Mine risk intelligent decision-making method and system

Abstract

The invention provides a mine risk intelligent decision-making method and system, and relates to the technical field of mine risk assessment, and the method comprises the steps: collecting real-time data, historical data, the type of the historical data, and a historical risk strategy of a mine; constructing a risk decision model, and training the risk decision model by using the historical data and the historical risk strategy to obtain a trained risk decision model; and inputting real-time data into the trained risk decision model to obtain a real-time risk strategy. According to the method, the abnormal real-time data is judged by constructing the risk decision model, and the conditional probability table of the Bayesian network sub-model is updated based on the conditional probability of the abnormal real-time data, so that the Bayesian network model is optimized, the adaptability of the Bayesian network model structure to a mine risk dynamic change environment is improved, and the safety of the mine risk dynamic change environment is improved. And the accuracy of a mine risk assessment result is further improved.

Description

Technical Field

[0001] The present application relates to the technical field of mine risk assessment, and in particular to a mine risk intelligent decision-making method and system. Background Art

[0002] Mining is an important pillar industry of the national economy, but its operating environment is complex and harsh, facing multiple risks such as geological disasters (such as landslides, gas explosions), equipment failures, human operating errors, and tailings dam breaches. Traditional mine risk management relies on manual experience judgment and static data records, and there are problems such as delayed risk identification, strong subjectivity in assessment, and extensive management and control measures. At the same time, as the mining depth increases, mines face extreme geological conditions such as high ground stress and high gas pressure; at the same time, dynamic operating scenarios exacerbate the randomness, coupling and complexity of risk factors. In addition, derivative risks such as tailings dam breaches and environmental pollution further increase the difficulty of mine risk management. In this context, there is an urgent need for accurate prediction and real-time dynamic management of mine risks.

[0003] In related technologies, historical data is used to construct a risk evolution topology to train the Bayesian network model, and the trained model is used to assess mine risks. However, the risk environment of mines is dynamically changing. With the progress of mine operations and equipment aging, the relationship between various risk factors may also change, resulting in changes in the mutual influence relationship between various mine risks. When the mine risk environment changes, the original risk evolution topology may not be able to accurately describe the relationship between the current risk factors, making the Bayesian network model unable to adapt to this dynamic change in time, resulting in the risk assessment results not being consistent with the actual situation, affecting the accuracy of the mine risk assessment results. Summary of the invention

[0004] In order to improve the accuracy of mine risk assessment results, the present application provides a mine risk intelligent decision-making method and system.

[0005] In the first aspect, the present application provides a method for intelligent decision-making of mine risks, which adopts the following technical solutions: A mine risk intelligent decision-making method, the method comprising: Data collection: collect real-time data, historical data, types of historical data and historical risk strategies of the mine; Model construction: construct a risk decision model, which includes an identification sub-model, a Bayesian network sub-model and a decision sub-model; Model training: Use historical data and historical risk strategies to train the risk decision model to obtain the trained recognition sub-model, the trained Bayesian network sub-model and the trained decision sub-model; First recognition: input the real-time data into the trained recognition sub-model to obtain abnormal real-time data, which is recorded as the first data, and integrate the types of abnormal data in the first data into the first combination; First calculation: based on the historical data, the first data and the first combination, the target conditional probability is calculated according to the conditional probability calculation formula; First judgment: determine whether the target conditional probability belongs to the conditional probability table of the trained Bayesian network sub-model: If yes, then the first evaluation step is performed; If not, then execute the first optimization step; First optimization: adding the target conditional probability to the conditional probability table of the trained Bayesian network sub-model to obtain a new Bayesian network sub-model, and updating the new Bayesian network sub-model to the trained Bayesian network sub-model; First evaluation: input the first data into the trained Bayesian network sub-model to obtain the second data; Strategy acquisition: Input the second data into the trained decision sub-model to obtain a real-time risk strategy.

[0006] By adopting the above technical scheme, based on the constructed risk decision model, the identification sub-model is used to determine whether there is abnormal real-time data, based on the abnormal real-time data, the abnormal data combination is obtained, and the conditional probability of the abnormal data combination is calculated based on the abnormal data combination. By judging whether the conditional probability of the abnormal data combination belongs to the conditional probability table of the trained Bayesian network sub-model, it is judged whether there is a change in the architecture of the Bayesian network model, and the conditional probability table of the Bayesian network sub-model is updated based on the judgment result to optimize the architecture of the Bayesian network model, so that it can adapt to the new risk environment and improve the accuracy of risk assessment. The present application uses the conditional probability of the abnormal data combination to judge whether the architecture of the Bayesian network model needs to be updated, and then dynamically optimizes the Bayesian network structure, which helps to improve the adaptability of the Bayesian network model structure to the dynamically changing environment of mine risks, improves the real-time, effectiveness and accuracy of the Bayesian network model, further improves the adaptability and accuracy of mine risk assessment, and enhances the immediacy of mine safety management.

[0007] Optionally, after performing the first identification step and before performing the first calculation step, the method further includes: First clustering: clustering analysis is performed on historical data using a clustering algorithm to obtain several clusters of historical data; Second calculation: calculate the Euclidean distance between the first data and the cluster center of each cluster, recorded as the first distance; First judgment: judging whether the first distances are all less than a preset distance threshold; If yes, then the first evaluation step is performed; If not, the first calculation step is performed.

[0008] By adopting the above technical solution, clustering algorithms are used to perform cluster analysis on historical data, and the similarity between real-time data and historical data is further judged by Euclidean distance, so as to avoid triggering model updates due to short-term fluctuations or noise, and improve the accuracy of abnormal data prediction. In addition, the Euclidean distance between real-time data and the cluster center of each historical cluster cluster is compared by clustering algorithm to determine whether to trigger model updates, which helps to reduce the possibility of misjudgment of data anomalies, helps to improve the adaptability of Bayesian network models to dynamic changes, improves the prediction accuracy of abnormal data, and at the same time, helps to optimize computing resources and improve the stability and adaptability of mine risk assessment.

[0009] Optionally, after the step of performing the first judgment and before the step of the first calculation, the method further includes: First processing: extracting data of the same type as in the first combination from the historical data, recorded as first extracted data, and extracting data of the same type as in the first combination from the first data, recorded as second extracted data; Third calculation: using a local outlier factor algorithm to calculate the local reachable density of the first extracted data and the second extracted data, and calculating the local outlier factor of the second extracted data based on the calculated local reachable density; Second judgment: judging whether the local outlier factors of the second extracted data are not lower than the preset local outlier factor threshold: If so, no action will be taken; If not, the type of the second extracted data that is higher than the preset local outlier factor threshold is recorded as a second combination, and the second combination is updated to the first combination.

[0010] By adopting the above technical solution, the local outlier factor of historical data and real-time data is calculated by the local outlier factor algorithm to judge the local abnormality of real-time data, and the type of abnormal data is updated based on the judgment result, which helps to accurately judge the dynamic change factors in the mine risk environment, reduces the possibility of misjudgment caused by a single or a small amount of abnormal data, and helps to improve the pertinence and accuracy of the Bayesian network structure adjustment, and improves the accuracy of risk assessment results. In addition, the LOF algorithm is used to pre-screen the type of abnormal data, and only the category of data with a high degree of outliers is updated, which reduces the possibility of executing the Bayesian network model update for all abnormal data, reduces unnecessary computing overhead, optimizes the allocation of computing resources, improves the operation efficiency of the model, and improves the accuracy and efficiency of mine risk assessment.

[0011] Optionally, based on the trained Bayesian network sub-model, obtaining the risk category of the historical data at the same timestamp, after performing the first discrimination step and before performing the first optimization step, further includes: First screening: sort the first distances from largest to smallest, and record the smallest first distance as the minimum distance; Second screening: record the cluster corresponding to the minimum distance as the first cluster, and record the risk category of the historical data corresponding to the first cluster as the first label; First labeling: labeling the first data with the first label as the risk category to obtain the labeled first data, which is recorded as new first data.

[0012] By adopting the above technical solution, based on cluster analysis and Euclidean distance sorting, the risk category most similar to the first data is matched from the historical data as the risk category of the first data, thereby improving the rationality of the risk category of the abnormal data. In addition, by dynamically labeling the risk category of the abnormal data, it is helpful to dynamically adjust the risk category mapping based on the labeled risk category of the first data when the change of the mine operating environment causes the risk category to evolve, so as to achieve the optimization of the Bayesian network model structure, which helps to improve the ability of the Bayesian network model to adapt to changes in the risk environment, so that the Bayesian network model can reflect the latest risk factor relationship and improve the accuracy of the risk assessment results.

[0013] Optionally, the risk decision model further includes a semi-Markov sub-model, and after performing the first labeling step and before performing the first optimization step, further includes: First extension: defining the state space of the semi-Markov sub-model based on the first data, the historical data and the risk category of the historical data; Fourth calculation: using the maximum likelihood estimation method to calculate the residence time of each state in the state space, and recording the calculated residence time as the first time; Matrix construction: Based on the first time, first data and historical data, the maximum likelihood estimation method is used to calculate the transition probability of each state in the state space, and the transition probability matrix is ​​constructed according to the calculated transition probability.

[0014] By adopting the above technical solution, a semi-Markov sub-model is introduced on the basis of the Bayesian network model, and the residence time distribution of different risk states is characterized based on the semi-Markov sub-model, thereby improving the description accuracy of the evolution of complex risk states. At the same time, by constructing a state transition probability matrix, the evolution law between different risk states is established, which helps the Bayesian network model to have time series self-adaptation capabilities, optimizes the time series correlation between data types, helps the Bayesian network model to adjust according to the dynamic changes of the mining operating environment, and improves the accuracy of risk prediction of the Bayesian network model.

[0015] Optionally, based on the trained Bayesian network sub-model, obtaining the edges between nodes in the Bayesian network sub-model, after executing the step of matrix construction and before executing the step of first optimization, further includes: Third judgment: Based on the risk category of the first combination and the first data, it is judged whether there is an edge between the nodes corresponding to the first combination in the Bayesian network sub-model: If so, no action will be taken; If not, based on the first combination, in the Bayesian network sub-model, adding edges between nodes corresponding to the first combination; Probability update: based on the first data, historical data and the transition probability matrix, the conditional probability between each node in the Bayesian network sub-model is calculated by using the maximum likelihood estimation method, and the conditional probability table is updated according to the calculation result to obtain an updated Bayesian network sub-model, and the updated Bayesian network sub-model is used as the new trained Bayesian network sub-model; In a first evaluation step, first data is input into the new trained Bayesian network sub-model.

[0016] By adopting the above technical solution, dynamically adding edges between nodes based on the first combination helps to update the topological structure of the Bayesian network model in real time, ensuring that the network structure of the Bayesian network model can reflect the association between the types of the latest data. In addition, the conditional probability table is updated based on the state transition probability matrix of the semi-Markov sub-model, which improves the accuracy of the probabilistic reasoning of the Bayesian network model, so that the Bayesian network model can not only adapt to the newly emerging state data, but also predict the evolution direction of risks based on the evolution trend of historical states combined with time factors, thereby improving the accuracy of risk assessment of the Bayesian network model.

[0017] Optionally, after the step of executing the policy acquisition step, the method further includes: Second processing: based on the second data, selecting a risk strategy corresponding to the risk category of the second data from a preset strategy library, recorded as a first risk strategy; Third processing: Based on the second data, the first risk strategy and the real-time risk strategy are simulated by using the Monte Carlo simulation method to obtain a risk indicator sequence for each risk strategy; Fourth process: define the reference risk indicator sequence of the risk strategy based on the ideal state; Fifth calculation: the correlation between the risk indicator sequence of each risk strategy and the reference risk indicator sequence calculated by the grey correlation analysis method, and the calculated correlation is recorded as the indicator correlation; Fifth processing: sort the obtained indicator correlations in ascending order, and select the risk strategy with the highest indicator correlation as the new real-time risk strategy.

[0018] By adopting the above technical scheme, the first risk strategy and the real-time risk strategy are simulated by Monte Carlo simulation to obtain the risk indicator sequence, and the reference risk indicator is constructed based on the ideal state. The grey correlation analysis method is used to analyze the correlation between the risk indicator sequence and the reference risk indicator sequence, and the advantages and disadvantages of different risk strategies are quantified. The strategy that best suits the current mine risk status is selected, which is helpful to dynamically optimize the risk strategy so that the mine risk strategy always matches the latest risk environment, improves the scientificity, accuracy and adaptability of the risk strategy selection, and enhances the dynamic response capability of mine safety management.

[0019] Optionally, the second data includes a risk level; after the step of executing the strategy acquisition, the method further includes: First setting: Set up the mine data based on the real-time risk strategy as test data; Second evaluation: input the test data into the trained Bayesian network sub-model to obtain the third data; Fourth judgment: judging whether the risk level of the third data is lower than the risk level of the second data: If so, the real-time risk strategy is executed; If yes, execute the first setting step; First setting: determine the optimization goal, which includes maximizing safety, minimizing resources, and minimizing costs; Function construction: construct the objective function based on the optimization goal; First definition: using an optimization algorithm to optimize the model parameters of the trained decision sub-model to obtain first model parameters, and using the first model parameters as new model parameters of the trained decision sub-model; Second definition: select a risk strategy corresponding to the risk category of the historical data from the preset strategy library as a new historical risk strategy; Second optimization: train the trained decision sub-model based on historical data, new historical risk strategies and new model parameters to obtain a new trained decision sub-model, update the new trained decision sub-model to the trained decision sub-model, and then execute the strategy acquisition step.

[0020] By adopting the above technical solution, the selected risk strategy is tested, and the effectiveness of the risk strategy is ensured by introducing risk level comparison. At the same time, the parameters of the decision sub-model are optimized based on the optimization algorithm and optimization target, which enhances the real-time and accuracy of the decision sub-model. In addition, the use of a cyclic optimization mechanism helps to improve the adaptability of the decision sub-model to the dynamically changing environment of the mine, improves the accuracy of the risk assessment of the risk decision model, and enhances the scientificity and reliability of the risk strategy of the risk decision model.

[0021] In the second aspect, the present application provides a mine risk intelligent decision-making system, which adopts the following technical solutions: A mine risk intelligent decision-making system, the system comprising: processor and memory, The memory stores program code; When the processor calls the program code in the memory, the steps of the method described in the first aspect are executed.

[0022] By adopting the above technical solution, based on the constructed risk decision model, it is determined whether there is abnormal real-time data, based on the abnormal real-time data, the abnormal data combination is obtained, and the conditional probability of the abnormal data combination is calculated based on the abnormal data combination to determine whether the architecture of the Bayesian network model needs to be updated, and then the Bayesian network model structure is dynamically optimized, which helps to improve the adaptability of the Bayesian network model structure to the dynamically changing environment of mine risks, improves the real-time, effectiveness and accuracy of the Bayesian network model, further improves the adaptability and accuracy of mine risk assessment, and enhances the immediacy of mine safety management.

[0023] In summary, the present application includes at least one of the following beneficial technical effects: 1. Based on the constructed risk decision model, the identification sub-model is used to determine whether there is abnormal real-time data, based on the abnormal real-time data, the abnormal data combination is obtained, and the conditional probability of the abnormal data combination is calculated based on the abnormal data combination. By judging whether the conditional probability of the abnormal data combination belongs to the conditional probability table of the trained Bayesian network sub-model, it is judged whether there is a change in the architecture of the Bayesian network model, and the conditional probability table of the Bayesian network sub-model is updated based on the judgment result to optimize the architecture of the Bayesian network model, so that it can adapt to the new risk environment and improve the accuracy of risk assessment. This application uses the conditional probability of abnormal data combinations to determine whether the architecture of the Bayesian network model needs to be updated, and then dynamically optimizes the Bayesian network structure, which helps to improve the adaptability of the Bayesian network model structure to the dynamically changing environment of mine risks, improves the real-time, effectiveness and accuracy of the Bayesian network model, further improves the adaptability and accuracy of mine risk assessment, and enhances the immediacy of mine safety management.

[0024] 2. Use clustering algorithms to perform cluster analysis on historical data, and further judge the similarity between real-time data and historical data through Euclidean distance, so as to avoid triggering model updates due to short-term fluctuations or noise, and improve the accuracy of abnormal data prediction. In addition, the clustering algorithm compares the Euclidean distance between real-time data and the cluster center of each historical cluster to determine whether to trigger model updates, which helps to reduce the possibility of misjudgment of data anomalies, helps to improve the adaptability of the Bayesian network model to dynamic changes, improves the prediction accuracy of abnormal data, and at the same time, helps to optimize computing resources and improve the stability and adaptability of mine risk assessment.

[0025] 3. A semi-Markov sub-model is introduced on the basis of the Bayesian network model. The residence time distribution of different risk states is characterized based on the semi-Markov sub-model, which improves the description accuracy of the evolution of complex risk states. At the same time, by constructing a state transition probability matrix, the evolution law between different risk states is established, which helps the Bayesian network model to have time series adaptability, optimizes the time series correlation between data types, helps the Bayesian network model to adjust according to the dynamic changes of the mining operating environment, and improves the accuracy of risk prediction of the Bayesian network model.

[0026] 4. Dynamically adding edges between nodes based on the first combination helps to update the topological structure of the Bayesian network model in real time, ensuring that the network structure of the Bayesian network model can reflect the association between the types of the latest data. In addition, the conditional probability table is updated based on the state transition probability matrix of the semi-Markov sub-model, which improves the accuracy of the probability reasoning of the Bayesian network model, so that the Bayesian network model can not only adapt to the newly emerging state data, but also predict the evolution direction of risks based on the evolution trend of historical states combined with time factors, thereby improving the accuracy of risk assessment of the Bayesian network model. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 is a flow chart of Example 1 of the present application; Figure 2 This is a flow chart of S41 data determination in Example 1 of the present application; Figure 3 is a flow chart of S61 model optimization of Example 2 of the present application; Figure 4 is a flowchart of S91 strategy optimization of Example 3 of the present application; Figure 5 This is a flowchart of S10 strategy verification of Example 4 of the present application. DETAILED DESCRIPTION

[0028] The following combination Figures 1 to 5 This application is described in further detail.

[0029] Embodiment 1: This embodiment discloses a method for intelligent decision-making of mine risks, such as Figure 1 As shown, the method includes: collecting real-time data, historical data, types of historical data and historical risk strategies of the mine; constructing a risk decision model, the risk decision model includes an identification sub-model, a Bayesian network sub-model and a decision sub-model; using historical data and historical risk strategies to train the risk decision model, to obtain a trained identification sub-model, a trained Bayesian network sub-model and a trained decision sub-model, inputting real-time data into the trained identification sub-model to obtain abnormal real-time data, and integrating the types of abnormal data in the abnormal real-time data into a first combination; calculating the target conditional probability of the first combination, and determining whether the target conditional probability belongs to the conditional probability table of the trained Bayesian network sub-model, if so, updating the Bayesian network sub-model, if not, not updating the Bayesian network sub-model, and then inputting the abnormal real-time data into the Bayesian network sub-model to obtain second data; inputting the second data into the trained decision sub-model to obtain a real-time risk strategy. This embodiment includes the following steps: S1 Data Collection: Collect the mine’s real-time data, historical data, historical data types, risk category labels for historical data, and historical risk strategies.

[0030] Real-time data includes equipment health data, mine environment data, employee operation data and production management data. Equipment health data includes equipment vibration data, equipment temperature data, equipment pressure data, equipment working status data and equipment current and voltage data; mine environment data includes temperature data, humidity data, wind speed data, wind direction data, precipitation data, gas concentration data, carbon dioxide concentration data, soil moisture data, gas temperature data in the mine, mine air flow data, blasting vibration data, pressure data in the mine, mine water level data, mine geological data and smoke concentration data; employee operation data includes employee daily operation record data and employee location data; production management data includes supplier information data, logistics and transportation data, procurement data, inventory data, production progress data, ore extraction data, ore extraction quality data, mining machinery load data and production efficiency data. The real-time risk data can be obtained through mine site monitoring systems, various sensors, equipment health monitoring platforms and other channels.

[0031] Historical data includes historical equipment health data, historical mine environment data, historical employee operation data and historical production management data. Historical equipment health data includes historical equipment vibration data, historical equipment temperature data, historical equipment pressure data, historical equipment working status data, historical equipment current and voltage data, etc. Historical mine environment data includes historical temperature data, historical humidity data, historical wind speed data, historical wind direction data, historical precipitation data, historical gas concentration data, historical carbon dioxide concentration data, historical soil moisture data, historical mine gas temperature data, historical mine air flow data, historical blasting vibration data, historical mine pressure data, historical mine water level data, historical mine geological data and historical smoke concentration data, etc. Historical employee operation data includes historical employee daily operation record data and historical employee location data; historical production management data includes historical supplier information data, historical logistics and transportation data, historical procurement data, historical inventory data, historical production progress data, historical ore extraction volume data, historical ore extraction quality data, historical mining machinery load data and historical production efficiency data.

[0032] The types of historical data include equipment vibration, equipment temperature, equipment pressure, equipment working status, equipment current and voltage, air temperature, humidity, wind speed, wind direction, precipitation, gas concentration, carbon dioxide concentration, soil moisture, gas temperature in the mine, mine air flow, blasting vibration, mine pressure, mine water level, mine geology, smoke concentration, employee daily operation records, employee location, supplier information, logistics and transportation, procurement, inventory, production progress, ore extraction volume, ore extraction quality, mining machinery load and production efficiency, etc.

[0033] In this embodiment, the real-time data is a set of equipment health data, mine environment data, employee operation data and production management data at the same timestamp, and the historical data is actually a historical data set, which contains several historical data at different times, and each historical data is a set of historical equipment health data, historical mine environment data, historical employee operation data and historical production management data at the same timestamp. For ease of understanding, some historical data are listed as examples, see Table 1.

[0034] Table 1 Historical data of mines serial number time Air temperature (℃) humidity(%) Wind speed (m / s) Precipitation (mm / h) Gas concentration (%) Carbon dioxide concentration (ppm) 1 2024.12.1 24 75 3.2 0 1.2 0.05 2 2024.12.4 20 68 2.5 5 1.0 0.04 3 2024.12.8 22 65 4.0 1 1.3 0.06 The risk category labels of historical data include equipment risks-electromechanical accidents, equipment risks-roof accidents, equipment risks-gas accidents, equipment risks-transportation accidents, equipment risks-blasting accidents, environmental risks-water accidents, environmental risks-fire accidents, environmental risks-air pollution, environmental risks-dust and geological disasters, personnel risks-object strikes, personnel risks-electromechanical injuries, personnel risks-vehicle injuries, personnel risks-falling from heights, personnel risks-electric shock injuries, personnel risks-lifting injuries, personnel risks-scalding injuries, personnel risks-poisoning and suffocation, production management risks-production scheduling errors, production management risks-inadequate safety management, production management risks-quality management issues and production management risks-non-compliance issues.

[0035] The historical risk strategy is a risk resolution strategy executed when historical data is at mine risk.

[0036] S2 model construction: Construct a risk decision model, which includes an identification sub-model, a semi-Markov sub-model, a Bayesian network sub-model and a decision sub-model.

[0037] The identification sub-model is an isolation forest model, and the decision sub-model uses a long short-term memory network model as a basic model.

[0038] S3 model training: Use historical data and historical risk strategies to train the risk decision model to obtain the trained recognition sub-model, trained semi-Markov sub-model, trained Bayesian network sub-model and trained decision sub-model.

[0039] The recognition sub-model is trained using historical data to obtain the trained recognition sub-model. The semi-Markov sub-model and the Bayesian network sub-model are trained using historical data and the risk category labels of the historical data to obtain the trained semi-Markov sub-model and the trained Bayesian network sub-model. The decision sub-model is trained using historical data and the corresponding historical risk strategy to obtain the trained policy sub-model.

[0040] S4 First Identification: Input the real-time data into the trained identification sub-model to obtain abnormal real-time data, which is recorded as the first data. The types of data with abnormalities in the first data are integrated into the first combination. For example, if the type of real-time data is (temperature, humidity, wind speed, wind direction, precipitation, gas concentration, carbon dioxide concentration), the real-time data is (15℃, 65%, 2.3m / s, 45 o , 0.5mm, 0.3%, 400ppm), input the real-time data into the trained recognition sub-model, and output the real-time data (15℃, 65%, 2.3m / s, 45 o, 0.5mm, 0.3%, 400ppm), the real-time data marked with abnormal labels are recorded as the first data, and then the types of abnormal data in the first data are integrated into the first combination, that is, (precipitation, gas concentration, carbon dioxide concentration) are recorded as the first combination. To simplify the description, only some types of real-time data are listed here. In this embodiment, the numerical setting of wind direction is 0 degrees north and increases in clockwise direction.

[0041] In this embodiment, when using historical data to train the recognition sub-model, each historical data is used as a feature to train the isolation forest to obtain a first isolation forest model. Then, each type of data in the historical data is extracted to train the isolation forest separately to obtain a plurality of second isolation models, and all the trained second isolation forest models are used as a recognition sub-model.

[0042] Before S4 first identification, the method further includes: inputting the real-time data as a data feature into the first isolation forest model: if the abnormal label of the output real-time data is "-1", the real-time data is recorded as new real-time data, and S4 first identification is performed. If the abnormal label of the output real-time data is "1", the real-time data is recorded as new first data, and S8 first evaluation is performed.

[0043] S41 Data identification: including S411 first clustering, S412 second calculation, S413 first judgment, S414 first processing, S415 third calculation, S416 second judgment, such as Figure 2 shown.

[0044] S411 First clustering: clustering analysis is performed on the historical data using a clustering algorithm (such as K-means, hierarchical clustering, DBSCAN, etc.) to obtain a number of clusters of the historical data. In this embodiment, the DBSCAN clustering algorithm is used to perform cluster analysis on the historical data to obtain a number of clusters of the historical data.

[0045] S412 Second calculation: using the Euclidean distance method to calculate the Euclidean distance between the first data and the cluster center of each cluster, recorded as the first distance. The first distance is a first distance set, including the Euclidean distance between the first data and the cluster center of each cluster.

[0046] The Euclidean distance formula is: .

[0047] in, Indicates the first data and the The Euclidean distance between the cluster centers of the clusters, Indicates the first data data, Indicates The cluster centers of the clusters data.

[0048] The DBSCAN clustering algorithm is used to perform cluster analysis on historical data to obtain clusters A, B, and C. The average value of all historical data in each cluster is calculated as the cluster center of each cluster. For example, cluster A contains 5 historical data. If the categories of these 5 historical data are (equipment temperature and equipment pressure), they are data 1 (20, 30), data 2 (22, 31), data 3 (21, 29), data 4 (19, 28), and data 5 (23, 32), then the cluster center of cluster A is ( , ) To simplify the description, only some types of historical data are listed here.

[0049] The Euclidean distance method is used to calculate the Euclidean distance between the first data and the cluster center of each cluster to obtain the corresponding first distance A, first distance B and first distance C. For example, the first data is (25, 35), and from the above example, it can be seen that the cluster center of cluster A is (21, 30), then , and the first distance A is .

[0050] S413 First determination: determining whether the first distances are all smaller than a preset distance threshold.

[0051] If yes, then execute S8 first evaluation.

[0052] If not, then execute S5 first calculation.

[0053] S414 First processing: extract the data of the type consistent with the first combination in the historical data, record it as the first extracted data, extract the data of the type consistent with the first combination in the first data, record it as the second extracted data. For example, if the first combination is (rainfall, gas concentration, carbon dioxide concentration), then extract the data of rainfall, gas concentration and carbon dioxide concentration in the historical data, record it as the first extracted data of the historical data.

[0054] S415 Third calculation: using a local outlier factor algorithm to calculate the local reachable density of the first extracted data and the second extracted data, and calculating the local outlier factor of the second extracted data based on the calculated local reachable density.

[0055] Define the data range of each type of data in the second extracted data, filter the historical data in the first extracted data based on the defined data range, filter out the data that meets each type in the first extracted data and is within the defined data range of each type of data, and record the filtered first extracted data as the new first extracted data.

[0056] The distance between each first extracted data and the second extracted data is calculated using the Euclidean distance method, and the reachable distance between the current first extracted data and the second extracted data is calculated using the reachable distance formula. The local reachable density of the second extracted data is calculated using the local reachable density formula.

[0057] The reachable distance formula is: .

[0058] in, Indicates The reachable distance between the first extracted data and the second extracted data, Indicates the second extracted data, Indicates First extract the data, represents the maximum Euclidean distance from the second extracted data to the first extracted data, Indicates the second extracted data to the First extract the Euclidean distance of the data.

[0059] The local reachable density calculation formula is: .

[0060] in, Indicates the number of first extracted data, represents the first extracted data within the data range of each type of data defined in the second extracted data, Represents the sum of the reachable distances between the second extracted data and all the first extracted data.

[0061] Based on the above steps, the local reachable density of all first extracted data is calculated, and the local outlier factor of the second extracted data is calculated according to the local outlier factor calculation formula.

[0062] The local outlier factor calculation formula is: .

[0063] in, Indicates First extract the local reachable density of the data.

[0064] S416 Second judgment: judging whether the local outlier factors of the second extracted data are not lower than a preset local outlier factor threshold.

[0065] If so, no processing is done.

[0066] If not, the type of the second extracted data that is higher than the preset local outlier factor threshold is recorded as the second combination, and the second combination is updated to the first combination. For example, according to the above calculation, the local outlier factors of the rainfall and gas concentration in the second extracted data are not lower than the preset local outlier factor threshold, and the local outlier factor of the carbon dioxide concentration is not higher than the preset local outlier factor threshold, then (rainfall, gas concentration) is updated to the first combination.

[0067] S5 first calculation: calculate the joint frequency and marginal probability of the first combination based on historical data and real-time data, and calculate the target conditional probability through the calculation formula of conditional probability.

[0068] In this embodiment, the calculation formula of conditional probability is: .

[0069] in, represents the joint probability of the first combination, represents the marginal probability of the first combination. represents the number of types in the first combination, Indicates the first combination The probability of data of categories appearing. For example, the type of real-time data is (temperature, humidity, gas concentration, carbon dioxide concentration), and the real-time data is (15℃, 65%, 0.3%, 400ppm), where the abnormal label of the real-time data is (1, -1, -1, 1), so the first data is (15℃, 65%, 0.3%, 400ppm) marked with abnormal labels (1, -1, -1, 1), so the first combination is (humidity, gas concentration). Based on historical data, the probability of the temperature being 15℃ is recorded as P1, the probability of the humidity being 65% is recorded as P2, the probability of the gas concentration being 0.3% is recorded as P3, and the probability of the carbon dioxide concentration being 400ppm is recorded as P4, then the joint probability of the first combination is calculated. , , calculate the marginal probability of the first combination , , then the target conditional probability of the first combination .

[0070] S6 first judgment: based on the obtained target conditional probability, traverse the conditional probability table of the trained Bayesian network sub-model to determine whether the target conditional probability belongs to the conditional probability table of the trained Bayesian network sub-model.

[0071] If yes, then execute S8 first evaluation.

[0072] If not, then execute S7 first optimization.

[0073] S7 first optimization: adding the target conditional probability to the conditional probability table of the trained Bayesian network sub-model to obtain a new conditional probability table, obtaining a new Bayesian network sub-model based on the obtained new conditional probability table, and updating the new Bayesian network sub-model to the trained Bayesian network sub-model.

[0074] S8 First evaluation: input the first data into the trained Bayesian network sub-model to obtain second data, wherein the second data includes a risk level.

[0075] S9 Strategy acquisition: Input the second data into the trained decision sub-model to obtain a real-time risk strategy. After obtaining the real-time risk strategy, the obtained real-time risk strategy is presented to the administrator, and the obtained real-time risk strategy is executed after the administrator confirms it.

[0076] In this embodiment, based on the constructed risk decision model, the identification sub-model is used to determine whether there is abnormal real-time data, based on the abnormal real-time data, the abnormal data combination is obtained, and the conditional probability of the abnormal data combination is calculated based on the abnormal data combination. By judging whether the conditional probability of the abnormal data combination belongs to the conditional probability table of the trained Bayesian network sub-model, it is judged whether there is a change in the architecture of the Bayesian network model, and the conditional probability table of the Bayesian network sub-model is updated based on the judgment result to optimize the architecture of the Bayesian network model, so that it can adapt to the new risk environment and improve the accuracy of risk assessment. The present application uses the conditional probability of the abnormal data combination to determine whether the architecture of the Bayesian network model needs to be updated, and then dynamically optimizes the Bayesian network structure, which helps to improve the adaptability of the Bayesian network model structure to the dynamically changing environment of mine risks, improves the real-time, effectiveness and accuracy of the Bayesian network model, further improves the adaptability and accuracy of mine risk assessment, and enhances the immediacy of mine safety management.

[0077] Embodiment 2: The difference from Embodiment 1 is that: The risk decision model also includes a semi-Markov sub-model. After executing the first judgment in S6 and before executing the first optimization in S7, it also includes: S61 model optimization. The S61 model optimization includes S611 first screening, S612 second screening, S613 first labeling, S614 first expansion, S615 fourth calculation, S616 matrix construction, S617 third judgment and S618 probability update. Figure 3 shown.

[0078] S611 First screening: sort the first distances in descending order, and record the smallest first distance as the minimum distance.

[0079] S612 Second screening: The cluster corresponding to the minimum distance is taken as the cluster closest to the first data, recorded as the first cluster, and the risk category of the historical data corresponding to the first cluster is recorded as the first label.

[0080] S613 First labeling: labeling the first data with the first label as the risk category, obtaining the labeled first data and recording it as new first data. The risk category of the cluster closest to the first data is used as the risk category of the first data.

[0081] S614 First expansion: Based on the first data, the historical data, and the risk categories of the historical data, define a state space of the semi-Markov sub-model.

[0082] S615 Fourth calculation: Obtain the state sequence observed at each moment, calculate the residence time of each state in the state space by using the maximum likelihood estimation method, and record the calculated residence time as the first time. The residence time is the duration of each state.

[0083] S616 Matrix construction: Based on the first time, the first data and the historical data, the maximum likelihood estimation method is used to calculate the transition probability of each state in the state space. For example, the number of times state b transitions to state j is counted, recorded as T, and the total number of transitions of state b is counted, recorded as K, then the probability of state b transitioning to state j is , and then the transition probability is optimized by the maximum likelihood estimation method to obtain the transition probability of state b to state j. Based on the calculated transition probabilities between all states, the transition probability matrix W is constructed, where W (b, j) represents the transition probability of state b to state j.

[0084] S617 Third judgment: Based on the categories in the first combination as the parent variable nodes of the Bayesian network sub-model, the risk categories of the first data are used as the child variable nodes of the Bayesian network sub-model to determine whether there is an edge between the nodes corresponding to the first combination in the Bayesian network sub-model.

[0085] If so, no processing is done.

[0086] If not, based on the first combination, in the Bayesian network sub-model, corresponding edges are added between the parent variable nodes of the Bayesian network sub-model.

[0087] For example, the first combination is (rainfall, gas concentration). Then, based on the first combination, check whether there is an edge between the rainfall node and the gas concentration node in the Bayesian network sub-model. If so, no processing is done. If not, in the Bayesian network sub-model, add a corresponding edge between the rainfall node and the gas concentration node in the Bayesian network sub-model.

[0088] S618 Probability update: Based on the first data, historical data and the transition probability matrix, the maximum likelihood estimation method is used to calculate the conditional probability between each node in the Bayesian network sub-model, and the conditional probability table is updated according to the calculation results to obtain an updated Bayesian network sub-model, and the updated Bayesian network sub-model is used as the new trained Bayesian network sub-model.

[0089] Based on the likelihood function defined for each node in the Bayesian network sub-model, it represents the probability of the child variable node taking different states given the state of the parent variable node. For example, let the child variable node be E and the parent variable node be , record the historical data and the first data as the node sample data set. There are H node samples in the node sample data set. Then the state of the parent variable node in the cth sample is , the state of the child variable node is , then the likelihood function .in, The parent variable node associated with the child variable node The number of nodes. By finding the extreme value of the likelihood function, the maximum likelihood estimate of the conditional probability between each node pair under the given node sample data is obtained. Based on the obtained maximum likelihood estimate of the conditional probability between each node pair, the conditional probability table is filled in to update the conditional probability table to obtain an updated conditional probability table. Based on the updated conditional probability table, an updated Bayesian network sub-model is obtained, and the updated Bayesian network sub-model is used as the new trained Bayesian network sub-model.

[0090] In S8 first evaluation, the first data is input into the new trained Bayesian network sub-model.

[0091] In this embodiment, a semi-Markov sub-model is introduced on the basis of the Bayesian network model, and the residence time distribution of different risk states is characterized based on the semi-Markov sub-model, which improves the description accuracy of the evolution of complex risk states. At the same time, by constructing a state transition probability matrix, the evolution law between different risk states is established, which helps the Bayesian network model to have time-series self-adaptation ability, optimizes the time-series correlation between the types of data, and helps the Bayesian network model to adjust according to the dynamic changes of the mine operation environment, thereby improving the accuracy of the risk prediction of the Bayesian network model. Based on the first combination, the edges between the nodes are dynamically increased, which helps to update the topological structure of the Bayesian network model in real time, ensuring that the network structure of the Bayesian network model can reflect the association between the types of the latest data. In addition, the conditional probability table is updated based on the state transition probability matrix of the semi-Markov sub-model, which improves the accuracy of the probability reasoning of the Bayesian network model, so that the Bayesian network model not only adapts to the newly emerged state data, but also can predict the evolution direction of the risk based on the evolution trend of the historical state combined with the time factor, thereby improving the accuracy of the risk assessment of the Bayesian network model.

[0092] Embodiment 3: The difference from Embodiment 1 is that: like Figure 4 As shown, after executing S9 policy acquisition, it also includes: S91 policy optimization, and S91 policy optimization includes S911 second processing, S912 third processing, S913 fourth processing, S914 fifth calculation and S915 fifth processing.

[0093] S911 Second processing: Based on the second data, a risk strategy corresponding to the risk category of the second data is selected from a preset strategy library, which is recorded as a first risk strategy. The preset strategy library is a pre-set risk strategy that contains risk strategies that can solve all mine risks.

[0094] S912 The third processing: Based on the second data, the first risk strategy and the real-time risk strategy are simulated by using the Monte Carlo simulation method to obtain a risk indicator sequence for each risk strategy.

[0095] Based on the numerical value of the second data as the numerical range, a Monte Carlo simulation method is used to construct mining scenarios of different scenarios, and the first risk strategy and the real-time risk strategy are simulated based on the mining scenarios of different scenarios, and the risk indicator sequence of the first risk strategy and the real-time risk strategy is obtained based on the simulation results, and the risk indicator sequence includes cost-effectiveness, resource consumption and recovery time. The risk indicator sequence of the first risk strategy is recorded as the first sequence, and the risk indicator sequence of the real-time risk strategy is recorded as the second sequence.

[0096] S913 Fourth process: define a reference risk indicator sequence based on a risk strategy under theoretical or ideal conditions, wherein the reference risk indicator sequence is minimum cost-effectiveness, minimum resource consumption, and minimum recovery time. Under ideal conditions, the values ​​of cost-effectiveness, resource consumption, and recovery time in the reference risk indicator sequence are substantially zero.

[0097] S914 Fifth calculation: Based on the reference risk indicator sequence, calculate the difference between the first sequence and each indicator in the reference risk indicator sequence, recorded as the first difference sequence, and calculate the difference between each indicator in the second sequence and the reference risk indicator sequence, recorded as the second difference sequence. Calculate the correlation coefficient between the first difference sequence and the reference risk indicator sequence to obtain the correlation coefficient of each indicator, recorded as the first coefficient, then calculate the average value of each correlation coefficient in the first coefficient to obtain the grey correlation degree, recorded as the first indicator correlation degree. Calculate the correlation coefficient between the second difference sequence and the reference risk indicator sequence to obtain the correlation coefficient of each indicator, recorded as the second coefficient, then calculate the average value of each correlation coefficient in the second coefficient to obtain the grey correlation degree, recorded as the second indicator correlation degree.

[0098] S915 Fifth processing: sort the obtained first indicator correlation and second indicator correlation in ascending order, and select the risk strategy with the highest indicator correlation as the new real-time risk strategy.

[0099] In this embodiment, Monte Carlo simulation is used to simulate the first risk strategy and the real-time risk strategy to obtain a risk indicator sequence, and a reference risk indicator is constructed based on the ideal state. The grey correlation analysis method is used to analyze the correlation between the risk indicator sequence and the reference risk indicator sequence, and the advantages and disadvantages of different risk strategies are quantified. The strategy that best suits the current mine risk status is selected, which helps to dynamically optimize the risk strategy so that the mine risk strategy always matches the latest risk environment, improves the scientificity, accuracy and adaptability of the risk strategy selection, and enhances the dynamic response capability of mine safety management.

[0100] Embodiment 4: The difference from Embodiment 1 is that: The second data includes the risk level. After executing S9 strategy acquisition, it also includes: S10 strategy verification, S10 strategy verification includes S101 first setting, S102 second evaluation, S103 fourth judgment, S104 first setting, S105 function construction, S106 first definition, S107 second definition and S108 second optimization, such as Figure 5 shown.

[0101] S101 First setting: Setting the data of the mine based on the real-time risk strategy as test data. According to the obtained real-time risk strategy, setting the data of the mine after executing the real-time risk strategy as test data.

[0102] S102 Second evaluation: input the test data into the trained Bayesian network sub-model for evaluation to obtain third data.

[0103] S103 Fourth judgment: judging whether the risk level of the third data is lower than the risk level of the second data.

[0104] If so, the resulting real-time risk strategy is executed.

[0105] If yes, then execute S104 first setting.

[0106] S104 First setting: Determine the optimization goal of the decision sub-model. The optimization goals of the decision sub-model include maximizing safety, minimizing resources, and minimizing costs.

[0107] S105 Function construction: Construct the objective function based on the optimization goal.

[0108] The objective function is: .

[0109] in, represents the objective function, represents the risk strategy, represents the security maximization optimization goal, represents the resource minimization optimization goal, represents the cost minimization optimization objective, , and represents the weight of each optimization objective, .

[0110] S106 First definition: Use an optimization algorithm to optimize the model parameters of the trained decision sub-model to obtain first model parameters, and use the first model parameters as new model parameters of the trained decision sub-model.

[0111] In this embodiment, the particle swarm optimization algorithm, the genetic algorithm and the simulated annealing algorithm can be used to optimize the model parameters of the decision sub-model. For example, the model parameters of the decision sub-model are optimized layer by layer by the particle swarm optimization algorithm, the genetic algorithm and the simulated annealing algorithm in sequence, and the optimized model parameters are recorded as the first model parameters, and the first model parameters are used as the new model parameters of the trained decision sub-model.

[0112] S107 Second definition: Select a risk strategy corresponding to the risk category of the historical data from the preset strategy library as a new historical risk strategy.

[0113] The preset strategy library is a pre-set risk strategy that can solve all mine risks. According to the risk category of each historical data, a corresponding risk strategy is selected from the preset strategy library for each historical data as a new risk strategy for the historical data to obtain a new historical risk strategy.

[0114] S108 Second Optimization: Model training is performed on the trained decision sub-model based on historical data, new historical risk strategies and new model parameters to obtain a new trained decision sub-model, and the new trained decision sub-model is updated to the trained decision sub-model, and then S9 strategy acquisition is executed. After executing S9 strategy acquisition, S10 strategy verification is executed until the risk level of the third data is less than the risk level of the second data. If the risk level of the third data is still less than the risk level of the second data within the preset cycle, the real-time data and all the output real-time risk strategies are sent to the administrator, and the administrator formulates the real-time risk strategy for the real-time data.

[0115] In this embodiment, the selected risk strategy is tested, and the effectiveness of the risk strategy is ensured by introducing risk level comparison. At the same time, the parameters of the decision sub-model are optimized based on the optimization algorithm and optimization target, which enhances the real-time and accuracy of the decision sub-model. In addition, the use of a cyclic optimization mechanism helps to improve the adaptability of the decision sub-model to the dynamically changing environment of the mine, improves the accuracy of the risk assessment of the risk decision model, and enhances the scientificity and reliability of the risk strategy of the risk decision model.

[0116] Embodiment 5: This embodiment discloses a mine risk intelligent decision-making system, the system comprising: processor and memory, The memory stores program code; When the processor calls the program code in the memory, the steps of the method described in the above embodiment are executed.

[0117] In this embodiment, based on the constructed risk decision model, it is determined whether there is abnormal real-time data, based on the abnormal real-time data, the abnormal data combination is obtained, and the conditional probability of the abnormal data combination is calculated based on the abnormal data combination to determine whether the architecture of the Bayesian network model needs to be updated, and then the Bayesian network model structure is dynamically optimized, which helps to improve the adaptability of the Bayesian network model structure to the dynamically changing environment of mine risks, improves the real-time, effectiveness and accuracy of the Bayesian network model, further improves the adaptability and accuracy of mine risk assessment, and enhances the immediacy of mine safety management.

[0118] The above are all preferred embodiments of the present application, and the protection scope of the present application is not limited thereto. Therefore, any equivalent changes made according to the structure, shape, and principle of the present application should be included in the protection scope of the present application.

Claims

1. A mine risk intelligent decision-making method, characterized in that: include: Data collection: collect real-time data, historical data, types of historical data and historical risk strategies of the mine; Model construction: construct a risk decision model, which includes an identification sub-model, a Bayesian network sub-model and a decision sub-model; Model training: Use historical data and historical risk strategies to train the risk decision model to obtain the trained recognition sub-model, the trained Bayesian network sub-model and the trained decision sub-model; First recognition: input the real-time data into the trained recognition sub-model to obtain abnormal real-time data, which is recorded as the first data, and integrate the types of abnormal data in the first data into the first combination; First calculation: based on the historical data, the first data and the first combination, the target conditional probability is calculated according to the conditional probability calculation formula; First judgment: determine whether the target conditional probability belongs to the conditional probability table of the trained Bayesian network sub-model: If yes, then the first evaluation step is performed; If not, then execute the first optimization step; First optimization: adding the target conditional probability to the conditional probability table of the trained Bayesian network sub-model to obtain a new Bayesian network sub-model, and updating the new Bayesian network sub-model to the trained Bayesian network sub-model; First evaluation: input the first data into the trained Bayesian network sub-model to obtain the second data; Strategy acquisition: Input the second data into the trained decision sub-model to obtain a real-time risk strategy.

2. The mine risk intelligent decision-making method according to claim 1 is characterized in that: After performing the first identification step and before performing the first calculation step, the method further includes: First clustering: clustering analysis is performed on historical data using a clustering algorithm to obtain several clusters of historical data; Second calculation: calculate the Euclidean distance between the first data and the cluster center of each cluster, recorded as the first distance; First judgment: judging whether the first distances are all less than a preset distance threshold; If yes, then the first evaluation step is performed; If not, the first calculation step is performed.

3. The mine risk intelligent decision-making method according to claim 2 is characterized in that: After executing the first judgment step and before executing the first calculation step, the method further includes: First processing: extracting data of the same type as in the first combination from the historical data, recorded as first extracted data, and extracting data of the same type as in the first combination from the first data, recorded as second extracted data; Third calculation: using a local outlier factor algorithm to calculate the local reachable density of the first extracted data and the second extracted data, and calculating the local outlier factor of the second extracted data based on the calculated local reachable density; Second judgment: judging whether the local outlier factors of the second extracted data are not lower than the preset local outlier factor threshold: If so, no action will be taken; If not, the type of the second extracted data that is higher than the preset local outlier factor threshold is recorded as a second combination, and the second combination is updated to the first combination.

4. The mine risk intelligent decision-making method according to claim 2 is characterized in that: Based on the trained Bayesian network sub-model, the risk category of the historical data at the same timestamp is obtained, after performing the first discrimination step and before performing the first optimization step, further comprising: First screening: sort the first distances from largest to smallest, and record the smallest first distance as the minimum distance; Second screening: record the cluster corresponding to the minimum distance as the first cluster, and record the risk category of the historical data corresponding to the first cluster as the first label; First labeling: labeling the first data with the first label as the risk category to obtain the labeled first data, which is recorded as new first data.

5. The mine risk intelligent decision-making method according to claim 4 is characterized in that: The risk decision model also includes a semi-Markov sub-model, which, after executing the first labeling step and before executing the first optimization step, further includes: First extension: defining the state space of the semi-Markov sub-model based on the first data, the historical data and the risk categories of the historical data; Fourth calculation: using the maximum likelihood estimation method to calculate the residence time of each state in the state space, and recording the calculated residence time as the first time; Matrix construction: Based on the first time, first data and historical data, the maximum likelihood estimation method is used to calculate the transition probability of each state in the state space, and the transition probability matrix is ​​constructed according to the calculated transition probability.

6. The mine risk intelligent decision-making method according to claim 5 is characterized in that: Based on the trained Bayesian network sub-model, obtaining the edges between nodes in the Bayesian network sub-model, after executing the step of matrix construction and before executing the step of first optimization, further comprising: Third judgment: Based on the risk category of the first combination and the first data, it is judged whether there is an edge between the nodes corresponding to the first combination in the Bayesian network sub-model: If so, no action will be taken; If not, based on the first combination, in the Bayesian network sub-model, adding edges between nodes corresponding to the first combination; Probability update: based on the first data, historical data and the transition probability matrix, the conditional probability between each node in the Bayesian network sub-model is calculated by using the maximum likelihood estimation method, and the conditional probability table is updated according to the calculation result to obtain an updated Bayesian network sub-model, and the updated Bayesian network sub-model is used as the new trained Bayesian network sub-model; In a first evaluation step, first data is input into the new trained Bayesian network sub-model.

7. The mine risk intelligent decision-making method according to claim 4 is characterized in that: After executing the strategy acquisition step, it also includes: Second processing: based on the second data, selecting a risk strategy corresponding to the risk category of the second data from a preset strategy library, recorded as a first risk strategy; Third processing: Based on the second data, the first risk strategy and the real-time risk strategy are simulated by using the Monte Carlo simulation method to obtain a risk indicator sequence for each risk strategy; Fourth process: define the reference risk indicator sequence of the risk strategy based on the ideal state; Fifth calculation: the correlation between the risk indicator sequence of each risk strategy and the reference risk indicator sequence calculated by the grey correlation analysis method, and the calculated correlation is recorded as the indicator correlation; Fifth processing: sort the obtained indicator correlations in ascending order, and select the risk strategy with the highest indicator correlation as the new real-time risk strategy.

8. The mine risk intelligent decision-making method according to claim 4 is characterized in that: The second data includes a risk level; After executing the strategy acquisition step, it also includes: First setting: Set up the mine data based on the real-time risk strategy as test data; Second evaluation: input the test data into the trained Bayesian network sub-model to obtain the third data; Fourth judgment: judging whether the risk level of the third data is lower than the risk level of the second data: If so, the real-time risk strategy is executed; If yes, execute the first setting step; First setting: determine the optimization goal, which includes maximizing safety, minimizing resources, and minimizing costs; Function construction: construct the objective function based on the optimization goal; First definition: using an optimization algorithm to optimize the model parameters of the trained decision sub-model to obtain first model parameters, and using the first model parameters as new model parameters of the trained decision sub-model; Second definition: select a risk strategy corresponding to the risk category of the historical data from the preset strategy library as a new historical risk strategy; Second optimization: train the trained decision sub-model based on historical data, new historical risk strategies and new model parameters to obtain a new trained decision sub-model, update the new trained decision sub-model to the trained decision sub-model, and then execute the strategy acquisition step.

9. A mine risk intelligent decision-making system, characterized in that: include: processor and memory, The memory stores program code; When the processor calls the program code in the memory, the steps of the method according to any one of claims 1 to 8 are executed.

Images