Coal mine rock burst occurrence probability dynamic speculation method

Abstract

The embodiment of the invention discloses a coal mine rock burst occurrence probability dynamic speculation method, and belongs to the technical field of coal mine danger prediction, and the method comprises the steps: carrying out the preprocessing of historical monitoring data of a target working face; using a comprehensive index method to calculate and quantify a danger index as the prior probability of Bayesian; dividing the target area into n sub-areas according to a preset step pitch; constructing a prediction system, wherein the prediction system comprises a prior layer and a likelihood layer; setting each monitoring layer of the likelihood layer as a 4-level discrete state, and determining a judgment threshold value of the 4-level discrete state; respectively judging the current discrete state of each monitoring layer of the likelihood layer according to the judgment threshold, and integrating the current states of all the monitoring layers; setting a probability transfer rule of an initial default dynamic state of each monitoring layer state of the likelihood layer; and calculating the posterior probability of rock burst occurrence in each monitoring state based on a Bayesian formula. According to the method, the engineering suitability, the dynamic flexibility and the field pertinence can be improved.

Description

Technical Field

[0001] This invention relates to the field of coal mine hazard prediction technology, specifically to a method for dynamically predicting the probability of coal mine rockburst occurrence. Background Technology

[0002] Rockbursts are a typical coal and rock dynamic disaster in deep coal mining, characterized by their sudden onset and high destructiveness, seriously threatening mine safety. With increasing mining depth and intensity, their frequency and severity rise significantly, becoming a core bottleneck for safe mining of deep resources. Therefore, developing accurate, reliable, and engineering-adaptable dynamic risk prediction technologies for rockbursts is crucial for achieving early warning and effective prevention and control of this disaster.

[0003] Currently, the industry has reached a technical consensus on "multi-source monitoring data fusion" and "combination of dynamic and static indicators." Two representative patent documents (D1: CN113673119B, D2: CN119624131A) demonstrate significant progress in this direction: Patent D1 constructs a "dynamic-static coupling evaluation" framework, fusing dynamic indicators such as microseismic activity and stress with static indicators such as geology and mining data using Bayesian methods, effectively improving the fusion degree of multi-system data and providing a path for multi-indicator collaborative evaluation. Patent D2 establishes a comprehensive evaluation system of "impact event probability - impact degree - impact range," systematically incorporating consequences such as personnel casualties and equipment damage into risk assessment for the first time, enriching the evaluation dimensions and enhancing the engineering reference value of the results.

[0004] Despite the positive progress made in existing technologies, the following technical bottlenecks still urgently need to be addressed in light of practical engineering applications: (1) The model structure is rigid and the engineering adaptability is poor: The indicator system and model structure of D1 and D2 are mostly rigid designs, which can only be adapted to mines with specific monitoring conditions. They cannot be flexibly adjusted according to the differences in on-site monitoring methods (such as old mines only having basic monitoring and newly built mines having added monitoring modules), resulting in poor universality.

[0005] (2) Lack of dynamic update mechanism makes it difficult to achieve "one policy for one mine": The rockburst pattern has significant "regional specificity", but the key parameters in the existing models (such as the warning threshold and prior probability of D1, and the normal standard range of D2) are mostly general preset values, and the probability transfer rules in their Bayesian networks are fixed. The existing technology has not established a mechanism for dynamically optimizing the above key parameters based on the mine's own multi-dimensional historical data (including rockburst cases), which makes it impossible for the model to "learn" through data accumulation, and the accuracy is difficult to improve with the mining process, so it is impossible to truly achieve effective prevention and control of "one policy for one mine".

[0006] (3) Threshold determination relies on experience and lacks objectivity: The hazard level threshold of D1 is based on experience and does not fully consider the differences in geology and mining conditions of different mines; D2 sets the normal range by means of data, and its data processing logic is flawed - the historical data used to construct the normal benchmark may itself be mixed with shock disturbance data, resulting in the benchmark threshold being "inherently distorted". Both lack objective statistical support based on actual data of specific mines.

[0007] (4) The application of Bayesian theory is not sufficiently scenario-based: D1 mainly uses Bayesian theory for the fusion of multi-model results, while D2's Bayesian network variables have a single dimension. Neither of them has refined the construction of the differential probability transmission logic of the top-level state of "impact occurred / did not occur" to the lower-level monitoring data state, making it difficult to fully reflect the actual correlation of risk evolution.

[0008] In summary, existing technologies have significant shortcomings in terms of engineering adaptability, dynamic flexibility, and on-site relevance. There is an urgent need to develop a method for predicting the dynamic occurrence probability of rockbursts that can adapt to different mine monitoring conditions, support dynamic updates based on historical data, determine thresholds based on objective statistics, and apply Bayesian theory in various scenarios, so as to further improve the scientificity and effectiveness of dynamic prevention and control of rockbursts in deep coal mines. Summary of the Invention

[0009] This invention provides a method for dynamically predicting the probability of rockbursts in coal mines, thereby improving engineering adaptability, dynamic flexibility, and on-site relevance.

[0010] To address the aforementioned technical problems, this application provides the following technical solution: A method for dynamically predicting the probability of rockbursts in coal mines, comprising: Step S1: Preprocess the historical monitoring data of the target working face; Step S2: Calculate the quantitative hazard index using the comprehensive index method in impact hazard assessment as the prior probability of Bayes; Step S3: Divide the target area into n sub-regions according to a preset step size; Step S4: Construct a prediction system, wherein the prediction system includes a prior layer and a likelihood layer, and the likelihood layer includes a basic layer and an extension layer; Step S5: Set each monitoring layer of the likelihood layer as a multi-level discrete state; except for the basic layer, the judgment threshold of the multi-level discrete state is determined by box statistical method for the other layers. Step S6: Determine the current discrete state of each monitoring layer of the likelihood layer according to the determination threshold, integrate the current states of all monitoring layers, and generate the likelihood layer monitoring evidence chain; Step S7: Set the initial default dynamic probability propagation rules for the states of each monitoring layer in the likelihood layer; Step S8: For each sub-region, calculate the posterior probability of rockburst based on the prior probability, the likelihood layer monitoring evidence chain, and the probability transfer rule, using the Bayesian formula.

[0011] In some embodiments, step S1 includes: If the historical monitoring data of the target working face is insufficient, supplement it with historical data of other working faces in the same mine with similar geological conditions; And / or, step S1 includes: All monitoring data from the period of the rockburst event and within 12 hours before and after it occurred were removed; Invalid data is removed simultaneously, and only reliable data verified on-site is retained.

[0012] In some embodiments, in step S4, the prior layer is used to quantify the basic probability of rockburst occurring at the target working face under the current geological mining conditions, the basic layer is used to quantify the basic probability of rockburst hazard in the current area at the target working face, and the extended layer is used to quantify the basic probability of the target working face exhibiting the state of the current monitoring layer.

[0013] In some embodiments, step S5 includes: Step S51: Calculate the theoretical maximum value of the dataset for each type of monitoring parameter in the statistical sample set; Step S52: Based on the increase in the monitored value relative to the theoretical maximum value, the likelihood layer, except for the basic layer, is divided into four discrete states: "none, weak, medium, and strong".

[0014] In some embodiments, the probability transfer rule in step S7 includes: The states of the prior layer include: rockburst occurred and rockburst did not occur; The initial probability of regional impact hazard in the basic layer adopts the principle of uniform distribution, with the initial default probability of each state set at 25%, ensuring that the total probability is 100%.

[0015] In some embodiments, in step S7, the probability transfer rule further includes: When a rockburst occurs, the probability distribution of the monitoring data status of each extended layer is not affected by the status of the parent node of the previous layer. It only follows the engineering rule that the proportion of strong danger signals is high when a rockburst occurs. That is, the initial default probability is uniformly set as: "None": 1%, "Weak": 2%, "Medium": 3%, "Strong": 94%, with only the probability being slightly adjusted (3% → 4%) from "Strong (parent node) to Medium (child node)". The probability distribution remains fixed. When a rock burst does not occur, the probability of the monitoring data status in each extended layer must be adjusted strictly according to the association rules between parent and child nodes. That is, the status of the parent node in the upper layer directly affects the probability of the monitoring data status in the lower layer. The association logic can be summarized into four priority principles: ① The states are completely identical: the danger levels of the parent and child nodes match, indicating the strongest correlation and the highest probability; ② Adjacent states: small difference in risk level, moderate correlation, and medium probability; ③ Status interval 1 level: The danger levels differ greatly, the correlation is weak, and the probability is low; ④ The states are completely opposite: the danger levels are completely opposite, the correlation is the weakest, and the probability is the lowest.

[0016] In some embodiments, step S8 includes: Step S81: Calculate the likelihood probability of a rockburst occurring; Step S82: Calculate the likelihood probability when a rockburst does not occur; Step S83: Calculate the evidence probability of the likelihood layer monitoring evidence chain according to the law of total probability; Step S84: Calculate the probability of rockburst occurrence, i.e., the posterior probability, based on Bayes' theorem and corrected by real-time monitoring evidence, according to Bayes' theorem.

[0017] In some embodiments, the method further includes: Step S9: Conduct data preparation and initial learning; The data reserve refers to the continuous collection and storage of complete data chains for each segment and sub-region, forming a continuously growing mine-specific database. When the accumulated data reaches a certain scale, initial learning is initiated. The initial learning involves using frequency statistics to perform statistical analysis on historical data in the database and updating the probability propagation rules in step S7 and the judgment threshold in step S5.

[0018] In some embodiments, the method further includes: Step S10: Calculate the posterior probability of the target region and match the hazard level; Step S10 includes: The posterior probabilities of each sub-region are combined to obtain the posterior probability of the target region. For the posterior probability of the target region, box statistical analysis is used to determine the theoretical maximum value. The posterior probability value is divided into 4 discrete states based on the increase of the posterior probability value relative to the theoretical maximum value.

[0019] In some embodiments, step S9 further includes: performing result verification and deep optimization; Among them, the result verification is to compare the hazard level judgment result output in step S10 with the actual impact situation on site one by one; The in-depth optimization: When there is a deviation between the hazard level judgment and the actual situation on site, this layer of optimization is initiated to make targeted adjustments to the deviations that the initial learning failed to correct, that is, to optimize the probability propagation rules in step S7 and the judgment threshold in step S5.

[0020] Based on the above technical solution, the present invention has at least the following beneficial effects: (1) Creative level: By combining Bayesian and box statistics in a three-layer deep synergy, we break through the technical understanding of "simple combination of the two", realize the innovation from "method superposition" to "mechanism synergy", and improve the matching accuracy of the model for rockburst risk; (2) Engineering application level: The "basic layer + extended layer" architecture is adapted to all mine monitoring conditions, the "data closed-loop update" realizes "one mine, one policy", and the "objective box-type threshold" reduces human experience interference, providing more scientific and reliable technical support for the prevention and control of rockburst in deep coal mines. Attached Figure Description

[0021] The accompanying drawings in this application are intended to supplement the textual description in the specification with graphics, and to further explain the technical solution of this application. They do not constitute an undue limitation on this application.

[0022] Figure 1 This is a flowchart illustrating the method for dynamically predicting the probability of rockbursts in coal mines according to the present invention. Figure 2 This is a direct translation of the box-shaped statistical elements in this invention. Detailed Implementation

[0023] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0024] In the description of this invention, it should be understood that the terms "center," "lateral," "longitudinal," "front," "rear," "left," "right," "upper," "lower," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limiting the scope of protection of this invention.

[0025] To address the four core shortcomings of existing technologies in terms of engineering adaptability, dynamic flexibility, threshold objectivity, and depth of Bayesian application, this case achieves precise solutions through the following technical features: (1) Regarding the problem of "fixed model structure and poor engineering adaptability": In step S4, this application constructs a likelihood layer monitoring system of "basic layer + extended layer". The basic layer only retains the regional impact hazard (adapting to the basic requirements of all rockburst mines). The extended layer can flexibly add or remove monitoring dimensions according to the actual monitoring capabilities of the mine (such as only stress monitoring, adding microseismic / displacement monitoring). For example, if there is no drill cuttings monitoring, the drill cuttings state layer is deleted; if ground sound monitoring is added, the ground sound state layer is inserted. There is no need to reconstruct the overall model architecture, which adapts to the differentiated monitoring conditions of old mines and new mines. (2) Regarding the lack of a dynamic update mechanism, making it difficult to achieve "one policy for one mine": This application constructs a closed-loop optimization mechanism through step S9 "data storage and data chain update" - after accumulating monitoring data from more than 100 segmented areas, the "threshold of box statistics" (step S5) and the "probability transfer rule of Bayesian likelihood layer" (step S7) are updated synchronously using continuous data accumulation and frequency statistics, so that the model can continuously "learn" its own data patterns as the mining process progresses, gradually adapt to the geological and mining characteristics of the target mine, and truly achieve "one policy for one mine"; (3) Regarding the issue that “the determination of the threshold depends on experience and lacks objectivity”: In step S1 of this application, “disturbance data of 12 hours before and after the occurrence of rockburst” is first removed to ensure that the sample used for box-type statistics is “pure normal mining data”; then, the theoretical maximum value (Q3+1.5×IQR) is calculated through box-type statistics in step S5, and the threshold is divided by “the increase of the monitoring value relative to the theoretical maximum value”. It is generated entirely based on the reliable data statistics of the target mine itself, avoiding the distortion of experience preset or benchmark data, and improving the objectivity of the threshold. (4) Regarding the “insufficient depth of application scenario of Bayesian theory”: Step S7 of this application is divided into two scenarios, “occurrence of rockburst” and “non-occurrence of rockburst”, and different probability transmission rules are designed. When the rockburst occurs, the engineering law of “strong danger signal dominates” (94% probability) is followed. When the rockburst does not occur, the logic of “correlation between the state of the parent node and the child node” (consistent > adjacent > interval > opposite) is followed. The probability correlation of “top-level state - lower-level monitoring data” is refined to fully reflect the actual law of rockburst risk evolution.

[0026] This invention provides a method for dynamically predicting the probability of rockbursts in coal mines based on the combined application of Bayesian theory and box statistics. Figure 1-2 As shown, it includes: Step S1: Preprocess the historical monitoring data of the target working face.

[0027] This step is for data preprocessing: (1) Data source: Historical monitoring data of the target working face shall be used first; if the data of the target working face is insufficient, historical data of other working faces in the same mine with similar geological conditions and mining technology may be used as supplementary data. (2) Data removal: Remove all monitoring data during and within 12 hours before and after the rockburst event. When a rockburst occurs, stress, displacement, microseismic data will show sudden anomalies (such as a sudden increase in stress or a surge in microseismic energy). Such data belongs to "impact disturbance data" and cannot reflect the monitoring characteristics under normal mining conditions. If it is included in the statistics, it will lead to distortion of the benchmark threshold. It is necessary to remove invalid data such as sensor failure at the same time and retain only reliable data verified on site to ensure the validity of the statistical sample.

[0028] Step S2: Calculate the quantitative risk index using the comprehensive index method as the prior probability of Bayes.

[0029] This step involves calculating the prior probability P(A): W can be directly obtained using the comprehensive index method in impact hazard assessment. t As P(A), that is, the prior probability of rockburst, 1-W t As P(¬A); the core logic of the comprehensive index method is to evaluate the degree of regional shock risk from a macro perspective: based on historical rockburst disaster cases, the influence of mining geology and mining technology factors is decomposed, and the factors are assigned weights through statistics and expert judgment, and then the quantitative risk index W is obtained by weighted summation. t This constitutes a regional hazard level classification model. From the perspective of the data source requirements for Bayesian prior probabilities, the two are highly compatible: the prior needs to be based on macroscopic cognition, historical patterns, or experience, while the comprehensive index method infers factors from historical disasters and integrates macroscopic patterns, its W... t After probabilistic calibration, it can be used as the prior probability basis for the initial cognitive level of Bayes.

[0030] Step S3: Divide the target area into n sub-regions according to the preset step size.

[0031] This step involves segmenting the target region: the target region can be labeled M, and divided into n parts according to a certain step size d. Each part is labeled M1, M2, M3...M in spatial order. n In practice, the step size d can be flexibly set as needed, such as 5m, 10m, 15m, etc. In this embodiment of the invention, the step size d is 5m.

[0032] S4: Construct a prediction system, wherein the prediction system includes a prior layer and a likelihood layer, and the likelihood layer includes a basic layer and an extension layer.

[0033] In this step, the prior layer is used to quantify the basic probability of rockburst occurring at the target working face under the current geological mining conditions; the basic layer is used to quantify the basic probability of rockburst hazard in the current area at the target working face; and the extended layers are used to quantify the basic probability of the target working face exhibiting the state of the current monitoring layer. This step uses a "basic layer + extended layer" architecture to construct the likelihood layer to adapt to the monitoring conditions of different mines. (1) Basic layer: only includes regional shock hazard (macro premise). The core basis for setting "regional shock hazard" as the first layer of the likelihood layer is that the occurrence of rockburst is directly related to the regional basic hazard - regional hazard is the macro premise of rockburst occurrence. Regardless of whether the shock event eventually occurs, it is strongly related to the basic hazard attributes of the region, so it should be included in the reasoning system first.

[0034] (2) Extended Layer: Selected based on the actual monitoring capabilities of the mine. If a microseismic monitoring system is equipped, a "microseismic state layer" (subdivided into the maximum energy value of a single microseismic event, the cumulative energy value of microseismic events, and the frequency of microseismic events) is added. If a drill cuttings monitoring system is equipped, a "drill cuttings state layer" (subdivided into the size of the drill cuttings) is added. The extended layer can be directly inserted after the basic layer without reconstructing the overall layer sequence. For each rockburst mine, the extended layer can be added or removed according to the level of monitoring capabilities of the mine.

[0035] When designing the second and subsequent hierarchical sequences, it was initially considered to set the subsequent hierarchical sequences (such as the relationship between "stress state, displacement state, and microseismic state" as parallel relationships) as parallel relationships. However, considering the computational characteristics of Bayesian trees, an ordered chain structure relationship was ultimately adopted for the following reasons: Even with a parallel relationship, the independence of each event (such as the correlation between stress and displacement) still needs to be verified separately during calculation, and the final probability derivation result is consistent with the derivation result of the ordered chain structure; moreover, the ordered chain structure can eliminate the independence verification step, while avoiding the omission of key information of parallel states, resulting in better computational efficiency and accuracy.

[0036] (3) Standard configuration: The inventors have set the standard configuration of the likelihood layer as shown in Table 1, which can be used as a reference, based on reliable on-site online monitoring methods.

[0037] Table 1 Standard Contents of Hierarchical Sequence Step S5: Set each monitoring layer of the likelihood layer to a 4-level discrete state; except for the basic layer, the judgment threshold of the multi-level discrete state is determined by box statistical method for the other layers.

[0038] In this step, a state threshold determination based on box pegged statistics is used. Specifically, for each monitoring layer of the likelihood layer (except for regional hazard, which is directly confirmed by the evaluation results in the identified impact hazard assessment report), the box pegged statistics method is used to determine the judgment thresholds for four discrete states: "none," "weak," "medium," and "strong." (1) Box-type statistical calculation: For each type of monitoring parameter (such as stress value, microseismic energy value) in the statistical sample set, calculate the theoretical maximum value of the dataset (Q3+1.5×IQR). (2) Delineation of discrete state thresholds: Based on the "increase in the monitored value relative to the theoretical maximum value", the likelihood layer is divided into 4 states except for the basic layer. For details, please refer to Table 2.

[0039] Table 2. 4-level discrete states To achieve objective quantitative determination of the discrete states of the likelihood layer ("none, weak, medium, strong"), this method introduces box-type statistics, such as... Figure 2 As shown, by statistically analyzing historical monitoring data, the evaluation thresholds for each state are determined to ensure the engineering rationality and data reliability of state determination. Box-type statistics describe the data distribution characteristics through eight core elements, the definitions of which are shown in Table 3, providing a basis for determining the "theoretical threshold".

[0040] Table 3. Eight core elements of box-type statistical methods Based on the eight elements of the box plotting method, the correspondence between the "theoretical threshold" and the "data validity interval" is further clarified, providing clear boundaries for the evaluation of "no / weak / medium / strong" states. The specific rules are shown in Table 4. Figure 2 As shown: Table 4. Correspondence between theoretical thresholds and data validity intervals Using only the "theoretical maximum value" (upper quartile, i.e., Q3 + 1.5IQR) obtained from box morphology statistics as the core benchmark, four discrete states are defined by the "increase in the monitored value relative to the theoretical maximum value," thereby achieving a quantitative and unified assessment of the state. The specific standards are shown in Table 5. Table 5 Specific Standards Step S6: Determine the current discrete state of each monitoring layer of the likelihood layer according to the determination threshold, integrate the current states of all monitoring layers, and generate the likelihood layer monitoring evidence chain.

[0041] This step generates the likelihood layer monitoring evidence chain: according to the threshold determined in step S5, the current discrete state of each monitoring layer in each likelihood layer is determined (e.g., a stress value increase of 30% is determined as "weak danger", and a micro-seismic single event energy increase of 80% is determined as "strong danger"); the current states of all monitoring layers are integrated to generate "likelihood layer monitoring evidence chain B" (e.g., "regional impact hazard medium + stress weak + displacement medium + micro-seismic strong").

[0042] Step S7: Set the initial default dynamic probability propagation rules for the states of each monitoring layer in the likelihood layer.

[0043] In this step, the initial default dynamic probability propagation rules for the states of each monitoring layer in the likelihood layer are set.

[0044] The representation of the likelihood layer states directly affects the computational efficiency and engineering applicability of the Bayesian tree. This application prioritizes discretization and sets rules around the probability of the discrete states to form a complete logical system.

[0045] Discretization can transform complex monitoring data into four distinct levels: "none, weak, medium, and strong". When initially applying Bayesian methods, probabilities can be assigned using historical data or engineering experience, significantly reducing the computational complexity of the overall mathematical model and making it easier for on-site personnel to understand and apply.

[0046] The specific rules for probability propagation are as follows: (1) Top layer (i.e., prior layer) state: rockburst occurred; rockburst did not occur. Assign path extension probability values ​​according to different top layer states.

[0047] (2) Likelihood layer first layer (i.e. basic layer): default probability of regional shock risk; as the starting layer of the likelihood layer, the initial probability of regional shock risk adopts the principle of uniform distribution. The initial default probability of each state is set to 25%, ensuring that the total probability is 100%, which provides a basis for the probability transmission of subsequent layers.

[0048] (3) Initial default probability of the second layer and beyond (stress / displacement / microseismic state, etc.): The initial probability of the monitoring data state (stress value magnitude, displacement value growth rate, etc.) of the second layer and beyond needs to be dynamically adjusted in combination with the state of the top layer (occurrence / non-occurrence of rockburst) and the state of the parent node of the previous layer (such as the first layer "regional rockburst hazard"), and is divided into two major scenarios: ①Scenario 1: Top-level status is "Rockburst Occurrence" When a rockburst is confirmed to occur, the probability distribution of the monitoring data status is not affected by the status of the parent node (regional rockburst hazard). It only follows the engineering principle that "a high proportion of strong hazard signals occur during rockbursts." The initial default probabilities are uniformly set as follows: "None": 1% (almost no hazard signals), "Weak": 2% (very few weak signals), "Medium": 3% (few medium signals), and "Strong": 94% (strong signals are absolutely dominant). Only the probability is slightly adjusted from "Strong" to "Medium" (3% → 4%). The core logic of this setting is that when a rockburst occurs, the monitoring data such as stress, displacement, and microseismic activity will inevitably show strong anomalies. The influence of the parent node (regional hazard) is covered by the strong dynamic response of the rockburst event, so the probability distribution remains fixed.

[0049] ②Scenario 2: Top-level status is "No rockburst occurred" When a rockburst is determined not to have occurred, the probability of the monitoring data status must be adjusted strictly according to the correlation rules between parent and child nodes. The status of the parent node at the upper level (such as "none / weak / medium / strong" for regional rockburst hazard) directly affects the probability of the monitoring data status at the lower level. The correlation logic can be summarized into four priority principles (from high to low priority): ① Completely consistent status (e.g., parent node "medium" → child node "medium"): The hazard levels of parent and child nodes match, the correlation is the strongest, and the probability is the highest; ② Adjacent status (e.g., parent node "weak" → child nodes "none" or "medium"): The difference in hazard levels is small, the correlation is secondary, and the probability is moderate; ③ Status interval of 1 level (e.g., parent node "none" → child node "medium"): The difference in hazard levels is large, the correlation is weak, and the probability is low; ④ Completely opposite status (e.g., parent node "none" → child node "strong"): The hazard levels are completely opposite, the correlation is the weakest, and the probability is the lowest.

[0050] Based on the above rules, a probability transfer table of "top-level state - parent node - child node" is formed, clarifying the probability allocation relationship of each level of state, as shown in Tables 6 and 7: Table 6. Probability Transmission Table for Rockburst Occurrence in Top-Level State Table 7. Probability Transmission Table for Top-Level State When Rockburst Has Not Occurred Step S8: For each sub-region, based on the prior probability, the likelihood layer monitoring evidence chain, and the probability transfer rule, calculate the posterior probability of rockburst occurrence under each monitoring state using Bayes' formula.

[0051] In this step, the posterior probability of rockburst (for each sub-region) is derived based on Bayes' theorem. That is, for region Mn, according to the probability propagation rule set in step S7, and combined with the "likelihood layer monitoring evidence B" generated in step S6, the two similarity probabilities are calculated respectively: (1) The likelihood probability P(B|A) of a rockburst. Based on the probability transfer table of Scenario 1, the state probabilities of each monitoring layer in the likelihood layer are multiplied together.

[0052] (2) The likelihood probability P(B|¬A) when the rockburst does not occur. Based on the probability transfer table of scenario 2, the state probabilities of each monitoring layer in the likelihood layer are multiplied together.

[0053] (3) Calculation of evidence probability and posterior probability ① Probability of evidence P(B): According to the total probability formula P(B)=P(B|A)×P(A)+P(B|¬A)×P(¬A), substitute P(A), P(¬A), P(B|A), P(B|¬A) to obtain the total probability of monitored evidence B; ② Posterior probability P(A|B): According to Bayes' theorem P(A|B)=[P(B|A)×P(A)] / P(B), the "probability of rockburst occurrence after correction based on real-time monitoring evidence" is calculated, which is the posterior probability. This probability reflects the current risk level of rockburst occurrence in the region.

[0054] In some embodiments of the present invention, the method further includes: Step S9: Data closed-loop update and model optimization.

[0055] (1) Data storage and initial learning (basic update layer) ① Data Reserve: Continuously collect and store the complete data chain for each segment area M (including monitoring data, calculated posterior probability P(A|B), and actual on-site risk conditions), forming a continuously growing mine-specific database. When the accumulated data reaches a certain scale (recommended ≥100 segment areas), initiate initial learning.

[0056] ②Preliminary learning / updating: Utilize frequency statistics to perform statistical analysis on historical data in the database.

[0057] a. Update the probability rule in step S7: For example, count the frequency of the "stress value" actually appearing in the state of "none, weak, medium, strong" when all "regional impact risk is medium", and update the corresponding initial default probability value in the S7 probability table with this frequency.

[0058] b. Update the box-type statistical threshold in step S5: When a large amount of new data is added, re-execute the box-type statistical calculation in step S5 to update the "theoretical maximum value" of various monitoring parameters, thereby indirectly updating the state thresholds of "none, weak, medium and strong".

[0059] (2) Result verification and deep optimization (closed-loop optimization layer) ① Result verification / comparison: Compare the hazard level judgment results output by S10 with the actual impact situation on site (such as signs of coal seam vibration).

[0060] ② Deep / Backtracking Optimization: When a "deviation between hazard level assessment and actual situation" occurs (for example, the system determines "no hazard" but slight impact signs appear on site), this layer of optimization is initiated to make targeted adjustments to the deviations that the initial learning failed to correct: a. Optimize the threshold boundary in step S5: Adjust the increase standard for the state division of specific monitoring indicators. For example, if the increase in "stress value" is between 20% and 25% (the original "no danger" range) in multiple missed cases, tighten the upper limit of the increase for "no danger" from ≤25% to ≤20% to improve the model's sensitivity to risk signals.

[0061] b. Optimize the probability rule in step S7: For "parent node-child node" state combinations that frequently result in misjudgments, correct their probability propagation values. For example, statistics show that when "weak stress" and "medium displacement" occur simultaneously, the frequency of subsequent impacts is very high, so increase the probability weight of "weak stress → medium displacement" in the S7 probability table.

[0062] In other embodiments of the present invention, the method further includes: Step S10: Calculate the posterior probability of the target region and match the danger level.

[0063] This step involves matching the hazard level (of the entire target area) using posterior probabilities. In practice, the posterior probability of each small region M is labeled P, with corresponding subscript numbers, thus P = [P1, P2, P3...P...]. n ] (n≥100, P1, P2, P3...P n The P(A|B) value for each segment region is calculated as follows: P increases continuously based on the data reserve. Using box-type statistical analysis, the dataset P is analyzed to determine the theoretical maximum value (Q3 + 1.5 × IQR). Based on the "increase in the posterior probability value relative to the theoretical maximum value", four levels of status are defined: No risk: increase ≤ 25% (corresponding to no obvious risk signal); Slight risk: 25% < increase ≤ 50% (corresponding to slight risk signal); Moderate risk: 50% < increase ≤ 75% (corresponding to moderate risk signal); Strong risk: increase > 75% (corresponding to severe risk signal).

[0064] This case addresses issues such as parameter rigidity and subjective thresholds by deeply coupling Bayesian methods with box statistics to form a three-layer collaborative mechanism, as detailed below: (1) Data preprocessing: Based on the "Bayesian and box statistics method", step S1 data preprocessing (removing disturbance data 12 hours before and after the impact + invalid sensor data) provides pure data from the same source: box statistics (step S5): ensure that the sample for calculating the "theoretical maximum value (Q3+1.5×IQR)" is pure normal mining data, and avoid the risk of high threshold and underreporting caused by abnormal impact data; (2) Parameter calculation: Combining the "basic layer (fixed area hazard) + extended layer (dimensions increase or decrease with monitoring capabilities)" architecture, a "threshold-probability" linkage is formed: ① Box statistics (step S5) provide a unified state threshold for Bayesian (step S7): For all monitoring dimensions (including the extended layer), the standard of "no / weak / medium / strong" is divided by "theoretical maximum value + increase" to ensure that Bayesian probability transmission has a unified benchmark; ② Expanding dimensions and adjusting synchronously: When adding / removing monitoring items (such as micro-seismic events, drill cuttings), the box statistic calculates the new indicator threshold synchronously, and the Bayesian method synchronously completes the "parent-child node probability rule" for that dimension (fitting the logic of impact occurrence / non-occurrence scenarios). (3) Result optimization: Combining step S9 "preliminary update + deep optimization" with step S10 posterior probability level division, a closed loop is formed: ① Box statistics for Bayesian result classification: The posterior probability P(A|B) is classified into risk levels according to "theoretical maximum value + increase" to improve the practicality of engineering; ② Backward optimization parameters based on Bayesian results: If the level determination deviates from the actual impact situation (e.g., underreporting), backtrack and adjust the box threshold (e.g., tighten the risk-free increase) and Bayesian probability rules (e.g., increase the probability of high-risk combinations).

[0065] In summary, the two form a "mutual support-verification-optimization" relationship: box statistics provide a standardized benchmark, while Bayesian methods provide the basis for result verification, enabling full-process iteration and solving the defects of parameter rigidity and subjective thresholds.

[0066] Implementation Cases The 2201 longwall face of a mining group. This mine experienced a moderate rockburst on May 10, 2023, at 14:00. Currently, it is only equipped with a basic "stress monitoring + displacement monitoring" system, and a microseismic monitoring system was added in March 2024. Previously, when using a fixed structural model for assessment, the newly added microseismic equipment was not effectively integrated into the original mathematical model, resulting in low utilization of monitoring data; moreover, the model parameters were based on industry-standard data, which deviated significantly from the actual conditions of this longwall face, namely "uneven coal seam hardness and significant fault influence," leading to evaluation results that did not match the on-site risks. Now, the method of this invention is used for dynamic probability prediction. The specific process is as follows, with each step (corresponding to steps S1-S10 above): S1. Data Preprocessing: ① Data Source: Stress and displacement monitoring data from the 2201 working face from June 2023 to February 2024 are prioritized; due to insufficient microseismic data (added in March 2024), microseismic data from the 2203 working face of the same mine (similar geological conditions and mining technology to 2201) from 2022 to 2023 are supplemented. ② Data Removal: Stress / displacement data from 8:00 to 20:00 on May 10, 2023 (12 hours before and after the impact) are removed; sensor malfunction data (such as stress values ​​constantly at 0, displacement values ​​jumping, etc.) are removed; finally, valid data are retained.

[0067] S2. Calculation of prior probability P(A): The impact hazard index Wt of the 2201 working face is calculated using the comprehensive index method. Then P(A) = 0.655 (65.5%); the prior probability that rockburst will not occur is P(¬A) = 1 - Wt = 0.345 (34.5%).

[0068] S3. Target Area Segmentation: The target area M is the haulage roadway of the 2201 working face (100m in length). Segmentation is based on a suggested step distance d=5m: Total number of segments n=100÷5=20 segments, labeled M1 (1-5m), M2 (6-10m), ..., M 20 (96-100m); Assume the current advance is to M8.

[0069] S4. Constructing a likelihood layer prediction system: Based on the monitoring conditions of "stress + displacement + new microseismic events" in the mine, a likelihood layer of "basic layer + extended layer" is constructed, as shown in Table 8: Table 8 Specific data for each sequence S5. Determination of state thresholds based on box-type statistics (taking "stress value magnitude" as an example only): (1) Calculate the theoretical maximum value of the key parameters: ①Quarters Q1 = 20 MPa (the dividing value of the first 75% after data sorting), Q3 = 28 MPa (the dividing value of the first 25%). ② Interquartile range (IQR) = Q3 - Q1 = 28 - 20 = 8 MPa; ③ Theoretical maximum value = Q3 + 1.5 × IQR = 28 + 1.5 × 8 = 40 MPa (the upper limit of reference after excluding extreme outliers).

[0070] (2) Divide the state threshold into 4 levels: ① No danger: Stress value increase ≤ 25% ② Slight risk: 25% < stress value increase ≤ 50% ③ Medium risk: 50% < stress value increase ≤ 75% ④ High risk: Stress value increase > 75% (3) Determination of stress value of section M8: The measured value is 26MPa. (26MPa-40MPa) / 40MPa=-35%, -35%≤25%, so the stress value of section M8 is not dangerous.

[0071] S6. Likelihood layer monitoring evidence chain B generation: Based on the threshold judgment result of S, integrate the discrete states of each monitoring layer in segment M8 to form an evidence chain: B = Regional impact hazard "medium" + stress value "none" + stress growth rate "none" + displacement value "none" + displacement growth rate "none" + microseismic single event energy "none" + microseismic cumulative energy "none" + microseismic event frequency "none".

[0072] S7. Likelihood probability calculation (based on the probability transfer rules in Tables 6 and 7): The likelihood probability P(B|A) when the shock occurs and the likelihood probability P(B|¬A) when the shock does not occur need to be calculated separately. The core logic is the product of the probabilities of the parent node state to the child node state (only the key steps are shown).

[0073] (1) Calculate P(B|A) (for rockburst scenarios, refer to Table 6) Table 6 Core Rule: When an impact occurs, the probability of the child node being in a "strong" state is absolutely dominant (94%), while the parent node has minimal influence. Probability propagation at key levels (parent → child): ① Layer 1 (in the region) → Layer 2-1 (no stress value): 3% (Table 2 "in the region → none"); ② Layer 2-1 (no stress value) → Layer 2-2 (no stress increase): 1% (Table 2 "None → None"); ③ Layer 2-2 (no stress increase) → Layer 3-1 (no displacement value): 1% (Table 2 "None → None"); ④ Layer 3-1 (no displacement value) → Layer 3-2 (no displacement growth rate): 1% (Table 2 "None → None"); ⑤ Layer 3-2 (no displacement growth rate) → Layer 4-1 (no microseismic energy): 1% (Table 2 "None → None"); ⑥ Layer 4-1 (no microseismic energy) → Layer 4-2 (no cumulative energy): 1% (Table 2 "None → None"); ⑦ Layer 4-2 (cumulative energy none) → Layer 4-3 (frequency none): 1% (Table 2 "none → none").

[0074] Multiplication: P(B|A) = 3% × 1% × 1% × 1% × 1% × 1% × 1% = 0.03 × 0.01 × 0.01 × 0.01 × 0.01 × 0.01 ≈ 3 × 10 -13 .

[0075] (2) Calculate P(B|¬A) (for scenarios where rockburst does not occur, refer to Table 7) Table 7 Core Rules: Strictly adhere to the "parent-child node state correlation" (consistent > adjacent > spaced > opposite). Select key levels for probability propagation (parent → child): ① Layer 1 (in the region) → Layer 2-1 (no stress value): 10% (Table 3 "in the region → none", 1 level interval, low probability); ② Layer 2-1 (no stress value) → Layer 2-2 (no stress increase): 40% (Table 3 "no → no", consistent state, highest probability); ③ Layer 2-2 (no stress increase) → Layer 3-1 (no displacement value): 40% (Table 3 "no → no", consistent state, highest probability); ④ Layer 3-1 (no displacement value) → Layer 3-2 (no displacement growth rate): 40% (Table 3 "no" → "no", consistent state, highest probability); ⑤ Layer 3-2 (no displacement growth) → Layer 4-1 (no microseismic energy): 40% (Table 3 "no → no", consistent state, highest probability); ⑥ Layer 4-1 (no microseismic energy) → Layer 4-2 (no cumulative energy): 40% (Table 3 "no → no", consistent state, highest probability); ⑦ Layer 4-2 (No cumulative energy) → Layer 4-3 (No frequency): 40% (Table 3 "No → No", consistent state, highest probability).

[0076] Multiplication: P(B|¬A) = 10% × 40% × 40% × 40% × 40% × 40% × 40% = 0.1 × 0.4 6 =0.1×0.004096=0.0004096.

[0077] S8. Calculation of posterior probability P(A|B) (Bayes' theorem): (1) Calculate the evidence probability P(B) (the law of total probability) P(B) = P(B|A) × P(A) + P(B|¬A) × P(¬A) Substituting the data: P(B) = 3 × 10 -13 ×0.655+0.0004096×0.345≈0+0.0001413=0.0001413; (2) Calculate the posterior probability P(A|B) (Bayes' Theorem) P(A|B) = [P(B|A) × P(A)] / P(B) Substituting the data: P(A|B) = [3 × 10] -13 [×0.655] / 0.0001413≈(1.965×10 -13 ) / 0.0001413≈1.39×10 -9 (i.e., 0.000000139%).

[0078] S9. Data Storage and Updates: (1) Data reserves and initial learning: ① The mine has accumulated complete data chains for 150 segmented areas across the 2201 and 2203 working faces, meeting the data reserve requirements.

[0079] ② Preliminary update of S7 rules: Through frequency statistics, the system found that in the stored data, when the "Regional Hazard" is "Medium", the historical frequency of "Stress Value" actually being "None" is 12%, which is higher than the initial default value (10%). Therefore, the system updated the probability of "Medium → None" in the S7 probability table (Table 3) from 10% to 12%.

[0080] ③ Initial update of S5 threshold: After adding new data, the theoretical maximum value of "stress value" is recalculated and updated from 40MPa to 42MPa. Its state thresholds of "none, weak, medium and strong" are also automatically adjusted accordingly.

[0081] (2) Result verification and in-depth optimization: ① Deviation detected: The system detected M 160 Segmentation: The posterior probability of decision S10 is 1.4 × 10⁻⁶. -9 The initial reading was "no danger," but three days later, slight vibrations of the coal seam were observed at the site, indicating a precursor to the impact.

[0082] ② Backtracking and optimizing the S5 threshold: Analysis of M 160 The data showed a "stress value increase" of 23%, which was initially classified as "none" according to the original S5 threshold (≤25% is considered no danger). Based on this and similar underreporting cases, the original threshold was deemed too lenient. Therefore, the standard for "no danger in stress value" increase is tightened from ≤25% to ≤20%.

[0083] ③ Backtracking and optimizing the S7 rule: Statistical analysis revealed that in multiple cases with precursory shocks, the actual frequency of the "weak stress → medium displacement" state combination was 30%, far higher than the initial probability (20%) in the S7 table. Therefore, the probability of this combination was increased from 20% to 30% to enhance the identification of the risk signal of this combination.

[0084] ④ Verify the optimization effect: Using the optimized S5 threshold and S7 rule, the subsequent M... 160 To M 170 Risk assessment was conducted segment by segment. Results showed that the accuracy of the hazard level assessment compared to the actual on-site risk improved from 88% to 96%. This signifies the success of the closed-loop optimization, the adoption of new parameters by the system, and the effective improvement of model accuracy.

[0085] S10, Posterior probability matching risk level: Calculate the posterior probability of the 150 pre-existing segments strictly according to the given formula (P=[1.2×10). -9 1.5×10 -9 ,...,4.2×10 -9 ]) Perform box type statistics: (1) Calculate key parameters: ①Q3=3.5×10 -9 (25th percentile, the cutoff value of the posterior probability in the top 25%) ②IQR = Q3 - Q1 = 3.5 × 10 -9 -2.1×10 -9 =1.4×10 -9 ; ③ Theoretical maximum value = Q3 + 1.5 × IQR = 3.5 × 10 -9 +1.5×1.4×10 -9 =5.6×10 -9 .

[0086] (2) Classification of risk levels (based on a given rate of increase formula): ① No danger: Price increase ≤ 25% (including negative values); ② Slight risk: 25% < increase ≤ 50%; ③Medium risk: 50% < increase ≤ 75%; ④ High risk: Price increase > 75%.

[0087] (3) Classification of segment M8: Posterior probability of segment M8 = 1.39 × 10 -9 The increase = (1.39 × 10) -9 -5.6×10 -9 ) / 5.6×10 -9 ×100%≈-75.2% (negative value) → According to the rules, it is judged as "no danger"; On-site prevention and control recommendations: maintain the regular monitoring frequency, no additional prevention and control measures are required.

[0088] In summary, the performance advantage of this case stems entirely from targeted improvements to the shortcomings of D1 and D2 technologies. Considering the differences in their technical characteristics, the performance can be explained from the following four core dimensions: (1) Model adaptability: This model can be flexibly adapted to different mine monitoring conditions, while D1 and D2 have poor adaptability. ①D1 / D2 defects: The "dynamic-static coupling evaluation" framework of D1 and the "impact event probability evaluation system" of D2 are both rigid structures (D1 is only suitable for microseismic + stress monitoring, while D2 has fixed monitoring index dimensions). They cannot be adjusted according to the differences in mine monitoring capabilities (such as old mines without microseismic monitoring and newly built mines with added ground sound monitoring), resulting in poor versatility (corresponding to D1 / D2 technical bottleneck 1). ② Advantages of this application: Through the “basic layer + extended layer” architecture in step S4 (the basic layer retains the regional hazards, and the extended layer increases or decreases the dimensions according to the monitoring capabilities; if there is no drill cuttings monitoring, the corresponding layer is deleted, and if microseismic monitoring is added, the microseismic layer is inserted), it can adapt to the differentiated monitoring conditions of old / new mines without reconstructing the model, and its adaptability is significantly better than D1 and D2.

[0089] (2) Parameter dynamism: This case can be continuously optimized based on its own data, while the D1 and D2 parameters are fixed and not updated. ① Parameter fixation defects of D1 / D2: D1: The warning threshold and Bayesian prior probability are both empirical / general presets, and the Bayesian probability propagation rules are fixed; it cannot be optimized using mine-specific impact cases and monitoring data, the parameters are not adjusted with the mining process, and the accuracy is not improved; D2: The normal standard range is only based on the initial data mean (impact disturbance data is not removed, and the baseline is easily distorted), and the impact impact degree quantification coefficient is manually preset; there is no dynamic update mechanism, and the parameters accumulate deviations from the actual working conditions after long-term use.

[0090] ② The dynamic optimization advantages of this case (based on Bayesian and boxed statistics methods): Through a two-layer mechanism of "initial update of data storage + in-depth optimization of results verification", the parameters are iterated throughout the entire lifecycle. a. First layer (preliminary update of step S9): After accumulating ≥100 segmented area data, optimize the box statistical threshold of step S5 (such as adjusting the stress without dangerous increase) and the Bayesian probability rule of step S7 (such as the probability of "no stress value in the area") using frequency statistics to adapt to the conventional laws of the mine. b. Second layer (result closed-loop optimization): Compare the hazard level in step S10 with the actual impact situation on site (such as missed / false alarms), backtrack and optimize the threshold in step S5 (such as tightening the stress increase from ≤25% to ≤20% when there is no hazard) and the probability rules in step S7 (such as increasing the probability of "weak stress → medium displacement"), and correct the deviation in special scenarios. The two layers form a closed loop of "data accumulation → initial update → verification → deep optimization", avoiding the accuracy decay caused by the "parameter solidification" of D1 and D2.

[0091] (3) Threshold objectivity: The thresholds in this case are generated based on pure data statistics, while the D1 and D2 thresholds are subjective / distorted.

[0092] ①D1 / D2 defects: The D1 hazard level threshold is based on empirical preset (e.g., 0≤P<0.25 is no hazard) and does not take into account the geological differences of the mine; D2 sets the normal range through the average of data, but does not remove impact disturbance data (e.g., abnormal data 12 hours before and after the impact), resulting in the "inherent distortion" of the benchmark threshold (corresponding to D1 / D2 technical bottleneck 3). ② Advantages of this case: Step S1 first removes "12-hour disturbance data before and after the impact + sensor failure data" to ensure that the sample used for box-type statistics in step S5 is "pure normal mining data"; then the theoretical maximum value is calculated by "Q3 + 1.5 × IQR", and the threshold is divided by "the increase of the monitored value relative to the theoretical maximum value". It is based entirely on the mine's own reliable data statistics, without experience interference. The objectivity of the threshold is far better than D1 and D2, which can reduce "normal data being misjudged as risk" (false alarm) or "risk data being missed as normal" (missed alarm).

[0093] (4) Risk assessment relevance: The application of Bayesian scenarios in this case is more in-depth, and D1 and D2 cannot reflect the actual risk association. ① D1 / D2 defects: D1 only uses Bayesian methods for multi-model result fusion and does not construct a differential probability transmission of "impact occurred / did not occur" to the lower-level data; D2 Bayesian network variables have a single dimension (only focusing on the probability of impact - degree of impact), which cannot reflect the risk evolution law in a refined way (corresponding to D1 / D2 technical bottleneck 4). ② Advantages of this case: Step S7 is divided into two scenarios, "impact occurred" and "impact did not occur", and the probability propagation rules are designed. When an impact occurs, the "strong danger signal dominates" (94% probability) which is consistent with the actual pattern of strong anomalies in data during an impact. When an impact does not occur, the "parent-child node state correlation" (consistent > adjacent > interval > opposite) is followed, which is consistent with the gradual change characteristics of risk signals under normal mining. This makes the probability reasoning more consistent with the actual risk evolution on site, and the degree of consistency is better than D1 and D2.

[0094] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for dynamically predicting the probability of rock burst occurrence in a coal mine, characterized in that, The method comprises the following steps: Step S1: preprocessing the historical monitoring data of the target working face; Step S2: calculating the quantitative danger index as the prior probability of Bayes by using the comprehensive index method in the impact danger evaluation; Step S3: dividing the target area into n sub-regions according to a preset step distance; Step S4: constructing a prediction system, wherein the prediction system comprises a prior layer and a likelihood layer, and the likelihood layer comprises a basic layer and an extension layer; Step S5: setting the monitoring layers of the likelihood layer as multi-level discrete states; except for the basic layer, the determination threshold of the multi-level discrete state is determined by using the box-type statistical method for the remaining layers; Step S6: determining the current discrete state of each monitoring layer of the likelihood layer according to the determination threshold, integrating the current states of all monitoring layers, and generating a likelihood layer monitoring evidence chain; Step S7: setting the initial default dynamic probability transfer rule of the state of each monitoring layer of the likelihood layer; Step S8: for each sub-region, calculating the impact ground pressure posterior probability based on the Bayes formula according to the prior probability, the likelihood layer monitoring evidence chain and the probability transfer rule.

2. The coal mine rock burst occurrence probability dynamic inference method according to claim 1, characterized in that, The step S1 comprises: If the historical monitoring data of the target working face is insufficient, the historical data of other working faces in the same mine and under similar geological conditions is supplemented; And / or, the step S1 comprises: All monitoring data during and within 12 hours before and after the impact ground pressure event is removed; Synchronously, invalid data is removed, and only reliable data verified on site is retained.

3. The coal mine rock burst occurrence probability dynamic inference method according to claim 1, characterized in that, In the step S4, the prior layer is used to quantify the basic probability of the occurrence of the impact ground pressure of the target working face under the current geological mining conditions, the basic layer is used to quantify the basic probability of the occurrence of the impact danger of the target working face in the current region, and the extension layer is used to quantify the basic probability of the occurrence of the state of the current monitoring layer of the target working face.

4. The coal mine rock burst occurrence probability dynamic inference method according to claim 1, characterized in that, The step S5 comprises: Step S51: calculating the theoretical maximum value of each type of monitoring parameter in the statistical sample set; Step S52: dividing each monitoring layer of the likelihood layer except the basic layer into four discrete states of "none, weak, medium and strong" according to the increase of the monitoring value relative to the theoretical maximum value as the determination basis.

5. The coal mine rock burst occurrence probability dynamic inference method according to claim 1, characterized in that, In the step S7, the probability transfer rule comprises: The state of the prior layer comprises: impact ground pressure occurrence and impact ground pressure non-occurrence; The initial probability of the regional impact danger in the basic layer adopts the uniform distribution principle, and the initial default probability of each state is set to 25%, ensuring that the probability sum is 100%.

6. The coal mine bump occurrence probability dynamic inference method according to claim 5, characterized in that, In the step S7, the probability transfer rule further comprises: When the impact ground pressure occurs, the probability distribution of the monitoring data state of each extension layer is not affected by the state of the parent node of the upper layer, and only follows the engineering law that the dangerous signal is strong and the proportion is high when the impact occurs, that is, the initial default probability is uniformly set to "none": 1%, "weak": 2%, "medium": 3%, and "strong": 94%, and the probability distribution remains fixed; When the impact ground pressure does not occur, the probability of the monitoring data state of each extension layer needs to strictly follow the correlation rules of the parent node and the child node, that is, the state of the parent node of the upper layer directly affects the probability of the monitoring data state of the next layer, and the correlation logic is summarized as four priority principles: ①State completely consistent: the parent-child node danger level matches, the relevance is the strongest, and the probability is the highest; ②State adjacent: the difference in danger level is small, the relevance is second, and the probability is medium; ③State interval 1 level: the difference in danger level is large, the relevance is weak, and the probability is low; ④State completely opposite: the danger level is completely opposite, the relevance is the weakest, and the probability is the lowest.

7. The coal mine bump occurrence probability dynamic inference method according to claim 6, characterized in that, The step S8 comprises: Step S81: calculating the likelihood probability when the rock burst occurs; Step S82: calculating the likelihood probability when the rock burst does not occur; Step S83: calculating the evidence probability of the likelihood layer monitoring evidence chain according to the total probability formula; Step S84: calculating the rock burst occurrence probability based on the real-time monitoring evidence, i.e., the posterior probability, according to the Bayes theorem.

8. The coal mine bump occurrence probability dynamic inference method according to claim 1, characterized in that, The method further comprises: Step S9: data reservation and preliminary learning; Wherein, the data reservation: continuously collects and stores the complete data chain of each segmented sub-region, forms a growing mine dedicated database, and when the cumulative data reaches a certain scale, the preliminary learning is started; The preliminary learning: using the frequency statistical method, the historical data in the database are statistically analyzed and the probability transfer rule of step S7 and the determination threshold of step S5 are updated.

9. The coal mine bump occurrence probability dynamic inference method according to claim 8, characterized in that, The method further comprises: Step S10: calculating the posterior probability of the target region and matching the danger level; Wherein, the step S10 comprises: Combining the posterior probabilities of each sub-region to obtain the posterior probability of the target region; for the posterior probability of the target region, the box-type statistical analysis is used to determine the theoretical maximum value; the rise of the posterior probability value relative to the theoretical maximum value is used as the determination basis to divide into multiple discrete states.

10. The coal mine bump occurrence probability dynamic inference method according to claim 9, characterized in that, The step S9 further comprises: result verification and depth optimization; Wherein, the result verification: comparing the danger level determination result output by step S10 with the actual rock burst occurrence on site one by one; The depth optimization: when the danger level determination deviates from the actual situation, the layer optimization is started, and the deviation that cannot be corrected by the preliminary learning is adjusted, i.e., the probability transfer rule of step S7 and the determination threshold of step S5 are optimized.

Images