Mine water distribution prediction method based on geophysics

Abstract

The invention provides a mine water distribution prediction method based on geophysics. The mine water distribution prediction method comprises the following steps: acquiring a time sequence signal of resistivity data acquired by a direct current resistivity method for a mine measuring area; carrying out feature extraction on the resistivity data time sequence signal of the mine measuring area obtained in the monitoring period to obtain feature information; constructing a weighted feature vector according to the feature information, and dividing according to a time window to form a structured time sequence matrix; according to the time sequence matrix, a mature prediction model is obtained through joint training, and an area where water disasters may occur is predicted; a visual monitoring and predicting working system is developed, and a monitoring and predicting result obtained by an underground system is reflected to a ground dispatching center in real time, so that ground personnel can conveniently carry out real-time emergency dispatching. Water inrush accidents possibly existing on the coal seam working face can be rapidly and accurately predicted, dynamic disaster early warning in coal mine production work is further achieved, and safe production is guaranteed. The method is used in the process of predicting the specific conditions of the mine water disasters.

Description

Technical Field

[0001] This invention relates to the field of coal mine safety, and more particularly to a geophysical method for predicting mine water distribution. Background Technology

[0002] Coal, as the "ballast" of my country's energy security, will not change its dominant position in the short term. However, with the shift of mining focus to deeper areas and increasingly complex geological conditions, coal mine safety faces severe challenges. Among these, the threat of water inrush in confined water mining is particularly prominent. High ground stress at depth makes the roof and floor strata of coal seams more prone to mining-induced fractures. These fractures connect with the highly water-rich Ordovician limestone aquifer, forming water-conducting channels. Once a water inrush is triggered, the volume, pressure, and force are enormous, easily leading to catastrophic accidents such as mine flooding and face inundation. Therefore, relying solely on traditional experience analysis is insufficient to address such hidden and dynamic disaster risks. There is an urgent need to develop an intelligent prediction method for mine water hazards that integrates multi-source geological information and provides real-time dynamic early warning. This method would enable accurate identification of water inrush channels and timely capture of disaster precursors, providing crucial technical support for the safe and efficient mining of deep coal resources. Summary of the Invention

[0003] To overcome the shortcomings of existing technologies, this invention provides a geophysical method for predicting mine water distribution, which solves the problem of being unable to predict the spatiotemporal changes of water-rich areas in the top and bottom strata of the working face in real time, effectively avoids the risk of mine water hazards, and reduces the cost of coal mine water control.

[0004] This invention is achieved through the following technical solution: a geophysical method for predicting mine water distribution, specifically including the following steps: Step S1: Obtaining the basic data for prediction: Geophysical data acquisition electrodes, detectors, sensors and other monitoring equipment are deployed in the upper and lower roadways and cut-off locations of the coal mining face. Periodic data acquisition is carried out according to the mining progress of the working face to build a geophysical database for predicting mine water distribution. Step S2, Geophysical Data Feature Extraction: A phased feature extraction strategy is designed, and feature-level fusion and spatiotemporal alignment mechanisms are used to achieve multi-source data fusion and extract the temporal and spatial features of the prediction data. Step S3: Preliminary establishment of the time series prediction model: Construct a hybrid architecture baseline model based on, but not limited to, LSTM-Transformer, with the aligned underlying data time series matrix as input; Step S4, Polynomial Fitting and Deep Learning Prediction: By training two heterogeneous model branches in parallel, namely, the safe data branch and the water leakage data branch, the model sensitivity problem under extreme sample imbalance is solved based on the Facolloss loss function and the dual-branch cross-attention training. Step S4, Post-processing and Result Optimization: Spatiotemporal smoothing and consistency verification are performed on the original prediction results; Step S5, Threshold Setting and Early Warning Mechanism: The output includes predicted values, risk scores, early warning signals, and visualizations. A multi-level design achieves a closed loop from data to decision-making. It is adjusted based on real-time multi-source data, environmental and traffic characteristics to adapt to environmental changes and mine data features. An early warning mechanism with three levels is set. Step S6, Deployment and Monitoring: Design the edge layer, cloud layer, and visualization interface; Edge layer: namely, road segment equipment, used to run self-encoders, calculate micro-motion time-domain characteristics, and cache 12 hours of raw data; Cloud layer: namely, the central server, which meets storage and computing needs; Visualization interface: displays risk scores, threshold lines, and historical comparison curves for each area of ​​the work surface; The visualization interface is combined with the thresholds.

[0005] As a preferred approach, the extraction of multi-source isomorphic data in step S1 specifically includes the following steps: Step S1-1: Extract features from DC resistivity sensor data: Step S1-1-1, Time Domain Features: (1) Rate of change of dynamic resistivity: Calculate the rate of change of resistivity for continuous time series data at the same measuring point. ; in, Let be the resistivity value at time t. ; ... (2) Sliding time window statistics: Set a sliding window and calculate the mean, standard deviation and other statistics of resistivity within the window to capture short-term fluctuations and long-term trends.

[0006] W represents the window width; Smooth noise and capture trends; if σ continues to increase, it indicates the development of rock fractures. (3) Time correlation analysis: By employing cross-correlation functions or dynamic time warping algorithms, the hysteresis correlation of resistivity time series at different measuring points is analyzed to identify water damage diffusion paths. The formula represents the lag correlation between measurement points A and B. If the maximum cross-correlation occurs... The number of days indicates that it takes 2 days for water to diffuse from point A to point B, which guides the warning time. (4) LSTM temporal feature extraction: Using LSTM-Transformer neural network, the periodicity and trend features of resistivity sequence are automatically extracted, solving the long-term dependency problem of traditional recurrent neural network, and outputting the hidden layer state as a high-order time feature. The core equation is as follows Automatically learns the long-term dependencies of resistivity sequences and outputs the hidden state. As a higher-order feature; Step S1-1-2, Spatial Features: (1) Calculation of resistivity spatial gradient: Calculate the gradients in the horizontal and vertical directions for two-dimensional / three-dimensional resistivity data at the same time. formula: in, The electrode spacing is set; (2) Spatial cluster analysis: Apply DBSCAN or K-means algorithms to cluster resistivity data according to spatial location and numerical value. Identify low-resistivity anomaly regions; The formula for the spatial variation index can be used: d represents the distance afterward, and N(d) represents the number of corresponding point pairs; (3) Three-dimensional resistivity volume feature extraction: For the three-dimensional inversion results, multiple slices are extracted along the strike / dip of the working surface, and the statistical characteristics of the resistivity of each slice are calculated. in, The volume occupied by the low-resistivity region. This represents the percentage of the area in the low-resistivity region. ; (4) U-Net spatial feature extraction: Resistivity data is input into a convolutional neural network (CNN), and multi-scale spatial features are extracted by an encoder. in, The kernel size is [size]. To make the weights trainable, features from local to global are extracted through multiple layers of convolution and pooling; Step S1-2: Implement multi-source data fusion using feature-level fusion and spatiotemporal alignment mechanisms. Spatiotemporal alignment of data for different working areas and different survey lines; Step S1-2-1, Time Synchronization (NTP Protocol): Unified timestamps via GPS / NTP protocol: Hardware configuration: Each resistivity meter has a built-in NTP client, a local NTP server is deployed, and time synchronization verification is performed. Calculate the clock offset: in: : Client request sending time : Server request reception time Server response time : Client response time, requiring all devices If the absolute value is less than 1ms, an alarm will be triggered. get Then, through the formula Obtain the calibrated standard time Step S1-2-2, Spatial Registration: The resistivity data from multiple measurement lines distributed in three-dimensional space are unified to a global coordinate system through coordinate transformation: the start / end point of each measurement line is recorded using GPS WGS84 coordinates, and a transformation from the local coordinate system of the measurement line to the global coordinate system is established. The coordinate transformation formula is as follows: Steps S1-3: Feature splicing: Step S1-3-1: Align the temporal and spatial characteristics of each sensor according to the time window to form a multidimensional input matrix: in, Where K is the number of survey lines, K is the feature dimension of each survey line, T is the length of the time window, and R is the set of real numbers, emphasizing the numerical characteristics of the input data. The following data are included in the two feature dimensions of each survey line: Time domain dimension: original resistivity ; indicates directly observed value; rate of change of resistivity ; indicates the rate of dynamic change; ; Indicates the overall trend; ; indicates local fluctuations; Spatial dimension: lateral gradient ; indicates a change in the direction of the survey line; vertical gradient Spatial variation index ; Indicates the spatial variation of resistivity in the coal mining face; the proportion of low-resistivity areas within the measurement area. ; indicates the degree of danger in the roof area; Step S1-3-2, Feature Normalization: Perform Z-Score standardization independently for each feature dimension: Where x represents the original data value; μ represents the center position of the mean dataset; and σ represents the standard deviation. This represents the standardized value.

[0007] As a preferred option, step S2 specifically includes the following steps: Step S2-1: Generate a high-dimensional temporal matrix after multi-source feature fusion: Step S2-1-1, Shape: , in; K represents the number of survey lines, K represents the feature dimension of each survey line, and T represents the length of the time window. Step S2-1-2, Normalization: Perform Z-Score standardization independently for each feature dimension: ;in, , Obtained based on the training set; Step S2-2: Generate the preliminary prediction model structure: LSTM regulates the flow of information through three gating structures: the forget gate, the input gate, and the output gate. Forgotten Gate: Decide how much old data to discard. Input Gate: Determine how much new data to store. Candidate memories: Generate the data to be written. Output gate: Control the amount of information currently output. in, , , : The resistivity feature vector at time t. Trainable parameter matrix; Step S2-3: Generate the memory update equation: in, Current state Hadamard product; Step S2-4, Output hidden state: Compress the memory content into the range [-1, 1].

[0008] As a preferred option, step S3 specifically includes the following steps: Step S3-1: Set up the branch architecture: Step S3-1-1, Security Branch Main Input data: security sample Loss function: Cross-entropy ; Step S3-1-2, seepage branch Aux Input data: seepage sample Loss function: Cross-entropy Optimization objective: Maximize the hazard detection rate; Step S3-2: Generate the joint loss function: Dynamic weights : Focal Loss parameters: Step S3-3, Compare Losses: in, ; Step S3-4, Feature Fusion: Step S3-4-1, Cross-attention: Allow the query on the safe branch to select key features from the seepage branch. ; ; Parameter matrix ; Step S3-4-2, Multi-head expansion: .

[0009] As a preferred option, step S4 specifically includes the following steps: Step S4-1, Timing Smoothing: Double exponential weighted average, Actual measurement ; Step S4-2, Spatial Consistency Check: For each survey line For each node, a spatial location is represented, and the edge weights depend on the combined distance and geological similarity. edge weight The first term on the right side of the equals sign is the distance term, and the second term is the geological term. , , Then, spectral convolution is used: Perform graph convolution correction; Finally, hard and soft constraints are applied to the model through geological rule injection. Hard constraints: , The representative survey line is located in a known impermeable layer area, so it is forcibly set to zero; Soft constraints: The mapping relationship between lithology and permeability probability is learned through MLP; Step S4-3, Dynamic Threshold Filtering: Adjust the warning threshold based on real-time data distribution. in, ; Step S4-4: Early Warning Event Generation: Warning = .

[0010] As a preferred option, step S5 specifically includes the following steps: Step S5-1, Risk Score Calculation: Step S5-1-1, Historical Baseline: ; Step S5-1-2, Real-time Risk Scoring: ; in, For real-time risk scoring, For the predicted water content at the current time step, The first in the past 30 days Predicted water content at each time step; Step S5-2: Set dynamic thresholds and early warning rules Step S5-2-1, Baseline Threshold: ;in, This represents the average value of the corresponding group of data. The variance of the corresponding group of data; Step S5-2-2, Environmental Correction Factor: When rainfall exceeds 50mm for 3 consecutive days: When the daily rise in groundwater level is greater than 10cm: ; Step S5-2-3, Multi-level Early Warning Rules: Level 1 monitoring level: Or the autoencoder malfunction persists for more than 2 hours; Level 2 Warning Level: Or Level 1 lasts for more than 12 hours; Level 3 Emergency Level: Or the RMS of the micro-motion signal increases by more than 50% week-on-week.

[0011] As a preferred option, step S6 specifically includes the following steps: Step S6-1, Design the edge layer: The edge side is responsible for real-time data preprocessing and lightweight anomaly detection.

[0012] Step S6-2, Design the cloud layer: The cloud is responsible for the centralized training, storage, and large-scale computation of the model.

[0013] Storage: Historical data is stored in time-series databases and spatial databases; Computation: The cloud constitutes the decision center of the system. Based on the cross-attention dual-branch training method described in this invention, it completes the model construction and training on the central server, it iteratively optimizes the model by aggregating real-time data uploaded from the edge layer, and finally outputs the prediction results that drive the visualization interface, forming a complete closed loop from edge perception to cloud decision-making.

[0014] Step S6-3: Generate a visual interface: Displays risk scores, threshold lines, and historical comparison curves for each region.

[0015] By employing the above technical solutions, this invention has the following beneficial effects compared to existing technologies: This invention extracts features from the time-series resistivity data of the monitoring area within a monitoring period to obtain feature information; compares the water inrush threshold and feature information to determine and mark the water inrush range of the monitoring area; constructs a weighted feature vector based on the feature information and divides it according to time windows to form a structured time-series matrix; it can obtain dynamic information of the entire working face from multiple monitoring lines in different monitoring areas of the coal mining face; it integrates the data of the monitoring areas to provide more comprehensive monitoring data and improve the reliability of water hazard early warning; based on the time-series matrix, a mature prediction model is obtained through joint training to predict the seepage range, and a set of visual monitoring and prediction systems is developed to reflect the monitoring and prediction results of the underground system to the ground dispatch center in real time; it not only improves the efficiency and accuracy of mine water hazard monitoring, but also significantly improves the safety and maintenance efficiency of normal production in the coal mining face through scientific early warning and timely response.

[0016] The mine water hazard monitoring and prediction technology based on the DC resistivity method proposed in this invention benefits from the significant electrical difference between water-rich rock masses and conventional rock masses during water inrush, as well as the sensitivity of the DC resistivity method to low-resistivity bodies, thus ensuring the accuracy of the prediction. The relatively simple construction scheme, the auxiliary operation of a large number of communication electronic devices, and the training and establishment of a high-precision prediction model ensure the efficiency of the prediction. The visualized prediction system provides a flexible ground scheduling scheme, making the monitoring and prediction results clearer, more transparent, and improving operability.

[0017] Additional aspects and advantages of the invention will become apparent in the following description or may be learned by practice of the invention. Attached Figure Description

[0018] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which: Fig. 1 This is a schematic diagram of the process of the present invention; Fig. 2 This is a schematic diagram of the mine water hazard prediction steps of the present invention; Fig. 3 This is a schematic diagram of the electronic device of the present invention. Detailed Implementation

[0019] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0020] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0021] The following is combined Figs. 1-3 The following is a detailed description of an embodiment of the present invention using a geophysical method for predicting mine water distribution as an example.

[0022] like Fig. 1 , Fig. 2 As shown, this invention proposes a geophysical-based method for predicting mine water distribution, characterized by the following steps: The process involves extracting temporal and spatial features from DC resistivity data of multiple survey lines within a certain range; standardizing the original feature data set and dynamically weighting the standardized feature data to construct a weighted feature vector; retrieving data from each survey line in multiple survey areas; extracting temporal and spatial features from the DC resistivity receiver data of each survey line; unifying the timestamps of sub-data in the multi-source feature data using the GPS / NTP protocol and spatially registering the sensor coordinates through coordinate transformation; aligning the temporal and spatial features of each sensor according to a time window to form a multi-dimensional input matrix; and standardizing each feature dimension individually to obtain a multi-source structured time-series matrix.

[0023] A multivariate time series matrix is ​​formed by collecting safety data from a large number of experiments and data on near-seepage conditions obtained from indoor simulations. The multi-source time series matrix and a preliminary prediction model are retrieved. A preliminary mature prediction model is obtained by adopting a dual-branch collaborative training architecture. A post-processing pipeline is used to improve the prediction accuracy, resulting in a more mature prediction model. The multi-source time series matrix is ​​input into the prediction model to obtain prediction information. The preset threshold is then obtained.

[0024] It is worth noting that the preset threshold is obtained in the following ways: Obtain historical feature information of historical floods, sort the historical feature information in chronological order, and divide the historical feature information into several groups according to a set period; Calculate the baseline threshold for each group: ;in, This represents the average value of each set of feature information. The variance of each set of feature information; Obtain the environmental correction factor, and adjust the threshold by calculating the product of the environmental correction factor and the baseline threshold to obtain the preset threshold.

[0025] In a specific implementation, the steps for using data for prediction are as follows: feature engineering and multi-source data fusion; initial establishment of the time series prediction model; dual-branch collaborative training; post-processing and result optimization; threshold setting and early warning mechanism; deployment and monitoring.

[0026] Step S1, Feature Engineering and Multi-Source Data Fusion: This involves extracting effective spatiotemporal features for the prediction model from multi-source isomorphic data, i.e., DC resistivity data from different survey lines in different survey areas. Feature-level fusion and spatiotemporal alignment mechanisms are used to achieve multi-source data fusion, supporting model prediction. A phased feature extraction strategy is designed based on sensor type and data characteristics to ensure the scientific validity and engineering practicality of the technical solution. For DC resistivity sensor data features, temporal and spatial features are extracted. The extraction of multi-source isomorphic data specifically includes the following steps: Step S1-1: Extract features from DC resistivity sensor data: Step S1-1-1, Time Domain Features: (1) Rate of change of dynamic resistivity (first-order difference): Calculate the rate of change of resistivity for continuous time series data (e.g., daily / weekly measurements) from the same measuring point. ; in, Let be the resistivity value at time t. ; for time intervals, the rate of change sequence of each measuring point is extracted as dynamic features; the rate of change of resistivity with time is calculated by difference to highlight the abrupt change points caused by seepage in the water-rich area.

[0027] (2) Sliding time window statistics: Set a sliding window and calculate the mean, standard deviation and other statistics of resistivity within the window to capture short-term fluctuations and long-term trends.

[0028] W represents the window width; Smooth noise and capture trends; if σ continues to increase (e.g., >3Ω·m), it indicates the development of rock strata fractures. (3) Time correlation analysis (cross-correlation function): Using cross-correlation functions or dynamic time warping (DTW) algorithms, the resistivity at different measurement points is analyzed. Lagged correlations between time series can be used to identify water hazard diffusion pathways. The formula represents the lag correlation between measurement points A and B. If the maximum cross-correlation occurs... The number of days indicates that it takes 2 days for water to diffuse from point A to point B, which guides the warning time. (4) LSTM temporal feature extraction: Using LSTM or Transformer neural networks, the periodicity, trend and other features of resistivity sequences are automatically extracted, and the hidden state is output as a higher-order time feature. The core equation is as follows Automatically learns the long-term dependencies of resistivity sequences and outputs the hidden state. As a higher-order feature; Step S1-1-2, Spatial Features: (1) Calculation of resistivity spatial gradient: Calculate the gradients in the horizontal and vertical directions for two-dimensional / three-dimensional resistivity data at the same time. Formula (taking the gradient in the x-direction, in two-dimensional discrete form as an example): in, For the set electrode spacing, the gradient extreme points correspond to the rock layer interface or water-bearing fracture boundary; (2) Spatial cluster analysis: Apply DBSCAN or K-means algorithms to cluster resistivity data according to spatial location and numerical value. Identify areas of low resistivity anomalies (potentially water-bearing areas); The formula for the spatial variation index can be used: d represents the distance afterward, and N(d) represents the number of corresponding point pairs; a gradual increase in γ(d) with d indicates homogeneous rock layers, while a sudden change indicates a water-bearing structure.

[0029] (3) Three-dimensional resistivity volume feature extraction: For the three-dimensional inversion results, extract multiple slices along the working surface direction / dip, and calculate the statistical characteristics of the resistivity of each slice (such as the area ratio of low resistivity regions). in, The volume occupied by the low-resistivity region. This represents the percentage of the area in the low-resistivity region. ; Set dangers according to the actual situation. , when requested Value greater than When this occurs, it indicates that the area under analysis is a high-risk area; (4) U-Net spatial feature extraction: Resistivity data is input into a convolutional neural network (CNN), and multi-scale spatial features are extracted by an encoder. in, The kernel size is [size]. To make the weights trainable, features from local to global are extracted through multiple layers of convolution and pooling; Step S1-2: Implement multi-source data fusion using feature-level fusion and spatiotemporal alignment mechanisms. Spatiotemporal alignment of data for different working areas and different survey lines; Step S1-2-1, Time Synchronization (NTP Protocol): Unified timestamps via GPS / NTP protocol: Hardware configuration: Each resistivity meter has a built-in NTP client (such as Chrony service), and a local NTP server is deployed (such as a GPS-synchronized NTP server with an accuracy of <1ms). Time synchronization verification: Calculate the clock offset: in: : The time when the client sends the request (local clock). : Server request reception time (server clock). Server response time (server clock). The client receives the response time (local clock), requiring all devices to do so. If the absolute value is less than 1ms, an alarm will be triggered. get Then, through the formula Obtain the calibrated standard time Step S1-2-2, Spatial Registration: The resistivity data from multiple measurement lines distributed in three-dimensional space are unified to a global coordinate system through coordinate transformation: the start / end point of each measurement line is recorded using GPS WGS84 coordinates, and a transformation from the local coordinate system of the measurement line to the global coordinate system is established. The coordinate transformation formula is as follows: Steps S1-3: Feature splicing: Step S1-3-1: Align the temporal and spatial characteristics of each sensor according to the time window to form a multidimensional input matrix: in, K represents the number of survey lines, K represents the feature dimension (time domain + space) of each survey line, T represents the length of the time window, and R represents the set of real numbers, emphasizing the numerical characteristics of the input data (continuous and computable). The following data are included in the two feature dimensions of each survey line: Time domain dimension: original resistivity ; indicates directly observed value; rate of change of resistivity ; indicates the rate of dynamic change; ; Indicates the overall trend; ; indicates local fluctuations; Spatial dimension: lateral gradient ; indicates a change in the direction of the survey line; vertical gradient Spatial variation index ; Indicates the spatial variation of resistivity in the coal mining face; the proportion of low-resistivity areas within the measurement area. ; indicates the degree of danger in the roof area; Step S1-3-2, Feature Normalization: Perform Z-Score standardization independently for each feature dimension: Where x represents the original data value; μ represents the center position of the mean dataset; and σ represents the standard deviation. This represents the standardized value.

[0030] Step S2: Preliminary establishment of the time series prediction model: A baseline model based on an LSTM-Transformer hybrid architecture is constructed, taking the time series matrix of resistivity data after alignment of multiple measurement lines as input, and outputting the probability of seepage; specifically, the following steps are included: Step S2-1: Generate a high-dimensional temporal matrix after multi-source feature fusion: Step S2-1-1, Shape: , in; K represents the number of survey lines, K represents the feature dimension (temporal domain + spatial domain) of each survey line, and T represents the length of the time window, such as 300 steps, which corresponds to a 5-minute window. Step S2-1-2, Normalization: Perform Z-Score standardization independently for each feature dimension: ;in, , Obtained based on the training set; Step S2-2: Generate the preliminary prediction model structure: LSTM regulates the flow of information through three gating structures: the forget gate, the input gate, and the output gate. Forgotten Gate: Decide how many old memories to discard. Input Gate: Determine how much new memory to store. Candidate memories: Generate the memory content to be written. Output gate: Control the amount of information currently output. in, , , : Current input (e.g., resistivity eigenvector at time t) Trainable parameter matrix Step S2-3: Generate the memory update equation: in, Current state (long-term memory) Hadamard product (element-by-element multiplication) Step S2-4, Output hidden state: Compress the memory content into the range [-1, 1].

[0031] Step S3, Dual-Branch Collaborative Training: Two heterogeneous model branches, namely a safety data expert and a seepage data expert, are trained in parallel. The Facol loss function addresses sample imbalance, and contrastive learning and dynamic weighted loss are used to solve the data imbalance problem caused by the limited number of seepage samples, thus improving the model's sensitivity to hazardous conditions. Specifically, this includes the following steps: Step S3-1: Set up the branch architecture: Step S3-1-1, Security Branch Main Input data: security sample Loss function: Cross-entropy; ; Step S3-1-2, seepage branch Aux Input data: seepage sample Loss function: Cross-entropy Optimization objective: Maximize the hazard detection rate; Step S3-2: Generate the joint loss function: Dynamic weights : Focal Loss parameters: Step S3-3, Contrast Loss (Feature Space Constraint): in, ; Step S3-4, Feature Fusion: Step S3-4-1, Cross-Attention: Allows the query in the safe branch to select key features from the seepage branch. ; ; Parameter matrix ; Step S3-4-2, Multi-head expansion: It is worth noting that, according to actual measurement data from coal mines, the best results are achieved when the number of heads H=8.

[0032] Step S4, Post-processing and Result Optimization: Integrating mathematical principles and engineering practice, the original prediction results are spatiotemporally smoothed and their consistency verified. The model is further optimized to ensure the reliability and interpretability of the early warning results in the spatiotemporal dimensions. Specifically, this includes the following steps: Step S4-1, Temporal Smoothing: Technical objective: Eliminate isolated noise in prediction; Technical method: Double exponential weighted average (DEWMA). Actual measurement ; Technical advantages: It responds to changes faster than simple moving averages while suppressing high-frequency noise; Step S4-2, Spatial Verification: Technical objective: To correct anomaly predictions using survey line topology relationships. Technical implementation steps: For each survey line For each node, a spatial location is represented, and the edge weights depend on the combined distance and geological similarity. edge weight The first term on the right side of the equals sign is the distance term, and the second term is the geological term. , , Then, spectral convolution (Chebyshev approximation) is used: Perform graph convolution correction; Finally, hard and soft constraints are applied to the model through geological rule injection. Hard constraints: , The representative survey line is located in a known impermeable layer area, so it is forcibly set to zero; Soft constraints: The mapping relationship between lithology and permeability probability is learned through MLP; Step S4-3, Adaptive Thresholding: Adjust the warning threshold based on real-time data distribution. in, ; Step S4-4, Warning Event Generation (Event Detection): Warning = .

[0033] Persistent detection: Avoiding transient false alarms Spatial aggregation: Adjacent A global early warning is only generated when the test line is triggered.

[0034] Step S5, Threshold Setting and Early Warning Mechanism: The output includes predicted values, risk scores, early warning signals, and visualizations. A multi-level design achieves a closed-loop process from data to decision-making. Adjustments are made based on real-time multi-source data, environmental and traffic characteristics to adapt to environmental changes and mine data features. A three-level early warning mechanism is set up to assist in decision-making and improve the efficiency and scientific rigor of addressing roof inrush water problems. Specifically, the following steps are included: Step S5-1, Risk Score Calculation: Step S5-1-1, Historical Baseline: ; Step S5-1-2, Real-time Risk Scoring: ; in, For real-time risk scoring, For the predicted water content at the current time step, The first in the past 30 days Predicted water content at each time step; Step S5-2: Set dynamic thresholds and early warning rules Step S5-2-1, Baseline Threshold: ;in, This represents the average value of the corresponding group of data. The variance of the corresponding group of data; Step S5-2-2, Environmental Correction Factor: When rainfall exceeds 50mm for 3 consecutive days: When the daily rise in groundwater level is greater than 10cm: ; Step S5-2-3, Multi-level Early Warning Rules: Level 1 monitoring level: Or the autoencoder malfunction persists for more than 2 hours; Level 2 Warning Level: Or Level 1 lasts for more than 12 hours; Level 3 Emergency Level: Or Level 2 lasts for more than 12 hours.

[0035] Step S6, Deployment and Monitoring: Design the edge layer, cloud layer, and visualization interface; Edge layer: This refers to the road segment equipment used to run the self-encoder, calculate micro-motion temporal characteristics, and cache 12 hours of raw data; Cloud layer: This refers to the central server, which handles storage and computation; Visualization interface: Displays risk scores, threshold lines, and historical comparison curves for each area of ​​the work surface; The combination of the visualization interface and thresholds can provide engineers with decision-making information and suggestions. Specifically, this includes the following steps: Step S6-1, Design the edge layer (mine equipment): The edge layer is responsible for real-time data preprocessing and lightweight anomaly detection.

[0036] Step S6-2: Design the cloud layer (central server): The cloud layer is responsible for the centralized training, storage, and large-scale computation of the model.

[0037] Storage: Historical data is stored in a time-series database (InfluxDB) and a spatial database (PostGIS); Computation: The cloud constitutes the decision center of the system. Based on the cross-attention dual-branch training method described in this manual, it completes the model construction and training on the central server, it iteratively optimizes the model by aggregating real-time data uploaded from the edge layer, and finally outputs the prediction results that drive the visualization interface, forming a complete closed loop from edge perception to cloud decision-making.

[0038] Step S6-3: Generate a visual interface: It displays risk scores, threshold lines, and historical comparison curves for each region, providing decision-making suggestions for engineers.

[0039] It is worth noting that this patent proposes the development of a visual monitoring and prediction system that reflects the monitoring and prediction data of the underground system to the ground dispatch center in real time. This invention transmits real-time monitoring data from the mine to cloud computing nodes using a transmission method suitable for the actual conditions of the work area. The data, inversion images, and prediction results obtained after cloud computing are transmitted to the ground dispatch room, where a multi-data interface is displayed through a front-end developed system, achieving a macroscopic response.

[0040] In the description of this invention, the term "a plurality of" refers to two or more. Unless otherwise explicitly defined, the terms "upper," "lower," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. The terms "connection," "installation," "fixing," etc., should be interpreted broadly. For example, "connection" can be a fixed connection, a detachable connection, or an integral connection; it can be a direct connection or an indirect connection through an intermediate medium. For those skilled in the art, the specific meaning of the above terms in this invention can be understood according to the specific circumstances.

[0041] In the description of this specification, the terms "one embodiment," "some embodiments," "specific embodiment," etc., refer to a specific feature, structure, material, or characteristic described in connection with that embodiment or example, which is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0042] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A geophysical method for predicting mine water distribution, characterized in that... Specifically, it includes the following steps: Step S1, Feature Engineering and Multi-Source Data Fusion: Spatiotemporal features are extracted from multi-source isomorphic data, i.e. DC resistivity data from different measurement lines in different measurement areas. Feature-level fusion and spatiotemporal alignment mechanisms are used to achieve multi-source data fusion to support model prediction. A phased feature extraction strategy is designed to extract the temporal and spatial features of DC resistivity sensor data. Step S2: Preliminary establishment of the time series prediction model: A baseline model based on the LSTM-Transformer hybrid architecture is constructed. The input is the time series matrix of resistivity data after multiple measurement lines are aligned, and the output is the probability of seepage. Step S3, Dual-branch Collaborative Training: Two heterogeneous model branches, namely, security data expert and seepage data expert, are trained in parallel. The sample imbalance is solved based on the Facol loss function, and contrastive learning and dynamic weighted loss are used. Step S4, Post-processing and Result Optimization: Spatiotemporal smoothing and consistency verification are performed on the original prediction results; Step S5, Threshold Setting and Early Warning Mechanism: The output includes predicted values, risk scores, early warning signals, and visualizations. A multi-level design achieves a closed loop from data to decision-making. It is adjusted based on real-time multi-source data, environmental and traffic characteristics to adapt to environmental changes and mine data features. An early warning mechanism with three levels is set. Step S6, Deployment and Monitoring: Design the edge layer, cloud layer, and visualization interface; Edge layer: namely, road segment equipment, used to run self-encoders, calculate micro-motion time-domain characteristics, and cache 12 hours of raw data; Cloud layer: namely, the central server, which meets storage and computing needs; Visualization interface: displays risk scores, threshold lines, and historical comparison curves for each area of ​​the work surface; The visualization interface is combined with the thresholds.

2. The method for predicting mine water distribution based on geophysics according to claim 1, characterized in that... The extraction of multi-source isomorphic data in step S1 specifically includes the following steps: Step S1-1: Extract features from DC resistivity sensor data: Step S1-1-1, Time Domain Features: (1) Rate of change of dynamic resistivity: Calculate the rate of change of resistivity for continuous time series data at the same measuring point. ； in, Let be the resistivity value at time t. ; ... (2) Sliding time window statistics: By setting a sliding window, the mean, standard deviation, and other statistical measures of resistivity within the window are calculated to capture short-term fluctuations and long-term trends. ； W represents the window width; Smooth noise and capture trends; if σ continues to increase, it indicates the development of rock fractures. (3) Time correlation analysis: Using cross-correlation functions or dynamic time warping algorithms, the resistivity at different measurement points is analyzed. Lagged correlations between time series can be used to identify water hazard diffusion pathways. The formula represents the lag correlation between measurement points A and B. If the maximum cross-correlation occurs... The number of days indicates that it takes 2 days for water to diffuse from point A to point B, which guides the warning time. (4) LSTM temporal feature extraction: Using an LSTM-Transformer neural network, features such as period and trend of resistivity sequence are automatically extracted, and the hidden state is output as a high-order time feature. The core equation is as follows Automatically learns the long-term dependencies of resistivity sequences and outputs the hidden state. As a higher-order feature; Step S1-1-2, Spatial Features: (1) Calculation of resistivity spatial gradient: Calculate the gradients in the horizontal and vertical directions for two-dimensional / three-dimensional resistivity data at the same time. formula: in, The electrode spacing is set; (2) Spatial cluster analysis: Apply DBSCAN or K-means algorithms to cluster resistivity data according to spatial location and numerical value. Identify low-resistivity anomaly regions; The formula for the spatial variation index can be used: d represents the distance afterward, and N(d) represents the number of corresponding point pairs; (3) Three-dimensional resistivity volume feature extraction: For the three-dimensional inversion results, multiple slices are extracted along the strike / dip of the working surface, and the statistical characteristics of the resistivity of each slice are calculated. in, The volume occupied by the low-resistivity region. This represents the percentage of the area in the low-resistivity region. ; (4) U-Net spatial feature extraction: Resistivity data is input into a convolutional neural network (CNN), and multi-scale spatial features are extracted by an encoder. in, The kernel size is [size]. To make the weights trainable, features from local to global are extracted through multiple layers of convolution and pooling; Step S1-2: Implement multi-source data fusion using feature-level fusion and spatiotemporal alignment mechanisms. Spatiotemporal alignment of data for different working areas and different survey lines; Step S1-2-1, Time Synchronization (NTP Protocol): Unified timestamps via GPS / NTP protocol: Hardware configuration: Each resistivity meter has a built-in NTP client, a local NTP server is deployed, and time synchronization verification is performed. Calculate the clock offset: in: : Client request sending time : Server request reception time Server response time : Client response time, requiring all devices If the absolute value is less than 1ms, an alarm will be triggered. get Then, through the formula Obtain the calibrated standard time Step S1-2-2, Spatial Registration: The resistivity data from multiple measurement lines distributed in three-dimensional space are unified to a global coordinate system through coordinate transformation: the start / end point of each measurement line is recorded using GPS WGS84 coordinates, and a transformation from the local coordinate system of the measurement line to the global coordinate system is established. The coordinate transformation formula is as follows: Steps S1-3: Feature splicing: Step S1-3-1: Align the temporal and spatial characteristics of each sensor according to the time window to form a multidimensional input matrix: in, Where K is the number of survey lines, K is the feature dimension of each survey line, T is the length of the time window, and R is the set of real numbers, emphasizing the numerical characteristics of the input data. The following data are included in the two feature dimensions of each survey line: Time domain dimension: original resistivity ; indicates directly observed value; rate of change of resistivity ; indicates the rate of dynamic change; ; Indicates the overall trend; ; indicates local fluctuations; Spatial dimension: lateral gradient ; indicates a change in the direction of the survey line; vertical gradient Spatial variation index ; Indicates the spatial variation of resistivity in the coal mining face; the proportion of low-resistivity areas within the measurement area. ; indicates the degree of danger in the roof area; Step S1-3-2, Feature Normalization: Perform Z-Score standardization independently for each feature dimension: Where x represents the original data value; μ represents the center position of the mean dataset; and σ represents the standard deviation. This represents the standardized value.

3. The method for predicting mine water distribution based on geophysics according to claim 1, characterized in that... Step S2 specifically includes the following steps: Step S2-1: Generate a high-dimensional temporal matrix after multi-source feature fusion: Step S2-1-1, Shape: , in; K represents the number of survey lines, K represents the feature dimension of each survey line, and T represents the length of the time window. Step S2-1-2, Normalization: Perform Z-Score standardization independently for each feature dimension: ;in, , Obtained based on the training set; Step S2-2: Generate the preliminary prediction model structure: LSTM regulates the flow of information through three gating structures: the forget gate, the input gate, and the output gate. Forgotten Gate: Decide how many old memories to discard. Input Gate: Determine how much new memory to store. Candidate memories: Generate the memory content to be written. Output gate: Control the amount of information currently output. in, , , : Current input (e.g., resistivity eigenvector at time t) Trainable parameter matrix Step S2-3: Generate the memory update equation: in, Current state :Hadamard product; Step S2-4, Output hidden state: Compress the memory content into the range [-1, 1].

4. The method for predicting mine water distribution based on geophysics according to claim 1, characterized in that... Step S3 specifically includes the following steps: Step S3-1: Set up the branch architecture: Step S3-1-1, Security Branch Main Input data: security sample Loss function: Cross-entropy; ; Step S3-1-2, seepage branch Aux Input data: seepage sample Loss function: Cross-entropy Optimization objective: Maximize the hazard detection rate; Step S3-2: Generate the joint loss function: Dynamic weights : Focal Loss parameters: Step S3-3, Compare Losses: in, ; Step S3-4, Feature Fusion: Step S3-4-1, Cross-attention: Allow the query on the safe branch to select key features from the seepage branch. Parameter matrix ; Step S3-4-2, Multi-head expansion: 。 5. The method for predicting mine water distribution based on geophysics according to claim 1, characterized in that... Step S4 specifically includes the following steps: Step S4-1, Timing Smoothing: Double exponential weighted average, Actual measurement ; Step S4-2, Spatial Consistency Check: For each survey line For each node, a spatial location is represented, and the edge weights depend on the combined distance and geological similarity. edge weight The first term on the right side of the equals sign is the distance term, and the second term is the geological term. , , Then, spectral convolution is used: Perform graph convolution correction; Finally, hard and soft constraints are applied to the model through geological rule injection. Hard constraints: , The representative survey line is located in a known impermeable layer area, so it is forcibly set to zero; Soft constraints: The mapping relationship between lithology and permeability probability is learned through MLP; Step S4-3, Dynamic Threshold Filtering: Adjust the warning threshold based on real-time data distribution. in, ; Step S4-4: Early Warning Event Generation: Warning = 。 6. The method for predicting mine water distribution based on geophysics according to claim 1, characterized in that... Step S5 specifically includes the following steps: Step S5-1, Risk Score Calculation: Step S5-1-1, Historical Baseline: ； Step S5-1-2, Real-time Risk Scoring: ; in, For real-time risk scoring, For the predicted water content at the current time step, The first in the past 30 days Predicted water content at each time step; Step S5-2: Set dynamic thresholds and early warning rules Step S5-2-1, Baseline Threshold: ;in, This represents the average value of the corresponding group of data. The variance of the corresponding group of data; Step S5-2-2, Environmental Correction Factor: When rainfall exceeds 50mm for 3 consecutive days: When the daily rise in groundwater level is greater than 10cm: ; Step S5-2-3, Multi-level Early Warning Rules: Level 1 monitoring level: Or the autoencoder malfunction persists for more than 2 hours; Level 2 Warning Level: Or Level 1 lasts for more than 12 hours; Level 3 Emergency Level: Or Level 2 lasts for more than 12 hours.

7. The method for predicting mine water distribution based on geophysics according to claim 1, characterized in that... Step S6 specifically includes the following steps: Step S6-1, Design the edge layer: The edge layer is responsible for real-time data preprocessing and lightweight anomaly detection; Step S6-2, Design the cloud layer: The cloud layer is responsible for the centralized training, storage, and large-scale computation of the model; Storage: Historical data is stored in time-series databases and spatial databases; Computation: The cloud constitutes the decision center of the system. Based on the cross-attention dual-branch training method described in this manual, it completes the model construction and training on the central server, it iteratively optimizes the model by aggregating real-time data uploaded from the edge layer, and finally outputs the prediction results that drive the visualization interface, forming a complete closed loop from edge perception to cloud decision-making. Step S6-3: Generate a visual interface: Displays risk scores, threshold lines, and historical comparison curves for each region.

Images