to comprehensively improve the efficiency and accuracy of rockburst monitoring and early warning, ultimately completing integrated decision-making and multimodal output, ensuring the overall performance and robustness of the system in dynamic environments. The introduction of a retrieval enhancement generation module into the rockburst knowledge question-answering agent facilitates the parsing of regulations and standards and provides case suggestions. Furthermore, the simultaneous existence of five intelligent agents—one for earthquake source mechanism intelligent analysis and early warning, one for daily reporting of shock hazard early warning analysis, one for rockburst knowledge Q&A, one for special events analysis of rockburst risk, and one for rockburst early warning and prevention auxiliary decision-making—allows each to independently handle its own main tasks. This scientifically decomposes tasks that traditionally rely on a single agent, significantly improving processing efficiency and reducing the computational demands of each agent. In particular, the simultaneous setup of five agents effectively avoids the system-wide failure that can occur when one agent malfunctions, as is common in traditional single-agent systems. In this system, even if one agent fails, the other four can still operate normally and output their respective task results, preserving most of the system's functionality and preventing significant losses from overall failure. This effectively ensures the timeliness and reliability of early warning and decision-making. This mechanism ensures that the output conforms to linguistic logic and is supported by domain knowledge, consistent with the document retrieval and knowledge fusion process of RAG in historical dialogues, thus enhancing decision credibility. Connecting the user input module to the knowledge-based question-answering AI allows for convenient support of natural language interaction. This enables the system to instantly parse user queries and generate dynamic suggestions based on real-time monitoring data, significantly improving emergency response efficiency.

[0008] This system adopts a modular design to build a rockburst intelligent question-and-answer and auxiliary decision-making system with solid professional knowledge, accurate response, and high reliability. Through an efficient communication mechanism, it enables various intelligent agents to work collaboratively in real time, ensuring seamless data transmission between modules during the monitoring and early warning of rockbursts. At the same time, by integrating multiple artificial intelligence agents, it realizes the intelligentization of the entire process of rockburst risk from perception, analysis, early warning to decision support, providing key technical support for mine safety.

[0009] To facilitate real-time data and information display, a result output and display module is also included. This module is connected to the multi-agent early warning and decision-making module, and includes five receiving channels and five corresponding display areas. The five receiving channels are respectively connected to the following intelligent agents: earthquake source mechanism intelligent analysis and early warning agent, shock hazard early warning analysis daily report agent, rockburst knowledge Q&A agent, rockburst risk special event analysis agent, and rockburst early warning and prevention auxiliary decision-making agent. The result output and display module is used to display earthquake risk early warning maps and reports in real time, a daily comprehensive safety situation analysis report integrating text and graphics, operational suggestions and preventive measures based on user questions, abnormal risk indices and causes of abnormal events, and pressure relief plans or prevention and control measures. Simultaneously, if no data is received from any channel within a set time, the corresponding intelligent agent's fault alarm information is directly displayed in the corresponding display area. This enables real-time online diagnostics for each intelligent agent.

[0010] As a preferred embodiment, the knowledge database includes mine safety regulations data, case literature data, historical monitoring data, expert experience and academic research results data, geological and mining environment data, and historical comprehensive analysis data; the retrieval enhancement generation module is used to dynamically retrieve relevant knowledge fragments from the knowledge database based on user input data and output retrieval results data.

[0011] This invention also provides a method for intelligent early warning and decision-making regarding rockbursts based on a large language model, employing an intelligent early warning and decision-making system for rockbursts based on a large language model, comprising the following steps: Step 1: Construct a domain-adaptive large model module; Using a general large language model as the base model, it is trained with a high-quality domain corpus built based on knowledge distillation technology, and low-rank adaptive technology is used for parameter fine-tuning to construct a domain adaptive large model module. Step 2: Construct a multi-agent early warning and decision-making module; S21: Collect data on mine safety regulations, case literature, historical monitoring data, expert experience and academic research results, geological and mining environment data, and historical comprehensive analysis data, and use specific vectorization methods to encode the text corpus into embedded vectors to construct a knowledge database; S22: Connects the intelligent agent for source mechanism analysis and early warning, the intelligent agent for daily report of shock hazard early warning analysis, the intelligent agent for knowledge Q&A of rockburst, the intelligent agent for special event analysis of rockburst risk, the intelligent agent for rockburst early warning and prevention auxiliary decision-making, and the knowledge database. In addition, a retrieval enhancement generation module is introduced into the intelligent agent for knowledge Q&A of rockburst to construct a multi-agent early warning and decision-making module. Step 3: Construct an intelligent early warning and decision-making system for rockbursts driven by a large language model; The system connects the domain adaptive large model module and the multi-agent early warning and decision-making module. A data preprocessing module is connected to the input end of the domain adaptive large model module, and the data preprocessing module is connected to the data acquisition module. A user input module is connected to the input end of the rockburst knowledge question answering agent, and a result output and display module is connected to the output end of the multi-agent early warning and decision-making module. This constructs a rockburst intelligent early warning and decision-making system driven by a large language model. Step 4: Online intelligent early warning and decision-making for rockburst; S41: Collect multi-source monitoring data from various monitoring nodes through the data acquisition module; wherein, the data acquisition module consists of multiple high-precision sensors deployed on multiple monitoring nodes in the area to be monitored. The high-precision sensors are used to collect various physical signals in real time, such as temperature, vibration or images. S42: The data preprocessing module performs noise reduction and standardization on the multi-source monitoring data, and extracts discriminative feature data from the processed multi-source monitoring data. Preferably, the data preprocessing module can be a data gateway, which is used for the collection and transmission of multi-source monitoring data, and also undertakes functions such as protocol conversion, data caching and filtering, standardization, and feature extraction. High-quality basic data can be obtained through filtering and standardization, and then discriminative feature vectors can be obtained from the original monitoring data through feature extraction, thereby reducing data dimensionality and highlighting potential trends. Feature extraction can be performed using methods such as time-frequency domain analysis. Further, preferably, statistical methods or machine learning algorithms can be combined to automatically identify and label abnormal equipment status and data outliers, so as to provide structured, high-quality data input for the subsequent domain adaptive large model module. S43: Using feature data as input data, the domain adaptive large model module is used for deep semantic understanding and professional reasoning, and outputs multi-source monitoring data, professional semantic parsing results, and risk reasoning results as the basis for early warning and decision-making. The risk reasoning results can include advanced cognitive results such as fault diagnosis and life prediction, so as to significantly improve the interpretability and reliability of the coefficients in professional scenarios. S44: When the user input module receives user input data, the retrieval enhancement generation module first dynamically retrieves relevant knowledge fragments from the knowledge database based on the user input data and outputs the retrieval results data. During this process, the retrieval of relevant information from the knowledge database and the output of retrieval results data through the RAG mechanism can further provide real-time and accurate upward and downward support for the multi-agent early warning and decision-making module. Then, the retrieval results data and the early warning and decision-making basic data are used together as input data. When the user input module does not receive user input data, the early warning and decision-making basic data are used as input data. Based on input data, the multi-agent early warning and decision-making module performs intelligent diagnosis, early warning and auxiliary decision-making for rockburst through the communication and negotiation mechanism of multi-agents. It outputs earthquake risk early warning map and report, daily safety situation comprehensive analysis report with integrated graphics and text, operation suggestions and preventive measures based on user questions, abnormal risk index and abnormal event causes, pressure relief plan or control measures to the result output and display module. Step 5: Output and Display of Results; After receiving the output result of the corresponding intelligent agent, the receiving channel of the result output and display module displays it in real time through the corresponding display area. If no output result is received by any receiving channel within a set time, the intelligent agent fault alarm is displayed in real time through the corresponding display area.

[0012] To ensure a simpler structure for the domain-adaptive large model while significantly reducing computational requirements while preserving performance, the process of constructing the domain-adaptive large model module in step one is as follows: S11: Collect and organize professional knowledge in the field of rockburst from a large amount of professional literature, academic papers, technical regulations, historical cases, and real-time monitoring data, and construct a structured domain corpus covering rockburst terminology, mine cases, and regulatory data. mine ; S12: Based on domain corpus D mine We extract internal knowledge representations from existing general-purpose large language models using knowledge distillation methods, and formulate knowledge transfer schemes based on domain characteristics to ultimately generate high-quality training datasets. The process of obtaining a high-quality training dataset is as follows; S12-1: Using the general large language model as the teacher model T and the model to be fine-tuned as the student model S, for the input sample x, the output probability of the teacher model T is obtained according to formula (1). The output probability of student model S is obtained according to formula (1). ; (1); (2); In the formula, Z T With Z S These are the logits for the teacher and student models, respectively, where T is the temperature coefficient. S12-2: Construct the distillation loss function according to formula (3) By minimizing the distillation loss function This enables the student model S to enhance its understanding of the semantics of rockburst terminology, cases, and procedures during the distillation process, and generates a high-quality training dataset. (3); In the formula, CE is the cross-entropy loss, KL is the Kullback-Leibler divergence, and α is the weighting coefficient; S13: Freeze the original parameters of the general large language model, inject trainable low-rank matrices only next to the key modules of the model, and reduce the amount of parameter updates through low-rank matrix decomposition; use the training dataset to fine-tune the parameters of the general large language model, so that while maintaining the general capabilities, the model focuses on training its ability to understand professional terms and knowledge in the field of rockburst, and obtain a domain-adaptive large model. The parameter fine-tuning process is as follows: Freeze the original parameter matrix. Only update the low-rank matrix. and According to formula (4), the new weight matrix after LoRA fine-tuning is obtained. ; (4); In the formula, Update the matrix for weights. , For rank, and ; S14: Evaluate the performance of the domain-adaptive large model and iteratively optimize the domain corpus, training dataset, and LoRA configuration parameters based on performance metrics.

[0013] As a preferred embodiment, in step S22 of step two, the process of retrieving the answer from the enhanced generation module based on the user query is as follows: S22-1: The user's query question Transform into a query vector ; S22-2: In the vector space, calculate the query vector according to formula (5). With each text vector Cosine similarity between ; (5); S22-3: Based on the calculated similarity score, filter out the top results most relevant to the current query from the entire knowledge database. document fragments Use contextual information as input to generate the model; S22-4: Calculate d for each document fragment i For a given query problem weight And obtain the probability of the final answer y according to formula (6); (6); In the formula, For a given user query and the entire collection of retrieved documents Under the given conditions, the overall probability of generating the final answer; S22-5: Obtain the final answer according to formula (7) ; (7).

[0014] To accurately output the warning level, in step four, S44, during the generation of the earthquake risk warning map and report, the intelligent source mechanism analysis and warning agent obtains the warning level by analyzing the source mechanism through earthquake waveform inversion. The process is as follows: S44-1: Obtain the seismic moment tensor according to formula (8) ; (8); In the formula, This refers to the rock mass shear modulus. For the sliding displacement component, For fault normal components; S44-2: Based on Seismic Moment Tensor The seismic energy release rate is calculated according to formula (9) based on the principal stress direction. ; (9); S44-3: Set a low energy threshold and high energy threshold The warning level is determined according to formula (10). ; (10); In order to accurately generate a daily security situation comprehensive analysis report integrating graphics and text, in step S44 of step four, during the process of generating the daily security situation comprehensive analysis report integrating graphics and text, the impact hazard early warning analysis daily report intelligent agent analyzes the trend changes through time series data, and first obtains the sliding window average value according to formula (11). Then, obtain the outliers according to formula (12). And based on outliers Perform anomaly detection; if it occurs consecutively... If so, the alarm mechanism will be triggered; (11); (12); In the formula, For threshold coefficient, The standard deviation is denoted as .

[0015] In order to accurately infer the causes of abnormal situations, in step S44 of step four, during the process of obtaining the abnormal risk index, the rockburst risk special event analysis agent first obtains the frequency according to formula (13). and with a set frequency threshold If a comparison is made, If so, it is judged as an anomaly, and then combined with geological parameters. and energy data The abnormal risk index is obtained according to formula (14). ; (13); (14); In the formula, α and β are two different weighting coefficients.

[0016] In order to accurately generate a pressure relief plan, in step S44 of step four, during the process of generating a pressure relief plan or prevention and control measures, the rockburst early warning and prevention and control auxiliary decision-making agent first performs multi-source data fusion according to formula (15) to obtain a comprehensive risk index. Then, the comprehensive risk index Compare with the set risk threshold ,like The pressure relief parameter adjustment ratio is obtained according to formula (16). ; (15); (16); In the formula, For seismic activity data; It is in a state of stress; Energy data; These are the weighting coefficients for seismic activity data; These are the weighting coefficients for the stress state; These are the weighting coefficients for the energy data; For safety factor; This is the expected energy value; This represents the current monitored energy value.

[0017] This invention provides an efficient and reliable intelligent early warning and prevention decision-making method for rockburst based on a large language model. First, it combines knowledge distillation technology and low-rank adaptive fine-tuning to construct a domain-adaptive large model module, significantly reducing training costs. Simultaneously, it avoids linguistic biases in specialized scenarios, enabling the output of professional semantic parsing results from multi-source monitoring data, providing interpretable intermediate results for subsequent risk warnings. Next, it integrates data such as mine safety regulations, case studies, and expert experience to construct a knowledge database, ensuring effective coverage of regulations, standards, historical events, and real-time monitoring, supporting comprehensive risk assessment. Finally, it introduces a retrieval enhancement generation module into the knowledge question-answering agent, enabling accurate and timely parsing of regulations and standards and providing case suggestions, ensuring the output conforms to linguistic logic and is supported by domain knowledge. Each agent in this system is designed for different tasks related to rockburst early warning and prevention. By encapsulating multiple agents that perform specific functions, and enabling these agents to negotiate and share information, tasks such as diagnosis, early warning, and scheduling are dynamically allocated. During execution, resource optimization and strategy collaboration are achieved, allowing the multi-agent early warning and decision-making module to comprehensively improve the efficiency and accuracy of rockburst monitoring and early warning, ultimately achieving integrated decision-making and multimodal output, ensuring the system's overall performance and robustness in dynamic environments. Furthermore, the user input module and result output display module facilitate convenient information input for users, supporting natural language queries and lowering the operational threshold. Simultaneously, the result output and display modules provide intuitive presentation of data such as graphs and reports, allowing on-site personnel to easily access the results. Furthermore, the simultaneous existence of five intelligent agents—one for earthquake source mechanism analysis and early warning, one for daily shock hazard early warning analysis, one for rockburst knowledge Q&A, one for rockburst risk special event analysis, and one for rockburst early warning and prevention auxiliary decision-making—allows each to independently handle its own main tasks. This scientifically decomposes tasks that traditionally rely on a single agent, significantly improving processing efficiency and reducing the computational demands of each agent. In particular, the simultaneous setup of five agents effectively avoids the system-wide failure that can occur when one agent malfunctions, as is common in traditional single-agent systems. In this system, even if one agent fails, the other four can still operate normally and output their respective task results, preserving most of the system's functionality and preventing significant losses from overall failure. This effectively ensures the timeliness and reliability of early warning and decision-making. Finally, the online intelligent early warning and decision-making process simultaneously generates diverse results such as earthquake risk early warning maps, daily safety reports, and prevention plans, meeting the needs of different scenarios.Meanwhile, by displaying the output results of the corresponding intelligent agents through the result output and display module, relevant personnel can not only accurately obtain the output results of the corresponding intelligent agents, but also promptly identify whether there is a single intelligent agent failure. This allows for timely repair of the faulty intelligent agents, thus avoiding major disasters caused by the inability to monitor properly. At the same time, this mechanism can work with five intelligent agents to effectively ensure continuous and reliable monitoring of rockbursts.

[0018] Compared with the prior art, the present invention has the following advantages: 1. Accuracy and professionalism: By integrating domain-adaptive large models and knowledge bases, it can ensure that risk identification and decision-making recommendations meet mine safety standards.

[0019] 2. Good real-time and dynamic performance: The entire process from data acquisition to early warning output is automated, which can support rapid response to emergencies.

[0020] 3. Interactivity and Explainability: The user input module and knowledge-based question-answering agent provide personalized support, while the early warning graph and reports enhance the transparency of decision-making.

[0021] 4. Cost-effectiveness: Low-rank adaptive fine-tuning and modular design reduce training and maintenance costs, making it suitable for large-scale deployment.

[0022] This method achieves accurate early warning and intelligent decision-making for rockburst risks through multi-source data fusion, domain adaptive models, multi-agent collaboration, and RAG technology. It significantly improves the automation level and decision quality of disaster prevention and control, effectively addresses the complexity and high-risk issues in mine safety production, and provides strong technical support for the intelligent construction and safety management of mines. Attached Figure Description

[0023] Figure 1 This is a principle block diagram of the system part of this invention; Figure 2 This is a flowchart of the intelligent agent for analyzing the earthquake source mechanism in this invention; Figure 3 This is a flowchart of the daily report generation intelligent agent in this invention; Figure 4 This is a flowchart of the knowledge-based question-answering intelligent agent in this invention; Figure 5 This is a flowchart of the special event analysis intelligent agent in this invention; Figure 6 This is a flowchart of the intelligent agent generated for decision support in this invention. Detailed Implementation

[0024] The invention will now be further described with reference to the accompanying drawings.

[0025] like Figures 1 to 6 As shown, the present invention provides an intelligent early warning and decision-making system for rockbursts based on a large language model, including a data acquisition module, a data preprocessing module, a domain adaptive large model module, a multi-agent early warning and decision-making module, and a user input module; The data acquisition module consists of multiple high-precision sensors, which are respectively arranged on multiple monitoring nodes in the area to be monitored to collect multi-source monitoring data in real time. The data preprocessing module is connected to the data acquisition module and is used to preprocess the multi-source monitoring data and extract features from the processed multi-source monitoring data to obtain feature data. The domain-adaptive large model module is connected to the data preprocessing module and is used to perform deep semantic understanding and professional reasoning based on feature data, and output multi-source monitoring data, professional semantic parsing results, and risk reasoning results. The multi-agent early warning and decision-making module is connected to the domain adaptive large model module. The multi-agent early warning and decision-making module includes an intelligent agent for source mechanism analysis and early warning, an intelligent agent for daily reports on shock hazard early warning analysis, an intelligent agent for answering questions about rockburst knowledge, an intelligent agent for analyzing special events related to rockburst risk, an intelligent agent for assisting in early warning and prevention of rockburst, and a knowledge database. These intelligent agents are interconnected, and the knowledge database is connected to each of them. The intelligent source mechanism analysis and early warning system is used to process and analyze earthquake waveform data, complete source location and moment tensor inversion, identify rupture types, and generate and output earthquake risk early warning maps and reports. The daily report system for shock hazard early warning analysis is used to process and analyze multi-source monitoring data, perform trend comparison and anomaly identification, and generate and output a daily comprehensive safety situation analysis report integrating text and graphics. The rockburst knowledge Q&A system includes a retrieval enhancement generation module, used to accurately interpret regulations and standards, answer professional questions, and provide case-based operational suggestions. The rockburst risk special event analysis system is used to automatically identify precursor signals such as sudden increases in microseismic frequency and high-energy events, and comprehensively analyze geological and mining conditions to obtain an abnormal risk index and infer the causes of rockbursts. The rockburst early warning and prevention auxiliary decision-making system is used to integrate multi-source monitoring data and geological and mining environment data to comprehensively assess the risk of rockbursts, and based on the assessment results, intelligently match and optimize prevention and control plans from the knowledge database, and generate and output pressure relief plans or prevention and control measures. The user input module is connected to the rockburst knowledge question-and-answer intelligent agent and is used to receive user input data.

[0026] To facilitate real-time data and information display, a result output and display module is also included. This module is connected to the multi-agent early warning and decision-making module, and includes five receiving channels and five corresponding display areas. The five receiving channels are respectively connected to the following intelligent agents: earthquake source mechanism intelligent analysis and early warning agent, shock hazard early warning analysis daily report agent, rockburst knowledge Q&A agent, rockburst risk special event analysis agent, and rockburst early warning and prevention auxiliary decision-making agent. The result output and display module is used to display earthquake risk early warning maps and reports in real time, a daily comprehensive safety situation analysis report integrating text and graphics, operational suggestions and preventive measures based on user questions, abnormal risk indices and causes of abnormal events, and pressure relief schemes or prevention and control measures. Simultaneously, if no data is received from any channel within a set time, the corresponding intelligent agent's fault alarm information is directly displayed in the corresponding display area. This enables real-time online diagnostics of each intelligent agent.

[0027] As a preferred embodiment, the knowledge database includes mine safety regulations data, case literature data, historical monitoring data, expert experience and academic research results data, geological and mining environment data, and historical comprehensive analysis data; the retrieval enhancement generation module is used to dynamically retrieve relevant knowledge fragments from the knowledge database based on user input data and output retrieval results data.

[0028] In this invention, high-precision sensors can collect multi-source monitoring data such as seismic waveforms, stress, and energy in real time. Combined with preprocessing and feature extraction techniques from the data preprocessing module, noise interference can be effectively eliminated, highlighting key risk signals. Through the setting of a domain-adaptive large model module, professional reasoning can be performed on the feature data to output risk analysis results. Semantic understanding integrates domain knowledge, improving the professionalism and accuracy of risk identification, avoiding the generality bias of traditional models, and ensuring accuracy in subsequent early warning and decision-making processes. For the multi-agent early warning and decision-making module, each agent is designed for different tasks related to rockburst early warning and prevention. By simultaneously encapsulating multiple agents performing specific functions, and enabling these agents to negotiate and share information, tasks such as diagnosis, early warning, and scheduling are dynamically allocated. Resource optimization and strategy collaboration are achieved during execution, allowing the multi-agent early warning and decision-making module to comprehensively improve the efficiency and accuracy of rockburst monitoring and early warning, ultimately completing integrated decision-making and multimodal output, ensuring the overall performance and robustness of the system in dynamic environments. The introduction of a retrieval enhancement generation module into the rockburst knowledge question-answering agent facilitates the parsing of regulations and standards and provides case suggestions. Furthermore, the simultaneous existence of five intelligent agents—one for earthquake source mechanism intelligent analysis and early warning, one for daily reporting of shock hazard early warning analysis, one for rockburst knowledge Q&A, one for special events analysis of rockburst risk, and one for rockburst early warning and prevention auxiliary decision-making—allows each to independently handle its own main tasks. This scientifically decomposes tasks that traditionally rely on a single agent, significantly improving processing efficiency and reducing the computational demands of each agent. In particular, the simultaneous setup of five agents effectively avoids the system-wide failure that can occur when one agent malfunctions, as is common in traditional single-agent systems. In this system, even if one agent fails, the other four can still operate normally and output their respective task results, preserving most of the system's functionality and preventing significant losses from overall failure. This effectively ensures the timeliness and reliability of early warning and decision-making. This mechanism ensures that the output conforms to linguistic logic and is supported by domain knowledge, consistent with the document retrieval and knowledge fusion process of RAG in historical dialogues, thus enhancing decision credibility. Connecting the user input module to the knowledge-based question-answering AI allows for convenient support of natural language interaction. This enables the system to instantly parse user queries and generate dynamic suggestions based on real-time monitoring data, significantly improving emergency response efficiency.

[0029] This system adopts a modular design to build a rockburst intelligent question-and-answer and auxiliary decision-making system with solid professional knowledge, accurate response, and high reliability. Through an efficient communication mechanism, it enables various intelligent agents to work collaboratively in real time, ensuring seamless data transmission between modules during the monitoring and early warning of rockbursts. At the same time, by integrating multiple artificial intelligence agents, it realizes the intelligentization of the entire process of rockburst risk from perception, analysis, early warning to decision support, providing key technical support for mine safety.

[0030] This invention also provides a method for intelligent early warning and decision-making regarding rockbursts based on a large language model, employing an intelligent early warning and decision-making system for rockbursts based on a large language model, comprising the following steps: Step 1: Construct a domain-adaptive large model module; A general-purpose large language model is adopted as the base model, with the Qwen2-7B model being the preferred choice. The model is trained using a high-quality domain corpus constructed based on knowledge distillation technology, enabling it to better adapt to the specific needs of the mining engineering field. To improve the model's efficiency and accuracy, Low-Rank Adaptation (LoRA) is employed for parameter fine-tuning. LoRA technology can finely adjust domain-specific model parameters while retaining most of the general model parameters, significantly improving computational efficiency and ensuring the model's efficient adaptability to the rockburst domain. This method constructs a domain-adaptive large model module, achieving deep adaptation of the model in the rockburst domain. This allows the model to learn a deep understanding and reasoning ability regarding rockburst-related terms (such as stress concentration factor, stress relief borehole parameters, energy release rate, etc.), thereby achieving accurate mastery of domain knowledge.

[0031] Step 2: Construct a multi-agent early warning and decision-making module; S21: Collect mine safety regulations data, case literature data, historical monitoring data, expert experience and academic research data, geological and mining environment data, and historical comprehensive analysis data. Then, use a specific vectorization method to encode the text corpus into embedded vectors to construct a structured domain knowledge database. This database uses a specific vectorization method (employing encoding models such as TF-IDF, BERT, or Word2Vec) to encode each document fragment in the text corpus. i Encode to form an embedding vector This enables effective support for fast retrieval and matching.

[0032] S22: Connects the intelligent agent for source mechanism analysis and early warning, the intelligent agent for daily report of shock hazard early warning analysis, the intelligent agent for knowledge Q&A of rockburst, the intelligent agent for special event analysis of rockburst risk, the intelligent agent for rockburst early warning and prevention auxiliary decision-making, and the knowledge database. In addition, a retrieval enhancement generation module is introduced into the intelligent agent for knowledge Q&A of rockburst to construct a multi-agent early warning and decision-making module. The introduction of the retrieval enhancement generation module further improves the system's responsiveness and accuracy. When processing user queries, this module dynamically retrieves the most relevant procedures, cases, and other knowledge fragments, and integrates them with content generated by the domain-adaptive large model module before inputting them into the subsequent generation process, ensuring the accuracy and timeliness of the answer. In this way, the system can maintain the reasoning ability of the domain-adaptive large model while ensuring the accuracy, timeliness, and domain relevance of the answer when responding to user questions. For example, if a user asks, "How should we handle the frequent microseismic events at the current working face?", the system will automatically retrieve relevant procedures, cases, and monitoring data, and combine them with real-time information to generate a comprehensive solution, ensuring the answer is professional and practical.

[0033] Step 3: Construct an intelligent early warning and decision-making system for rockbursts driven by a large language model; The system connects the domain adaptive large model module and the multi-agent early warning and decision-making module. A data preprocessing module is connected to the input end of the domain adaptive large model module, and the data preprocessing module is connected to the data acquisition module. A user input module is connected to the input end of the rockburst knowledge question answering agent, and a result output and display module is connected to the output end of the multi-agent early warning and decision-making module. This constructs a rockburst intelligent early warning and decision-making system driven by a large language model. Step 4: Online intelligent early warning and decision-making for rockburst; S41: Collect multi-source monitoring data from each monitoring node through the data acquisition module; S42: The data preprocessing module performs noise reduction and standardization on the multi-source monitoring data, and extracts discriminative feature data from the processed multi-source monitoring data; S43: Using feature data as input data, it utilizes the domain-adaptive large model module for deep semantic understanding and professional reasoning, and outputs multi-source monitoring data, professional semantic parsing results, and risk reasoning results as the basis for early warning and decision-making. S44: When the user input module receives user input data, it first dynamically retrieves relevant knowledge fragments from the knowledge database based on the user input data through the retrieval enhancement generation module and outputs the retrieval result data; then, it uses the retrieval result data and the early warning and decision-making basis data together as input data; when the user input module does not receive user input data, it uses the early warning and decision-making basis data as input data. Based on input data, the multi-agent early warning and decision-making module performs intelligent diagnosis, early warning and auxiliary decision-making for rockburst through the communication and negotiation mechanism of multi-agents. It outputs earthquake risk early warning map and report, daily safety situation comprehensive analysis report with integrated graphics and text, operation suggestions and preventive measures based on user questions, abnormal risk index and abnormal event causes, pressure relief plan or control measures to the result output and display module. Step 5: Output and Display of Results; After receiving the output result of the corresponding intelligent agent, the receiving channel of the result output and display module displays it in real time through the corresponding display area. If no output result is received by any receiving channel within a set time, the intelligent agent fault alarm is displayed in real time through the corresponding display area.

[0034] In this process, the intelligent analysis and early warning agent for seismic sources automatically processes seismic waveform data from monitoring equipment, completes source location and moment tensor inversion, and then identifies rupture types. By analyzing the data, the agent can generate seismic risk early warning maps and, based on historical and real-time monitoring data, automatically generate risk warning reports for seismic events, providing timely safety warnings for mine operations. The daily rockburst early warning analysis agent integrates data from multiple sensors (such as microseismic, stress, and temperature sensors) for trend comparison and anomaly identification. Through data mining technology, the agent can automatically generate a comprehensive daily safety situation analysis report integrating text and graphics, providing timely safety assessments for mine managers and predicting potential rockburst risks. The rockburst knowledge Q&A agent is specifically designed to handle professional questions regarding rockburst prevention, technical specifications, regulations, and standards. By accurately interpreting technical procedures and operation manuals, the agent can provide case-based operational suggestions and preventative measures based on user questions, helping mine workers quickly solve practical problems. The rockburst knowledge Q&A agent can automatically categorize questions according to type, such as: operation procedures, case reasoning, and safety control. The rockburst risk special event analysis agent is responsible for automatically identifying precursor signals such as sudden increases in microseismic frequency and high-energy events, and conducting a comprehensive analysis in conjunction with the mine's geological conditions and mining environment. Through multi-dimensional data analysis, the rockburst risk special event analysis agent can obtain a comprehensive risk index and infer potential rockburst triggers, providing early warning signals for mine managers to take preventive measures and avoid accidents. The rockburst early warning and prevention auxiliary decision-making agent combines multi-source sensing data (such as pressure sensors, temperature sensors, etc.) and mining environment information to comprehensively assess the risk of rockburst. Based on the assessment results, the agent can intelligently match and optimize prevention and control plans from the prevention and control measures knowledge base, and generate specific pressure relief plans or prevention and control measures, providing decision support for mine management.

[0035] To ensure a simpler structure for the domain-adaptive large model while significantly reducing computational requirements while preserving performance, the process of constructing the domain-adaptive large model module in step one is as follows: S11: Collect and organize professional knowledge in the field of rockburst from a large amount of professional literature, academic papers, technical regulations, historical cases, and real-time monitoring data, and construct a structured domain corpus covering rockburst terminology, mine cases, and regulatory data. mine ; S12: Based on domain corpus D mine We extract internal knowledge representations from existing general-purpose large language models using knowledge distillation methods, and formulate knowledge transfer schemes based on domain characteristics to ultimately generate high-quality training datasets. The process of obtaining a high-quality training dataset is as follows; S12-1: Using the general large language model as the teacher model T and the model to be fine-tuned as the student model S, for the input sample x, the output probability of the teacher model T is obtained according to formula (1). The output probability of student model S is obtained according to formula (1). ; (1); (2); In the formula, Z T With Z S These are the logits for the teacher and student models, respectively, where T is the temperature coefficient. S12-2: Construct the distillation loss function according to formula (3) By minimizing the distillation loss function This enables the student model S to enhance its understanding of semantics in areas such as rockburst terminology, cases, and procedures during the distillation process, and generates a high-quality training dataset. (3); In the formula, CE is the cross-entropy loss, KL is the Kullback-Leibler divergence, and α is the weighting coefficient; S13: Freeze the original parameters of the general-purpose large language model, injecting only trainable low-rank matrices around key modules such as the model's attention mechanism, and reduce the amount of parameter updates through low-rank matrix factorization; fine-tune the parameters of the general-purpose large language model using the training dataset, so that while maintaining the model's general capabilities, it focuses on training the model's ability to understand professional terminology and knowledge in the field of rockburst, thus obtaining a domain-adaptive large model; this method significantly reduces computational complexity while preserving the performance of the original model. The parameter fine-tuning process is as follows: Freeze the original parameter matrix. Only update the low-rank matrix. and According to formula (4), the new weight matrix after LoRA fine-tuning is obtained. ; (4); In the formula, Update the matrix for weights. , For rank, and ; S14: Evaluate the performance of the domain-adaptive large model and iteratively optimize the domain corpus, training dataset, and LoRA configuration parameters based on performance metrics.

[0036] As a preferred embodiment, in step S22 of step two, the process of retrieving the answer from the enhanced generation module based on the user query is as follows: S22-1: The user's query question Transform into a query vector ; S22-2: In the vector space, calculate the query vector according to formula (5). With each text vector Cosine similarity between ; (5); S22-3: Based on the calculated similarity score, filter out the top results most relevant to the current query from the entire knowledge database. document fragments Use contextual information as input to generate the model; S22-4: Calculate d for each document fragment i For a given query problem weight And obtain the probability of the final answer y according to formula (6); where, Obtained through similarity normalization, this strategy ensures that the output conforms to both linguistic logic and domain knowledge support.

[0037] (6); In the formula, For a given user query and the entire collection of retrieved documents Under the given conditions, the overall probability of generating the final answer; S22-5: Obtain the final answer according to formula (7) ; (7).

[0038] To accurately output the warning level, in step four, S44, during the generation of the earthquake risk warning map and report, the intelligent source mechanism analysis and warning agent obtains the warning level by analyzing the source mechanism through earthquake waveform inversion. The process is as follows: S44-1: Obtain the seismic moment tensor according to formula (8) ; (8); In the formula, This refers to the rock mass shear modulus. For the sliding displacement component, For fault normal components; S44-2: Based on Seismic Moment Tensor The seismic energy release rate is calculated according to formula (9) based on the principal stress direction. ; (9); S44-3: Set a low energy threshold and high energy threshold The warning level is determined according to formula (10). ; (10); In order to accurately generate a daily security situation comprehensive analysis report integrating graphics and text, in step S44 of step four, during the process of generating the daily security situation comprehensive analysis report integrating graphics and text, the impact hazard early warning analysis daily report intelligent agent analyzes the trend changes through time series data, and first obtains the sliding window average value according to formula (11). Then, obtain the outliers according to formula (12). And based on outliers Perform anomaly detection; if it occurs consecutively... If so, the alarm mechanism will be triggered; (11); (12); In the formula, For threshold coefficient, The standard deviation is denoted as .

[0039] In order to accurately infer the causes of abnormal situations, in step S44 of step four, during the process of obtaining the abnormal risk index, the rockburst risk special event analysis agent first obtains the frequency according to formula (13). and with a set frequency threshold If a comparison is made, If so, it is judged as an anomaly, and then combined with geological parameters. and energy data The abnormal risk index is obtained according to formula (14). ; (13); (14); In the formula, α and β are two different weighting coefficients.

[0040] In order to accurately generate a pressure relief plan, in step S44 of step four, during the process of generating a pressure relief plan or prevention and control measures, the rockburst early warning and prevention and control auxiliary decision-making agent first performs multi-source data fusion according to formula (15) to obtain a comprehensive risk index. Then, the comprehensive risk index Compare with the set risk threshold ,like The pressure relief parameter adjustment ratio is obtained according to formula (16). ; (15); (16); In the formula, For seismic activity data; It is in a state of stress; Energy data; These are the weighting coefficients for seismic activity data; These are the weighting coefficients for the stress state; These are the weighting coefficients for the energy data; For safety factor; This is the expected energy value; This represents the current monitored energy value.

[0041] This invention provides an efficient and reliable intelligent early warning and prevention decision-making method for rockburst based on a large language model. First, it combines knowledge distillation technology and low-rank adaptive fine-tuning to construct a domain-adaptive large model module, significantly reducing training costs. Simultaneously, it avoids linguistic biases in specialized scenarios, enabling the output of professional semantic parsing results from multi-source monitoring data, providing interpretable intermediate results for subsequent risk warnings. Next, it integrates data such as mine safety regulations, case studies, and expert experience to construct a knowledge database, ensuring effective coverage of regulations, standards, historical events, and real-time monitoring, supporting comprehensive risk assessment. Finally, it introduces a retrieval enhancement generation module into the knowledge question-answering agent, enabling accurate and timely parsing of regulations and standards and providing case suggestions, ensuring the output conforms to linguistic logic and is supported by domain knowledge. Each agent in this system is designed for different tasks related to rockburst early warning and prevention. By encapsulating multiple agents that perform specific functions, and enabling these agents to negotiate and share information, tasks such as diagnosis, early warning, and scheduling are dynamically allocated. During execution, resource optimization and strategy collaboration are achieved, allowing the multi-agent early warning and decision-making module to comprehensively improve the efficiency and accuracy of rockburst monitoring and early warning, ultimately achieving integrated decision-making and multimodal output, ensuring the system's overall performance and robustness in dynamic environments. Furthermore, the user input module and result output display module facilitate convenient information input for users, supporting natural language queries and lowering the operational threshold. Simultaneously, the result output and display modules provide intuitive presentation of data such as graphs and reports, allowing on-site personnel to easily access the results. Furthermore, the simultaneous existence of five intelligent agents—one for earthquake source mechanism analysis and early warning, one for daily shock hazard early warning analysis, one for rockburst knowledge Q&A, one for rockburst risk special event analysis, and one for rockburst early warning and prevention auxiliary decision-making—allows each to independently handle its own main tasks. This scientifically decomposes tasks that traditionally rely on a single agent, significantly improving processing efficiency and reducing the computational demands of each agent. In particular, the simultaneous setup of five agents effectively avoids the system-wide failure that can occur when one agent malfunctions, as is common in traditional single-agent systems. In this system, even if one agent fails, the other four can still operate normally and output their respective task results, preserving most of the system's functionality and preventing significant losses from overall failure. This effectively ensures the timeliness and reliability of early warning and decision-making. Finally, the online intelligent early warning and decision-making process simultaneously generates diverse results such as earthquake risk early warning maps, daily safety reports, and prevention plans, meeting the needs of different scenarios.Meanwhile, by displaying the output results of the corresponding intelligent agents through the result output and display module, relevant personnel can not only accurately obtain the output results of the corresponding intelligent agents, but also promptly identify whether there is a single intelligent agent failure. This allows for timely repair of the faulty intelligent agents, thus avoiding major disasters caused by the inability to monitor properly. At the same time, this mechanism can work with five intelligent agents to effectively ensure continuous and reliable monitoring of rockbursts.

[0042] Compared with the prior art, the present invention has the following advantages: 1. Accuracy and professionalism: By integrating domain-adaptive large models and knowledge bases, it can ensure that risk identification and decision-making recommendations meet mine safety standards.

[0043] 2. Good real-time and dynamic performance: The entire process from data acquisition to early warning output is automated, which can support rapid response to emergencies.

[0044] 3. Interactivity and Explainability: The user input module and knowledge-based question-answering agent provide personalized support, while the early warning graph and reports enhance the transparency of decision-making.

[0045] 4. Cost-effectiveness: Low-rank adaptive fine-tuning and modular design reduce training and maintenance costs, making it suitable for large-scale deployment.

[0046] This method achieves accurate early warning and intelligent decision-making for rockburst risks through multi-source data fusion, domain adaptive models, multi-agent collaboration, and RAG technology. It significantly improves the automation level and decision quality of disaster prevention and control, effectively addresses the complexity and high-risk issues in mine safety production, and provides strong technical support for the intelligent construction and safety management of mines.

Claims

1. A rockburst intelligent early warning and decision-making system driven by a large language model, characterized in that, It includes a data acquisition module, a data preprocessing module, a domain-adaptive large model module, a multi-agent early warning and decision-making module, and a user input module; The data acquisition module consists of multiple high-precision sensors, which are respectively arranged on multiple monitoring nodes in the area to be monitored to collect multi-source monitoring data in real time. The data preprocessing module is connected to the data acquisition module and is used to preprocess the multi-source monitoring data and extract features from the processed multi-source monitoring data to obtain feature data. The domain-adaptive large model module is connected to the data preprocessing module and is used to perform deep semantic understanding and professional reasoning based on feature data, and output multi-source monitoring data, professional semantic parsing results, and risk reasoning results. The multi-agent early warning and decision-making module is connected to the domain adaptive large model module. The multi-agent early warning and decision-making module includes an intelligent agent for source mechanism analysis and early warning, an intelligent agent for daily reports on shock hazard early warning analysis, an intelligent agent for answering questions about rockburst knowledge, an intelligent agent for analyzing special events related to rockburst risk, an intelligent agent for assisting in early warning and prevention of rockburst, and a knowledge database. These intelligent agents are interconnected, and the knowledge database is connected to each of them. The intelligent source mechanism analysis and early warning system is used to process and analyze earthquake waveform data, complete source location and moment tensor inversion, identify rupture types, and generate and output earthquake risk early warning maps and reports. The intelligent daily report system for shock hazard early warning analysis is used to process and analyze multi-source monitoring data, perform trend comparison and anomaly identification, and generate and output a daily comprehensive safety situation analysis report integrating text and graphics. The intelligent question-and-answer system for rockburst knowledge includes a retrieval enhancement generation module, used to parse regulations and standards, answer professional questions, and provide case-based operational suggestions. The intelligent system for analyzing special events of rockburst risk is used to automatically identify sudden increases in microseismic frequency and precursor signals of high-energy events, and comprehensively analyze geological and mining conditions to obtain an abnormal risk index and infer the causes of rockburst. The intelligent system for rockburst early warning and prevention and control auxiliary decision-making is used to integrate multi-source monitoring data and geological and mining environment data to comprehensively assess the risk of rockburst, and based on the assessment results, intelligently match and optimize prevention and control plans from the knowledge database, and generate and output pressure relief plans or prevention and control measures. The user input module is connected to the rockburst knowledge question-and-answer intelligent agent and is used to receive user input data.

2. The intelligent early warning and decision-making system for rockburst based on a large language model as described in claim 1, characterized in that, It also includes a result output and display module, which is connected to the multi-agent early warning and decision-making module. The module has five receiving channels and five display areas corresponding to the five receiving channels. The five receiving channels are respectively connected to the intelligent agent for source mechanism analysis and early warning, the intelligent agent for daily report of shock hazard early warning analysis, the intelligent agent for knowledge Q&A of rockburst, the intelligent agent for special event analysis of rockburst risk, and the intelligent agent for early warning and prevention and control of rockburst. The results output and display module is used to display earthquake risk early warning maps and reports in real time, daily comprehensive safety situation analysis reports integrating graphics and text, operation suggestions and preventive measures based on user questions, abnormal risk indices and causes of abnormal events, pressure relief plans or control measures. At the same time, if no data is received in any channel within a set time, the corresponding intelligent agent fault alarm information is directly displayed in the corresponding display area.

3. A rockburst intelligent early warning and decision-making system based on a large language model as described in claim 1 or 2, characterized in that, The knowledge database includes mine safety regulations data, case literature data, historical monitoring data, expert experience and academic research results data, geological and mining environment data, and historical comprehensive analysis data; the retrieval enhancement generation module is used to dynamically retrieve relevant knowledge fragments from the knowledge database based on user input data and output retrieval results data.

4. A method for intelligent early warning and decision-making of rockbursts based on a large language model, employing the intelligent early warning and decision-making system for rockbursts based on a large language model as described in any one of claims 1 to 3, characterized in that, Includes the following steps: Step 1: Construct a domain-adaptive large model module; Using a general large language model as the base model, it is trained with a high-quality domain corpus built based on knowledge distillation technology, and low-rank adaptive technology is used for parameter fine-tuning to construct a domain adaptive large model module. Step 2: Construct a multi-agent early warning and decision-making module; S21: Collect data on mine safety regulations, case literature, historical monitoring data, expert experience and academic research results, geological and mining environment data, and historical comprehensive analysis data, and use specific vectorization methods to encode the text corpus into embedded vectors to construct a knowledge database; S22: Connects the intelligent agent for source mechanism analysis and early warning, the intelligent agent for daily report of shock hazard early warning analysis, the intelligent agent for knowledge Q&A of rockburst, the intelligent agent for special event analysis of rockburst risk, the intelligent agent for rockburst early warning and prevention auxiliary decision-making, and the knowledge database. In addition, a retrieval enhancement generation module is introduced into the intelligent agent for knowledge Q&A of rockburst to construct a multi-agent early warning and decision-making module. Step 3: Construct an intelligent early warning and decision-making system for rockbursts driven by a large language model; The system connects the domain adaptive large model module and the multi-agent early warning and decision-making module. A data preprocessing module is connected to the input end of the domain adaptive large model module, and the data preprocessing module is connected to the data acquisition module. A user input module is connected to the input end of the rockburst knowledge question answering agent, and a result output and display module is connected to the output end of the multi-agent early warning and decision-making module. This constructs a rockburst intelligent early warning and decision-making system driven by a large language model. Step 4: Online intelligent early warning and decision-making for rockburst; S41: Collect multi-source monitoring data from each monitoring node through the data acquisition module; S42: The data preprocessing module performs noise reduction and standardization on the multi-source monitoring data, and extracts discriminative feature data from the processed multi-source monitoring data; S43: Using feature data as input data, it utilizes the domain-adaptive large model module for deep semantic understanding and professional reasoning, and outputs multi-source monitoring data, professional semantic parsing results, and risk reasoning results as the basis for early warning and decision-making. S44: When the user input module receives user input data, it first uses the retrieval enhancement generation module to dynamically retrieve relevant knowledge fragments from the knowledge database based on the user input data, and then outputs the retrieval results data. The search results data and the early warning and decision-making basis data are used together as input data; when the user input module does not receive user input data, the early warning and decision-making basis data are used as input data. Based on input data, the multi-agent early warning and decision-making module performs intelligent diagnosis, early warning and auxiliary decision-making for rockburst through the communication and negotiation mechanism of multi-agents. It outputs earthquake risk early warning map and report, daily safety situation comprehensive analysis report with integrated graphics and text, operation suggestions and preventive measures based on user questions, abnormal risk index and abnormal event causes, pressure relief plan or control measures to the result output and display module. Step 5: Output and Display of Results; After receiving the output result of the corresponding intelligent agent, the receiving channel of the result output and display module displays it in real time through the corresponding display area. If no output result is received by any receiving channel within a set time, the intelligent agent fault alarm is displayed in real time through the corresponding display area.

5. The intelligent early warning and decision-making method for rockburst based on a large language model as described in claim 4, characterized in that, In step one, the process of constructing the domain-adaptive large model module is as follows: S11: Collect and organize professional knowledge in the field of rockburst from a large amount of professional literature, academic papers, technical regulations, historical cases, and real-time monitoring data, and construct a structured domain corpus covering rockburst terminology, mine cases, and regulatory data. mine ; S12: Based on domain corpus D mine We extract internal knowledge representations from existing general-purpose large language models using knowledge distillation methods, and formulate knowledge transfer schemes based on domain characteristics to ultimately generate high-quality training datasets. The process of obtaining a high-quality training dataset is as follows; S12-1: Using the general large language model as the teacher model T and the model to be fine-tuned as the student model S, for the input sample x, the output probability of the teacher model T is obtained according to formula (1). The output probability of student model S is obtained according to formula (1). ; （1）； （2）； In the formula, Z T With Z S These are the logits for the teacher and student models, respectively, where T is the temperature coefficient. S12-2: Construct the distillation loss function according to formula (3) By minimizing the distillation loss function This enables the student model S to enhance its understanding of the semantics of rockburst terminology, cases, and procedures during the distillation process, and generates a high-quality training dataset. （3）； In the formula, CE is the cross-entropy loss, KL is the Kullback-Leibler divergence, and α is the weighting coefficient; S13: Freeze the original parameters of the general large language model, inject trainable low-rank matrices only next to the key modules of the model, and reduce the amount of parameter updates through low-rank matrix decomposition; use the training dataset to fine-tune the parameters of the general large language model, so that while maintaining the general capabilities, the model focuses on training its ability to understand professional terms and knowledge in the field of rockburst, and obtain a domain-adaptive large model. The parameter fine-tuning process is as follows: Freeze the original parameter matrix. Only update the low-rank matrix. and According to formula (4), the new weight matrix after LoRA fine-tuning is obtained. ; （4）； In the formula, Update the matrix for weights. , For rank, and ; S14: Evaluate the performance of the domain-adaptive large model and iteratively optimize the domain corpus, training dataset, and LoRA configuration parameters based on performance metrics.

6. The intelligent early warning and decision-making method for rockburst based on a large language model as described in claim 4, characterized in that, In step S22 of step two, the process of retrieving the answer from the enhanced generation module based on the user query is as follows: S22-1: The user's query question Transform into a query vector ; S22-2: In the vector space, calculate the query vector according to formula (5). With each text vector Cosine similarity between ; （5）； S22-3: Based on the calculated similarity score, filter out the top results most relevant to the current query from the entire knowledge database. document fragments Use contextual information as input to generate the model; S22-4: Calculate d for each document fragment i For a given query problem weight And obtain the probability of the final answer y according to formula (6); （6）； In the formula, For a given user query and the entire collection of retrieved documents Under the given conditions, the overall probability of generating the final answer; S22-5: Obtain the final answer according to formula (7) ; （7）。 7. The intelligent early warning and decision-making method for rockburst based on a large language model as described in claim 4, characterized in that, In step S44 of step four, during the generation of earthquake risk warning maps and reports, the intelligent agent for source mechanism analysis and early warning obtains the warning level by analyzing the source mechanism through earthquake waveform inversion. The process is as follows: S44-1: Obtain the seismic moment tensor according to formula (8) ; （8）； In the formula, This refers to the rock mass shear modulus. For the sliding displacement component, For fault normal components; S44-2: Based on Seismic Moment Tensor The seismic energy release rate is calculated according to formula (9) based on the principal stress direction. ; （9）； S44-3: Set a low energy threshold and high energy threshold The warning level is determined according to formula (10). ; （10）。 8. The intelligent early warning and decision-making method for rockburst based on a large language model as described in claim 4, characterized in that, In step S44 of step four, during the process of generating the daily comprehensive analysis report of the security situation integrating graphics and text, the intelligent agent of the impact hazard early warning analysis daily report analyzes the trend changes through time series data, and first obtains the sliding window average value according to formula (11). Then, obtain the outliers according to formula (12). And based on outliers Perform anomaly detection; if it occurs consecutively... If so, the alarm mechanism will be triggered; （11）； （12）； In the formula, For threshold coefficient, The standard deviation is denoted as .

9. The intelligent early warning and decision-making method for rockburst based on a large language model as described in claim 4, characterized in that, In step S44 of step four, during the process of obtaining the abnormal risk index, the rockburst risk special event analysis agent first obtains the frequency according to formula (13). and with a set frequency threshold If a comparison is made, If so, it is judged as an anomaly, and then combined with geological parameters. and energy data The abnormal risk index is obtained according to formula (14). ; （13）； （14）； In the formula, α and β are two different weighting coefficients.

10. The intelligent early warning and decision-making method for rockburst based on a large language model as described in claim 4, characterized in that, In step S44 of step four, during the process of generating a pressure relief plan or prevention and control measures, the rockburst early warning and prevention and control auxiliary decision-making agent first performs multi-source data fusion according to formula (15) to obtain a comprehensive risk index. Then, the comprehensive risk index Compare with the set risk threshold ,like The pressure relief parameter adjustment ratio is obtained according to formula (16). ; （15）； （16）； In the formula, For seismic activity data; It is in a state of stress; Energy data; These are the weighting coefficients for seismic activity data; These are the weighting coefficients for the stress state; These are the weighting coefficients for the energy data; For safety factor; This is the expected energy value; This represents the current monitored energy value.

Images