Fault water disaster sensing and danger relieving operation method and system
By employing three-dimensional spatial grid segmentation and multimodal fusion technology, the problem of multi-source data collaborative perception and intelligent early warning collaboration in monitoring water hazards in fault zones of deep mines has been solved, enabling high-precision risk perception and emergency response operations, and ensuring mine safety.
Patent Information
- Application Number
- CN202511583204.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-31
- Publication Date
- 2026-01-23
AI Technical Summary
Existing fault water hazard monitoring technologies in deep mines suffer from insufficient multi-source data collaborative perception and weak intelligent collaboration in early warning and hazard mitigation, resulting in fragmented risk perception, reduced early warning accuracy, and low hazard mitigation efficiency.
A standardized data model is constructed by employing three-dimensional spatial grid segmentation and multimodal fusion technology. By calculating the coupling coefficients of microseismic, electrical, and hydrological data, and combining knowledge reasoning, acoustic alarms, and three-dimensional visual modalities, the spatiotemporal synchronization of multi-source data and risk level determination are achieved. A reward function update early warning model is designed, and drilling parameters and grouting schemes are customized.
It achieves spatiotemporal synchronization of multi-source data and continuity of risk perception, improves the accuracy of early warning and the pertinence of mitigation measures, reduces the risk of false alarms and missed alarms, and ensures the safety of deep mine mining.
Smart Images

Figure CN121390899A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of mine disaster prevention and control technology, and in particular to a method and system for sensing and mitigating fault-induced water hazard disasters. Background Technology
[0002] As coal mining in my country extends to deeper levels, the coupling effect of high ground pressure, high water pressure, and complex fault structures becomes increasingly prominent. Water inrush disasters caused by faults connecting aquifers have become a core risk threatening safe coal mine production. While existing fault water hazard prevention and control technologies have formed a basic framework from monitoring to early warning to mitigation, key technological bottlenecks remain in the face of the complex geological environment and dynamically changing risk situation in deep mines, making it difficult to meet the actual needs of precise prevention and efficient management.
[0003] First, the collaborative perception and full-process risk characterization capabilities of multi-source monitoring data are insufficient. Current fault flood monitoring relies heavily on independent monitoring methods such as microseismic, electrical, and hydrological methods. Each type of monitoring data corresponds to a single dimension of information, such as rock mass fracturing, water migration, and hydrological dynamics, lacking a collaborative perception framework that can link the evolutionary logic of flood hazards. The failure to effectively integrate various data sources to reflect the entire process of flood disasters—from rock mass fracturing creating water-conducting channels to water transport along these channels—results in fragmented and discontinuous risk perception, creating hidden dangers for subsequent early warning and mitigation efforts.
[0004] Secondly, the level of intelligent coordination between early warning and mitigation is weak. On the one hand, the core parameters of existing early warning models are mostly preset manually and fixed over a long period of time, making it impossible to dynamically optimize them based on actual early warning effects and changes in the geological environment. This results in poor model adaptability and a tendency for early warning accuracy to decrease after long-term operation. On the other hand, mitigation operations in high-risk areas are mostly based on macro-regional judgments, failing to deeply match the precisely identified risk characteristics and lacking targeted operational plan design. This leads to poor coordination between early warning and mitigation, making it difficult to guarantee mitigation efficiency and governance effectiveness.
[0005] Therefore, it is necessary to solve the technical problems of multi-source data collaborative perception, early warning and emergency response intelligent collaboration, and build a fault water hazard prevention and control system adapted to the complex environment of deep mines. Summary of the Invention
[0006] To address the aforementioned problems, this invention provides a method and system for detecting and mitigating fault-related water hazard disasters.
[0007] In a first aspect, the present invention provides a method for sensing and mitigating fault-related water hazard disasters, which employs the following technical solution: A method for detecting and mitigating fault-related water hazard disasters includes: The monitoring area is defined and divided into three-dimensional spatial grids. Microseismic monitoring data, electrical resistivity monitoring data, and hydrological monitoring data of the monitoring area are obtained, and a standardized data model is constructed. Calculate the coupling coefficient of the data in each grid, set multi-level risk thresholds, determine the risk level of the coupling coefficient of each grid, and visualize and render grids with different risk levels in a three-dimensional geological model. Based on the determined grid risk level, a knowledge reasoning modality, an acoustic alarm modality, and a 3D visual modality are constructed. The original outputs of each modality are normalized, and the normalization results are weighted and fused according to the preset confidence weights to obtain the final risk value. The final risk value is compared with the preset warning threshold to trigger the corresponding level of warning. Based on the corresponding level of the warning grid, the grid with high risk warning is identified, and the drilling verification and grouting sealing and hazard relief process is initiated. An early warning model is constructed, defining a state space that includes monitoring features, model performance indicators, and geological context. The early warning threshold and confidence weight of each grid are used as adjustment objects. A reward function based on early warning effect and action stability is designed, and the model is updated based on the reward function.
[0008] Furthermore, the construction of the standardized data model includes Microseismic sensors, electrical resistivity electrodes, and hydrological sensors are connected to a unified time server. Event energy, P-wave / S-wave velocity ratio, and event density characteristics are extracted from microseismic monitoring data. Apparent resistivity, apparent resistivity change rate, and spontaneous potential characteristics are extracted from electrical resistivity monitoring data. Water pressure and water temperature data are extracted from hydrological monitoring data. Data of different categories are aligned by timestamp to form spatiotemporally synchronized data blocks. The spatiotemporally synchronized data blocks are encapsulated into a standardized data model. The standardized data model includes a public metadata segment and a proprietary data body. The public metadata segment includes a unique device identifier, sensor three-dimensional coordinates, data acquisition timestamps, and data types. The proprietary data body includes a proprietary data body for microseismic data, a proprietary data body for electrical resistivity data, and a proprietary data body for hydrological data. The proprietary data bodies are encapsulated according to their categories.
[0009] Further, the calculation of the coupling coefficient of the data in each grid includes: The number of microseismic events within the grid is counted, and the rate of change of apparent resistivity within the grid is calculated. The geometric center of the low resistivity anomaly region in the electrical resistivity test was determined using a clustering algorithm, and the minimum Euclidean distance from the grid center to the geometric center was calculated. ; The coupling coefficient for each grid cell is calculated using the following formula: ,in, The coupling coefficient of the mesh. This represents the number of microseismic events within the grid. The apparent resistivity change rate within the grid. This represents the minimum distance from the grid center to the geometric center of the low-resistivity anomaly region in the electrical resistivity test. It is a smoothing constant. A unique index for the grid location.
[0010] Furthermore, the process of setting multi-level risk thresholds, determining the risk level of the coupling coefficient of each grid, and visualizing and rendering grids of different risk levels in a three-dimensional geological model includes: Determine the monitoring area Range of directions ( Assign a unique index to each grid. And set the side length of the three-dimensional spatial mesh to L; Calculate the center coordinates of each grid cell to obtain the three-dimensional center coordinates of each grid cell. ; Set a low-risk threshold Medium risk threshold High-risk threshold And satisfy ,in, There is no water guiding channel. As a potential water-conducting fracture, A water diversion channel has been formed. A sudden water inrush occurred; In a three-dimensional geological model, according to The value determines the color rendering of the grid; Furthermore, the knowledge reasoning modality is constructed, including: Construct a knowledge graph containing entities and relationships related to fault-related water hazards; The entity and relationship of the knowledge graph are stored using a graph database. Multiple risk inference rules are defined using a graph query language. The inference rules are set based on entity attributes and the relationships between entities. Real-time monitoring data is matched with the inference rules, and the number of successfully matched rules is counted. Each inference rule is assigned a weight, which is set based on the importance of the inference rule to the risk assessment. The sum of the weights of all inference rules is 1. When the inference rules are completely matched, the matching score is 1. When the inference rules are not completely matched, the matching score is 0. Substitute the values into the formula to calculate the original risk confidence level. The formula is: ,in, The number of rules to be matched. For rule weights, The rule matching score is 1 for a perfect match and 0 for an otherwise perfect match. The original risk confidence score is then normalized.
[0011] Furthermore, the acoustic alarm mode is constructed, including: Obtain the frequency spectrum of each microseismic event and extract the frequency corresponding to the maximum amplitude in the spectrum as the dominant frequency. Mapping the dominant frequency from the effective frequency band of micro-vibrations to the audible frequency band that the human ear is sensitive to. The energy of microseismic events is mapped to the actual alarm volume using a piecewise linear function. When the mesh coupling coefficient is greater than the medium risk threshold At that time, the mapped audible sound waves and voice prompts are played, and the actual alarm volume is used as the raw output of the acoustic mode, and the raw output is normalized.
[0012] Furthermore, constructing the three-dimensional visual modality includes: Based on a three-dimensional spatial grid, a four-dimensional feature tensor is constructed. The four-dimensional feature tensor includes multiple feature dimensions, including grid coupling coefficient, cumulative energy of microseismic events, apparent resistivity change rate, microseismic event density, P-wave / S-wave velocity ratio, and distance from the grid to the fault. A three-dimensional convolutional neural network model architecture is constructed. Its input layer receives a four-dimensional feature tensor, and the output layer outputs the risk level probability of each grid through the Softmax function. The risk level probability includes the probability of no water channel, the probability of potential water-conducting fracture, the probability of water channel, and the probability of water inrush. The probability representing a high-risk level is selected as the original output of the three-dimensional visual modality, where the high-risk level is the formed water channel, and the original output is normalized.
[0013] Furthermore, the step of weighting and fusing the normalization results according to preset confidence weights to obtain the final risk value, comparing the final risk value with a preset warning threshold, and triggering a corresponding level of warning includes: Multimodal reliability weights are set to assign reliability weights to the normalized knowledge reasoning modal output, acoustic alarm modal output, and 3D vision modal output; The normalized outputs of the three modalities are weighted and fused together using the following formula: , in, This is the final risk value. , , These are the reliability weights for the knowledge reasoning modality, the acoustic alarm modality, and the 3D vision modality, respectively. , , These are the outputs after modal normalization, respectively; Set multi-level fusion early warning thresholds, compare the final risk value with the multi-level early warning thresholds, and trigger the corresponding level of early warning based on the comparison results.
[0014] Furthermore, the construction of the early warning model defines a state space that includes monitoring features, model performance indicators, and geological context, including: The monitoring features include the mean of the coupling coefficient, the standard deviation of the feature parameters, the maximum energy of the microseismic event, and the average water discharge within the set monitoring period. The model performance indicators include the accuracy, false alarm rate, and missed alarm rate of the most recent warnings. The geological and environmental context includes the geological structure complexity, working face advance speed, water pressure change rate, current warning level, and hazard mitigation effect score. The original values of each dimension feature are obtained. The original values of the monitoring features are obtained by statistically analyzing the monitoring data within the set monitoring period. The original values of the model performance indicators are obtained by comparing the results of the most recent warnings with the actual situation. The original values of the geological and environmental context are obtained by collecting the corresponding geological and operational environment parameters. The original values of all the obtained dimension features are combined by category to form a multi-dimensional feature vector. The multi-dimensional feature vectors are subjected to Z-score normalization, which maps each feature to a preset interval. The normalization method is as follows: ,in, For the first 3D features at time step Standardized value For the first 3D features at time step The original value, For the first Historical mean of the dimensional feature For the first Historical standard deviation of dimensional features; The standardized multi-dimensional feature vector is the defined state space that includes monitoring features, model performance indicators, and geological context.
[0015] Furthermore, the design is based on a reward function that considers both the early warning effect and the stability of the action, and the model is updated based on this reward function, including: Design a reward function, which is expressed by the formula: , in, For accurate early warning indication functions, the value must be 1. This is a false alarm indicator function; a false alarm is set to 1. This is a missed detection indicator function; the value is 1 for missed detections. , , The penalty coefficient is... The penalty coefficient for the action. Let be the norm of the action vector; Based on the current state Select Action Observe the new state after execution. And get rewards , will experience tuples ( , , , Stored in the experience buffer; The PPO algorithm is adopted to calculate the action advantage through generalized advantage estimation, update the policy network by minimizing the pruning objective function, and update the state value network by minimizing the mean squared error loss.
[0016] Secondly, an intelligent auction system includes: The data acquisition module is configured to set the monitoring area and perform three-dimensional spatial grid segmentation, acquire microseismic monitoring data, electrical resistivity monitoring data, and hydrological monitoring data of the monitoring area, and construct a standardized data model. The risk level determination module is configured to calculate the coupling coefficient of the data in each grid, set multi-level risk thresholds, determine the risk level of the coupling coefficient of each grid, and visualize and render grids with different risk levels in a three-dimensional geological model. The multimodal fusion early warning module is configured to construct a knowledge reasoning modality, an acoustic alarm modality, and a three-dimensional visual modality based on the determined grid risk level. The original output of each modality is normalized, and the normalization results are weighted and fused according to the preset confidence weights to obtain the final risk value. The final risk value is compared with the preset early warning threshold to trigger the corresponding level of early warning. The hazard mitigation module is configured to identify high-risk grids based on the corresponding level of early warning and initiate drilling verification and grouting sealing hazard mitigation procedures. The early warning model update module is configured to construct an early warning model, define a state space that includes monitoring features, model performance indicators and geological context, use the early warning threshold and confidence weight of each grid as adjustment objects, design a reward function based on early warning effect and action stability, and update the model based on the reward function.
[0017] Thirdly, the present invention provides a computer-readable storage medium storing a plurality of instructions adapted to be loaded and executed by a processor of a terminal device as described in the method for sensing and mitigating fault-related flood disasters.
[0018] Fourthly, the present invention provides a terminal device, including a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, the instructions being adapted to be loaded and executed by the processor to provide the method for sensing and mitigating fault-induced flood disasters.
[0019] In summary, the present invention has the following beneficial technical effects: 1. This invention proposes a method and system for sensing and mitigating fault-related water hazards. Addressing the problem that traditional monitoring methods rely on independent operation of microseismic, electrical, and hydrological data, making it difficult to correlate them with the evolutionary logic of water hazards, this invention constructs a standardized data model to achieve spatiotemporal synchronization of multi-source data. Combined with three-dimensional spatial grid segmentation, multi-source data can be mapped to each grid cell, reflecting the water hazard process of rock fractures creating water-conducting channels and water sources migrating along these channels. This solves the problem of discontinuous risk perception in traditional monitoring and enables visualization through rendering.
[0020] 2. This invention addresses the problems of traditional single-modal early warning systems being susceptible to interference and providing incomplete information. It constructs professional early warning capabilities through multiple modalities. The knowledge reasoning modality is based on a knowledge graph in the field of fault-induced flooding, using logical rules to associate key risk characteristics such as fault activation and resistivity changes. The acoustic alarm modality converts microseismic signals into frequency bands perceptible to the human ear. The three-dimensional visual modality intuitively presents a grid-level risk distribution heatmap. The original outputs of each modality are weighted and fused to obtain the final risk value. Through multi-modal collaborative verification, the limitations of single-modality systems can be corrected, and the advantages of each modality can be combined to reduce the risk of false alarms and missed alarms, making early warning information easier to understand and respond to.
[0021] 3. This invention addresses the problem of traditional hazard mitigation operations relying on macroscopic regional assessments and lacking effective integration with early warning information. It targets high-risk grid centers with excessive coupling coefficients, customizing drilling parameters and grouting schemes to match hazard mitigation measures with the characteristic depth of water-conducting channels, avoiding the blindness of traditional macroscopic borehole layout. Simultaneously, the early warning level directly guides the preparation and implementation of hazard mitigation operations, ensuring safety in deep mine mining. Attached Figure Description
[0022] Figure 1 This is a flowchart of a fault-based water hazard detection and mitigation operation method according to Embodiment 1 of the present invention; Figure 2 This is a flowchart illustrating the calculation of the coupling coefficient of data in each grid according to Embodiment 1 of the present invention; Figure 3 This is a flowchart of the multimodal early warning model structure of Embodiment 1 of the present invention; Figure 4 This is a flowchart of the emergency response process in Embodiment 1 of the present invention. Detailed Implementation
[0023] The present invention will be further described in detail below with reference to the accompanying drawings.
[0024] Example 1 Reference Figure 1 This embodiment of a fault-based flood disaster sensing and mitigation method includes: S1. Define the monitoring area and perform three-dimensional spatial grid segmentation to obtain microseismic monitoring data, electrical resistivity monitoring data, and hydrological monitoring data of the monitoring area, and construct a standardized data model.
[0025] The construction of the standardized data model includes: S11. Connect the microseismic sensor, electrical resistivity electrode, and hydrological sensor to a unified time server. Extract event energy, P-wave / S-wave velocity ratio, and event density characteristics from the microseismic monitoring data. Extract apparent resistivity, apparent resistivity change rate, and spontaneous potential characteristics from the electrical resistivity monitoring data. Extract water pressure and water temperature data from the hydrological monitoring data. Align different types of data according to timestamps to form spatiotemporally synchronized data blocks.
[0026] To achieve accurate acquisition of microseismic, electrical resistivity, and hydrological data, a monitoring system was deployed according to S11 requirements and connected to a unified time reference. Deployment of microseismic monitoring system: In the track roadway, transport roadway and cut-out eye of the working face, microseismic monitoring system and network parallel electrical resistivity monitoring system are jointly deployed according to the principle of spatial gridding; the microseismic monitoring system consists of 16 to 32 high-frequency three-component detectors with a sampling frequency ≥1kHz, and the detector spacing is 100-150 meters.
[0027] Data acquisition: The P-wave / S-wave velocity ratio was obtained through a microseismic monitoring system. Event energy E (unit: J), event frequency N (times / hour), and event density (times / m³); P-wave / S-wave velocity ratio ( The event energy E (unit: J), event frequency N (times / hour), and event density (≥5 events / m³ per unit volume is considered high density) were analyzed by measuring the P-wave velocity ratio. An abnormal increase in event energy and frequency can identify the upwelling of confined water within the fractured area. A sharp increase in event energy and frequency reflects an intensified degree of rock mass damage. Event density can characterize fracture clusters, and high-value areas (e.g., ≥5 events / m³) often indicate the formation of potential water-conducting channels. Therefore, microseismic monitoring, by tracking the spatiotemporal evolution of rock mass fractures in real time, enables dynamic perception and early warning of the water inrush channel formation process.
[0028] Deployment of electrical resistivity monitoring system: Electrode selection: The electrical resistivity monitoring system consists of no fewer than 64 electrodes with an electrode spacing of 6-10 meters; during deployment, ensure that the microseismic detector and the electrical resistivity electrodes are spatially adjacent and share the same coordinate reference.
[0029] Data acquisition involves obtaining the relative rate of change of apparent resistivity through an electrical resistivity monitoring system. (Unit: % / day), natural potential gradient (ΔSP, unit: mV / m), primary potential gradient (ΔPFP, unit: mV / m), excitation current gradient (ΔEC, unit: mA / m) indices; Deployment of hydrological monitoring system: For sensor selection, four YSJ-6 water pressure sensors (measuring range 0-8MPa, accuracy 0.01MPa) and three SW-1 water temperature sensors (measuring range 0-60℃, accuracy 0.1℃) were installed.
[0030] For deployment, two water pressure sensors were installed in boreholes near the F8 fault, and two were installed in low-lying areas of the working face; the water temperature sensor was installed in the same borehole as the water pressure sensor to collect water temperature data synchronously.
[0031] Data acquisition: The sensor transmits data via RS485 bus, with a data acquisition interval of 15 minutes, recording changes in water pressure and temperature in real time.
[0032] The time synchronization control connects the microseismic sensor, electrical electrode, hydrological sensor, and data acquisition station to the NTP time server, configuring a time synchronization period of 10s to ensure that the time deviation of all devices is ≤1ms.
[0033] S12. The spatiotemporally synchronized data blocks are encapsulated into a standardized data model. The standardized data model includes a public metadata segment and a proprietary data body. The public metadata segment includes a unique device identifier, sensor three-dimensional coordinates, data acquisition timestamp, and data type. The proprietary data body includes a microseismic data proprietary data body, an electrical resistivity data proprietary data body, and a hydrological data proprietary data body. The proprietary data bodies are encapsulated according to their categories.
[0034] Multi-source data feature extraction extracts key parameters from the raw data collected by various monitoring systems according to the feature type S11, as follows: For microseismic monitoring data, extract event energy and the P-wave / S-wave velocity ratio (P / S / S / W / C / W). Event density (the cumulative number of events in the grid where the event is located within 1 hour), for electrical resistivity monitoring data, extract apparent resistivity, relative rate of change of apparent resistivity, and spontaneous potential; for hydrological monitoring data, extract water pressure (borehole water pressure near the fault in the same period) and water temperature (water temperature in the same period).
[0035] Outlier Removal: To ensure data quality, outlier data is removed according to the following rules: Microseismic data: Event energy E < 15 J (system noise threshold) and P-wave / S-wave velocity ratio are excluded. For abnormal events with a wave velocity ratio <1.2 or >2.5 (exceeding the normal range for rock mass fracture), electrical resistivity data should be excluded. <1Ω m (metallic interference) or >2000Ω m (insulation interference), relative rate of change of apparent resistivity | |>50% (measurement error) outliers, hydrological data: remove outliers with water pressure <0MPa or >8MPa (outside the sensor range) and water temperature <10℃ or >40℃ (outside the normal downhole water temperature range).
[0036] Data spatiotemporal synchronization is processed as follows: Apache Kafka is used to enable real-time access to multi-source data, including geological, hydrological, monitoring, and video stream data, which are then pushed to corresponding Kafka topics, such as topic_microseismic (microseismic data topic) and topic_resistivity (resistivity data topic). Spark Streaming is used for stream processing to achieve time synchronization and alignment, data cleaning, format standardization, and key correlation calculations for multi-source data. For example, the coordinates of microseismic events within the same time window are spatially matched with low-resistivity anomaly areas by electrical resistivity analysis, and the coupling coefficient K is calculated in real time. Alternatively, hydrological pressure mutations and water outflow alarm signals from video analysis can be fused. The data fusion latency is less than 1 second, providing crucial real-time data support for multimodal intelligent early warning of water hazards.
[0037] (1) Time Synchronization Alignment: Due to the different generation and transmission delays of data from various sensors, a unified device timestamp is applied to each data entry before it is accessed by Kafka. Spark Streaming consumption employs near real-time alignment based on a processing time window. For example, a 5-10 second sliding window is set, treating data from different topics such as `topic_microseismic` and `topic_resistivity` that fall within the same window as a single batch for joint processing.
[0038] (2) Data cleaning: Remove outliers that are significantly beyond the physical range (such as negative resistivity or excessive micro-vibration caused by blasting and depressurization).
[0039] (3) Format standardization: After cleaning, data with different structures are converted into a unified standardized data model, which includes common metadata segments (such as device_id, timestamp, data_type) and proprietary data bodies (such as the microseismic_waveform array or resistance_value). This step provides a unified interface for subsequent correlation calculations.
[0040] (4) Key correlation calculation: Within a time-aligned window, data from different topics are connected using a common key (such as borehole number). For example, the energy of microseismic events in the same area is correlated with the resistivity drop value, and the fused risk index is finally output.
[0041] The time-space synchronized data blocks are encapsulated into a standardized data model according to the S12 structure. The common metadata segment encapsulates the common metadata segment into the common attributes of all types of data, and unifies the format.
[0042] The model storage and retrieval system stores the encapsulated standardized data model in a MySQL database and exposes it to the outside world through an API interface. This supports real-time data retrieval for modules such as coupling coefficient calculation and multimodal construction in subsequent steps, enabling data to be encapsulated once and reused on multiple devices.
[0043] Reference Figure 2 S2. Calculate the coupling coefficient of the data in each grid, set multi-level risk thresholds, determine the risk level of the coupling coefficient of each grid, and visualize and render the grids with different risk levels in the three-dimensional geological model. The calculation of the coupling coefficient of the data in each grid includes: S21. Count the number of microseismic events within the grid. Calculate the rate of change of apparent resistivity within the grid.
[0044] The number of microseismic events within a grid is counted based on the most recent 24 hours. Microseismic events meeting the condition of "event energy E≥15J (system effective event threshold)" are extracted from the microseismic data volume of the standardized data model and indexed by grid. Count the quantity one by one : If the three-dimensional coordinates of the microseismic event If an event falls within the spatial range of a certain grid (X∈[Xmin+(i-1)L,Xmin+iL], Y∈[Ymin+(j-1)L,Ymin+jL], Z∈[Zmin+(k-1)L,Zmin+kL]), then the event is determined to belong to that grid.
[0045] The apparent resistivity change rate within the grid is calculated by extracting real-time apparent resistivity within the grid from the proprietary data volume of the electrical resistivity data in the standardized data model. Combined with the average stable apparent resistivity (background value) at the initial monitoring stage ), calculate the rate of change of apparent resistivity using the following formula. .
[0046] Background value Determine the arithmetic mean of all apparent resistivity data collected at 30-minute intervals within the grid during the initial monitoring phase, and the real-time value. Extract the average apparent resistivity of the grid within the statistical period (the most recent 24 hours).
[0047] S22. Determine the geometric center of the low resistivity anomaly region using a clustering algorithm, and calculate the minimum Euclidean distance from the grid center to the geometric center. ; The anomaly determination criteria are based on the electrical resistivity response characteristics of the Taiyuan Formation limestone aquifer in this mine, and the relative rate of change of apparent resistivity is set. A value <-15% indicates a low-resistivity anomaly (reflecting a decrease in resistivity caused by water seepage).
[0048] Clustering algorithm selection and parameters: The DBSCAN (Density-based Noise Application Spatial Clustering) algorithm was used to cluster electrical resistivity data points that met the anomaly criteria throughout the entire monitoring area, and the center coordinates of each grid were extracted. Calculate the coordinates of the geometric center using the arithmetic mean method. : Calculation formula: , , (n is the number of abnormal grids).
[0049] The minimum Euclidean distance D from the grid center to the anomaly center is calculated using the following formula: .
[0050] S23. The coupling coefficient for each grid is calculated using the following formula: ,in, The coupling coefficient of the mesh. This represents the number of microseismic events within the grid. The apparent resistivity change rate within the grid. This represents the minimum distance from the grid center to the geometric center of the low-resistivity anomaly region in the electrical resistivity test. This is a smoothing constant, with a value of 1. A unique index for the grid location.
[0051] The process of setting multi-level risk thresholds, determining the risk level of the coupling coefficient of each grid, and visualizing and rendering grids of different risk levels in a 3D geological model includes: Calculate the coupling coefficient, where the smoothing constant C should be chosen to avoid a denominator of zero (taking C=1m based on the grid side length and the distribution of the anomaly area).
[0052] S24. Determine the monitoring area in Range of directions ( ) Assign a unique index to each grid And set the side length of the three-dimensional spatial grid to L.
[0053] Grid side length: L=5m, grid accuracy is 5m×5m×5m, which is set based on the geological complexity and monitoring accuracy requirements of the working face, taking into account both computational efficiency and identification accuracy.
[0054] S25. Calculate the center coordinates of each grid cell to obtain the three-dimensional center coordinates of each grid cell. .
[0055] S26. Set a low-risk threshold Medium risk threshold High-risk threshold And satisfy ,in, There is no water guiding channel. As a potential water-conducting fracture, A water diversion channel has been formed. A sudden water inrush occurred.
[0056] The risk level of each grid is determined by comparing its K value with a threshold according to the rule in claim S26. Low risk: ≤10 No water guiding channel; Medium risk: 10 ≤20 potential water-conducting fractures; High risk: 20 ≤30 has formed a water guiding channel; Extreme risks: > A sudden flood occurred.
[0057] S27. In a three-dimensional geological model, according to The value determines the color rendering of the grid.
[0058] The three-dimensional geological model is based on the geological exploration report (borehole columnar section, seismic exploration profile) and measured data of the mine working face. The three-dimensional geological model is constructed using drawing software. The model includes geological bodies such as coal seam, roof and floor strata, F8 fault, and Taiyuan Formation limestone aquifer. The coordinate system is consistent with the coordinate system of the monitoring area (mine independent coordinate system). The model accuracy reaches 0.5m.
[0059] Linking the grid with the geological model involves indexing all grid cells. By associating the grid with the spatial location of the 3D geological model, it ensures that each grid is accurately superimposed on the corresponding geological body (such as coal seam, fault fracture zone, aquifer), thus achieving a visual fusion of geological background and grid risk.
[0060] The risk level color rendering uses the grid coloring function of Surfer 17.0 software to color-render the grid according to the risk level. The rendering rules and visual identifications are as follows: Low risk (K ≤ 10): sky blue (RGB: 135, 206, 235), corresponding to no water-conducting channels. Most areas of the coal seam in the model are sky blue; Medium risk (10 < K ≤ 20): lemon yellow (RGB: 255, 250, 205), corresponding to potential water-conducting fissures. The transition area between the coal seam and the fault in the model is lemon yellow; High risk (20 < K ≤ 30): orange red (RGB: 255, 140, 0), corresponding to formed water-conducting channels. Local areas near the F8 fault in the model are orange red; Extreme risk (K > 30): dark red (RGB: 220, 20, 60), corresponding to water inrush. The grid group with index (110, 25, 9) in the hanging wall of the F8 fault in the model is dark red. After the rendering is completed, a "3D thermal map of the working face fault water disaster risk" is generated and displayed in real time through a mine 3D visualization platform (such as the KJ95N type).
[0061] Refer to Figure 3 , S3. Based on the determined grid risk level, construct a knowledge inference modality, an acoustic alarm modality, and a 3D visual modality. Normalize the original outputs of each modality, and weighted-fuse the normalized results according to the pre-set confidence weights to obtain the final risk value. Compare the final risk value with the pre-set warning threshold to trigger the corresponding level of warning; Among them, constructing the knowledge inference modality includes: S311. Construct an entity and relationship knowledge graph containing fault water disaster.
[0062] The inference modality is based on knowledge graph technology to construct a semantic network containing the relationship between faults - aquifers - working faces, store entity relationships using the Neo4j graph database, and use the Cypher query language to implement risk inference (such as "IF fault activation AND resistivity decrease THEN high conduction risk"). Based on logical inference large models (such as expert systems and knowledge graphs), extract abnormal feature combinations of water disaster evolution, for example: Abnormal increase + frequent microseismic events in a specific frequency band + continuous decrease in resistivity" is a strong feature signal of water body approaching.
[0063] Construction tool: Train a mine water disaster domain dictionary (including professional terms such as "water-conducting channel", "low-resistance anomaly", "microseismic dense area", etc.) based on the BERT pre-trained model, and use the PyTorch framework of Python to automatically extract entities and relationships to form a structured knowledge graph.
[0064] The knowledge semantic understanding system for complex coal mine environments refers to the automatic extraction of professional knowledge from unstructured texts such as geological reports, regulations, and case studies using natural language processing technology to form a knowledge graph. The system has the ability to semantically fuse multi-source heterogeneous data, and can associate real-time monitoring data with professional knowledge, understand the equivalent semantics of "sudden drop in resistivity" and "water seepage". Based on the knowledge graph and rule-based reasoning, it can achieve deep-level situational perception and intelligent linkage, not only identifying abnormal areas, but also interpreting the mechanisms behind the abnormalities, automatically inferring the water inrush path and risk level, and providing intelligent support for emergency response decisions.
[0065] S312. Utilize a graph database to store entities and relationships of a knowledge graph, define multiple sets of risk inference rules through a graph query language, the inference rules are set based on entity attributes and relationships between entities, match real-time monitoring data with the inference rules, and count the number of successfully matched rules; S313. Assign a weight to each inference rule. The weight is set based on the importance of the inference rule to the risk judgment. The sum of the weights of all inference rules is 1. The matching score is 1 when the inference rules are completely matched and 0 when the inference rules are not completely matched.
[0066] Based on the importance of the rules in risk assessment (Rule 1, associated with faults and connected aquifers, has the highest risk; Rule 2, associated with multiple anomalies, has the second highest risk; Rule 3, associated with only a single anomaly, has the lowest importance), weights are assigned and the sum of the weights is ensured to be 1: Rule 1 Weight (The assessment of extreme risks is of the highest importance); Rule 2 Weight (High-risk assessment, secondary importance); Rule 3 Weight (Medium risk assessment, lowest importance).
[0067] S314. Substitute the values into the formula to calculate the original risk confidence level. The formula is: ,in, The number of rules to be matched. For rule weights, The rule matching score is 1 for a perfect match and 0 for an otherwise perfect match. The original risk confidence score is then normalized.
[0068] Rule 1 and Rule 2 match exactly, therefore =1、 =1; Rule 3 not matched. =0. =0.6×1+0.3×1+0.1×0=0.9, using the minimum and maximum normalization method, in historical monitoring Minimum value (Unmatched rules), maximum value (All rules match).
[0069] The normalization formula is: Substituting this into the equation, we get R1′=(0.9-0) / (1-0)=0.9 (range [0,1]).
[0070] The construction of the acoustic alarm mode includes: S321. Obtain the frequency spectrum of each microseismic event and extract the frequency corresponding to the maximum amplitude in the spectrum as the dominant frequency.
[0071] S322. Map the dominant frequency from the effective frequency band of micro-vibration to the audible frequency band that the human ear is sensitive to.
[0072] S323. Map the energy of microseismic events to the actual alarm volume using a piecewise linear function.
[0073] For grid index The micro-seismic events within the past hour were analyzed, and the original vibration signals (sampling frequency 1 kHz) were subjected to Fast Fourier Transform (FFT) using the SciPy library in Python to obtain the frequency spectrum ranging from 0 to 500 Hz.
[0074] Dominant frequency extraction: The frequency corresponding to the maximum amplitude in the frequency spectrum is the dominant frequency. ; The effective frequency band for micro-vibrations is 50~200Hz (excluding low-frequency mechanical interference and high-frequency electromagnetic noise), while the audible frequency band for the human ear is 500~2000Hz. The linear mapping formula is as follows: , in =50Hz =200Hz (Boundary of effective frequency band for microseismic events) =500Hz =2000Hz (the boundary of the audible frequency band); Divide the microseismic energy E into three segments, corresponding to different alarm volumes V (unit: dB): (1) E<15J (effective event threshold): V=30dB (background notification sound, no alarm meaning); (2) 15J≤E≤200J (medium risk energy range): V=30+(E-15)×(60-30) / (200-15) (60dB is the medium risk warning sound); (3) E>200J (high-risk energy range): V=60+(E-200)×(85-60) / (500-200) (85dB is the extreme risk alarm tone, 500J is the maximum monitoring energy of the system). S324. When the mesh coupling coefficient is greater than the medium risk threshold At that time, the mapped audible sound waves and voice prompts are played, and the actual alarm volume is used as the raw output of the acoustic mode, and the raw output is normalized.
[0075] Alarm triggering condition: the mesh coupling coefficient Once the triggering conditions are met, the system will play audible sound waves through the downhole broadcasting equipment and simultaneously play a voice prompt: "A high-energy microseismic event has occurred in the grid near the xx fault in the xx working face. The risk level of water damage is high. Please strengthen patrols."
[0076] Raw output determination: The actual alarm volume is used as the raw output of the acoustic mode. .
[0077] Normalization: The minimum-maximum normalization method is used. =30dB (minimum volume), =85dB (maximum alarm volume), normalization formula is: Substituting into the equation, we get R2′=(66.67-30) / (85-30)=36.67 / 55≈0.667 (range [0,1]).
[0078] The construction of the three-dimensional visual modality includes: S331. Based on a three-dimensional spatial grid, a four-dimensional feature tensor is constructed. The four-dimensional feature tensor includes multiple feature dimensions, including grid coupling coefficient, cumulative energy of microseismic events, apparent resistivity change rate, microseismic event density, P-wave / S-wave velocity ratio, and distance from the grid to the fault.
[0079] Based on the three-dimensional spatial grid from step S2, a four-dimensional feature tensor (Channels, D, H, W) is constructed, with each dimension defined as follows: Channels (Characteristic Channels): There are 6 in total, namely, grid coupling coefficient K, cumulative energy of microseismic events (the sum of the energies of all microseismic events within the grid in the last 24 hours), and apparent resistivity change rate. Microseismic event density (number of microseismic events in the grid within the last hour / m³), P-wave and S-wave velocity ratio The distance from the grid to the fault.
[0080] S332. Construct a three-dimensional convolutional neural network model architecture, whose input layer receives a four-dimensional feature tensor, and whose output layer outputs the risk level probability of each grid through the Softmax function. The risk level probability includes the probability of no water channel, the probability of potential water-conducting fracture, the probability of water channel, and the probability of water inrush.
[0081] The visual modality trains the microseismic-electrical method coupling data cube based on a three-dimensional convolutional neural network (3D-CNN) and outputs a risk zoning heat map (red: high risk, K>30; orange: medium risk, 10<K≤30; yellow: low risk, K≤10).
[0082] The three-dimensional convolutional neural network (3D-CNN) is an end-to-end deep learning model, mainly composed of an input layer, multiple 3D convolutional-pooling modules, a fully connected layer, and an output layer. The input data is a four-dimensional tensor (Channels, D, H, W), where the number of channels Channels contains multi-modal features such as the number of microseismic events, energy release, and electrical resistivity. D is the burial depth, H is the height, and W is the width. The network core uses a 3D convolutional kernel (such as 3×3×3) for feature extraction, and its operation is defined as: , where W is the convolutional kernel weight, I is the input tensor, and b is the bias, represents the output feature map, , , is the index of the output feature map, representing all elements in the burial depth, height, and width directions, and determining the current calculated output position. represents the element traversal index of the convolutional kernel, which is used for feature extraction operations. k represents the size of the 3D convolutional kernel in each spatial dimension (burial depth, height, width) (i.e., the side length of the convolutional kernel), and is indexed with the grid cells described above Although the same letter is used, the meaning of the letter is different.
[0083] After convolution, the ReLU activation function (f(x)=max(0,x)) is used to introduce non-linearity. The pooling layer uses max pooling (kernel size 2×2×2) to gradually compress the size of the feature map and enhance translation invariance. After multiple convolution-pooling operations, the feature tensor is compressed into a one-dimensional feature vector of length 256 by the global average pooling layer, and then input into a classifier composed of two fully connected layers. The first fully connected layer contains 128 neurons and uses the ReLU activation; the second output layer contains 4 neurons, corresponding to four risk categories (blue, yellow, orange, red) respectively. Finally, the classification probability is output through the Softmax function, and the calculation formula is: , where zc is the activation value of the c-th neuron in the output layer, and zj is the activation value of the j-th neuron in the output layer.
[0084] S333. Select the probability representing the high-risk level as the original output of the three-dimensional visual modality. The high-risk level is that a water-conducting channel has been formed, and the original output is normalized.
[0085] High-risk probability determination: According to the claims, "a water channel has been formed" is set as a high-risk level, and the probability of 0.82 corresponding to this level is extracted as the original output of the 3D visual modality. .
[0086] Normalization: The min-max normalization method is used, and historical output is processed. Minimum value =0 (no high-risk probability), maximum value =1 (high risk probability 100%), the normalization formula is: Substituting into =(0.82-0) / (1-0)=0.82 (range [0,1]).
[0087] The process involves weighting and fusing the normalized results according to pre-set confidence weights to obtain a final risk value, comparing the final risk value with a pre-set warning threshold, and triggering a corresponding level of warning, including: S341. Set multimodal reliability weights to assign reliability weights to the normalized knowledge reasoning modal output, acoustic alarm modal output, and 3D vision modal output.
[0088] Based on the early warning reliability of each modality (3D visual modality directly outputs risk probability, with the highest reliability; knowledge reasoning modality is based on rule logic, followed by acoustic alarm modality, which only provides auxiliary prompts and has the lowest reliability), reliability weights are assigned by combining domain expert experience, and the sum of the weights must be 1: Knowledge reasoning modal weights =0.4; Acoustic alarm modal weights =0.4; 3D visual modal weights =0.2.
[0089] S342. The normalized outputs of the three modes are weighted and fused according to their respective weights, using the following formula: , in, This is the final risk value. , , These are the reliability weights for the knowledge reasoning modality, the acoustic alarm modality, and the 3D vision modality, respectively. , , These are the outputs after modal normalization, respectively.
[0090] Substitute into the formula: , =0.4×0.9+0.4×0.667+0.2×0.82≈0.79 (range [0,1]).
[0091] S343. Set a multi-level fusion early warning threshold, compare the final risk value with the multi-level early warning threshold, and trigger an early warning of the corresponding level based on the comparison result; Data from 40 early warning cases at the mine from 2019 to 2024 were collected, and the final risk value was calculated. Correspondence with actual flood conditions: When the value is ≤0.2, 100% of cases are risk-free (no water channel). 0.2< When the value is ≤0.4, 88% of the cases are of medium risk (potential water-conducting fractures). 0.4 When the value is ≤0.7, 92% of the cases are high-risk (a water diversion channel has been formed). When the value is greater than 0.7, 95% of the cases are considered to be at extreme risk (either a sudden water inrush occurs or conditions for a sudden water inrush are present). Based on the fault risk characteristics of the working face, a multi-level fusion early warning threshold was set after calibration: =0.2 (low-risk threshold) =0.4 (medium risk threshold) =0.7 (high-risk threshold) =0.9 (extreme risk threshold).
[0092] The multimodal early warning mechanism constructed in this embodiment is an intelligent system based on multi-source information fusion, parallel analysis, and collaborative decision-making. Its core lies in breaking through the traditional single alarm mode. Through the mutual verification and supplementation of reasoning, sound, and visual modalities, it achieves three-dimensional and highly reliable perception of precursors to fault-related flooding. Multimodal cross-validation greatly reduces false alarms and missed alarms caused by distortion from a single data source or model misjudgment. It satisfies the cognitive habits of different roles (experts, managers, and field workers), ensuring that early warning information is quickly and accurately understood. The combination of the temporal prediction capability of the visual modality and the logical reasoning capability of the reasoning modality enables early warning of disaster trends, not just the current state.
[0093] S344. When the final risk value is less than or equal to the low-level warning threshold, a low-risk warning is triggered, and routine monitoring is maintained; when the final risk value is greater than the low-level warning threshold but less than or equal to the medium-level warning threshold, a medium-risk warning is triggered, and the frequency of monitoring data collection is increased; when the final risk value is greater than the medium-level warning threshold but less than or equal to the higher-level warning threshold, a high-risk warning is triggered, and preliminary preparations for hazard mitigation operations are initiated; when the final risk value is greater than the higher-level warning threshold, an extreme risk warning is triggered, the flood emergency plan is activated, and personnel in the danger zone are guided to evacuate.
[0094] Low risk warning ( ≤0.2): Maintain routine monitoring, with data acquisition intervals of 1 hour, 30 minutes, and 15 minutes for microseismic, electrical, and hydrological data, respectively, without requiring additional intervention; Medium risk warning (0.2 < ≤0.4): Increase monitoring frequency, reduce the microseismic acquisition interval to 30 minutes, electrical resistivity tomography to 15 minutes, and hydrological data to 5 minutes, and have technical personnel analyze risk change trends daily; High-risk warning (0.4 < l≤0.7): Initiate preliminary preparations for the emergency response operation, allocate 2 ZDY4000L drilling rigs and 50t of cement-water glass dual-liquid slurry, delineate the drilling site location (X=5500m, Y=2715m), and further reduce the monitoring frequency to 15 minutes for micro-vibration, 10 minutes for electrical resistivity tomography, and 3 minutes for hydrology. Extreme risk warning ( >0.9): Immediately activate the water hazard emergency plan, notify personnel within 500m of the F8 fault via underground broadcast and personnel positioning system to evacuate along the preset evacuation route (track roadway → main inclined shaft), shut off power supply to the area (except for emergency lighting), coordinate the grouting team to be on standby 24 hours a day, and report to the mine-level emergency command center at the same time.
[0095] Reference Figure 4 S4. Based on the corresponding level of the warning grid, identify the grid with high risk warning and initiate the drilling verification and grouting sealing and hazard relief process; First, a standardized, targeted drilling verification and hazard mitigation operation procedure is established. This procedure follows the principles of "precise positioning, rapid verification, and efficient hazard mitigation" and is divided into four core stages: (1) Precise Targeting and Positioning Stage: Based on the core coordinates of the risk anomaly area determined by the microseismic-electrical coupling analysis, the system automatically generates a borehole design document, which clarifies the azimuth, dip angle, depth and target stratum of the borehole, ensuring that each borehole points directly to the potential water-conducting channel and avoiding blind borehole layout; (2) Drilling verification and real-time identification stage: During the drilling process, it is required to record the changes in lithology in detail. More importantly, the borehole television technology is used to perform real-time imaging and identification of the fractures and karst development revealed by the borehole. Combined with the real-time monitoring data of borehole water (water volume, water pressure, water temperature), the location is quickly identified on site as a water inrush channel. (3) Targeted grouting and sealing stage: If it is verified that it is indeed a water-conducting channel, the grouting and hazard-relief operation will be started immediately. Based on the characteristics of the water outlet (clear water or turbid water), water volume and water pressure, the grouting materials (such as cement-based, chemical grout), proportions and processes (such as intermittent grouting, pressure-controlled grouting) will be intelligently matched to achieve precise treatment of "one hole, one policy"; (4) Effect inspection and closed-loop stage: After the grouting is completed, the diffusion radius of the grouting and the plugging effect are inspected through re-drilling water pressure tests or by using borehole transient electromagnetic method (TEM) to ensure the formation of an effective water-blocking barrier and complete the closed-loop management from "detection - verification - treatment - inspection".
[0096] Secondly, a real-time supervision method for standardized operations based on computer vision is constructed. To address the non-standardization of manual operations in the above process, high-definition cameras are deployed at key perspectives in the drill site, and the YOLOv7 object detection and deep learning algorithms are integrated to conduct visual monitoring and intelligent analysis of the entire operation process: (1) Supervision of drilling parameter compliance: The angle of the drill pipe is identified in real time, and the deviation between its angle and the designed angle is calculated through image algorithms to ensure the accuracy requirement of ±2°. The number of casings lowered is automatically counted to verify whether it is consistent with the design. (2) Supervision of key processes and safety: Key process nodes such as hole opening, casing lowering, grouting pipeline connection, and hole sealing are identified, and whether the process is compliant and whether there are any skipped steps are judged through sequential action analysis. At the same time, the status of the hole mouth is continuously monitored. Once abnormalities such as "water outflow" and "hole gushing" are visually recognized, the highest-level alarm is immediately triggered. (3) Supervision of personnel behavior and safety: Automatically identify whether the operators wear safety helmets properly and whether non-staff members enter the dangerous area, etc., to improve the safety management level.
[0097] From all the grids in the full monitoring area, screen the grids whose final risk values in step S3 meet "0.4 < ≤ 0.7". The coupling coefficient K of the above grids is in the range of 20 < K ≤ 30, corresponding to the risk state of "water-conducting channels have been formed", and they are determined as the core target areas for drilling verification and grouting plugging.
[0098] Drill site layout and equipment selection: One standardized drill site is arranged at the projection position of the high-risk grid group in the working face transportation lane, and the ZDY4000L full-hydraulic mine roadway drill is selected.
[0099] Sectional grouting construction: The first section (hole depth 68 - 72m, water-conducting channel section): Low-pressure slow infiltration grouting is adopted. The grouting pressure is gradually increased from 1.0 MPa to 2.2 MPa (0.85 times the water pressure of the aquifer), the grouting flow rate is controlled at 50 - 80 L / min, and the grouting is continued for 40 min, injecting 12 m³ of slurry. The second section (hole depth 60 - 68m, fracture-developed section): The grouting pressure is increased to 2.5 MPa, the flow rate is 80 - 120 L / min, and the grouting is continued for 30 min, injecting 8 m³ of slurry. Grouting effect verification: 72 hours after grouting, a water pressure test was conducted on borehole Z2 (test pressure 2.0 MPa, lasting 30 min). The leakage rate was monitored to be ≤0.3 L / min, indicating that the sealing was qualified. At the same time, the coupling coefficient K=12.5 (reduced to the medium risk range) of the corresponding grid (109,25,9) of this borehole was re-measured, and the final risk value was determined. =0.38 (dropped below the medium-risk warning threshold), crisis resolution process completed.
[0100] S5. Construct an early warning model, define a state space that includes monitoring features, model performance indicators and geological context, use the early warning threshold and confidence weight of each grid as the adjustment objects, design a reward function based on early warning effect and action stability, and update the model based on the reward function.
[0101] The aforementioned early warning model is constructed by defining a state space that includes monitoring features, model performance indicators, and geological context, including: S51. The monitoring features include the mean of the coupling coefficient, the standard deviation of the feature parameters, the maximum energy of the microseismic event and the average water output during the set monitoring period. The model performance indicators include the accuracy, false alarm rate and missed alarm rate of the most recent warnings. The geological and environmental context includes the geological structure complexity, working face advance speed, water pressure change rate, current warning level and hazard mitigation effect score. The reinforcement learning module (i.e., the early warning model) integrated in this embodiment endows the early warning system with the ability to dynamically optimize. This module treats the adjustment of parameters of the early warning model (such as the threshold of the coupling coefficient K and multimodal weights) as the actions of the agent. After each early warning is issued, the system receives reward feedback based on subsequent actual drilling verification or disaster conditions: accurate early warnings are rewarded positively, while false alarms or missed alarms are rewarded negatively. Through algorithms such as policy gradient, the agent continuously learns from these feedbacks and autonomously adjusts its decision-making strategy, thereby finding the optimal parameter combination through continuous interaction. For example, the model may discover that fine-tuning the K threshold from 25 to 23 in the current mining area can significantly reduce missed alarms without significantly increasing false alarms. This closed-loop learning mechanism enables the system to proactively adapt to dynamic changes in geological conditions, achieving autonomous and continuous evolution of the early warning model and continuously improving the accuracy and reliability of early warnings.
[0102] The state space contains 12 features, defined as follows: Monitoring characteristics (4 dimensions): ① Mean coupling coefficient (arithmetic mean of grid K values in the last 24 hours); ② Standard deviation of coupling coefficient (dispersion of grid K values in the last 24 hours); ③ Maximum energy of microseismic events (maximum microseismic energy within the grid in the last 24 hours); ④ Average water output (average water inflow of drilling verification boreholes, 0 when no boreholes are drilled). Model performance metrics (3 dimensions): ① Early warning accuracy (the percentage of accurate early warnings out of the last 10 warnings); ② False alarm rate (the percentage of false alarms out of the last 10 warnings); ③ Missed warning rate (the percentage of no warnings out of the last 10 actual risks). Geological and environmental context (5 dimensions): ① Geological structural complexity ("simple=1, medium=2, complex=3", F8 fault area is 3); ② Working face advance speed (average daily advance distance, unit m / d); ③ Water pressure change rate (maximum water pressure change in the last 24 hours, unit MPa / h); ④ Current warning level ("low=1, medium=2, high=3, extreme=4"); ⑤ Risk mitigation effect score (risk reduction rate after grouting, range 0~100 points).
[0103] S52. Obtain the original values of each dimension feature, wherein the original values of the monitoring features are obtained by statistically analyzing the monitoring data within the set monitoring period, the original values of the model performance indicators are obtained by comparing the results of the most recent warnings with the actual situation, and the original values of the geological and environmental context are obtained by collecting corresponding geological and operational environment parameters. The original values of all the obtained dimension features are combined by category to form a multi-dimensional feature vector.
[0104] S53. The multi-dimensional feature vector is subjected to Z-score standardization, mapping each feature to a preset interval. The standardization method is as follows: ,in, For the first 3D features at time step Standardized value For the first 3D features at time step The original value, For the first Historical mean of the dimensional feature For the first Historical standard deviation of the dimensional feature.
[0105] S54. The standardized multi-dimensional feature vector is the defined state space that includes monitoring features, model performance indicators, and geological context.
[0106] The design is based on a reward function that considers both the early warning effect and the stability of the action. The model is updated based on this reward function, including: S55. Design a reward function, which is expressed by the formula: , in, For accurate early warning indication functions, the value must be 1. This is a false alarm indicator function; a false alarm is set to 1. This is a missed detection indicator function; the value is 1 for missed detections. , , The penalty coefficient is... The penalty coefficient for the action. Let be the norm of the action vector.
[0107] Based on the importance of early warning, we set α=10 (reward for accurate early warning), β=5 (penalty for false alarm), γ=20 (penalty for missed alarm), and λ=0.1 (penalty for smooth action).
[0108] S56. Based on the current state Select Action Observe the new state after execution. And get rewards , will experience tuples ( , , , The data is stored in an experience buffer. This design enables the system to adapt to early warning needs under different geological conditions, continuously optimize decision-making strategies through continuous learning, and ultimately achieve a steady improvement in the accuracy and reliability of early warnings.
[0109] S57. The PPO algorithm is adopted to calculate the action advantage through generalized advantage estimation, update the policy network by minimizing the pruning objective function, and update the state value network by minimizing the mean squared error loss. Both the policy network (output action probability distribution) and the state value network (output state value) adopt a 3-layer fully connected neural network (64 and 32 hidden layer neurons, ReLU activation function) to minimize the mean squared error loss and improve the accuracy of model warning.
[0110] This embodiment achieves precise positioning of water-conducting channels (grid accuracy 5m³) through a microseismic-electrical resistivity spatial coupling model (K-value algorithm), significantly improving the accuracy of fault water hazard perception. Multimodal early warning integrates logical reasoning, acoustic prompts, and visual visualization, significantly improving early warning accuracy. Computer vision (YOLOv7) enables standardized supervision of the entire drilling operation process, replacing manual inspections and improving geophysical exploration and drilling efficiency. The platform has self-learning and semantic understanding capabilities, achieving an upgrade from "data-driven" to "knowledge-driven," contributing to mine geological transparency and dynamic disaster risk management.
[0111] Microseismic events directly reflect the location of rock mass fractures but are insensitive to water content; electrical resistivity methods can effectively detect water-rich anomalies through resistivity differences. Combining the two methods can pinpoint the location of water inrush channels. This embodiment primarily involves manually spatially overlaying the monitoring results from the two methods, allowing for targeted monitoring of specific areas within the structure and foundation to effectively track the spatiotemporal evolution of water hazards.
[0112] Example 2 This embodiment provides a fault-based flood disaster sensing and mitigation system, including: The data acquisition module is configured to set the monitoring area and perform three-dimensional spatial grid segmentation, acquire microseismic monitoring data, electrical resistivity monitoring data, and hydrological monitoring data of the monitoring area, and construct a standardized data model. The risk level determination module is configured to calculate the coupling coefficient of the data in each grid, set multi-level risk thresholds, determine the risk level of the coupling coefficient of each grid, and visualize and render grids with different risk levels in a three-dimensional geological model. The multimodal fusion early warning module is configured to construct a knowledge reasoning modality, an acoustic alarm modality, and a three-dimensional visual modality based on the determined grid risk level. The original output of each modality is normalized, and the normalization results are weighted and fused according to the preset confidence weights to obtain the final risk value. The final risk value is compared with the preset early warning threshold to trigger the corresponding level of early warning. The hazard mitigation module is configured to identify high-risk grids based on the corresponding level of early warning and initiate drilling verification and grouting sealing hazard mitigation procedures. The early warning model update module is configured to construct an early warning model, define a state space that includes monitoring features, model performance indicators and geological context, use the early warning threshold and confidence weight of each grid as adjustment objects, design a reward function based on early warning effect and action stability, and update the model based on the reward function.
[0113] A computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of a terminal device of the method for sensing and mitigating fault-flood disasters.
[0114] A terminal device includes a processor and a computer-readable storage medium, the processor being used to implement various instructions; the computer-readable storage medium being used to store multiple instructions, the instructions being adapted to be loaded and executed by the processor to provide a method for sensing and mitigating fault-related flood disasters.
[0115] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.
Claims
1. A method for sensing and mitigating fault-related water hazard disasters, characterized in that, include: The monitoring area is defined and divided into three-dimensional spatial grids. Microseismic monitoring data, electrical resistivity monitoring data, and hydrological monitoring data of the monitoring area are obtained, and a standardized data model is constructed. Calculate the coupling coefficient of the data in each grid, set multi-level risk thresholds, determine the risk level of the coupling coefficient of each grid, and visualize and render grids with different risk levels in a three-dimensional geological model. Based on the determined grid risk level, a knowledge reasoning modality, an acoustic alarm modality, and a 3D visual modality are constructed. The original outputs of each modality are normalized, and the normalization results are weighted and fused according to the preset confidence weights to obtain the final risk value. The final risk value is compared with the preset warning threshold to trigger the corresponding level of warning. Based on the corresponding level of the warning grid, the grid with high risk warning is identified, and the drilling verification and grouting sealing and hazard relief process is initiated. An early warning model is constructed, defining a state space that includes monitoring features, model performance indicators, and geological context. The early warning threshold and confidence weight of each grid are used as adjustment objects. A reward function based on early warning effect and action stability is designed, and the model is updated based on the reward function.
2. The method for sensing and mitigating fault-related water hazard disasters according to claim 1, characterized in that, The calculation of the coupling coefficient of the data in each grid includes: The number of microseismic events within the grid is counted, and the rate of change of apparent resistivity within the grid is calculated. The geometric center of the low resistivity anomaly region in the electrical resistivity test was determined using a clustering algorithm, and the minimum Euclidean distance from the grid center to the geometric center was calculated. ; The coupling coefficient for each grid cell is calculated using the following formula: ,in, The coupling coefficient of the mesh. This represents the number of microseismic events within the grid. The apparent resistivity change rate within the grid. This represents the minimum distance from the grid center to the geometric center of the low-resistivity anomaly region in the electrical resistivity test. It is a smoothing constant. A unique index for the grid location.
3. The method for fault-based water hazard detection and mitigation operations according to claim 1, characterized in that, The visualization and rendering of grids with different risk levels in a 3D geological model includes: Determine the monitoring area The range of directions, assigning a unique index to each grid. And set the side length of the three-dimensional spatial mesh to L; Calculate the center coordinates of each grid cell to obtain the three-dimensional center coordinates of each grid cell. ; Set a low-risk threshold Medium risk threshold High-risk threshold And satisfy ,in, There is no water guiding channel. As a potential water-conducting fracture, A water diversion channel has been formed. A sudden water inrush occurred; In a three-dimensional geological model, according to The value determines the color rendering of the grid.
4. The method for sensing and mitigating fault-related water hazard disasters according to claim 1, characterized in that, Constructing the knowledge reasoning modality includes: Construct a knowledge graph containing entities and relationships related to fault-related water hazards; The entity and relationship of the knowledge graph are stored using a graph database. Multiple risk inference rules are defined using a graph query language. The inference rules are set based on entity attributes and the relationships between entities. Real-time monitoring data is matched with the inference rules, and the number of successfully matched rules is counted. Each inference rule is assigned a weight, which is set based on the importance of the inference rule to the risk assessment. The sum of the weights of all inference rules is 1. When the inference rules are completely matched, the matching score is 1. When the inference rules are not completely matched, the matching score is 0. Substitute the values into the formula to calculate the original risk confidence level. The formula is: ,in, The number of rules to be matched. For rule weights, The rule matching score is 1 for a perfect match and 0 for an otherwise perfect match. The original risk confidence score is then normalized.
5. The method for sensing and mitigating fault-related water hazard disasters according to claim 1, characterized in that, Constructing the acoustic alarm mode includes: Obtain the frequency spectrum of each microseismic event and extract the frequency corresponding to the maximum amplitude in the spectrum as the dominant frequency. Map the dominant frequency from the effective frequency band of micro-vibrations to the audible frequency band that the human ear is sensitive to; The energy of microseismic events is mapped to the actual alarm volume using a piecewise linear function; When the mesh coupling coefficient is greater than the medium risk threshold At that time, the mapped audible sound waves and voice prompts are played, and the actual alarm volume is used as the raw output of the acoustic mode, and the raw output is normalized.
6. The method for sensing and mitigating fault-related water hazard disasters according to claim 1, characterized in that, Constructing the three-dimensional visual modality includes: Based on a three-dimensional spatial grid, a four-dimensional feature tensor is constructed. The four-dimensional feature tensor includes multiple feature dimensions, including grid coupling coefficient, cumulative energy of microseismic events, apparent resistivity change rate, microseismic event density, P-wave / S-wave velocity ratio, and distance from the grid to the fault. A three-dimensional convolutional neural network model architecture is constructed. Its input layer receives a four-dimensional feature tensor, and the output layer outputs the risk level probability of each grid through the Softmax function. The risk level probability includes the probability of no water channel, the probability of potential water-conducting fracture, the probability of water channel, and the probability of water inrush. The probability representing a high-risk level is selected as the original output of the three-dimensional visual modality, where the high-risk level is the formed water channel, and the original output is normalized.
7. The method for fault-based water hazard detection and mitigation operations according to claim 1, characterized in that, The process involves weighting and fusing the normalized results according to pre-set confidence weights to obtain a final risk value, comparing the final risk value with a pre-set warning threshold, and triggering a corresponding level of warning, including: Multimodal reliability weights are set to assign reliability weights to the normalized knowledge reasoning modal output, acoustic alarm modal output, and 3D vision modal output; The normalized outputs of the three modalities are weighted and fused together using the following formula: , in, This is the final risk value. , , These are the reliability weights for the knowledge reasoning modality, the acoustic alarm modality, and the 3D vision modality, respectively. , , These are the outputs after modal normalization, respectively; Set multi-level fusion early warning thresholds, compare the final risk value with the multi-level early warning thresholds, and trigger the corresponding level of early warning based on the comparison results.
8. The method for sensing and mitigating fault-related water hazard disasters according to claim 1, characterized in that, The aforementioned early warning model is constructed by defining a state space that includes monitoring features, model performance indicators, and geological context, including: The monitoring features include the mean of the coupling coefficient, the standard deviation of the feature parameters, the maximum energy of the microseismic event, and the average water discharge within the set monitoring period. The model performance indicators include the accuracy, false alarm rate, and missed alarm rate of the most recent warnings. The geological and environmental context includes the geological structure complexity, working face advance speed, water pressure change rate, current warning level, and hazard mitigation effect score. The original values of each dimension feature are obtained. The original values of the monitoring features are obtained by statistically analyzing the monitoring data within the set monitoring period. The original values of the model performance indicators are obtained by comparing the results of the most recent warnings with the actual situation. The original values of the geological and environmental context are obtained by collecting the corresponding geological and operational environment parameters. The original values of all the obtained dimension features are combined by category to form a multi-dimensional feature vector. The multi-dimensional feature vectors are subjected to Z-score normalization, which maps each feature to a preset interval. The normalization method is as follows: ,in, For the first 3D features at time step Standardized value For the first 3D features at time step The original value, For the first Historical mean of the dimensional feature For the first Historical standard deviation of dimensional features; The standardized multi-dimensional feature vector is the defined state space that includes monitoring features, model performance indicators, and geological context.
9. The method for sensing and mitigating fault-related water hazard disasters according to claim 1, characterized in that, The design is based on a reward function that considers both the early warning effect and the stability of the action. The model is updated based on this reward function, including: Design a reward function, which is expressed by the formula: , in, For accurate early warning indication functions, the value must be 1. This is a false alarm indicator function; a false alarm is set to 1. This is a missed detection indicator function; the value is 1 for missed detections. , , The penalty coefficient is... The penalty coefficient for the action. Let be the norm of the action vector; Based on the current state Select Action Observe the new state after execution. And get rewards , will experience tuples ( , , , Stored in the experience buffer; The PPO algorithm is adopted to calculate the action advantage through generalized advantage estimation, update the policy network by minimizing the pruning objective function, and update the state value network by minimizing the mean squared error loss.
10. A fault-based flood disaster sensing and mitigation system, characterized in that, The method for fault-based flood disaster sensing and mitigation as described in any one of claims 1-9 includes: The data acquisition module is configured to set the monitoring area and perform three-dimensional spatial grid segmentation, acquire microseismic monitoring data, electrical resistivity monitoring data, and hydrological monitoring data of the monitoring area, and construct a standardized data model. The risk level determination module is configured to calculate the coupling coefficient of the data in each grid, set multi-level risk thresholds, determine the risk level of the coupling coefficient of each grid, and visualize and render grids with different risk levels in a three-dimensional geological model. The multimodal fusion early warning module is configured to construct a knowledge reasoning modality, an acoustic alarm modality, and a three-dimensional visual modality based on the determined grid risk level. The original output of each modality is normalized, and the normalization results are weighted and fused according to the preset confidence weights to obtain the final risk value. The final risk value is compared with the preset early warning threshold to trigger the corresponding level of early warning. The hazard mitigation module is configured to identify high-risk grids based on the corresponding level of early warning and initiate drilling verification and grouting sealing hazard mitigation procedures. The early warning model update module is configured to construct an early warning model, define a state space that includes monitoring features, model performance indicators and geological context, use the early warning threshold and confidence weight of each grid as adjustment objects, design a reward function based on early warning effect and action stability, and update the model based on the reward function.
Citation Information
Patent Citations
Method for judging alarm threshold value for water disaster micro-earthquake
CN113605984A
Mine multi-dimensional fusion intelligent monitoring and early warning system method
CN119393187A
Coal mine geological disaster and environment early warning method based on multi-source monitoring data fusion
CN119494536A
Hidden structure detection system for intelligent monitoring, early warning and prevention of mine water disasters
CN120254952A
Coal mine disaster early warning method and system based on geological model, and storage medium
CN120412198A
Cited By
Intelligent decision-making system and method for bed separation grouting opportunity based on sky-ground cooperation
CN121598034A