Rock burst intelligent early warning and prevention system and method based on large language model

Abstract

The invention discloses an intelligent rock burst early warning and prevention system and method based on a large language model, and the system is characterized in that a rock burst monitoring and early warning module comprises a high-precision sensor group, a rock burst prediction module and a rock burst danger level grading module, and is used for obtaining a grading result of a rock burst danger level; the anti-scour pressure relief measure self-adaptive optimization decision module is used for outputting optimal pressure relief measures and pressure relief parameters; the man-machine interaction module based on the large language model is used for performing engineering interpretation and man-machine interaction; the method comprises the following steps: predicting a rock burst danger early warning result by using the rock burst prediction module based on preorder multi-mode monitoring data; obtaining a grading result of an impact danger grade by utilizing a rock burst danger grade grading module; outputting an optimal pressure relief measure and a pressure relief parameter by using an anti-scour pressure relief measure self-adaptive optimization decision module; and a man-machine intelligent interaction module based on a large language model is used for providing intelligent decision support. According to the system and the method, the rock burst prevention and control scheme can be accurately and intelligently realized in advance.

Description

Technical Field

[0001] The present invention belongs to the cross - technical field of mine safety engineering and artificial intelligence, and particularly relates to a rock burst intelligent early warning and prevention system and method based on a large language model. Background Technique

[0002] As the coal mining depth in China has exceeded the 1000 - meter mark, the in - situ stress environment borne by the roadway surrounding rock has generally exceeded 30 MPa, resulting in an annual growth rate of 18.7% in the accidents of surrounding rock instability in deep mines in the past five years. In mines with a tendency to rock burst, traditional roadway stability assessment methods face three major technical bottlenecks: First, the lack of a multi - source data fusion mechanism. Key parameters such as geological structures, microseismic sequences, and support states are still in an isolated analysis state. There is a lack of intelligent algorithms to analyze the non - linear correlation characteristics of multi - source data. Existing AI prediction models are difficult to interpret the analysis results, which is difficult for mine - related management and maintenance personnel to understand. Second, the decision - making of pressure relief measures dominated by manual experience lags seriously. The current assessment mode still relies heavily on manual experience for judgment to a large extent, and it is unable to effectively capture millisecond - level micro - fracture signals (information loss rate of 28%), resulting in the proposed pressure relief measures being difficult to solve on - site dangerous situations in a timely and accurate manner. Third, the existing early warning and prevention systems lack interaction capabilities. Most current early warning systems are passive early warnings, unable to answer technicians' questions based on natural language interaction and optimize themselves according to technicians' conversations.

[0003] Therefore, there is an urgent need to provide a surrounding rock state evaluation system for multi - scale data fusion and a decision - support platform with engineering interpretability, establish an adaptive optimization mechanism for rock burst prevention and support parameters, and form a closed - loop control system of "monitoring - evaluation - regulation" to effectively ensure the safe production operation of mines. Summary of the Invention

[0004] Aiming at the problems existing in the above - mentioned prior art, the present invention provides a rock burst intelligent early warning and prevention system and method based on a large language model. The system has a simple structure and a high degree of intelligence. It can intelligently classify and predict the risk levels of rock bursts based on multi - source monitoring data collected in real - time, and can output the optimal pressure relief measures and pressure relief parameters based on the classification and prediction results of rock burst risk levels. At the same time, it can give an engineering explanation for the output pressure relief measures and pressure relief parameters, and can answer technicians' questions based on natural language interaction. The method has a high degree of intelligence and ideal interaction performance, and can realize a more accurate, intelligent, advanced, interpretable and self - optimizing rock burst prevention and control plan.

[0005] The present invention provides a rock burst intelligent early warning and prevention system based on a large language model, including a rock burst monitoring and early warning module, an adaptive optimization decision - making module for rock burst prevention and pressure relief measures, and a human - machine interaction module based on a large language model;

[0006] The rock burst monitoring and early warning module includes a high-precision sensor group, a rock burst prediction module, and a rock burst danger level classification module;

[0007] The high-precision sensor group includes a variety of high-precision sensors, which are respectively installed on multiple monitoring nodes in the mine and are used to collect multi-source monitoring data in real time; the rock burst prediction module is used to predict the rock burst danger level based on the multi-source monitoring data and output the rock burst danger early warning result; the rock burst danger level classification module is used to obtain the impact danger degree based on the static data of the geological structure and mining information during the mining period, and is used to obtain the classification result of the rock burst danger level based on the impact danger degree and the rock burst danger early warning result;

[0008] The self-adaptive optimization decision-making module for anti-burst pressure relief measures is used to dynamically optimize the decision-making of the current impact danger level pressure relief measures based on the classification result of the impact danger level and output the optimal pressure relief measures and pressure relief parameters;

[0009] The human-computer interaction module based on the large language model is used to receive the optimal pressure relief measures and pressure relief parameters, and conduct engineering interpretations on the rock burst prevention and control professional terms in the optimal pressure relief measures and pressure relief parameters, and output and display the optimal pressure relief measures and pressure relief parameters and the interpretation information; at the same time, when there is technical consultation information input by the user, it is used to retrieve relevant contexts based on the mine anti-burst knowledge base, then form intelligent decision-making information based on the retrieval results and the technical consultation information, and finally output and display the intelligent decision-making information.

[0010] As an optimization, the variety of high-precision sensors includes microseismic sensors, ground sound sensors, rock mass stress sensors, and electromagnetic sensors.

[0011] In the present invention, by installing a variety of high-precision sensors on multiple monitoring nodes in a mine, it is convenient to collect multi-source monitoring data in real time, and then it is convenient to provide multi-angle information for the prediction of the risk level of rock burst, which is beneficial to ensuring the accurate prediction of the rock burst risk warning result. Through the setting of the rock burst prediction module, it is convenient to use the built-in rock burst prediction model to efficiently and accurately predict the rock burst warning result. Through the setting of the rock burst risk level grading module, it is convenient to independently evaluate the degree of impact risk, and can combine the rock burst prediction module to output the rock burst risk level grading result. In this way, the comprehensive prediction of the degree of rock burst risk can be realized. Through the setting of the anti-rock burst pressure relief measure adaptive optimization decision-making module, it is convenient to output the optimal pressure relief measure and pressure relief parameters based on the grading result of the impact risk level. Through the setting of the human-computer interaction module based on the large language model, not only can it facilitate the engineering interpretation of the optimal pressure relief measure and pressure relief parameters, so that relevant technical personnel can better and more clearly understand the optimal pressure relief measure and pressure relief parameters, but also it can answer the questions of technical personnel based on natural language interaction, significantly improving the interaction ability of the system.

[0012] The system has a simple structure and high intelligence. It can intelligently classify and predict the risk level of rock burst based on the multi-source monitoring data collected in real time, and can output the optimal pressure relief measure and pressure relief parameters based on the grading prediction result of the rock burst risk level. At the same time, it can conduct engineering interpretation on the output pressure relief measure and pressure relief parameters, and can answer the questions of technical personnel based on natural language interaction.

[0013] The present invention also provides a rock burst intelligent warning and prevention method based on a large language model, including the following steps:

[0014] Step 1: Construct a rock burst monitoring and warning module, and use the rock burst monitoring and warning module to obtain the grading result of the impact risk level during mining;

[0015] A1: Construct a multi-modal data set;

[0016] a11: Use a variety of high-precision sensors installed on multiple monitoring nodes in multiple different mines to collect multi-modal signals in historical rock burst events, and obtain original multi-modal data based on the multi-modal signals;

[0017] a12: Preprocess the original multi-modal data to construct a precursor pattern sequence, and through experts, according to the characteristics of different application mining areas and the actual risk situation of rock burst events, convert the precursor pattern sequence into the corresponding grade form, divide the corresponding rock burst risk level labels, and then construct a massive multi-modal data set through the precursor pattern sequence and the corresponding labels. Then, divide the multi-modal data set into a training set, a test set, and a validation set according to a set ratio;

[0018] A2: Construct a rock burst prediction module, and use the rock burst prediction module to predict the results of rock burst danger early warning;

[0019] a21: Construct an initial rock burst prediction model based on the CNN-LSTM architecture, train the initial rock burst prediction model using the training set, and optimize the model by minimizing the error. After training, obtain the rock burst prediction model;

[0020] a22: Use a variety of high-precision sensors installed on multiple monitoring nodes in the mine to collect real-time pre-order multi-modal monitoring data during mining;

[0021] a23: Take the pre-order multi-modal monitoring data as input data, input it into the rock burst prediction model, use the rock burst prediction model for prediction, and output the rock burst danger early warning result W p ;

[0022] A3: Construct a rock burst danger level classification module, and use the rock burst danger level classification module to obtain the classification results of the impact danger level;

[0023] a31: Based on the static data of the geological structure and mining information during mining, use the comprehensive index method to analyze the impact danger and obtain the impact danger degree W q ;

[0024] a32: Use the CRITIC algorithm to dynamically divide the two index weights of the rock burst danger early warning result W p and the impact danger degree W q to obtain the rock burst danger degree RL of the prediction time period, and obtain the classification results of the impact danger level based on the rock burst danger degree RL;

[0025] Step 2: Construct an adaptive optimization decision-making module for rock burst prevention and pressure relief measures, and use the adaptive optimization decision-making module for rock burst prevention and pressure relief measures to output the optimal pressure relief measures and pressure relief parameters;

[0026] Based on the classification results of the real-time input impact danger level, adopt a reinforcement learning model and an adaptive optimization mechanism, according to the preset constraint conditions, and based on the early warning information of rock burst events and the corresponding prevention and control measure data stored in the historical database, dynamically realize the optimization decision of the pressure relief measures for the current impact danger level, and optimize the pressure relief parameters. Finally, output the optimal pressure relief measures and pressure relief parameters according to the current rock burst danger level; As a preference;

[0027] Meanwhile, after each pressure relief prevention and control measure is implemented, multi-source monitoring data for multiple construction cycles of the pressure relief prevention and control measure is collected, and based on the multi-source monitoring data, the implementation effect of the current pressure relief prevention and control measure is evaluated. Then, based on the evaluation result, the prevention and control measure data in the historical database is modified to obtain the optimal pressure relief measure through adaptive adjustment;

[0028] Step 3: Construct a human-machine intelligent interaction module based on the large language model, and use the human-machine intelligent interaction module based on the large language model to output the intelligent decision-making result of the pressure relief measure;

[0029] C1: Construct a natural language interaction core based on the Qwen2.5 model fine-tuned by LoRA, and introduce the RAG technology to retrieve relevant technical documents, regulations, and case data from the mine rock burst prevention knowledge base built locally as the generation basis to form a human-machine intelligent interaction module;

[0030] C2: Based on the real-time input optimal pressure relief measure and pressure relief parameters, use the Qwen2.5 model fine-tuned by LoRA to accurately analyze the optimal pressure relief measure and pressure relief parameters, and conduct engineering interpretations on the rock burst prevention and control professional terms in the optimal pressure relief measure and pressure relief parameters, and output and display the optimal pressure relief measure and pressure relief parameters as well as the interpretation information;

[0031] When there is technical consultation information input by the user, first extract the key intent and entity information in the technical consultation information through parsing, then match relevant document fragments from the mine rock burst prevention knowledge base through semantic similarity calculation. Then, the most relevant context retrieval results are screened out by the re-ranking module, and then the retrieval results and the technical consultation information input by the user are jointly input into the Qwen2.5 model fine-tuned by LoRA for the inference generation of context-aware answers to form intelligent decision-making information. Finally, the intelligent decision-making information is output and displayed.

[0032] Furthermore, in order to ensure high prediction accuracy, in a12 of Step 1, the construction process of the precursor pattern sequence is as follows:

[0033] a12-1: Use a specific time window to process the multi-source monitoring data set D into j time window multi-source monitoring data sets, and statistically calculate the maximum energy e of the microseismic data in each time window max , average energy e avg , variance energy e σ , frequency f, online maximum stress value σ max and the maximum value of drill cuttings and obtain the kth time window multi-source monitoring data set D according to formula (1) k ;

[0034]

[0035] a12-2: Construct the precursor pattern sequence P based on the multi-source monitoring dataset D, as shown in formula (2), where the a-th precursor pattern sequence is p a As shown in formula (3);

[0036] P = [p0, p1,..., p l (2);

[0037] p a = [D a×b , D a×b+1 , D a×b+2 ,..., D a×b+c-1 (3);

[0038] In the formula, l is the number of precursor pattern sequences, b is the sampling step size, and c is the sequence length.

[0039] Furthermore, to ensure that the structure of the rock burst prediction module is relatively simple and, at the same time, to ensure the prediction accuracy, in a21 of step one, the structure of the initial rock burst prediction model based on the CNN-LSTM architecture is as follows:

[0040] The first layer is the input layer, which is used to receive the input data and input it to the second layer;

[0041] The second layer is the CNN layer, which is used to perform convolution, pooling, and activation function processing on the input data in sequence. By sliding the one-dimensional convolution kernel along the time axis, the local spatial correlation features in the time series data are extracted, and then the data is unfolded into one-dimensional data by using the unfolding operation;

[0042] The third layer is a combined layer of multiple LSTMs, which is used to extract the features of the one-dimensional data output by the second layer through a triple gating mechanism;

[0043] The fourth layer is the attention mechanism layer, which is used to enhance the key time step information output by the third layer through dynamic weight allocation and improve the focusing ability on the microseismic time series features;

[0044] The fifth layer is the output layer, which uses a fully connected layer to learn the distribution features and realizes the output of the rock burst danger warning result.

[0045] Furthermore, to accurately obtain the degree of rock burst danger, in a31 of step one, the degree of rock burst danger W is obtained according to formula (4) q ;

[0046]

[0047] In the formula, W geologe is the geological structure data, which includes the historical number of rock bursts W1 in the same coal seam 1 , the current actual mining depth W1 2, the distance W1 between the hard thick rock stratum in the overlying fissure zone and the coal seam 3 , the characteristic parameter W1 of the roof rock stratum thickness 4 , the degree of tectonic stress concentration in the mining area W1 5 , the uniaxial compressive strength of coal W1 6 and the elastic energy index of coal W1 7 ; W mine is mining information, which includes the pressure relief degree W2 of the protective layer 1 , the horizontal distance W2 between the working face and the coal pillar left by the mining of the upper protective layer 2 , the relationship with the adjacent goaf W2 3 , the working face length W2 4 , the width of the sectional coal pillar W2 5 , the thickness of the coal left at the bottom W2 6 , the distance from the goaf when driving towards the goaf W2 7 , the distance from the goaf when advancing towards the goaf W2 8 , the distance from the fault W2 9 , the distance from the fold W2 10 and the distance from the coal seam phase transition zone W2 11 .

[0048] Further, in order to accurately and efficiently obtain the result of the rock burst danger level, in a32 of step one, the rock burst danger degree RL in the prediction time period is obtained, and the calculation process of obtaining the classification result of the rock burst danger level based on the rock burst danger degree RL is as follows:

[0049] a32-1: Normalize the data according to formula (5) to eliminate the dimension difference;

[0050]

[0051] Wherein, x ij is the jth index of the qth sample, μ j , σ j are the mean value and variance respectively;

[0052] a32-2: Calculate the standard deviation S according to formula (6) j , and measure the volatility of each index through the standard deviation to reflect the discrimination ability of each index for the overall data;

[0053]

[0054] a32-3: Calculate the correlation coefficient R according to formula (7) j , and use the correlation coefficient R j to evaluate the redundancy between indicators to avoid repeated weighting of highly correlated indicators;

[0055]

[0056] In the formula, r jk is the Pearson correlation coefficient between index j and index k;

[0057] a32 - 4: Combine the contrast intensity and conflict, and calculate the comprehensive weight α according to formula (8) j ;

[0058]

[0059] In the formula, C j is the comprehensive information score of index j, and C j = S j × R j ;

[0060] a32 - 5: Calculate the rock burst danger level RL during the prediction period according to formula (9);

[0061]

[0062] a32 - 6: First, normalize the rock burst danger level RL to the interval [0, 1], and then obtain the classification result of the rock burst danger level according to the value range of the normalized RL.

[0063] Furthermore, in order to accurately evaluate the implementation effect of the current pressure relief prevention and control measures in a quantitative way, in step two, the specific process of evaluating the implementation effect of the current pressure relief prevention and control measures is as follows:

[0064] B1: Real - time collect multi - source monitoring data of multiple construction cycles of the pressure relief measures, and perform standardization processing on the monitoring data using Z - score. Obtain the Z - score standard value z i ;

[0065]

[0066] In the formula, μ is the average value within the collection time window of each type of monitoring data, ξ is the standard deviation within the collection time window, which is used to reflect the degree of data dispersion,

[0067] B2: Based on the multi - source monitoring data that has been standardized, obtain the microseismic event energy release rate E rate , obtain the spatial density f rate according to formula (7), obtain the stress change rate ▽σ according to formula (8), obtain the critical value S of the drill cuttings amount critical according to formula (9), and obtain the change rate ΔS of the drill cuttings amount according to formula (10);

[0068]

[0069] Wherein, E i is the energy of the microseismic events in the monitoring area within the time window, and T is the time window;

[0070]

[0071] Wherein, f is the frequency of microseismic events within the time window, and S is the monitoring area within the time window;

[0072]

[0073] Wherein, σ pre , σ post are the stress values before and after pressure relief respectively, and L is the spacing between measurement points;

[0074] S critical = K·S0 (9);

[0075] Wherein, S0 is the reference value of the drill cuttings amount under the normal stress of the coal body, and K is the geological influence coefficient, taking 1.5 - 2.0;

[0076]

[0077] B3: Obtain the pressure relief effect evaluation index P of the current pressure relief measure according to formula (11) score ;

[0078]

[0079] Wherein, α, β, χ, and δ are the weight indexes of the microseismic event energy, spatial density, stress value, and drill cuttings amount respectively;

[0080] B4: Normalize the pressure relief effect evaluation index P score to the interval [0, 1], and then evaluate the effect of the pressure relief measure according to the value range of the normalized P score .

[0081] As an optimization, in a12 of step one, the multi-modal data set is divided into a training set, a test set, and a validation set according to a ratio of 8:1:1.

[0082] The present invention provides a method for intelligent early warning and prevention of rock bursts based on large language models. First, multi-modal data obtained from historical rock burst events is used, and then a precursor pattern sequence is constructed based on the multi-modal data. At the same time, the rock burst danger level labels of the precursor pattern sequence are divided by experts, and then a large amount of multi-modal data sets are constructed based on the precursor pattern sequence and the corresponding labels. On this basis, the initial rock burst prediction model based on the CNN-LSTM architecture is trained to obtain a rock burst prediction model with high prediction accuracy and high prediction efficiency. Furthermore, in the actual application process, only a shorter acquisition time is required, and shorter pre-order multi-modal monitoring data than that required for conventional early warning can be obtained to efficiently and accurately obtain the early warning result of rock burst danger. On this basis, the rock burst danger degree is independently evaluated by combining the geological structure data and mining information data during mining, and the contribution ratios of the rock burst danger early warning result and the rock burst danger degree to the comprehensive index are allocated through a weight division method, so that the grading result of the rock burst danger level can be efficiently and accurately obtained. Thus, the danger of rock bursts can be predicted and discovered earlier than traditional early warning systems and methods, which is conducive to taking early pressure relief prevention measures earlier and greatly improving the safety factor of coal mining operations. Based on the grading result of the real-time obtained rock burst danger level, a reinforcement learning model and an adaptive optimization mechanism are adopted. According to the preset constraints and based on the early warning information and corresponding prevention measure data of rock burst events stored in the historical database, the optimization decision of the current pressure relief measure for the rock burst danger level is dynamically realized, and the pressure relief parameters are optimized, which can ensure the timeliness and accuracy of the pressure relief measures. At the same time, it can ensure that the occurrence probability of rock burst accidents is minimized. At the same time, after each pressure relief prevention measure is executed, the implementation effect of the current pressure relief prevention measure is evaluated, and the prevention measure data in the historical database can be adaptively adjusted. Thus, the data in the historical database can be corrected and updated in real time, which is conducive to obtaining the optimal pressure relief measures and pressure relief parameters in each decision-making process. At the same time, a closed-loop control strategy of monitoring-evaluation-regulation is realized. A human-machine intelligent interaction module is constructed based on the Qwen 2.5 model fine-tuned by LoRA, and the RAG technology is introduced. Not only can the model have the ability to accurately understand the professional terms and engineering scenarios of rock burst prevention and control, but also after obtaining the optimal pressure relief measures and pressure relief parameters of the real-time input, it can carry out engineering interpretations, so that relevant personnel can more clearly and thoroughly understand the execution details of the pressure relief measures, which is conducive to ensuring the accurate implementation of the pressure relief measures and further ensuring the safety of coal mining operations. At the same time, relevant technical documents can also be retrieved from the mine rock burst prevention knowledge base constructed locally through the RAG technology as the basis for generating consultation answers, and at the same time, it is ensured that the output answer content meets both technical specifications and language naturalness, thus realizing intelligent decision support with both professional accuracy and natural interaction experience.

[0083] This method has a high degree of intelligence and ideal interaction performance. By introducing advanced multi-modal data fusion technology, the parsing ability of the Qwen 2.5 model with LoRA fine-tuning, and an adaptive optimization decision-making mechanism, it solves the technical bottleneck that most current rock burst warning systems are passive warnings, unable to achieve early warning, unable to answer technicians' questions based on natural language interaction, and unable to self-optimize according to technicians' conversations. Using this method can achieve early evaluation of surrounding rock conditions with multi-scale data fusion and intelligent decision support with engineering interpretability, and establish an adaptive optimization mechanism for rock burst prevention and support parameters, forming a closed-loop control process of "monitoring - evaluation - regulation" to effectively ensure the safe production operation of the mine and can achieve a more accurate, intelligent, early, interpretable, and self-optimized rock burst prevention and control plan. BRIEF DESCRIPTION OF THE DRAWINGS

[0084] Figure 1 is a schematic structural diagram of the rock burst prediction model in the present invention;

[0085] Figure 2 is a schematic structural diagram of the human-machine intelligent interaction module based on the large language model in the present invention;

[0086] Figure 3 is a schematic block diagram of the system part in the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0087] The present invention will be further described below with reference to the accompanying drawings.

[0088] As Figures 1 to 3 shown, the present invention provides a rock burst intelligent warning and prevention system based on a large language model, including a rock burst monitoring and warning module, an adaptive optimization decision-making module for rock burst prevention and pressure relief measures, and a human-machine interaction module based on a large language model;

[0089] The rock burst monitoring and warning module includes a high-precision sensor group, a rock burst prediction module, and a rock burst danger level classification module;

[0090] The high-precision sensor group includes a variety of high-precision sensors, which are respectively installed on multiple monitoring nodes in the mine for real-time collection of multi-source monitoring data; the rock burst prediction module is used to predict the rock burst danger level based on multi-source monitoring data and output a rock burst danger warning result; the rock burst danger level classification module is used to obtain the impact danger degree based on the static data of geological structures and mining information during mining, and is used to obtain the classification result of the rock burst danger level based on the impact danger degree and the rock burst danger warning result;

[0091] The anti-bumping pressure relief measure adaptive optimization decision-making module is used to dynamically optimize and make decisions on the current bumping risk level pressure relief measures based on the classification results of the bumping risk level, and output the optimal pressure relief measures and pressure relief parameters;

[0092] The human-machine interaction module based on the large language model is used to receive the optimal pressure relief measures and pressure relief parameters, and conduct engineering interpretations on the bumping prevention and control professional terms in the optimal pressure relief measures and pressure relief parameters, output and display the optimal pressure relief measures and pressure relief parameters as well as the interpretation information; meanwhile, when there is technical consultation information input by the user, it is used to retrieve relevant contexts based on the mine bumping prevention knowledge base, and then form intelligent decision-making information based on the retrieval results and the technical consultation information, and finally output and display the intelligent decision-making information.

[0093] As a preference, the multiple high-precision sensors include microseismic sensors, ground sound sensors, rock mass stress sensors, and electromagnetic sensors. In this way, it is convenient to collect microseismic energy quantity, ground sound data, stress data, and electromagnetic data. Through the high-frequency collection of multi-dimensional data, it can provide multi-angle information for the accurate prediction of bumping.

[0094] In the present invention, by installing multiple high-precision sensors at multiple monitoring nodes in the mine, it is convenient to collect multi-source monitoring data in real time, and then it is convenient to provide multi-angle information for the prediction of the bumping risk level, which is beneficial to ensuring the accurate prediction of the bumping risk warning result. Through the setting of the bumping prediction module, it is convenient to use the built-in bumping prediction model to efficiently and accurately predict the bumping warning result. Through the setting of the bumping risk level classification module, it is convenient to independently evaluate the bumping risk degree, and combine with the bumping prediction module to output the bumping risk level classification result. In this way, the comprehensive prediction of the bumping risk degree can be realized. Through the setting of the anti-bumping pressure relief measure adaptive optimization decision-making module, it is convenient to output the optimal pressure relief measures and pressure relief parameters based on the classification result of the bumping risk level. Through the setting of the human-machine interaction module based on the large language model, not only can it facilitate the engineering interpretation of the optimal pressure relief measures and pressure relief parameters, so that relevant technical personnel can better and more clearly understand the optimal pressure relief measures and pressure relief parameters, but also it can answer the questions of technical personnel based on natural language interaction, significantly improving the interaction ability of the system.

[0095] This system has a simple structure and a high degree of intelligence. It can intelligently classify and predict the bumping risk level based on the real-time collected multi-source monitoring data, and can output the optimal pressure relief measures and pressure relief parameters based on the classification prediction result of the bumping risk level. At the same time, it can conduct engineering interpretations on the output pressure relief measures and pressure relief parameters, and can answer the questions of technical personnel based on natural language interaction.

[0096] The present invention also provides a method for intelligent early warning and prevention of rock bursts based on large language models, including the following steps:

[0097] Step 1: Construct a rock burst monitoring and early warning module, and use the rock burst monitoring and early warning module to obtain the classification results of the rock burst danger levels during mining and excavation;

[0098] A1: Construct a multimodal dataset;

[0099] a11: Collect multimodal signals in historical rock burst events using a variety of high-precision sensors installed on multiple monitoring nodes in multiple different mines, and obtain the original multimodal data based on the multimodal signals;

[0100] a12: Preprocess the original multimodal data to construct a precursor pattern sequence, and through experts, according to the characteristics of different application mining areas and the actual danger of rock burst events, convert the precursor pattern sequence into the corresponding grade form, divide the corresponding rock burst danger level labels, and then construct a massive multimodal dataset through the precursor pattern sequence and the corresponding labels. Then, divide the multimodal dataset into a training set, a test set, and a validation set according to a set ratio for the training of the subsequent rock burst prediction model. Among them, the construction criteria for the rock burst danger level labels are shown in Table 1, and the specific values can also be adaptively modified according to the specific situation of the coal mine;

[0101] Table 1: Construction criteria for rock burst danger level labels

[0102]

[0103] A2: Construct a rock burst prediction module, and use the rock burst prediction module to predict the rock burst danger warning results;

[0104] a21: Construct an initial rock burst prediction model based on the CNN-LSTM architecture, use the training set to train the initial rock burst prediction model, and optimize the model by minimizing the error. After training, obtain the rock burst prediction model;

[0105] a22: Use a variety of high-precision sensors installed on multiple monitoring nodes in the mine to collect real-time pre-sequence multimodal monitoring data during mining and excavation;

[0106] a23: Use the pre-sequence multimodal monitoring data as input data, input it into the rock burst prediction model, and use the rock burst prediction model to predict, output the rock burst danger probability result, and obtain the rock burst danger warning result W p ;

[0107] A3: Construct a classification module for the danger level of rock burst, and use the classification module for the danger level of rock burst to obtain the classification result of the rock burst danger level;

[0108] a31: Based on the static data of geological structures and mining information during mining, use the comprehensive index method to analyze the rock burst danger and obtain the rock burst danger level W q , because, the rock burst danger level can be independently evaluated;

[0109] a32: To further comprehensively evaluate the danger level of rock burst, through the weight division method, divide the contribution ratio of the rock burst danger early warning result W p and the rock burst danger level W q in the comprehensive index, so as to achieve the comprehensive prediction of the danger level; specifically, use the CRITIC algorithm to dynamically divide the rock burst danger early warning result W p and the rock burst danger level W q of the two index weights. In this way, through the method of comparing the intensity and conflict of comprehensive indexes, the adaptive adjustment of the weights between the two indexes of multi-source monitoring dynamic data and geological and mining static data is realized, and then the rock burst danger level RL of the prediction time period is obtained. And based on the rock burst danger level RL, the classification result of the rock burst danger level is obtained;

[0110] Step 2: Construct an adaptive optimization decision module for anti-rock burst pressure relief measures, and use the adaptive optimization decision module for anti-rock burst pressure relief measures to output the optimal pressure relief measures and pressure relief parameters;

[0111] During the actual operation of the mine, in order to realize the optimization decision of pressure relief measures and be able to respond to environmental changes and changes in the new rock burst danger level in real time, continuously update the rock burst danger level based on real-time data feedback, adjust the anti-rock burst pressure relief measures according to the new rock burst danger level, and at the same time, the reinforcement learning model and the adaptive optimization mechanism can be dynamically optimized according to new data in each mine operation. The specific process is as follows:

[0112] Based on the classification results of the impact hazard level of real-time input, using a reinforcement learning model and an adaptive optimization mechanism, according to the preset constraint conditions, and based on the early warning information of rock burst events stored in the historical database and the corresponding prevention and control measure data, using reinforcement learning (RL) or optimization algorithms (such as genetic algorithms, particle swarm optimization, etc.) to dynamically realize the optimization decision of the pressure relief measures for the current impact hazard level, and optimize the pressure relief parameters. Finally, according to the current rock burst hazard level, the optimal pressure relief measures and pressure relief parameters are output; as an option, when the classification result of the impact hazard level is no rock burst hazard, no measures are taken; when the classification result of the impact hazard level is weak rock burst hazard, the method of large-diameter borehole pressure relief is adopted to prevent the local concentration of stress to maintain the stability of coal and rock, and the spacing and length of the large-diameter boreholes are determined; when the classification result of the impact hazard level is medium rock burst hazard, the methods of large-diameter borehole pressure relief, roof pre-splitting blasting or coal body pressure relief blasting are adopted to reduce the concentration degree of high-stress areas to effectively block the stress transmission path, and the spacing and length of the large-diameter boreholes, the coupling coefficient and angle of roof blasting, the spacing and charge amount of coal body pressure relief blasting holes, etc. are determined; when the classification result of the impact hazard level is strong rock burst hazard, a combination of large-diameter borehole pressure relief, roof pre-splitting blasting and coal body pressure relief blasting is adopted to quickly release the accumulated energy and reduce the stress concentration degree to avoid the occurrence of sudden dynamic disasters, and the spacing and length of the large-diameter boreholes, the coupling coefficient and angle of roof blasting, the spacing and charge amount of coal body pressure relief blasting holes, etc. are determined;

[0113] Among them, the early warning information of rock burst events stored in the historical database and the corresponding prevention and control measure data include the pressure relief measures, support schemes and their effects taken during each impact. The constraint conditions are set according to factors such as the safety standards of the mine, equipment capabilities, and personnel operation restrictions. As an option, in order to ensure the timeliness and accuracy of the pressure relief measures, the maximum safety index of the mine during rock burst occurrence, the minimum response time of the prevention and control measures, and the maximum pressure relief effect and support stability can be defined. The main goal is to minimize the occurrence probability of rock burst accidents;

[0114] At the same time, after each pressure relief prevention and control measure is executed, multi-source monitoring data for multiple construction cycles of the pressure relief prevention and control measure are collected (if the pressure relief construction time is T, the time window data with a sampling step of 3T are collected, including the monitoring data within time T before and after the implementation of the pressure relief measure), such as microseismic energy e, frequency f, stress value σ, and drill cuttings Based on multi-source monitoring data, evaluate the implementation effect of the current pressure relief prevention and control measures, and then modify the prevention and control measure data in the historical database based on the evaluation results, so as to obtain the optimal pressure relief measures through adaptive adjustment. In this way, it can ensure that under different impact hazard levels, the output pressure relief measures and pressure relief parameters can maintain the optimal configuration;

[0115] In this way, by combining all data feedback and decision-making results, a closed-loop control decision-making process of "monitoring - evaluation - regulation" can be realized.

[0116] Step 3: Use the open-source model Qwen2.5 as the base model to construct a human-machine intelligent interaction module based on the large language model, and use the human-machine intelligent interaction module based on the large language model to output the intelligent decision-making results of the pressure relief measures;

[0117] Qwen2.5 is the latest generation of large language model (LLM) developed by the Qwen team, aiming to support diverse artificial intelligence tasks through powerful language understanding, generation, and reasoning capabilities. This model uses a high-quality dataset of up to 1.8 trillion multilingual tokens during the pre-training stage, which is a significant improvement compared to the 700 billion tokens of the previous model Qwen2, covering multiple fields such as common sense, professional knowledge, mathematics, and programming. Qwen2.5 provides versions with various parameter scales, including open-weight models such as 0.5B, 1.5B, 3B, 7B, 14B, 32B, and 72B, as well as proprietary models Qwen2.5-Turbo and Qwen2.5-Plus based on Mixture of Experts (MoE). This method selects the Qwen2.5-7B version as the base model for fine-tuning to balance computational efficiency and performance requirements, suitable for the real-time requirements of the intelligent early warning and prevention and control system for rock burst.

[0118] Qwen2.5 continues the decoder architecture based on Transformer and incorporates multiple technical optimizations to improve performance. Among them, Grouped Query Attention (GQA) is the core optimization mechanism. By grouping queries and sharing key-value pairs, it significantly improves the utilization efficiency of the key-value cache (KV Cache). While retaining the ability of the attention mechanism to capture long-range dependencies, GQA makes the model perform excellently in processing complex long-sequence tasks by reducing redundant calculations and memory occupancy. Its calculation formula is: GQA(Q, K, V) = Concat(head1, head2, …, head G )W O , where for each group g (a total of G groups), the calculation method is: Among them, Q g is the query matrix of the g-th group; K is the shared key matrix; V is the shared value matrix; d kis the dimension of the key, used to scale the dot product to prevent the vanishing gradient problem; W o is the output projection matrix.

[0119] In addition, Qwen 2.5 adopts the SwiGLU (Switched Gated Linear Unit) activation function in the architecture design to significantly enhance the model's non-linear expression ability. Compared with traditional activation functions (such as ReLU or GELU), SwiGLU can more flexibly capture the complex relationships between input features by introducing a gating mechanism and the dynamic adjustment characteristics of the Swish function, thereby improving the model's performance in language understanding, generation, and reasoning tasks. In the intelligent early warning and prevention of rock bursts, the non-linear modeling ability of SwiGLU helps to more accurately analyze the potential patterns in ground pressure data and improve the generation accuracy of early warning signals and the reasoning effect of prevention strategies. Its calculation formula is: SwiGLU(x) = σ(Swish(xW1 + b1)) ⊙ (xW2 + b2), where the Swish function is defined as: Swish(z) = z·σ(z), x is the input vector with dimension (N, d in ); W1, W2 are two independent linear transformation weight matrices with dimensions (d in , d hidden ), respectively, and b1, b2 are the corresponding bias vectors with dimension (d hidden ); is the sigmoid function for gating; ⊙ represents element-wise multiplication (Hadamard product); d in is the input layer dimension, and d hidden is the hidden layer dimension.

[0120] At the same time, Qwen 2.5 uses Rotary Positional Embeddings (RoPE) to encode the position information in the sequence. RoPE embeds the position information in the form of a rotation matrix into the attention mechanism, not only retaining the spatial perception ability of traditional position encoding but also enhancing the model's ability to model long-sequence dependencies by dynamically adjusting the encoding frequency. Its formula is: where pos is the position in the sequence; i is the dimension index; d is the dimension of the word embedding.

[0121] Qwen 2.5, with its large-scale pre-training and optimized architecture design, performs excellently in aspects such as language understanding, long text generation, structured data analysis, and instruction following, especially having significant advantages in mathematical reasoning and programming tasks. In its pre-training stage, a high-quality dataset of up to 1.8 trillion tokens was used, combined with techniques such as Grouped Query Attention (GQA), SwiGLU activation function, and RoPE positional encoding for optimization, enabling the model to efficiently handle complex tasks. However, due to the large parameter scale of Qwen 2.5, the Qwen 2.5 - 7B adopted in this method has 7 billion parameters. If all parameters are fine-tuned in full, not only a huge amount of computing resources and training data are required, but also the performance may decline due to overfitting or unstable training. To adapt to the specific requirements of the rockburst intelligent early warning and prevention system, such as real-time analysis of ground pressure data, early warning signal generation, and prevention strategy reasoning, this method adopts the LoRA (Low-Rank Adaptation) fine-tuning technique. LoRA achieves efficient task adaptation by adding a low-rank update matrix to the pre-trained weights, only fine-tuning a small number of additional parameters while keeping the original weights frozen, and at the same time retaining the powerful capabilities of the model in general tasks. This method significantly reduces the computational cost and memory requirements, ensuring that the system can quickly respond to rockburst risks in resource-constrained environments, improving the efficiency of real-time early warning and prevention decision-making. Its core formula is: W′ = W + ΔW, where ΔW = AB, W is the original pre-trained weight matrix; W′ is the fine-tuned weight matrix; ΔW is the low-rank update matrix; A is a low-rank matrix with dimension (d, r); B is a low-rank matrix with dimension (r, k); r is the rank, which is much smaller than d and k.

[0122] When constructing the natural language interaction module of the rockburst intelligent early warning and prevention system, a mechanism based on Retrieval-Augmented Generation (RAG) was designed to ensure that the system can generate answers that meet technical specifications and are natural and fluent when responding to users' technical consultations. This method selects a high-quality pre-trained embedding model: GTE, which has shown excellent performance in Chinese text processing and can effectively capture the semantic features and syntactic structures of language. When the user inputs a technical consultation request (such as "How to evaluate the rockburst risk level of a certain mine"), the system first retrieves relevant document fragments from the local knowledge base through semantic similarity calculation. The local knowledge base includes an industry standard library, an expert experience library, and a historical case library in the field of mine rockburst prevention, covering professional terms, engineering specifications, and technical cases of rockburst prevention. The calculation of semantic similarity uses the cosine similarity method, comparing the problem vector input by the user with the vector of the document fragments in the knowledge base. Its formula is: Among them, q is the embedded vector of the user's question, generated by the embedding model; d is the embedded vector of the document fragment in the knowledge base; · represents the dot product of vectors; || represents the L2 norm of the vector. Through this formula, the system preliminarily screens out several document fragments with the closest semantics to the user's question. Subsequently, a re-ranking module is introduced to optimize the preliminary retrieval results. The re-ranking scores based on the context relevance and technical authority between the document fragment and the question. The scoring function can be expressed as: Score(d) = α·Similarity(q, d) + β·Authority(d), where α and β are weighting coefficients used to balance semantic similarity and document authority; Authority(d) is the technical authority score of the document. After screening out the most relevant context, it is input into the fine-tuned Qwen2.5-7B model together with the user's question. The model utilizes its powerful context awareness ability and combines the retrieved technical documents to generate answers, thereby improving the understanding ability and response accuracy for the scenario of rock burst prevention and control.

[0123] The specific process of using the human-machine intelligent interaction module based on the large language model to output the intelligent decision result of the pressure relief measure is as follows:

[0124] C1: Build the natural language interaction core based on the Qwen2.5 model fine-tuned by LoRA, and perform parameter fine-tuning by combining the professional corpus in the field of mine rock burst prevention, so that the model has the accurate understanding ability of the professional terms and engineering scenarios of rock burst prevention and control. Introduce the RAG (Retrieval-Augmented Generation) technology to retrieve relevant technical documents, regulations and case data from the locally built mine rock burst prevention knowledge base as the basis for generation, so as to enhance the accuracy and professionalism of the answer, and form a human-machine intelligent interaction module; among them, the mine rock burst prevention knowledge base includes an industry standard library, an expert experience library and a historical case library;

[0125] C2: Based on the optimal pressure relief measures and pressure relief parameters input in real time, use the Qwen2.5 model fine-tuned by LoRA to accurately analyze the optimal pressure relief measures and pressure relief parameters, and conduct engineering interpretations on the professional terms of rock burst prevention and control in the optimal pressure relief measures and pressure relief parameters, and output and display the optimal pressure relief measures and pressure relief parameters as well as the interpretation information;

[0126] When there is technical consultation information input by the user, first extract the key intent and entity information in the technical consultation information through parsing, then match relevant document fragments from the mine rock burst prevention knowledge base through semantic similarity calculation, and then, after screening by the re-ranking module, the most relevant context retrieval results are selected, and then the retrieval results and the technical consultation information input by the user are jointly input into the Qwen2.5 model fine-tuned by LoRA for the inference generation of context-aware answers, ensuring that the output content meets technical standardization and language naturalness, and then forming intelligent decision-making information. Finally, the intelligent decision-making information is output and displayed.

[0127] The human-machine intelligent interaction module based on large language models significantly improves the understanding ability of professional terms and engineering scenarios through fine-tuning the LoRA adaptation layer for the mine safety field. At the same time, the introduction of RAG ensures that the generated suggestions always comply with the latest technical specifications, and finally realizes intelligent decision-making support with both professional accuracy and natural interaction experience.

[0128] To ensure high prediction accuracy, in a12 of step one, the construction process of the precursor pattern sequence is as follows:

[0129] a12-1: Process the multi-source monitoring data set D into j time-window multi-source monitoring data sets using a specific time window, and statistically calculate the maximum energy e, max average energy e, avg variance energy e, σ frequency f, the online maximum value of stress σ, max and the maximum value of drill cuttings, and obtain the kth time-window multi-source monitoring data set D according to formula (1); k ;

[0130]

[0131] a12-2: Construct the precursor pattern sequence P according to the multi-source monitoring data set D, as shown in formula (2), where the a-th precursor pattern sequence is p, a as shown in formula (3);

[0132] P = [p0, p1,..., p l (2);

[0133] p a = [D a×b , D a×b+1 , D a×b+2 ,..., D a×b+c-1 (3);

[0134] In the formula, l is the number of precursor pattern sequences, b is the sampling step size, and c is the sequence length.

[0135] To ensure that the structure of the rock burst prediction module is relatively simple and, at the same time, to ensure prediction accuracy, in a21 of step one, the initial rock burst prediction model mainly includes a CNN, an LSTM, and an attention layer. By successively mining the hidden features between microseismic data through each layer, and the layers work together to finally realize the prediction of the occurrence probability of rock burst grades and the impact medium risk level. The structure of the initial rock burst prediction model based on the CNN-LSTM architecture is as follows:

[0136] The first layer is the input layer, which is used to receive input data and input it to the second layer;

[0137] The second layer is the CNN layer, which is used to perform convolution, pooling, and activation function processing on the input data in sequence. By sliding a one-dimensional convolution kernel along the time axis, local spatial correlation features in the time series data are extracted to make up for the deficiency of LSTM in modeling spatial components, and then the unfolding operation is used to unfold the data into one-dimensional data;

[0138] Among them, the convolution operation formula is: Y = W * X + b, where W is the convolution kernel weight matrix, X is the input data, b is the bias term, and * represents the convolution operation. Different time-span feature patterns are captured through multi-scale convolution kernels, and the parameter quantity is reduced by using the weight sharing mechanism.

[0139] Among them, the formula for the average pooling function of the pooling operation is: Among them, x i is the value of the i-th pixel within the window, and w is the size of the pooling window. The pooling operation can compress the feature dimension, retain significant activation values, and enhance translational robustness.

[0140] Among them, the laser function uses the ReLU activation function, and the formula is: ReLU(x) = max(0, x), so as to suppress the vanishing gradient through the ReLU activation function and maintain time series continuity;

[0141] This design enables the high-dimensional features output by the CNN layer to have both spatial correlation and time evolution attributes, providing a composite feature expression for subsequent time series prediction.

[0142] The third layer is a combined layer of multiple LSTMs, which is used to extract the features of the one-dimensional data output by the second layer through a triple gating mechanism. Its core calculation process includes the gating mechanism and the multi-layer structure;

[0143] The gating mechanism includes an input gate, a forget gate, and an output gate. The input gate controls the inflow of new information into the cell state, combines the current input with the previous hidden state, and generates an update weight through the Sigmoid function. Among them, the formula of the Sigmoid function is: The forget gate determines whether to retain or discard historical information in the cell state to avoid the vanishing gradient caused by long-term dependence. The output gate generates the final hidden state based on the current cell state and transmits it to the next time step.

[0144] In the multi-layer structure, a stacked design is adopted, and the hidden state of the previous layer is used as the input of the next layer to extract high-order time series patterns layer by layer.

[0145] The fourth layer is the attention mechanism layer, which is used to strengthen the key time step information output by the third layer through dynamic weight allocation, and improve the focusing ability on the microseismic time series features. Its core processing process is as follows:

[0146] Input feature mapping: The hidden state sequence (including multi-granularity time features) output by the last layer of LSTM is input into the fully connected layer and mapped into an attention score vector through a learnable parameter matrix; among them, the score vector formula output by each hidden layer is: S i = tan(W * H i + b i ), and then weighted summation is performed based on the score vector. The formula for the result C i after weighted summation is: C i = ∑ i＝0 α i H i , where α i is the weight coefficient, α i = Softmax(S i ), H i is the output of the last layer of LSTM hidden layer, and Softmax is the activation function;

[0147] Weight generation: Apply the Softmax function to normalize the score vector to generate a probability distribution representing the importance of each time step. The larger the weight value, the higher the contribution degree of the corresponding time step to the prediction result;

[0148] Context vector synthesis: Based on the weights, weighted summation is performed on the hidden state sequence to generate a context vector that fuses the global temporal dependence relationship and serves as the input to the final prediction module.

[0149] The fifth layer is the output layer, which uses the fully connected layer to learn the distribution characteristics and realizes the output of the outburst risk warning result.

[0150] The outburst prediction model uses the mean absolute error (MAE) and root mean square error (RMSE) as evaluation indicators to realize the output of the outburst risk warning result W p based on multi-source monitoring data, and quantitatively analyzes the prediction effect of the model. Among them, the formula for the mean absolute error (MAE) is: The formula for the root mean square error (RMSE) is where z is the number of prediction samples, is the predicted value of the z-th sample, is the true value of the z-th sample;

[0151] In order to accurately obtain the outburst risk level, in a31 of step one, the outburst risk level W q is obtained according to formula (4);

[0152]

[0153] In the formula, W geologeGeological structure data, which includes the historical occurrence times W1 of rock bursts in the same coal seam 1 , the actual mining depth W1 at the current stage 2 , the distance W1 from the hard thick rock strata (thickness greater than 5m, UCS≥20MPa) in the overlying fractured zone to the coal seam 3 , the characteristic parameter of the roof rock thickness W1 4 , the degree of tectonic stress concentration in the mining area W1 5 , the uniaxial compressive strength of coal W1 6 and the elastic energy index of coal W1 7 ; W mine is mining information, which includes the pressure relief degree W2 of the protective layer 1 , the horizontal distance W2 from the working face to the coal pillar left by the upper protective layer mining 2 , the relationship with the adjacent goaf W2 3 , the working face length W2 4 , the width of the sectional coal pillar W2 5 , the thickness of the bottom coal left W2 6 , the distance W2 from the heading face to the goaf when driving towards the goaf 7 , the distance W2 from the advancing face to the goaf when advancing towards the goaf 8 , the distance W2 from the fault 9 , the distance W2 from the fold 10 and the distance W2 from the coal seam phase transition zone 11 .

[0154] The division criteria of geological structure and mining information data using the comprehensive index method are shown in Table 2 and Table 3. Among them, the specific standards of some factors can be modified accordingly according to the actual situation;

[0155] Table 2: Division criteria of geological structure under the influence of geological data (W geologe ) using the comprehensive index method

[0156]

[0157]

[0158] Table 3: Division criteria of geological structure under the influence of mining data (W mine ) using the comprehensive index method

[0159]

[0160]

[0161] In order to accurately and efficiently obtain the result of the rock burst danger level, in a32 of Step 1, the rock burst danger level RL during the prediction time period is obtained, and the calculation process for obtaining the classification result of the rock burst danger level based on the rock burst danger level RL is as follows:

[0162] a32-1: Normalize the data according to formula (5) to eliminate the dimension difference;

[0163]

[0164] where x ij is the j-th index of the q-th sample, and μ j , σ j are the mean and variance respectively;

[0165] a32-2: Calculate the standard deviation S j according to formula (6), and measure the volatility of each index through the standard deviation to reflect the discrimination ability of each index for the overall data; the larger S j is, the greater the amount of information contained in index j, and a larger weight should be assigned.

[0166]

[0167] a32-3: Calculate the correlation coefficient R j according to formula (7), and use the correlation coefficient R j to evaluate the redundancy between indicators to avoid repeated weighting of highly correlated indicators;

[0168]

[0169] where r jk is the Pearson correlation coefficient between index j and index k; the smaller R j is, the stronger the correlation between index j and other indicators, the lower the conflict, and the more the same information is reflected, and the weight assigned to this index should be reduced.

[0170] a32-4: Combine the contrast intensity and conflict, and calculate the comprehensive weight α j according to formula (8);

[0171]

[0172] where C j is the comprehensive information score of index j, and C j = S j × R j ;

[0173] a32-5: Calculate the rock burst danger level RL during the prediction time period according to formula (9);

[0174]

[0175] a32 - 6: First, normalize the rock burst danger level RL to the interval [0, 1] as shown in Table 4, and then obtain the classification result of the rock burst danger level according to the value range of the normalized RL.

[0176] Table 4: Rock burst danger level

[0177]

[0178] In order to accurately evaluate the implementation effect of the current pressure relief prevention and control measures through a quantitative method, in step two, the specific process of evaluating the implementation effect of the current pressure relief prevention and control measures is as follows:

[0179] B1: Real - time collect multi - source monitoring data of multiple construction cycles of the pressure relief measures. The multi - source monitoring data includes data such as the stress value of in - situ stress, micro - seismic energy, frequency, and drill cuttings volume. And use Z - score to standardize the monitoring data, and obtain the Z - score standard value z according to formula (5) i , to eliminate the dimension difference;

[0180]

[0181] In the formula, μ is the average value within the acquisition time window of each type of monitoring data, ξ is the standard deviation within the acquisition time window, which is used to reflect the degree of data dispersion,

[0182] B2: Based on the multi - source monitoring data that has been standardized, propose evaluation indicators and methods for micro - seismic, stress, and drill cuttings respectively. For micro - seismic data, compare from micro - seismic energy and frequency. Specifically, obtain the micro - seismic event energy release rate E according to formula (6) rate , obtain the spatial density f according to formula (7) rate , obtain the stress change rate ▽σ according to formula (8), obtain the critical value S of drill cuttings volume according to formula (9) critical , obtain the drill cuttings volume change rate ΔS according to formula (10);

[0183]

[0184] In the formula, E i is the micro - seismic event energy in the monitoring area within the time window, and T is the time window;

[0185]

[0186] In the formula, f is the micro - seismic event frequency within the time window, and S is the monitoring area within the time window (including the area before and after pressure relief);

[0187]

[0188] In the formula, σ pre , σ post are the stress values before and after pressure relief respectively, and L is the spacing between measurement points; when ▽σ decreases, it indicates that the pressure relief measure is effective.

[0189] S critical = K·S0 (9);

[0190] In the formula, S0 is the reference value of the drill cuttings amount under the normal stress of the coal body, and K is the geological influence coefficient, taking 1.5 - 2.0;

[0191]

[0192] B3: Obtain the pressure relief effect evaluation index P of the current pressure relief measure according to formula (11) score ;

[0193]

[0194] In the formula, α, β, χ, δ are the weight indexes of the microseismic event energy, spatial density, stress value, and drill cuttings amount respectively. The initial weight ratio is determined by AHP to be 0.25:0.25:0.25:0.25;

[0195] B4: Normalize the pressure relief effect evaluation index P score to the interval [0, 1], and then evaluate the effect of the pressure relief measure according to the value range of the normalized P score as shown in Table 5, so as to evaluate the pressure relief effect of the pressure relief measure.

[0196] Table 5: Pressure relief effect level

[0197]

[0198] As an optimization, in a12 of step one, the multi-modal data set is divided into a training set, a test set, and a validation set according to the ratio of 8:1:1.

[0199] The present invention provides a method for intelligent early warning and prevention and control of rock bursts based on large language models. First, multi-modal data obtained from historical rock burst events is used, and then a precursor pattern sequence is constructed based on the multi-modal data. At the same time, the rock burst danger level labels of the precursor pattern sequence are divided by experts, and then a massive multi-modal data set is constructed based on the precursor pattern sequence and the corresponding labels. On this basis, the initial rock burst prediction model based on the CNN-LSTM architecture is trained to obtain a rock burst prediction model with high prediction accuracy and high prediction efficiency. Furthermore, in the actual application process, only a shorter acquisition time is required, and shorter pre-sequence multi-modal monitoring data than that required for conventional early warning can be obtained to efficiently and accurately obtain the early warning result of rock burst danger. On this basis, the degree of rock burst danger is independently evaluated by combining geological structure data and mining information data during the mining process, and the contribution ratio of the rock burst danger early warning result and the degree of rock burst danger to the comprehensive index is allocated through a weight division method, so as to efficiently and accurately obtain the classification result of the rock burst danger level. Thus, the danger of rock bursts can be predicted and discovered earlier than traditional early warning systems and methods, which is conducive to taking early pressure relief prevention and control measures earlier, and greatly improving the safety factor of coal mining operations. Based on the classification result of the real-time obtained rock burst danger level, a reinforcement learning model and an adaptive optimization mechanism are adopted. According to the preset constraints, and based on the early warning information and corresponding prevention and control measure data of rock burst events stored in the historical database, the optimal decision-making of the current pressure relief measures for the rock burst danger level is dynamically realized, and the pressure relief parameters are optimized, which can ensure the timeliness and accuracy of the pressure relief measures. At the same time, it can ensure that the occurrence probability of rock burst accidents is minimized. At the same time, after each pressure relief prevention and control measure is executed, the implementation effect of the current pressure relief prevention and control measure is evaluated, and the prevention and control measure data in the historical database can be adaptively adjusted. Thus, the data in the historical database can be corrected and updated in real time, which is conducive to obtaining the optimal pressure relief measures and pressure relief parameters in each decision-making process. At the same time, a closed-loop control strategy of monitoring-evaluation-regulation is realized. A human-machine intelligent interaction module is constructed based on the Qwen2.5 model fine-tuned by LoRA, and the RAG technology is introduced. Not only can the model have the ability to accurately understand the professional terms and engineering scenarios of rock burst prevention and control, but also it can perform engineering interpretations after obtaining the optimal pressure relief measures and pressure relief parameters of the real-time input, so that relevant personnel can understand the implementation details of the pressure relief measures more clearly and thoroughly, which is conducive to ensuring the accurate implementation of the pressure relief measures and further ensuring the safety of coal mining operations. At the same time, relevant technical documents can also be retrieved from the mine rock burst prevention knowledge base constructed locally through the RAG technology as the basis for generating consultation answers, and at the same time, it is ensured that the content of the output answers meets both technical specifications and the naturalness of the language, thus realizing intelligent decision support with both professional accuracy and natural interaction experience.

[0200] This method has a high degree of intelligence and ideal interaction performance. By introducing advanced multi-modal data fusion technology, the parsing ability of the Qwen 2.5 model with LoRA fine-tuning, and an adaptive optimization decision-making mechanism, it solves the technical bottleneck that most current rock burst early warning systems are passive early warnings, unable to achieve early warnings, unable to answer technicians' questions based on natural language interaction and self-optimize according to technicians' conversations. Using this method can achieve early evaluation of surrounding rock state with multi-scale data fusion and intelligent decision support with engineering interpretability, and establish an adaptive optimization mechanism for anti-burst support parameters, forming a closed-loop control process of "monitoring - evaluation - regulation" to effectively ensure the safe production operation of the mine and achieve a more accurate, intelligent, early, interpretable and self-optimizing rock burst prevention and control plan.

Claims

1. An intelligent rock burst early warning and prevention system based on large language models, characterized in that, It includes a rock burst monitoring and early warning module, an adaptive optimization decision-making module for rock burst prevention and pressure relief measures, and a human-computer interaction module based on a large language model; The rock burst monitoring and early warning module includes a high-precision sensor group, a rock burst prediction module, and a rock burst danger level classification module; The high-precision sensor group includes a variety of high-precision sensors. The variety of high-precision sensors are respectively installed on multiple monitoring nodes in the mine and are used to collect multi-source monitoring data in real time. The rock burst prediction module is used to predict the rock burst danger level based on the multi-source monitoring data and output the rock burst danger warning result. The rock burst danger level classification module is used to obtain the impact danger degree based on the static data of the geological structure and mining information during the mining period, and is used to obtain the classification result of the rock burst danger level based on the impact danger degree and the rock burst danger warning result; The adaptive optimization decision-making module for rock burst prevention and pressure relief measures is used to dynamically realize the optimization decision of the pressure relief measures for the current rock burst danger level based on the classification result of the rock burst danger level and output the optimal pressure relief measures and pressure relief parameters; The human-computer interaction module based on the large language model is used to receive the optimal pressure relief measures and pressure relief parameters, and conduct engineering interpretations on the rock burst prevention and control professional terms in the optimal pressure relief measures and pressure relief parameters, and output and display the optimal pressure relief measures and pressure relief parameters as well as the interpretation information. At the same time, when there is technical consultation information input by the user, it is used to retrieve relevant contexts based on the mine rock burst prevention knowledge base, then form intelligent decision-making information based on the retrieval results and the technical consultation information, and finally output and display the intelligent decision-making information.

2. The intelligent rock burst early warning and prevention system based on large language model according to claim 1, characterized in that, The variety of high-precision sensors include microseismic sensors, ground sound sensors, rock mass stress sensors, and electromagnetic sensors.

3. An intelligent warning and prevention method for rock burst based on large language models, characterized in that, It includes the following steps: Step 1: Construct a rock burst monitoring and early warning module, and use the rock burst monitoring and early warning module to obtain the classification result of the rock burst danger level during the mining period; A1: Construct a multi-modal data set; a11: Use a variety of high-precision sensors installed on multiple monitoring nodes in multiple different mines to collect multi-modal signals in historical rock burst events, and obtain the original multi-modal data based on the multi-modal signals; a12: Preprocess the original multi-modal data to construct a precursor pattern sequence. Then, according to the characteristics of different application mining areas and the actual danger situation of rock burst events by experts, convert the precursor pattern sequence into the corresponding grade form, divide the corresponding rock burst danger level labels, and then construct a massive multi-modal data set through the precursor pattern sequence and the corresponding labels. Then, divide the multi-modal data set into a training set, a test set, and a validation set according to a set ratio; A2: Construct a rock burst prediction module, and use the rock burst prediction module to predict the rock burst danger warning result; a21: Construct an initial rock burst prediction model based on the CNN-LSTM architecture, train the initial rock burst prediction model using the training set, and optimize the model by minimizing the error. After training, obtain the rock burst prediction model; a22: Real-time collect the pre-order multi-modal monitoring data during mining using a variety of high-precision sensors installed on multiple monitoring nodes in the mine; a23: Input the previous multi-modal monitoring data as input data into the rock burst prediction model, and use the rock burst prediction model for prediction to output the rock burst danger warning result W p ; A3: Construct a rock burst danger level classification module, and use the rock burst danger level classification module to obtain the classification results of the rock burst danger level; a31: Analyze the rockburst hazard using the comprehensive index method based on the static data of geological structures and mining information during mining, and obtain the rockburst hazard level W q ; a32: Dynamically divide the warning results W of rock burst hazards using the CRITIC algorithm p and the degree of impact hazard W q for two types of index weights, obtain the degree of rock burst hazard RL during the prediction period, and based on the degree of rock burst hazard RL, obtain the classification results of the impact hazard level; Step 2: Construct an adaptive optimization decision-making module for rock burst prevention and pressure relief measures, and use the adaptive optimization decision-making module for rock burst prevention and pressure relief measures to output the optimal pressure relief measures and pressure relief parameters; Based on the classification results of the real-time input rock burst danger level, adopt a reinforcement learning model and an adaptive optimization mechanism. According to the preset constraint conditions, and based on the early warning information of rock burst events and the corresponding prevention and control measure data stored in the historical database, dynamically realize the optimization decision-making of the pressure relief measures for the current rock burst danger level, and optimize the pressure relief parameters. Finally, output the optimal pressure relief measures and pressure relief parameters according to the current rock burst danger level; At the same time, after each execution of the pressure relief prevention and control measures, collect the multi-source monitoring data of multiple construction cycles of the pressure relief prevention and control measures, evaluate the implementation effect of the current pressure relief prevention and control measures based on the multi-source monitoring data, and then modify the prevention and control measure data in the historical database based on the evaluation effect to obtain the optimal pressure relief measures through adaptive adjustment; Step 3: Construct a human-machine intelligent interaction module based on a large language model, and use the human-machine intelligent interaction module based on a large language model to output the intelligent decision-making results of the pressure relief measures; C1: Construct a natural language interaction core based on the Qwen2.5 model fine-tuned by LoRA, and introduce the RAG technology to retrieve relevant technical documents, regulations and case data from the locally constructed mine rock burst prevention knowledge base as the generation basis to form a human-machine intelligent interaction module; C2: Based on the real-time input optimal pressure relief measures and pressure relief parameters, use the Qwen2.5 model fine-tuned by LoRA to accurately analyze the optimal pressure relief measures and pressure relief parameters, and give an engineering explanation of the rock burst prevention and control professional terms in the optimal pressure relief measures and pressure relief parameters, and output and display the optimal pressure relief measures and pressure relief parameters as well as the explanation information; When there is technical consultation information input by the user, first extract the key intent and entity information in the technical consultation information by parsing, then match relevant document fragments from the mine rock burst prevention knowledge base through semantic similarity calculation, and then, after screening by the re-ranking module, select the most relevant context retrieval results, and then input the retrieval results and the technical consultation information input by the user into the Qwen2.5 model fine-tuned by LoRA for the inference generation of context-aware answers to form intelligent decision-making information. Finally, output and display the intelligent decision-making information.

4. The intelligent warning and prevention method for rock burst based on large language model according to claim 3, wherein In a12 of Step 1, the construction process of the precursor mode sequence is as follows: a12-1: Process the multi-source monitoring data set D into j time-window multi-source monitoring data sets using a specific time window, and statistically calculate the maximum energy e, average energy e, variance energy e, frequency f, online maximum stress value σ, and maximum drill cuttings volume of the microseismic data for each time window. Then, obtain the k-th time-window multi-source monitoring data set D according to formula (1). max and average energy e avg and variance energy e σ and frequency f, online maximum stress value σ max and maximum drill cuttings volume and obtain the k-th time-window multi-source monitoring data set D according to formula (1). k ; a12-2: Construct the precursor pattern sequence P based on the multi-source monitoring data set D as shown in formula (2), where the a-th precursor pattern sequence is p a As shown in formula (3); P = [p0, p1,..., p l (2); p a = [D a×b , D a×b+1 , D a×b+2 ,..., D a×b+c-1 (3); In the formula, l is the number of precursor mode sequences, b is the sampling step, and c is the sequence length.

5. The intelligent warning and prevention method for rock burst based on large language model according to claim 3, wherein, In a21 of Step 1, the structure of the initial rock burst prediction model based on the CNN-LSTM architecture is as follows: The first layer is the input layer, which is used to receive the input data and input it to the second layer; The second layer is the CNN layer, which is used to perform convolution, pooling, and activation function processing on the input data in sequence. By sliding a one-dimensional convolution kernel along the time axis, local spatial correlation features in the time series data are extracted, and then the data is unfolded into one-dimensional data using an unfolding operation; The third layer is a combined layer of multiple LSTMs, which is used to extract the features of the one-dimensional data output by the second layer through a triple gating mechanism; The fourth layer is the attention mechanism layer, which is used to enhance the key time step information output by the third layer through dynamic weight allocation, and improve the focusing ability on the microseismic time series features; The fifth layer is the output layer, which uses a fully connected layer to learn the distribution features and realizes the output of the outburst risk warning result.

6. The intelligent warning and prevention method for rock burst based on large language model according to claim 3, characterized in that In a31 of Step 1, the impact hazard level W is obtained according to Formula (4). q ; Where, W geologe is geological structure data, which includes the historical occurrence times of rock burst W1 in the same coal seam 1 , the actual mining depth W1 at the current stage 2 , the distance between the hard thick rock stratum in the overlying fissure zone and the coal seam W1 3 , the characteristic parameter of roof rock stratum thickness W1 4 , the degree of tectonic stress concentration in the mining area W1 5 , the uniaxial compressive strength of coal W1 6 and the elastic energy index of coal W1 7 ; W mine Mining information, which includes the pressure relief degree of the protective layer Horizontal distance between the working face and the coal pillar left after the mining of the upper protective layer Relationship with the adjacent goaf Length of the working face Width of the sectional coal pillar Thickness of the bottom coal left Distance from the goaf when driving towards the goaf Distance from the goaf when advancing towards the goaf Distance from the fault Distance from the fold And distance from the coal seam phase transition zone 7. The intelligent warning and prevention method for rock burst based on large language model according to claim 3, wherein In a32 of step one, the outburst risk level RL of the prediction time period is obtained, and the calculation process of the classification result of the outburst risk level based on the outburst risk level RL is as follows: a32-1: Normalize the data according to formula (5) to eliminate the dimension difference; where x ij is the j-th index of the q-th sample, μ j , σ j are the mean and variance respectively; a32-2: Calculate the standard deviation S according to formula (6) j and measure the volatility of each indicator through the standard deviation to reflect the discrimination ability of each indicator for the overall data; a32-3: Calculate the correlation coefficient R according to formula (7) j , and use the correlation coefficient R j to evaluate the redundancy between evaluation indicators to avoid double weighting of highly correlated indicators; where r jk is the Pearson correlation coefficient between index j and index k; a32-4: Combine the contrast intensity and conflict, and calculate the comprehensive weight α according to formula (8) j ; where C j is the comprehensive information score of index j, and C j = S j × R j ; a32-5: Calculate the outburst risk level RL of the prediction time period according to formula (9); a32-6: First, normalize the outburst risk level RL to the interval [0,1], and then obtain the classification result of the outburst risk level according to the value range of the normalized RL.

8. The intelligent warning and prevention method for rock burst based on large language model according to claim 3, characterized in that In step two, the specific process of evaluating the implementation effect of the current pressure relief prevention and control measures is as follows: B1: Collect multi-source monitoring data of multiple construction cycles of pressure relief measures in real time, perform standardization processing on the monitoring data using Z-score, and obtain the Z-score standardized value z according to formula (10). i ; where μ is the average value within the acquisition time window of each type of monitoring data, ξ is the standard deviation within the acquisition time window, which is used to reflect the degree of data dispersion, B2: Based on the multi-source monitoring data that has been standardized, obtain the microseismic event energy release rate E according to formula (11) rate , obtain the spatial density f according to formula (12) rate , obtain the stress change rate according to formula (13) Obtain the critical value S of the drill cuttings amount according to formula (14) critical , obtain the change rate ΔS of the drill cuttings amount according to formula (15); where E i is the energy of microseismic events in the monitoring area within the time window, and T is the time window; In the formula, f is the frequency of microseismic events within the time window, and S is the monitoring area within the time window; Where, σ pre and σ post are the stress values before and after pressure relief respectively, and L is the spacing between measurement points; S critical = K·S0 (14); In the formula, S0 is the reference value of the drill cuttings amount under the normal stress of the coal body, and K is the geological influence coefficient, taking 1.5 - 2.0; B3: Obtain the pressure relief effect evaluation index P of the current pressure relief measure according to formula (16) score ; In the formula, α, β, X, and δ are the weight indicators of the microseismic event energy, spatial density, stress value, and drill cuttings amount respectively; B4: Normalize the pressure relief effect evaluation index P score to the range of [0, 1], and then evaluate the effect of the pressure relief measure according to the value range of the normalized P score ​ 9. The intelligent warning and prevention method for rock burst based on large language model according to claim 3, characterized in that, In a12 of step one, the multi-modal dataset is divided into a training set, a test set, and a validation set according to the ratio of 8:1:1.

Images