A mine water disaster early warning method and system based on big data analysis

Abstract

The application discloses a mine water disaster early warning method and system based on big data analysis, relates to the technical field of big data analysis, and comprises the following steps: collecting mine water disaster multi-source associated data, generating a mine water disaster associated state sequence; performing a LiNGAM causal discovery algorithm, constructing a double-root cause causal skeleton; constructing an improved N-HiTS model, forming a water state prediction trajectory; calculating water disaster causal contribution and operational causal contribution according to connection strength transmission level activation amount; performing competitive matching, forming a causal contribution closed path; screening the causal contribution closed path, forming an active water disaster disaster-causing chain; deducing the water disaster causal contribution arrival time, determining the water disaster state and outputting the early warning result. Through the introduction of the LiNGAM causal discovery algorithm and the improved N-HiTS model, the application realizes mine water disaster root cause identification, disaster-causing channel tracking, water gushing trend prediction and precise early warning driven by abnormal transmission time lag.

Description

Technical Field

[0001] This invention relates to the field of big data analytics, and in particular to a method and system for early warning of mine water hazards based on big data analytics. Background Technology

[0002] Mine water hazards are a common safety risk in coal mining, mainly related to factors such as confined aquifer water, water accumulation in old workings, fault fissures, water-conducting channels, and mining disturbances. Before a water hazard occurs, changes in monitoring parameters such as water pressure, water level, water inflow, micro-vibrations, and resistivity are usually observed. Traditional mine water hazard monitoring systems typically deploy pressure gauges, water level gauges, flow meters, and geophysical monitoring equipment in aquifers, boreholes, working faces, and water tanks to continuously collect hydrogeological and production operation data. Anomalies are then identified using threshold judgment, trend analysis, and statistical analysis. Early warnings are issued when monitored parameters exceed set ranges or when continuous changes reach alarm conditions, assisting in mine drainage scheduling and personnel safety management.

[0003] As the types and frequency of mine monitoring data increase, existing technologies are beginning to utilize big data analytics, machine learning, and time series forecasting methods to perform correlation analysis on data such as aquifer water pressure, borehole water pressure, working face water inflow, water tank level, drainage volume, and mining operations, and to predict water inflow trends based on historical variation patterns. However, the time scales of change for different monitoring parameters vary, and production operations such as mining progress, pump start-up and shutdown, and blasting can cause significant fluctuations in conditions. Existing methods tend to mix production disturbances with water hazard anomalies in their analysis, and they struggle to accurately represent the transmission process between changes in water sources, channel changes, and the working face's water response, resulting in insufficient accuracy in determining the source of water hazards, locating anomalies, and determining early warning times.

[0004] Therefore, how to provide a mine water hazard early warning method and system based on big data analysis is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a mine water hazard early warning method and system based on big data analysis. This invention fully utilizes multi-source monitoring data from mines, the LiNGAM causal discovery algorithm, and the improved N-HiTS model to describe in detail the causal analysis and time-series prediction process between water source anomalies, production operations, channel status, and water reception status. By constructing a dual-root-cause causal framework, it distinguishes water hazard factors from production operation disturbances. It uses causal path erasure counterfactual, minimum sufficient path inversion, path time delay interpolation, and conjugate back-feed closed-loop processing to extract water source, channel, and water reception activation trajectories. Combined with causal contribution competitive matching, it identifies active water hazard disaster chains and deduces the anomaly transmission time delay and the arrival time of water hazard causal contributions. This achieves mine water hazard source identification, abnormal channel location, water inrush trend prediction, and early warning, and has the advantages of strong resistance to production disturbances, accurate disaster chain identification, and high early warning timeliness.

[0006] A mine water hazard early warning method based on big data analysis according to an embodiment of the present invention includes: Collect multi-source correlation data of mine water hazards, and perform preprocessing on the multi-source correlation data of mine water hazards to generate a mine water hazard correlation state sequence; Based on the state change features extracted from the mine water hazard associated state sequence, the LiNGAM causal discovery algorithm is executed to retain the causal edges from the water source root cause and the production operation root cause to the water receiving state through the channel state, and to construct a dual-root cause causal skeleton. An improved N-HiTS model is constructed, and multi-rate sampling, hierarchical residual decomposition and hierarchical interpolation are performed on the variable sequence in the dual-root cause causal skeleton to extract the water source activation trajectory, channel activation trajectory, water receiving activation trajectory and hierarchical activation amount, and the water receiving activation trajectory is extrapolated to form the water receiving state prediction trajectory. Along the causal direction of the dual-root cause causal framework, the causal contribution of water damage and the operational causal contribution are calculated according to the activation amount of the connection strength transmission hierarchy. The causal contribution of water damage and the causal contribution of operation are competitively matched with the hierarchical activation amount to form a closed causal contribution path; Based on the temporal sequence and hydrogeological connection of the water source activation trajectory, channel activation trajectory and water-receiving activation trajectory, causal contribution closed paths are screened to form active water hazard disaster chains and determine the anomaly transmission time lag. Based on the time delay of abnormal transmission along the active water hazard disaster chain, the arrival time of the causal contribution of water hazard is deduced, the predicted inflow volume in the water-receiving state prediction trajectory is read, the water hazard state is determined, and the early warning result is output.

[0007] Optionally, the multi-source correlation data of mine water hazards specifically includes aquifer water pressure data, water level data of old workings water accumulation area, borehole water pressure data, microseismic energy data, apparent resistivity data, working face water inflow data, water tank water level data, water pump drainage data, working face advance data, water pump start and stop data, blasting operation data, hydrogeological connection data, and data acquisition timestamps.

[0008] Optionally, the preprocessing of the multi-source associated data of mine water hazards specifically includes time alignment, missing data completion, anomaly removal, spatial correlation, and normalization.

[0009] Optionally, constructing the dual-root cause causal framework includes: Based on the data collection timestamp, the mine water hazard associated state sequence is extracted, and the state difference, rate of change and cumulative change at adjacent collection times are calculated to obtain the state change characteristics. The aquifer water pressure and the water level in the old workings water accumulation area are classified as water source root variables, the working face advancement, water pump start-up and shutdown and blasting operations are classified as production operation root variables, the borehole water pressure, micro-vibration energy and apparent resistivity are classified as channel state variables, and the working face water inflow, water tank water level and water pump drainage are classified as water receiving state variables. The causal analysis sample is composed according to the data acquisition timestamp. The LiNGAM causal discovery algorithm is executed on the causal analysis samples to determine the causal order of variables and calculate the connection strength between variables, thereby obtaining candidate causal edges. Perform self-service resampling on candidate causal edges, and retain causal edges that have an occurrence rate of 80% and an absolute value of connection strength of 0.20, pointing from water source root cause or production operation root cause to channel state and from channel state to water receiving state. By connecting the retained causal edges according to the causal direction, associating hydrogeological data with data acquisition timestamps, a dual-root cause causal framework is constructed with water source root cause and production operation root cause as the starting node, channel status as the intermediate node, and water reception status as the ending node.

[0010] Optionally, the extrapolated water-activated trajectory forming the water-receiving state prediction trajectory includes: An improved N-HiTS model is constructed, which includes a causal path erasure counterfactual layer, a minimum sufficient path inversion layer, a path delay interpolation layer, and a conjugate back-injection closed-loop layer. The causal path erasure counterfactual layer extracts candidate causal paths and corresponding variable sequences from the mine water hazard associated state sequence, forming a path retention sequence and a path erasure sequence that sets the current path connection strength to 0; Hierarchical sampling, residual fitting, and timestamp back-interpolation were performed on the path-preserved sequence, path-erased sequence, and water-receiving state variable sequence at 1, 2, and 4 sampling intervals to obtain the path-preserved water-receiving layer component, the path-erased water-receiving layer component, and the actual water-receiving layer component. The path-specific activation amount was obtained by subtracting the path-preserved water-receiving layer component and the path-erased water-receiving layer component. The actual water-receiving layer components are reconstructed using the minimum sufficient path inversion layer superimposed with path-specific activation. The path combinations with the smallest reconstruction deviation ratio (not exceeding 20%) and the fewest number of paths are selected. The path combinations with a reconstruction deviation ratio exceeding 20% ​​after deleting any path are determined as the minimum sufficient disaster-causing path set. The path delay interpolation layer calculates the activation start time difference between adjacent nodes in the minimum sufficient disaster-causing path set to obtain the path propagation delay. It then allocates path-specific activation amounts according to the connection strength and translates and interpolates along the causal direction to form the initial water source activation trajectory, initial channel activation trajectory, initial water receiving activation trajectory, and initial hierarchical activation amount. Finally, it extends and superimposes the initial water receiving activation trajectory to form the initial water receiving state prediction trajectory. The conjugate back-feeding closed-loop layer back-feeds the initial water-receiving state prediction trajectory to the channel state and water source root cause in the opposite direction of the path transmission time delay, forming the channel back-feeding trajectory and the water source back-feeding trajectory. Calculate the closed-loop deviation between the channel return trajectory and the water source return trajectory and the corresponding initial activation trajectory. Delete paths whose closed-loop deviation exceeds 20% of the cumulative absolute value of the corresponding initial activation trajectory. Re-execute the minimum sufficient path inversion and path time delay interpolation until the minimum sufficient disaster-causing path set remains unchanged. Determine the final iteration results as the water source activation trajectory, channel activation trajectory, water receiving activation trajectory, hierarchical activation amount, and water receiving state prediction trajectory. The improved N-HiTS model was trained by weighting the prediction error of water state, the reconstruction error of hierarchical components, and the deviation of conjugate back-projection. The parameters of the causal path erasure counterfactual layer, the minimum sufficient path inversion layer, the path delay interpolation layer, and the conjugate back-projection closed loop layer were optimized. The improved N-HiTS model was determined to have converged when the change in loss was less than 0.001 and the change rate of closed loop deviation was less than 1% after 10 consecutive rounds of verification.

[0011] Optionally, the calculation of the causal contribution of water damage and the causal contribution of operation includes: Read the causal edges, connection strength and hierarchical activation amount in the dual-root cause causal skeleton, and associate the aquifer water pressure and the water level in the old working water accumulation area as water source root causes, and associate the working face advancement, water pump start and stop and blasting operation as production operation root causes, and associate the channel status and water receiving status. The connection strength pointing to the same node is normalized, and the hierarchical activation amount of the upstream node is allocated according to the normalized connection strength, and then passed to the channel state and the water-receiving state in sequence along the causal direction. The hierarchical activation amounts of water source root causes and production operation root causes transmitted along each causal edge to the channel state and water-receiving state are summarized to obtain the causal contribution of water damage and the causal contribution of operation.

[0012] Optionally, forming a causal contribution closed path includes: Based on the data collection timestamp and hierarchical alignment of the causal contribution of water hazards, the operational causal contribution, and the hierarchical activation amount, the proportions of the causal contribution of water hazards and the operational causal contribution are calculated to obtain the water hazard competition weight and the operational competition weight. The activation amount of each level is allocated according to the competition weight of water hazards and the competition weight of operations. The activation amount of water hazards and the activation amount of operations are obtained. The deviation ratio between the activation amount of water hazards and the causal contribution of water hazards and the deviation ratio between the activation amount of operations and the causal contribution of operations are calculated. The causal contribution of water hazards and the causal contribution of operations with a deviation ratio of no more than 20% are retained. Connect the root cause nodes, channel state nodes, and water-receiving state nodes corresponding to the retained water hazard causal contribution or operational causal contribution along the dual root cause causal framework, associate the water hazard matching activation amount or operational matching activation amount, and form a closed path of causal contribution.

[0013] Optionally, the process of forming an active flood-causing chain and determining the abnormal transmission time delay includes: Based on the data collection timestamp, the causal contribution closed path is aligned with the water source activation trajectory, channel activation trajectory and water receiving activation trajectory. The start time of water source activation, the start time of channel activation and the start time of water receiving activation are extracted. The causal contribution closed path with the activation start time increasing sequentially and satisfying the hydrogeological connection relationship is retained. From the preserved closed causal contribution paths, root cause nodes, channel state nodes, and water-receiving state nodes that contain the causal contribution of water hazards are selected, and connected sequentially according to hydrogeological connection relationships and causal directions to form an active water hazard disaster chain. The time difference between the activation start time of the water source and the activation start time of the channel, and between the activation start time of the channel and the activation start time of the water receiving are calculated along the active flood disaster chain, and the time differences are accumulated to obtain the abnormal transmission time delay of the active flood disaster chain.

[0014] Optionally, determining the state of flood damage and outputting early warning results includes: Read the activation start time of the water source in the active water disaster chain, decompose the abnormal transmission time delay into the water source to channel time delay and the channel to water receiving time delay, and accumulate them segment by segment to obtain the arrival time of the water disaster causal contribution to the channel state and the water receiving state. According to the arrival time of the causal contribution of water damage to the water-receiving state, the predicted water inflow is read from the predicted trajectory of the water-receiving state, and the water inflow at the working face, the water pump discharge and the water level in the water tank are read according to the current data collection timestamp, and the water level change in the water tank at adjacent collection times is calculated. The predicted water inflow is compared with the water inflow at the working face and the drainage volume of the water pump. The water hazard status is determined by combining the water level change in the water tank. The results of the active water hazard disaster chain, the arrival time of the causal contribution of the water hazard, the predicted water inflow, and the early warning results corresponding to the water hazard status are output.

[0015] A mine water hazard early warning system based on big data analysis according to an embodiment of the present invention includes the following modules: The data acquisition module is used to collect and preprocess multi-source correlation data of mine water hazards and generate a mine water hazard correlation state sequence. The causal skeleton construction module is used to extract state change features and execute the LiNGAM causal discovery algorithm to construct a dual-root cause causal skeleton. The water hazard activation prediction module is used to extract the water source activation trajectory, channel activation trajectory, water receiving activation trajectory and hierarchical activation amount by improving the N-HiTS model, and form the water receiving state prediction trajectory. The causal contribution calculation module is used to pass hierarchical activation along the double root cause causal framework to obtain the causal contribution of water damage and the causal contribution of operation. The contribution competition matching module is used to compete and match the causal contribution of water damage, the causal contribution of operation, and the hierarchical activation amount to form a closed causal contribution path; The disaster chain identification module is used to combine the time sequence of activation trajectories and hydrogeological connections to screen closed paths of causal contribution, form active water hazard disaster chains and determine the time lag of anomaly transmission. The early warning output module is used to deduce the arrival time of the causal contribution of water damage, determine the state of water damage, and output early warning results.

[0016] The beneficial effects of this invention are: This invention proposes a mine water hazard early warning method and system based on big data analysis. By performing unified correlation analysis on aquifer water pressure, water level in old workings water accumulation area, borehole water pressure, microseismic energy, apparent resistivity, working face water inflow, water tank water level, water pump drainage, and production operation data, and using the LiNGAM causal discovery algorithm to construct a dual root cause causal framework, the root cause of water source is distinguished from the root cause of production operation. This can reduce the impact of normal production disturbances such as working face advancement, water pump start-up and shutdown, and blasting operations on the judgment of water hazard anomalies, and improve the accuracy of water hazard root cause identification.

[0017] This invention improves the N-HiTS model by performing path erasure counterfactual, minimum sufficient path inversion, path delay interpolation, and conjugate back-loop processing on candidate causal paths. It extracts water source activation trajectories, channel activation trajectories, and water-receiving activation trajectories, and forms a predicted water-receiving state trajectory. This allows water source changes, channel responses, and water-receiving responses at different time scales to be correlated according to actual transmission delays. At the same time, it filters out causal contributions that are inconsistent with actual hierarchical activation changes through causal contribution competition matching. Combined with hydrogeological connectivity, it forms an active water hazard disaster chain, thereby determining the abnormal transmission delay and the arrival time of water hazard causal contributions. This enables collaborative judgment of water hazard sources, abnormal transmission channels, water inrush trends, and early warning times, improving the accuracy, anti-disturbance capability, and timeliness of mine water hazard early warning. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a mine water hazard early warning method based on big data analysis proposed in this invention; Figure 2 This is a schematic diagram of the structure of the improved N-HiTS model for a mine water hazard early warning method based on big data analysis proposed in this invention. Figure 3 This is a schematic diagram of the structure of a mine water hazard early warning system based on big data analysis proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figure 1 and Figure 2 A mine water hazard early warning method based on big data analysis includes: Collect multi-source correlation data of mine water hazards, and perform preprocessing on the multi-source correlation data of mine water hazards to generate a mine water hazard correlation state sequence; Based on the state change features extracted from the mine water hazard associated state sequence, the LiNGAM causal discovery algorithm is executed to retain the causal edges from the water source root cause and the production operation root cause to the water receiving state through the channel state, and to construct a dual-root cause causal skeleton. An improved N-HiTS model is constructed, and multi-rate sampling, hierarchical residual decomposition and hierarchical interpolation are performed on the variable sequence in the dual-root cause causal skeleton to extract the water source activation trajectory, channel activation trajectory, water receiving activation trajectory and hierarchical activation amount, and the water receiving activation trajectory is extrapolated to form the water receiving state prediction trajectory. Along the causal direction of the dual-root cause causal framework, the causal contribution of water damage and the operational causal contribution are calculated according to the activation amount of the connection strength transmission hierarchy. The causal contribution of water damage and the causal contribution of operation are competitively matched with the hierarchical activation amount to form a closed causal contribution path; Based on the temporal sequence and hydrogeological connection of the water source activation trajectory, channel activation trajectory and water-receiving activation trajectory, causal contribution closed paths are screened to form active water hazard disaster chains and determine the anomaly transmission time lag. Based on the time delay of abnormal transmission along the active water hazard disaster chain, the arrival time of the causal contribution of water hazard is deduced, the predicted inflow volume in the water-receiving state prediction trajectory is read, the water hazard state is determined, and the early warning result is output.

[0021] In this embodiment, the multi-source correlation data of mine water hazards specifically includes aquifer water pressure data, old working water accumulation area water level data, borehole water pressure data, microseismic energy data, apparent resistivity data, working face water inflow data, water tank water level data, water pump drainage data, working face advance data, water pump start and stop data, blasting operation data, hydrogeological connection data, and data acquisition timestamps.

[0022] In this embodiment, the preprocessing of multi-source associated data of mine water hazards specifically includes time alignment, missing data completion, anomaly removal, spatial correlation, and normalization.

[0023] In this embodiment, constructing a dual-root-cause causal framework includes: Based on the data acquisition timestamps, the mine water hazard-related state sequence is extracted, and the state difference, rate of change, and cumulative change between adjacent acquisition times are calculated to obtain the state change characteristics, including: The state difference, rate of change, and cumulative change at adjacent acquisition times are calculated as follows: According to the data acquisition timestamps, the normalized values ​​corresponding to the aquifer water pressure, old working water level, borehole water pressure, microseismic energy, apparent resistivity, working face water inflow, water tank level, water pump drainage, working face advance, water pump start-up and shutdown, and blasting operation in the mine water hazard associated state sequence are read sequentially from early to late. For each variable, the normalized value at the current acquisition time is subtracted from the normalized value at the previous acquisition time to obtain the state difference. The state difference is divided by the time difference between the current data acquisition timestamp and the previous data acquisition timestamp to obtain the rate of change. Starting from the initial acquisition time of the extracted mine water hazard associated state sequence, the state difference values ​​of each adjacent acquisition time are accumulated to obtain the cumulative change at the current acquisition time. The state difference, rate of change, and cumulative change of each variable at the same data acquisition timestamp are jointly determined as the state change characteristics. Aquifer water pressure and water level in the old workings water accumulation area are classified as water source root variables; working face advancement, pump start-up and shutdown, and blasting operations are classified as production operation root variables; borehole water pressure, micro-vibration energy, and apparent resistivity are classified as channel state variables; and working face water inflow, water tank level, and pump discharge are classified as water receiving state variables. A causal analysis sample is then formed based on the data acquisition timestamps, where: The causal analysis samples were composed based on the data collection timestamps, specifically: According to the same data acquisition timestamp, the state change characteristics corresponding to aquifer water pressure, old goaf water level, working face advancement, pump start-up and shutdown, blasting operation, borehole water pressure, micro-vibration energy, apparent resistivity, working face water inflow, water tank level, and pump drainage are read separately. The state change characteristics corresponding to the same data acquisition timestamp are recorded in a fixed order of aquifer water pressure, old goaf water level, working face advancement, pump start-up and shutdown, blasting operation, borehole water pressure, micro-vibration energy, apparent resistivity, working face water inflow, water tank level, and pump drainage. The records corresponding to each acquisition time are then arranged from earliest to latest according to the data acquisition timestamp to form a causal analysis sample. The LiNGAM causal discovery algorithm is applied to the causal analysis samples to determine the causal order of variables and calculate the connection strength between variables, thereby obtaining candidate causal edges, where: Determining the causal order of variables and calculating the connection strength between variables involves the following steps: One variable is selected from the causal analysis sample in sequence. The selected variable is then matched with each variable in the causal analysis sample other than the selected variable according to the data collection timestamp. The average value of the state change characteristics of the selected variable is subtracted from each state change characteristic of the selected variable. The average value of the corresponding state change characteristics of the variables compared with the selected variable is subtracted from each state change characteristic of the variables compared with the selected variable. The two differences with the same data collection timestamp are multiplied and summed, and then divided by the sum of the squares of the differences of the selected variable to obtain the fitting coefficient. The difference between each state change characteristic of the selected variable and the average value of the state change characteristics of the selected variable is multiplied by the fitting coefficient, and then added to the average value of the state change characteristics of the variables compared to obtain the fitting value for the corresponding data collection timestamp. The fitting value is subtracted from the state change characteristics of the variables compared to obtain the fitting residual. The square of the difference between each state change characteristic of the selected variable and the average value is multiplied by the corresponding fitting residual and summed. At the same time, the square of each fitting residual is multiplied by the difference between the corresponding state change characteristic and the average value of the selected variable and summed. The absolute values ​​of the two summed results are added to obtain the dependence value between the selected variable and the fitting residual. The variables are selected by summing all their dependencies and choosing the variable with the smallest sum as the first variable in the causal order. Variables whose order has been determined are excluded from the causal analysis sample. The remaining variables are then used to determine the next variable using the same process of calculating the fitting coefficient, fitting residual, and dependency value, until all variables are arranged to obtain the causal order. The current variable and the input variables arranged before it are read sequentially according to the causal order. A least squares equation is established for each input variable. The left side of the equation is the sum of the products of the current input variable and the state change characteristics of each input variable, multiplied by the corresponding regression coefficient. The right side of the equation is the sum of the products of the current input variable and the state change characteristics of the current variable. All least squares equations are solved simultaneously to obtain the regression coefficients. The regression coefficients are used as the connection strength from the corresponding input variable to the current variable, forming candidate causal edges. Perform self-sampling on candidate causal edges, retaining causal edges where the occurrence rate reaches 80% and the absolute value of the connection strength reaches 0.20, pointing from the water source root cause or production operation root cause to the channel state, and from the channel state to the water receiving state. Perform bootstrap resampling on candidate causal edges, specifically as follows: Sample records with replacement are drawn from the causal analysis sample, with the same number of records as the causal analysis sample. Each draw forms a resampled causal analysis sample, resulting in a total of 1000 resampled causal analysis samples. For each resampled causal analysis sample, candidate causal edges and their corresponding connection strengths are obtained through the processes of calculating fitting coefficients, calculating fitting residuals, comparing dependency values, determining the causal order of variables, and calculating connection strength. The number of times the same candidate causal edge appears in the 1000 resampled causal analysis samples is counted, and the occurrence ratio is obtained by dividing the occurrence ratio by 1000. The connection strengths corresponding to the same candidate causal edge each time it appears are added together and divided by the number of occurrences of the same candidate causal edge to obtain the average connection strength. Causal edges that have an occurrence ratio of 80% and an absolute value of average connection strength of 0.20, pointing from water source root cause or production operation root cause to channel state and from channel state to water receiving state, are retained. By connecting the retained causal edges according to the causal direction, and associating hydrogeological data with data acquisition timestamps, a dual-root-cause causal framework is constructed, with water source root cause and production operation root cause as the starting node, channel status as the intermediate node, and water reception status as the ending node. The causal direction refers to the direction of variable transmission from the water source root cause or the production operation root cause to the channel state and from the channel state to the water receiving state. The water source root cause corresponds to the aquifer water pressure and the water level in the old working water accumulation area. The production operation root cause corresponds to the working face advancement, water pump start-up and shutdown and blasting operations. The channel state corresponds to the borehole water pressure, micro-vibration energy and apparent resistivity. The water receiving state corresponds to the working face water inflow, water tank water level and water pump drainage. A dual-root-cause causal framework is constructed, with water source root cause and production operation root cause as the starting node, channel state as the intermediate node, and water reception state as the ending node, specifically as follows: Water source root cause nodes are established based on aquifer water pressure and old working water level; production operation root cause nodes are established based on working face advancement, pump start / stop, and blasting operations; channel status nodes are established based on borehole water pressure, microseismic energy, and apparent resistivity; and water receiving status nodes are established based on working face water inflow, water tank level, and pump discharge. The starting variable, ending variable, and average connection strength corresponding to the retained causal edges are read and preserved. The water source root cause nodes or production operation root cause nodes are connected to the channel status nodes, and the channel status nodes are connected to the water receiving status nodes according to the direction of the retained causal edges. Then, the hydrogeological connection relationships between the corresponding water source location, borehole location, working face location, and water tank location are read from the hydrogeological connection data. Causal edges that do not have hydrogeological connection relationships between the starting and ending locations are deleted. The data acquisition start time stamp and data acquisition end time stamp corresponding to the retained causal edges are associated with the corresponding causal edges to obtain a dual-root cause causal skeleton with water source root cause and production operation root cause as the starting node, channel status as the intermediate node, and water receiving status as the ending node.

[0024] In this embodiment, the extrapolation of the water-receiving activation trajectory to form a water-receiving state prediction trajectory includes: An improved N-HiTS model is constructed, comprising a causal path erasure counterfactual layer, a minimum sufficient path inversion layer, a path delay interpolation layer, and a conjugate back-projection closed-loop layer, wherein: The improved N-HiTS model is constructed as follows: By adding candidate causal path retention and candidate causal path erasure processing to the multi-rate sampling input of the traditional N-HiTS model, a causal path erasure counterfactual layer is obtained, so that the same candidate causal path can enter the multi-rate sampling and hierarchical residual decomposition process in the retained connection state and the erased connection state, respectively. By adding path-specific activation comparison and minimum path combination screening to the hierarchical residual output of the traditional N-HiTS model, a minimum sufficient path inversion layer is obtained, so that the hierarchical residual decomposition results are inverted and screened according to candidate causal paths to form a minimum sufficient disaster-causing path set. By adding path propagation delay and connection strength constraints to the hierarchical interpolation module of the traditional N-HiTS model, a path delay interpolation layer is obtained, which adjusts the interpolation time position according to the causal direction and path propagation delay in the minimum sufficient disaster-causing path set. By adding reverse projection of the water-receiving state prediction trajectory and path screening feedback processing to the prediction output of the traditional N-HiTS model, a conjugate projection closed-loop layer is obtained, which allows the water-receiving state prediction trajectory to be projected back to the channel state and water source root cause in the opposite direction of the path transmission time delay, and the projection result is fed back to the minimum sufficient path inversion layer. By combining the causal path erasure counterfactual layer, the minimum sufficient path inversion layer, the path delay interpolation layer, and the conjugate back-injection closed-loop layer, an improved N-HiTS model is obtained. This model retains the multi-rate sampling, hierarchical residual decomposition, and hierarchical interpolation structure of the traditional N-HiTS model, and integrates candidate causal path erasure, minimum sufficient path screening, path propagation delay constraints, and back-injection verification processing. The causal path erasure counterfactual layer extracts candidate causal paths and corresponding variable sequences from the mine water hazard associated state sequence, forming a path-preserving sequence and a path-erasing sequence that sets the current path connection strength to 0, wherein: Causal path erasure counterfactual layer includes: Path reader: Reads candidate causal paths, corresponding variable sequences, and connection strengths; Path copier: Copies the sequence of variables and connection strength corresponding to candidate causal paths; Connection strength eraser: retains one set of connection strengths to form a path retention sequence, and sets another set of current path connection strengths to 0 to form a path erasure sequence; Residual Differentiator: Performs multi-rate residual processing on the path-preserved sequence and the path-erased sequence, compares the corresponding water-receiving layer components, and outputs the path-specific activation amount; The extraction of candidate causal paths and corresponding variable sequences is as follows: Read the continuous causal edges from the dual-root cause causal skeleton, starting from the water source root cause or production operation root cause, passing through the channel state and reaching the water receiving state. Determine each continuous connection as a candidate causal path. According to the water source root cause or production operation root cause, channel state and water receiving state in the candidate causal path, read the variable values ​​corresponding to the same data collection timestamp from the mine water hazard associated state sequence, and arrange them from the earliest to the latest according to the data collection timestamp to form a corresponding variable sequence. Associate the connection strength of the corresponding variable sequence with the causal edge in the candidate causal path one by one and keep the connection strength unchanged to form a path preservation sequence. Copy the path preservation sequence, set the connection strength of the water source root cause or production operation root cause pointing to the channel state and the channel state pointing to the water receiving state in the current candidate causal path to 0, while keeping the variable values ​​and data collection timestamp unchanged to form a path erasure sequence. Hierarchical sampling, residual fitting, and timestamp back interpolation were performed on the path-preserved sequence, path-erased sequence, and water-receiving state variable sequence at sampling intervals of 1, 2, and 4 to obtain the path-preserved water-receiving layer components, the path-erased water-receiving layer components, and the actual water-receiving layer components. The path-specific activation value was obtained by subtracting the path-preserved water-receiving layer components from the path-erased water-receiving layer components. The execution of hierarchical sampling, residual fitting, and timestamp back interpolation is as follows: First, the path retention sequence retains all collected values ​​at one collection interval and forms the first-level input sequence according to the data collection timestamp. The first-level input sequence is fed into the first-level residual block of N-HiTS. When the model is initially established, the connection weights are assigned according to a normal distribution with a mean of 0 and a standard deviation of 0.01, and the bias is assigned to 0. During training iteration, the connection weights and bias are updated. The values ​​in the first-level input sequence are multiplied by the corresponding connection weights, summed, and the bias is added to obtain the backtracking coefficients. The backtracking coefficients are placed at equal intervals from the start data collection timestamp to the end data collection timestamp of the first-level input sequence. For the data collection timestamp between two adjacent backtracking coefficients, the time ratio is obtained by dividing the time difference between the current data collection timestamp and the previous backtracking coefficient timestamp by the time difference between the two adjacent backtracking coefficient timestamps. The difference between the values ​​of the two adjacent backtracking coefficients is multiplied by the time ratio and added to the value of the previous backtracking coefficient to obtain the backtracking fitting value corresponding to the current data collection timestamp. Subtract the backfit value corresponding to the same data acquisition time stamp from the sampled value corresponding to each data acquisition time stamp in the first-level input sequence to obtain the residual value corresponding to each data acquisition time stamp. Arrange the residual values ​​according to the data acquisition time stamp to form the first-level residual sequence. From the first-level residual sequence, retain data once every 1 acquisition time at 2 acquisition intervals and obtain the second-level backfit value and the second-level residual sequence in the same way. Then, from the second-level residual sequence, retain data once every 3 acquisition time at 4 acquisition intervals and obtain the third-level backfit value in the same way. The backfit values ​​corresponding to the water-receiving state variables in the three levels are restored to the original data acquisition timestamps and arranged according to the water-receiving state variables, sampling levels, and data acquisition timestamps to obtain the path-preserved water-receiving level components. The same processing is then performed on the path-erasing sequence to obtain the path-erased water-receiving level components. The same processing is then performed on the water-receiving state variable sequence to obtain the actual water-receiving level components. According to the same water-receiving state variables, the same sampling levels, and the same data acquisition timestamps, the path-preserved water-receiving level components are subtracted from the path-erased water-receiving level components to obtain the path-specific activation amount. The minimum sufficient path inversion layer superimposed path-specific activation amount reconstructs the actual water-receiving layer components. Path combinations with the smallest reconstruction deviation ratio (not exceeding 20%) and the fewest number of paths are selected. Path combinations whose reconstruction deviation ratio exceeds 20% after deleting any path are determined as the minimum sufficient disaster-causing path set, where: The minimum sufficient path inversion layer includes: Path activation overlay: Overlays path-specific activation values ​​based on water-receiving state variables, sampling levels, and data acquisition timestamps; Reconstruction Deviation Calculator: Compares the superposition results with the actual water-receiving layer components and calculates the reconstruction deviation ratio; Path combination filter: Filters path combinations with a reconstruction deviation ratio of no more than 20% and the fewest number of paths; Path deletion validator: Deletes candidate causal paths one by one and recalculates the reconstruction deviation ratio to determine the minimum set of sufficiently catastrophic paths; The minimum sufficient path inversion layer superposition path-specific activation amount reconstructs the actual water-receiving layer components, specifically: The path-specific activations selected into the same path combination are aligned according to the water-receiving state variable, sampling level, and data acquisition timestamp. The path-specific activations with the same water-receiving state variable, sampling level, and data acquisition timestamp are added one by one. The results of the addition are arranged according to the same water-receiving state variable, the same sampling level, and the same data acquisition timestamp, so that the working face water inflow corresponds to the working face water inflow, the water tank level corresponds to the water tank level, and the water pump discharge corresponds to the water pump discharge, thus obtaining the reconstructed actual water-receiving level component corresponding to the path combination. The path combinations with the smallest reconstruction deviation ratio (not exceeding 20%) and the fewest number of paths are selected as follows: The reconstructed actual water-receiving layer components are subtracted from the actual water-receiving layer components one by one according to the same water-receiving state variables, the same sampling level, and the same data acquisition timestamp, and the absolute values ​​are taken. All absolute values ​​are summed to obtain the reconstruction deviation. The absolute values ​​of the corresponding values ​​in the actual water-receiving layer components are summed to obtain the total actual components. The reconstruction deviation is divided by the total actual components to obtain the reconstruction deviation ratio. First, the path combination containing only one candidate causal path is calculated. Then, the number of candidate causal paths in the path combination is increased sequentially. The process stops when the number of paths with a reconstruction deviation ratio of no more than 20% appears for the first time. The path combination with the smallest reconstruction deviation ratio is selected from the path combinations with the same number of paths. The determination of the minimum sufficient disaster-causing path set is as follows: One candidate causal path is deleted sequentially from the path combination with the smallest number of paths and a reconstruction deviation ratio not exceeding 20%. The path-specific activation values ​​corresponding to the remaining candidate causal paths are summed up item by item and the reconstruction deviation ratio is calculated in the same way. The deletion calculation is performed once for each candidate causal path in the path combination. When the reconstruction deviation ratio exceeds 20% after each deletion of one candidate causal path, the candidate causal paths in the path combination before deletion are determined as the minimum sufficient disaster-causing path set. The path delay interpolation layer calculates the activation start time difference between adjacent nodes in the minimum sufficient disaster-causing path set to obtain the path propagation delay. Path-specific activation quantities are allocated according to connection strength and then translated and interpolated along the causal direction to form the initial water source activation trajectory, initial channel activation trajectory, initial water reception activation trajectory, and initial hierarchical activation quantity. The initial water reception activation trajectory is then extended and superimposed to form the initial water reception state prediction trajectory, where: The path delay interpolation layer includes: Activation time detector: Determines the activation start time of water source root causes or production operation root causes, channel status, and water receiving status; Path delay calculator: Calculates the path propagation delay from root cause to channel state and from channel state to water-receiving state; Activate the allocation interpolator: allocate path-specific activation amounts according to connection strength, and adjust the corresponding data acquisition timestamps and execution times based on path propagation delays; Trajectory predictor: generates initial water source activation trajectory, initial channel activation trajectory, initial water receiving activation trajectory and initial hierarchical activation amount, and extends the initial water receiving activation trajectory to form the initial water receiving state prediction trajectory; The calculation of the activation start time difference between adjacent nodes in the minimum sufficient disaster-causing path set is as follows: Read the variable sequences corresponding to the water source root cause or production operation root cause, channel status, and water receiving status according to the causal connection order in the minimum sufficient disaster-causing path set. Take the value corresponding to the first data collection timestamp of each variable sequence as the benchmark value, calculate the absolute difference between the values ​​corresponding to the remaining data collection timestamps and the benchmark value, and determine 20% of the maximum absolute difference as the activation judgment value. The data collection timestamp when the absolute difference first reaches the activation judgment value is determined as the activation start time. Subtract the activation start time of the water source root cause or production operation root cause from the activation start time of the channel status to obtain the path transmission delay from the root cause to the channel status. Subtract the activation start time of the channel status from the activation start time of the water receiving status to obtain the path transmission delay from the channel status to the water receiving status. Path-specific activation values ​​are assigned based on connection strength and then shifted and interpolated along the causal direction, specifically as follows: For each candidate causal path in the minimum sufficient disaster-causing path set, the absolute value of the connection strength from the water source root cause or production operation root cause to the current channel state is divided by the sum of the absolute values ​​of the connection strength from the same water source root cause or production operation root cause to all channel states in the minimum sufficient disaster-causing path set to obtain the root cause to channel allocation ratio. The path-specific activation quantity is multiplied by the root cause to channel allocation ratio to obtain the channel state activation quantity. The data acquisition timestamp of the path-specific activation quantity is added to the path transmission delay from the root cause to the channel state to obtain the timestamp corresponding to the channel state activation quantity. Then, the absolute value of the connection strength from the current channel state to the current water receiving state is divided by the sum of the absolute values ​​of the connection strength from the same channel state to all water receiving states in the minimum sufficient disaster-causing path set to obtain the channel to water receiving allocation ratio. The activation amount of the channel state is multiplied by the channel-to-water allocation ratio to obtain the activation amount of the water receiving state. The time delay of the path from the channel state to the water receiving state is added to the timestamp corresponding to the activation amount of the channel state to obtain the timestamp corresponding to the activation amount of the water receiving state. When the timestamp corresponding to the activation amount is the same as the original data acquisition timestamp, the corresponding activation amount is directly recorded. When it is between two adjacent original data acquisition timestamps, the activation amount is allocated to the two original data acquisition timestamps according to the time ratio between the two original data acquisition timestamps. The activation amounts corresponding to the water source root cause, channel state and water receiving state in each candidate causal path are arranged according to the data acquisition timestamp to form the initial water source activation trajectory, the initial channel activation trajectory and the initial water receiving activation trajectory. Then the activation amounts corresponding to the same variable, the same sampling level and the same data acquisition timestamp are added to obtain the initial level activation amount. The initial water-activated trajectory is extended and superimposed to form the initial water-activated state prediction trajectory, specifically: For each candidate causal path, the activation values ​​of the working face water inflow, water tank level, and pump discharge corresponding to the last two data acquisition timestamps of the initial water-receiving activation trajectory are read. The difference between the two activation values ​​is divided by the corresponding time difference to obtain the unit time activation change. Starting from the last data acquisition timestamp, four extension times are set according to the original acquisition interval. The unit time activation change is multiplied by the time difference between each extension time and the last data acquisition timestamp, and then added to the last activation value to obtain the extended activation value of each candidate causal path at each extension time. This extends the path pointing to the same water-receiving state. The activation values ​​of the extended variables with the same extension time are added together to obtain the prediction increment of the corresponding water-receiving state variable. The working face water inflow, water level in the water tank and water pump drainage corresponding to the last data collection timestamp of the mine water hazard associated state sequence are read as the prediction starting value. The prediction increment of the first extension time is added to the prediction starting value to obtain the first prediction value. Then the prediction increment of the second extension time is added to the first prediction value to obtain the second prediction value. The third and fourth prediction values ​​are obtained in sequence. The prediction values ​​are arranged according to the extension time to form the initial water-receiving state prediction trajectory. The conjugate feedback closed-loop layer, following the opposite direction of the path propagation time delay, feeds back the initial water-receiving state prediction trajectory to the channel state and the water source root cause, forming the channel feedback trajectory and the water source feedback trajectory, wherein: The conjugate loop-closed layer includes: Predictive Increment Extractor: Extracts the predicted increment of water reception from the predicted trajectory of the initial water reception state; Time-delay return device: Calculates the channel return activation amount and water source return activation amount in reverse according to the path transmission time delay, forming the channel return trajectory and water source return trajectory; Closed-loop deviation calculator: Compares the re-projection trajectory with the corresponding initial activation trajectory and calculates the closed-loop deviation; Path Feedback: Delete candidate causal paths with a closed-loop deviation exceeding 20% ​​and re-execute path inversion and time-delay interpolation until the minimum set of sufficiently catastrophic paths remains unchanged; The initial water-receiving state prediction trajectory is then projected back onto the channel state and water source root cause, specifically as follows: The difference between adjacent predicted values ​​in the initial water-receiving state prediction trajectory is calculated to obtain the water-receiving prediction increment corresponding to each extended time. For candidate causal paths pointing to the same water-receiving state variable, the sum of the absolute values ​​of the activation amount of the initial water-receiving activation trajectory of the current candidate causal path is divided by the sum of the absolute values ​​of the activation amounts of all corresponding candidate causal paths to obtain the prediction increment allocation ratio of the current candidate causal path. The water-receiving prediction increment is multiplied by the prediction increment allocation ratio to obtain the water-receiving return activation amount of the current candidate causal path. The channel state is obtained by subtracting the path transmission delay from the channel state to the water receiving state from the timestamp corresponding to the water receiving activation quantity. The channel activation quantity is obtained by dividing the water receiving activation quantity by the channel-to-water receiving allocation ratio. The root cause is obtained by subtracting the path transmission delay from the root cause to the channel state from the timestamp corresponding to the channel state. The water source root cause is obtained by dividing the channel activation quantity by the root cause-to-channel allocation ratio. The activation quantities that fall between two adjacent original data acquisition timestamps are allocated to the two original data acquisition timestamps according to the time ratio between the two original data acquisition timestamps. The channel activation quantity and the water source activation quantity are arranged according to the data acquisition timestamps to form the channel activation trajectory and the water source activation trajectory. Calculate the closed-loop deviation between the channel return trajectory and the water source return trajectory and the corresponding initial activation trajectory. Delete paths whose closed-loop deviation exceeds 20% of the cumulative absolute value of the corresponding initial activation trajectory. Re-execute the minimum sufficient path inversion and path delay interpolation until the minimum sufficient disaster-causing path set remains unchanged. The final iteration results are determined as the water source activation trajectory, channel activation trajectory, water reception activation trajectory, hierarchical activation amount, and water reception state prediction trajectory, where: The closed-loop deviation between the channel return trajectory and the water source return trajectory and the corresponding initial activation trajectory is calculated as follows: For each candidate causal path, the activation amount corresponding to the same data collection timestamp of the channel return trajectory is subtracted from that of the initial channel activation trajectory, and the absolute value is taken. The activation amount corresponding to the same data collection timestamp of the water source return trajectory is subtracted from that of the initial water source activation trajectory, and the absolute value is taken. All absolute values ​​are added together to obtain the closed-loop deviation. The absolute values ​​of the activation amounts in the initial channel activation trajectory and the initial water source activation trajectory are then taken and summed to obtain the cumulative absolute value of the initial activation trajectory. The closed-loop deviation is compared with 20% of the cumulative absolute value of the initial activation trajectory, and candidate causal paths with a closed-loop deviation exceeding 20% ​​of the cumulative absolute value of the initial activation trajectory are deleted. Re-execute the minimum sufficient path inversion and path delay interpolation, specifically as follows: After deleting candidate causal paths with a closed-loop deviation exceeding 20%, the path-specific activation values ​​corresponding to the retained candidate causal paths are added one by one according to the same water-receiving state variables, sampling levels, and data acquisition timestamps to reconstruct the actual water-receiving level components. The reconstruction deviation ratio between the reconstruction result and the actual water-receiving level components is calculated. Path combinations with a reconstruction deviation ratio not exceeding 20% ​​are selected from the fewest to the most numerous paths. Candidate causal paths in the selected path combinations are deleted one by one, and the reconstruction deviation ratio is recalculated. Path combinations with a reconstruction deviation ratio exceeding 20% ​​after deleting any candidate causal path are determined as the new minimum sufficient disaster-causing path set. Then, recalculate the path propagation delays from the root cause to the channel state and from the channel state to the water-receiving state in the new minimum sufficient disaster-causing path set. Reallocate the path-specific activation quantities according to the connection strength. Determine the corresponding timestamps of the channel state and the water-receiving state by adding the corresponding path propagation delays to the original data acquisition timestamps. Distribute the activation quantities between two adjacent original data acquisition timestamps according to the time ratio. Re-form the initial water source activation trajectory, initial channel activation trajectory, initial water-receiving activation trajectory, initial hierarchical activation quantity, and initial water-receiving state prediction trajectory. Then, re-execute the incremental allocation of water-receiving prediction, reverse deduction of path propagation delays, and calculation of closed-loop deviation. Stop the iteration when the minimum sufficient disaster-causing path set obtained in two consecutive iterations contains the exact same candidate causal path. Determine the initial water source activation trajectory, initial channel activation trajectory, initial water-receiving activation trajectory, initial hierarchical activation quantity, and initial water-receiving state prediction trajectory obtained when the iteration stops as the water source activation trajectory, channel activation trajectory, water-receiving activation trajectory, hierarchical activation quantity, and water-receiving state prediction trajectory, respectively. An improved N-HiTS model was trained using a weighted sum of water state prediction error, hierarchical component reconstruction error, and conjugate re-projection deviation. The parameters of the causal path erasure counterfactual layer, minimum sufficient path inversion layer, path delay interpolation layer, and conjugate re-projection closed-loop layer were optimized. The improved N-HiTS model was considered convergent when the change in loss was less than 0.001 and the rate of change in closed-loop deviation was less than 1% after 10 consecutive rounds of verification. The N-HiTS model is improved by training a weighted sum of water state prediction error, hierarchical component reconstruction error, and conjugate retracement deviation. Specifically: The mine water hazard associated state sequence is arranged from earliest to latest according to the data collection timestamps. The 80% mark of the total number of data collection timestamps is used as the dividing point. The data before the dividing point is used as training data, and the data after the dividing point is used as validation data. In the training data, data from 24 consecutive collection times are read as one model input sample, and the data from the next four collection times are read as the corresponding prediction targets. The starting position is moved backward according to one collection time to form training samples. The training samples are input into the improved N-HiTS model to obtain the water state prediction trajectory, reconstruct the actual water level components, channel return trajectory, and water source return trajectory. The water supply prediction error is obtained by subtracting the actual values ​​corresponding to the same time in the prediction target from the water inflow, water level in the reservoir, and water pump discharge at each prediction time in the water supply prediction trajectory, taking the absolute value, and averaging the results. The level component reconstruction error is obtained by subtracting the reconstructed actual water supply level component from the actual water supply level component according to the same water supply state variable, sampling level, and data acquisition timestamp, taking the absolute value, and averaging the results. The conjugate re-projection deviation is obtained by subtracting the channel return trajectory from the initial channel activation trajectory and the water source return trajectory from the initial water source activation trajectory according to the same data acquisition timestamp, taking the absolute value, and averaging the results. The training loss is obtained by multiplying the water supply prediction error, level component reconstruction error, and conjugate re-projection deviation by 0.5, 0.3, and 0.2 respectively. The trainable parameters in the improved N-HiTS model are updated based on the training loss with an initial learning rate of 0.001. After each parameter update of all training samples, the validation data is used as the input of the improved N-HiTS model for 24 consecutive acquisition times, and the next 4 acquisition times are used as the prediction targets. The data is then moved one acquisition time at a time to form validation samples. For each validation sample, the water state prediction error, hierarchical component reconstruction error, and conjugate re-projection deviation are calculated. These are then multiplied by 0.5, 0.3, and 0.2 respectively and added together to obtain the loss of a single validation sample. The losses of all validation samples are added together and divided by the number of validation samples to obtain the validation loss of the current round. When the absolute difference between the validation loss of 10 consecutive rounds and the validation loss of the previous round is less than 0.001, and the absolute value of the difference between the conjugate re-projection deviations of two adjacent rounds divided by the conjugate re-projection deviation of the previous round is less than 1%, the parameter update is stopped and the improved N-HiTS model is considered to have converged.

[0025] In this embodiment, the calculation of the causal contribution of water damage and the operational causal contribution includes: Read the causal edges, connection strength, and hierarchical activation in the dual-root-cause causal framework, and map the aquifer water pressure and the water level in the old workings water accumulation area as water source root causes, and map the working face advancement, water pump start-up and shutdown, and blasting operations as production operation root causes, and associate them with the channel status and water receiving status, where: Channel state refers to borehole water pressure, micro-vibration energy, and apparent resistivity located between the water source root cause or production operation root cause and the water receiving state in the dual root cause causal framework. Water receiving status refers to the working face water inflow, water tank level, and water pump drainage volume in the receiving channel state causal transmission of the dual-root cause-effect framework. The connection strengths pointing to the same node are normalized, and the hierarchical activation amounts of upstream nodes are allocated according to the normalized connection strengths, which are then sequentially propagated to the channel state and the water-receiving state along the causal direction, wherein: The hierarchical activation amount of upstream nodes is allocated according to the normalized connection strength, specifically as follows: For the same channel state, the absolute values ​​of the connection strengths of all water source root causes and production operation root causes pointing to the current channel state are taken and summed. The absolute value of the connection strength of each causal edge is divided by the summation result to obtain the corresponding normalized connection strength. The hierarchical activation amount of the current water source root cause or production operation root cause at the same sampling level and data acquisition timestamp is multiplied by the corresponding normalized connection strength to obtain the hierarchical activation amount transmitted to the current channel state. Then, for the same water receiving state, the connection strengths of all channel states pointing to the current water receiving state are normalized in the same way. The hierarchical activation amount that has been transmitted to the current channel state is multiplied by the normalized connection strength of the current channel state pointing to the water receiving state to obtain the hierarchical activation amount transmitted to the water receiving state along the current causal path. The hierarchical activation amounts of water source root causes and production operation root causes transmitted along each causal edge to the channel state and water-receiving state are summarized to obtain the causal contribution of water damage and the causal contribution of operation, where: The causal contributions of water damage and operational causal contributions are obtained as follows: According to the same sampling level and data acquisition timestamp, the activation amounts of the aquifer water pressure and the water level in the old workings water accumulation area are transferred to the same channel state and water-receiving state along each causal path, and the water hazard causal contribution of the corresponding channel state and water-receiving state is added respectively. The activation amounts of the working face advancement, water pump start-up and shutdown, and blasting operations are transferred to the same channel state and water-receiving state along each causal path, and the operational causal contribution of the corresponding channel state and water-receiving state is added respectively. The water hazard causal contribution and operational causal contribution are recorded according to the channel state or water-receiving state, sampling level, and data acquisition timestamp.

[0026] In this embodiment, forming a causal contribution closed path includes: Based on the data collection timestamp and hierarchical alignment of the causal contribution of water hazards, the operational causal contribution, and the hierarchical activation amount, the proportions of the causal contribution of water hazards and the operational causal contribution are calculated to obtain the water hazard competition weight and the operational competition weight, where: The water hazard competition weights and operational competition weights are obtained as follows: According to the same channel status or water-receiving status, the same sampling level and the same data collection timestamp, the causal contribution of water damage, the causal contribution of operation and the activation amount of the level are corresponding to the causal contribution of water damage and the causal contribution of operation. The absolute value of the causal contribution of water damage is added to the absolute value of the causal contribution of operation to obtain the total causal contribution. The absolute value of the causal contribution of water damage is divided by the total causal contribution to obtain the water damage competition weight. The absolute value of the causal contribution of operation is divided by the total causal contribution to obtain the operation competition weight. The sum of the water damage competition weight and the operation competition weight at the same position is 1. The activation amounts are allocated hierarchically based on the competition weights for water hazards and operations, resulting in water hazard matching activation amounts and operation matching activation amounts. The deviation ratios between the water hazard matching activation amounts and the causal contributions of water hazards, and between the operation matching activation amounts and the causal contributions of operations, are calculated. Water hazard causal contributions and operation causal contributions with deviation ratios not exceeding 20% ​​are retained. The activation amount for each level is allocated according to the competition weights for water hazards and operations, specifically as follows: Based on the same channel status or water-receiving status, the same sampling level and the same data collection timestamp, the hierarchical activation quantity is multiplied by the water hazard competition weight to obtain the water hazard matching activation quantity, and the hierarchical activation quantity is multiplied by the operation competition weight to obtain the operation matching activation quantity, so that the same hierarchical activation quantity corresponds to the water hazard causal contribution caused by the water source root cause and the operation causal contribution caused by the production operation root cause, respectively. The deviation ratios of the water hazard matching activation amount and the causal contribution of water hazards, as well as the deviation ratios of the operational matching activation amount and the operational causal contribution, are calculated as follows: Subtract the water hazard matching activation quantity from the water hazard causal contribution corresponding to the same channel state or water receiving state, the same sampling level, and the same data acquisition timestamp, and take the absolute value. Then divide the absolute value of the water hazard causal contribution to obtain the water hazard deviation ratio. Subtract the operation matching activation quantity from the corresponding operation causal contribution, and take the absolute value. Then divide the absolute value of the operation causal contribution to obtain the operation deviation ratio. Keep the water hazard causal contribution and the operation causal contribution with an operation deviation ratio of no more than 20% and the corresponding matching activation quantity, respectively. Connect the root cause nodes, channel state nodes, and water-receiving state nodes corresponding to the retained water hazard causal contributions or operational causal contributions along the dual-root-cause causal framework, and associate them with water hazard matching activation quantities or operational matching activation quantities to form a closed causal contribution path, wherein: The causal contribution closed path is formed as follows: From the dual-root cause causal framework, read the causal edge from the water source root cause to the water receiving state via the channel state. Connect the continuous causal edges that retain the causal contribution of water damage and the corresponding water damage matching activation amount in both the channel state and the water receiving state. Form a closed path of water damage causal contribution by combining the water source root cause, channel state, water receiving state, causal contribution of water damage, and matching activation amount. Then, read the causal edge from the production operation root cause to the water receiving state via the channel state. Connect the continuous causal edges that retain the causal contribution of operation and the corresponding operation matching activation amount in both the channel state and the water receiving state. Form a closed path of operation causal contribution by combining the production operation root cause, channel state, water receiving state, causal contribution of operation, and matching activation amount. And determine the two types of paths together as the causal contribution closed path.

[0027] In this embodiment, the step of forming an active flood-causing chain and determining the abnormal transmission time delay includes: Based on the data collection timestamps, the causal contribution closed paths are aligned with the water source activation trajectory, channel activation trajectory, and water reception activation trajectory. The start times of water source activation, channel activation, and water reception activation are extracted. Causal contribution closed paths with sequentially increasing activation start times that satisfy hydrogeological connectivity are retained, among which: Hydrogeological connectivity refers to the continuous hydraulic connection between the water source root corresponding to the water body, the channel state corresponding to the location, and the water receiving state corresponding to the location recorded in the hydrogeological connectivity data. The causal contribution closed paths that retain sequentially increasing activation start times and satisfy hydrogeological connectivity relationships are as follows: For each causal contribution closed path, the start time of water source activation, the start time of channel activation, and the start time of water reception activation are read according to the same data collection timestamp. The causal contribution closed paths where the start time of water source activation is earlier than the start time of channel activation and the start time of channel activation is earlier than the start time of water reception activation are retained. Then, the hydrogeological connection data is read, and the causal contribution closed paths where there is a hydrogeological connection relationship between the root cause of water source and channel state, and between channel state and water reception state are retained. From the preserved causal contribution closed paths, root cause nodes, channel state nodes, and water-receiving state nodes that contain causal contributions to water hazards are selected. These nodes are then connected sequentially according to hydrogeological connections and causal directions to form an active water hazard disaster chain, in which: The active flood disaster chain is formed, specifically as follows: Delete the paths containing only operational causal contributions from the retained causal contribution closed paths, retain the paths containing water hazard causal contributions, read the water source root cause, channel state and water reception state corresponding to the path, connect them according to the causal direction of water source root cause pointing to channel state and channel state pointing to water reception state, and retain the continuous paths where both connections satisfy the hydrogeological connection relationship, and form the active water hazard disaster chain by connecting the water source root cause, channel state, water reception state and corresponding water hazard causal contribution. The time differences from the start time of water source activation to the start time of channel activation, and from the start time of channel activation to the start time of water reception activation, are calculated along the active flood hazard chain. These time differences are then summed to obtain the abnormal transmission time delay of the active flood hazard chain. The abnormal transmission time lag of the active flood disaster chain was obtained, specifically: For each active flood-causing chain, the abnormal transmission delay from the water source to the channel is obtained by subtracting the activation start time of the channel from the activation start time of the channel. The abnormal transmission delay from the channel to the water receiving state is obtained by subtracting the activation start time of the receiving state from the activation start time of the receiving state. Then, the abnormal transmission delay from the water source to the channel and the abnormal transmission delay from the channel to the water receiving state are added together to obtain the abnormal transmission delay of the active flood-causing chain from the root cause of the water source to the receiving state through the channel state.

[0028] In this embodiment, determining the state of flood damage and outputting early warning results includes: Read the activation start time of the water source in the active flood disaster chain, decompose the anomaly transmission time delay into the water source to channel time delay and the channel to water receiving time delay, and accumulate them segment by segment to obtain the arrival time of the flood causal contribution to the channel state and the water receiving state, where: The anomaly transmission time delay is decomposed into the time delay from the water source to the channel and the time delay from the channel to the water receiving point, and then accumulated segment by segment, specifically: The time difference between the activation start time of the channel and the activation start time of the water source in the active water disaster chain is read as the water source to channel time lag. The time difference between the activation start time of water reception and the activation start time of the channel is read as the channel to water reception time lag. The water source activation start time is added to the water source to channel time lag to obtain the arrival time of the water disaster causal contribution to the channel state. Then, the channel to water reception time lag is added to the arrival time of the channel state to obtain the arrival time of the water disaster causal contribution to the water reception state. Based on the arrival time of the causal contribution of water damage to the water-receiving state, the predicted inflow is read from the predicted trajectory of the water-receiving state. Then, according to the current data acquisition timestamp, the inflow at the working face, the pump discharge, and the water level in the water tank are read. The change in water level in the water tank at adjacent acquisition times is calculated, where: The calculation of the water level change in the reservoir at adjacent data collection times is as follows: Read the current water level of the reservoir according to the current data collection timestamp, and then read the water level of the reservoir corresponding to the previous data collection timestamp. Subtract the water level of the reservoir corresponding to the previous data collection timestamp from the current water level to obtain the change in water level. A change in water level greater than 0 indicates that the water level has risen, and a change in water level less than 0 indicates that the water level has fallen. The predicted water inflow is compared with the water inflow at the working face and the drainage volume of the pumps. The water level change in the water tank is then used to determine the flood hazard status. The resulting warnings include the active flood hazard chain, the arrival time of the flood hazard's causal contribution, the predicted water inflow, and the corresponding flood hazard status. The state of flood damage is determined by combining changes in the water level of the reservoir, specifically as follows: The difference between the predicted inflow and the current inflow at the working face is calculated. The difference is divided by the current inflow to obtain the inflow growth rate. The predicted inflow is compared with the current pump discharge rate. When the inflow growth rate does not exceed 20% and the predicted inflow does not exceed the current pump discharge rate, it is considered a normal state. When the inflow growth rate exceeds 20% or the predicted inflow exceeds the current pump discharge rate, it is considered a state of concern. When the inflow growth rate exceeds 20%, the predicted inflow exceeds the current pump discharge rate, and the water level change in the water tank exceeds 0.10 meters, it is considered a flood warning state. The corresponding flood state, the active flood hazard chain, the arrival time of the flood causal contribution channel and the water receiving state, and the predicted inflow are all written into the warning result.

[0029] refer to Figure 3 A mine water hazard early warning system based on big data analysis includes the following modules: The data acquisition module is used to collect and preprocess multi-source correlation data of mine water hazards and generate a mine water hazard correlation state sequence. The causal skeleton construction module is used to extract state change features and execute the LiNGAM causal discovery algorithm to construct a dual-root cause causal skeleton. The water hazard activation prediction module is used to extract the water source activation trajectory, channel activation trajectory, water receiving activation trajectory and hierarchical activation amount by improving the N-HiTS model, and form the water receiving state prediction trajectory. The causal contribution calculation module is used to pass hierarchical activation along the double root cause causal framework to obtain the causal contribution of water damage and the causal contribution of operation. The contribution competition matching module is used to compete and match the causal contribution of water damage, the causal contribution of operation, and the hierarchical activation amount to form a closed causal contribution path; The disaster chain identification module is used to combine the time sequence of activation trajectories and hydrogeological connections to screen closed paths of causal contribution, form active water hazard disaster chains and determine the time lag of anomaly transmission. The early warning output module is used to deduce the arrival time of the causal contribution of water damage, determine the state of water damage, and output early warning results.

[0030] Example 1: To verify the feasibility of the present invention in practice, the present invention was applied to a mine water hazard simulation monitoring scenario with an aquifer, a water-filled old goaf, a borehole monitoring area, a coal mining face, a water tank, and drainage equipment. In the scenario, the aquifer is hydrogeologically connected to the borehole monitoring area and the working face through a fracture zone, while the water-filled old goaf is not directly hydrogeologically connected to the current working face. At the same time, working face advancement, water pump start-up and shutdown, and blasting operations are set up so that production disturbances and water hazard anomalies overlap during continuous operation.

[0031] The system continuously collects data on aquifer water pressure, water level in the old workings water accumulation area, borehole water pressure, microseismic energy, apparent resistivity, working face water inflow, water tank level, pump discharge, working face advance, pump start / stop, blasting operations, and data acquisition timestamps. During stable operation, the aquifer water pressure is maintained at 2.18 to 2.21 MPa, the borehole water pressure at 0.73 to 0.76 MPa, the working face water inflow at 24.6 to 25.5 cubic meters per hour, the water tank level at 1.33 to 1.36 meters, and the pump discharge at 38.0 to 39.2 cubic meters per hour. During the preprocessing, a borehole water pressure acquisition value was missing, with the previous values ​​being 0.82 MPa and 0.86 MPa respectively. 0.84 MPa was added. An isolated microseismic value of 38,100 joules appeared at an acquisition location where no blasting operation had occurred, with the previous values ​​being 1,860 joules and 1,930 joules respectively. After being removed, 1,895 joules was added. Subsequently, time alignment, spatial correlation and normalization were completed to form a mine water hazard related state sequence.

[0032] The improved N-HiTS model was trained using 8640 consecutive acquisition times, of which 6912 were used as training data (80%) and 1728 were used as validation data (20%). 24 consecutive acquisition times were used as model inputs, and the next 4 acquisition times were used as prediction targets. After shifting by 1 acquisition time, 6885 training samples and 1701 validation samples were formed. In a production disturbance training sample, blasting operations increased the microseismic energy from 1820 joules to 7480 joules, but the aquifer water pressure only changed from 2.19 MPa to 2.21 MPa, and the working face water inflow only changed from 25.0 cubic meters per hour to 26.3 cubic meters per hour. In a water hazard sample, the aquifer water pressure increased from 2.20 MPa to 2.54 MPa, the borehole water pressure increased laggingly to 1.08 MPa, the apparent resistivity decreased from 117 ohm-meters to 84 ohm-meters, and the working face water inflow increased from 25.4 cubic meters per hour to 37.8 cubic meters per hour. The subsequent four predicted targets were 40.2, 43.1, 46.5, and 49.3 cubic meters per hour, respectively. In the training loss, the weights for water state prediction error, hierarchical component reconstruction error, and conjugate retrieval deviation are 0.5, 0.3, and 0.2, respectively, with a sum of 1. After 10 consecutive rounds of validation, the loss decreased from 0.0416 to 0.0386, with the change in adjacent rounds all below 0.001. The rate of change of conjugate retrieval deviation was less than 1%, indicating that the model had converged.

[0033] During the simulation of the improved N-HiTS model, a blasting operation occurred first, and the microseismic energy increased from 1810 joules to 7420 joules. However, the aquifer water pressure remained at 2.21 to 2.22 MPa, and the water inflow at the working face increased only from 25.8 cubic meters per hour to 26.4 cubic meters per hour before falling back to 26.0 cubic meters per hour. Subsequently, the aquifer water pressure continuously increased from 2.21 MPa to 2.24, 2.29, 2.35, 2.43, 2.51, 2.54 and 2.55 MPa, the water level in the old workings water accumulation area remained at 4.62 to 4.65 meters, the borehole water pressure gradually increased from 0.76 MPa to 1.07 MPa, the apparent resistivity decreased from 116 ohm-meters to 82 ohm-meters, and the working face water inflow gradually increased from 25.8 cubic meters per hour to 35.6 cubic meters per hour, forming a state in which changes in the water source preceded changes in the channel, and changes in the channel preceded the water supply response.

[0034] After 1000 self-sampling iterations, the LiNGAM causal discovery algorithm showed that the causal edge from aquifer water pressure to borehole water pressure appeared 926 times (92.6% occurrence rate) with an average connection strength of 0.63; the causal edge from aquifer water pressure to apparent resistivity appeared 887 times (88.7% occurrence rate) with an average absolute connection strength of 0.46; the causal edge from borehole water pressure to working face water inflow appeared 914 times (91.4% occurrence rate) with an average connection strength of 0.59; and the causal edge from apparent resistivity to working face water inflow appeared 869 times (86.9% occurrence rate) with an average absolute connection strength of 0.43. All these results met the retention criteria of 80% and 0.20. The causal edge from blasting operations to microseismic energy appeared in 95.4% of the time; therefore, the production operation relationship remains in the dual-rooted causal framework and participates in subsequent competition.

[0035] After performing connection retention and connection erasure on candidate paths from aquifer water pressure to working face water inflow via borehole water pressure, path-specific activation values ​​of 0.236, 0.095, and 0.035 were obtained at the three sampling levels, totaling 0.366. The path-specific activation value of the path from aquifer water pressure to working face water inflow via apparent resistivity totaled 0.290, while the cumulative activation value of blast-related paths was only 0.076. When reconstructing the actual water-receiving layer components using only the two water source paths, the reconstruction deviation ratios were 24.6% and 29.8%, respectively. After both paths were used together, the reconstruction deviation ratio decreased to 10.9%. After deleting either path, the reconstruction deviation ratio exceeded 20% again. Therefore, the two water source paths were determined to be the minimum sufficient disaster-causing path set.

[0036] The maximum absolute change in aquifer water pressure is 0.34 MPa, and the 20% activation threshold is 0.068 MPa. The aquifer water pressure becomes active when it increases to 2.29 MPa. The corresponding activation threshold for borehole water pressure is 0.062 MPa, and it becomes active when it increases to 0.83 MPa, which is two sampling intervals later than the aquifer water pressure. The working face water inflow is based on 25.8 cubic meters per hour, with a maximum absolute change of 16.9 cubic meters per hour and a 20% activation threshold of 3.38 cubic meters per hour. The aquifer water pressure becomes active when it increases to 30.8 cubic meters per hour, which is two sampling intervals later than the borehole water pressure. Therefore, the complete anomaly transmission time lag is 4 sampling intervals.

[0037] The connection strengths of the aquifer water pressure to the borehole water pressure and apparent resistivity are 0.63 and 0.46, respectively, which, after normalization, are 0.578 and 0.422, respectively, and their sum is 1. In the water supply prediction stage, the current working face inflow is 31.6 cubic meters per hour. The predicted increments for the four future data acquisition locations are 3.4, 3.9, 4.4, and 4.9 cubic meters per hour, respectively. These increments are accumulated to form predicted inflows of 35.0, 38.9, 43.3, and 48.2 cubic meters per hour. After conjugate re-implementation, the closed-loop deviation rates for the two water source paths are 9.23% and 11.7%, respectively, both below 20%. The closed-loop deviation rate for the blasting-related path is 33.8%, therefore, the blasting-related path is deleted.

[0038] At a water-receiving response acquisition location, the causal contribution of water hazard is 0.338, the causal contribution of operation is 0.082, and the total is 0.420. This yields a water hazard competition weight of 0.805 and an operation competition weight of 0.195, summing to 1. The current level activation amount is 0.410. After allocation, the water hazard matching activation amount is 0.330, and the operation matching activation amount is 0.080, still summing to 0.410. The corresponding deviation ratios are 2.37% and 2.44%, respectively, both below 20%. Combining the hydrogeological connectivity and the temporal sequence of activation of water source, channel, and water-receiving state, an active water hazard disaster chain is ultimately formed, whereby aquifer water pressure is transmitted to the working face water inflow via borehole water pressure and apparent resistivity.

[0039] When the actual water inflow at the working face is 31.6 cubic meters per hour, the water pump discharge is 39.5 cubic meters per hour, and the water level in the water tank is 1.40 meters, this invention predicts that the water inflow at the working face will increase to 48.2 cubic meters per hour. Subsequently, the actual water inflow at the working face reaches 34.8, 38.2, 42.6, and 47.9 cubic meters per hour, respectively, with absolute errors of 0.2, 0.7, 0.7, and 0.3 cubic meters per hour from the predicted values. The water level in the water tank increases from 1.40 meters to 1.46, 1.57, 1.72, and 1.91 meters. When the change in water level in adjacent water tanks reaches 0.11 meters and the predicted water inflow exceeds the discharge of 39.5 cubic meters per hour, the system outputs a water hazard warning status.

[0040] Traditional methods employ a combination of fixed thresholds and slip trend analysis, common in mine monitoring. Alarm thresholds are set for aquifer pressure, microseismic energy, working face water inflow, and water level in the reservoir. Warnings are issued based on the degree to which monitored values ​​exceed limits and their continuous changing trends. In the blasting segment, the traditional method generates an anomaly alert when microseismic energy reaches 7420 joules, but the working face water inflow subsequently recovers, resulting in a false alarm. During the water hazard evolution process, the traditional method requires waiting for the working face water inflow to continuously reach 42.6 and 47.9 cubic meters per hour before issuing a water hazard warning. This invention issues a warning when the working face water inflow reaches 38.2 cubic meters per hour, the water level in the reservoir changes by 0.11 meters, and the predicted water inflow exceeds the drainage capacity—two data collection intervals earlier than the traditional method.

[0041] The invention was validated using 48 independent simulated segments, including 24 water hazard events and 24 non-water hazard events. The invention correctly identified 23 water hazard events (95.8%), while the traditional method correctly identified 20 (83.3%). The invention generated one false alarm for non-water hazard events (4.2%), while the traditional method generated six false alarms (25.0%). The invention achieved a 91.7% accuracy rate for source water location and an 87.5% accuracy rate for anomaly channel location, with an average absolute prediction error of 2.3 cubic meters per hour for working face water inflow. In contrast, the traditional method achieved a 66.7% accuracy rate for approximate source water location and a 58.3% accuracy rate for approximate anomaly channel location, with an average absolute prediction error of 5.8 cubic meters per hour for working face water inflow. This demonstrates that the invention can improve the accuracy of source water hazard identification, anomaly transmission location, water inflow trend prediction, and early warning even under conditions of operational disturbances.

[0042] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A mine water hazard early warning method based on big data analysis, characterized in that, include: Collect multi-source correlation data of mine water hazards, and perform preprocessing on the multi-source correlation data of mine water hazards to generate a mine water hazard correlation state sequence; Based on the state change features extracted from the mine water hazard associated state sequence, the LiNGAM causal discovery algorithm is executed to retain the causal edges from the water source root cause and the production operation root cause to the water receiving state through the channel state, and to construct a dual-root cause causal skeleton. An improved N-HiTS model is constructed, and multi-rate sampling, hierarchical residual decomposition and hierarchical interpolation are performed on the variable sequence in the dual-root cause causal skeleton to extract the water source activation trajectory, channel activation trajectory, water receiving activation trajectory and hierarchical activation amount, and the water receiving activation trajectory is extrapolated to form the water receiving state prediction trajectory. Along the causal direction of the dual-root cause causal framework, the causal contribution of water damage and the operational causal contribution are calculated according to the activation amount of the connection strength transmission hierarchy. The causal contribution of water damage and the causal contribution of operation are competitively matched with the hierarchical activation amount to form a closed causal contribution path; Based on the temporal sequence and hydrogeological connection of the water source activation trajectory, channel activation trajectory and water-receiving activation trajectory, causal contribution closed paths are screened to form active water hazard disaster chains and determine the anomaly transmission time lag. Based on the time delay of abnormal transmission along the active water hazard disaster chain, the arrival time of the causal contribution of water hazard is deduced, the predicted inflow volume in the water-receiving state prediction trajectory is read, the water hazard state is determined, and the early warning result is output.

2. The mine water hazard early warning method based on big data analysis according to claim 1, characterized in that, The specific multi-source correlation data of mine water hazards includes aquifer water pressure data, old working water accumulation area water level data, borehole water pressure data, microseismic energy data, apparent resistivity data, working face water inflow data, water tank water level data, water pump drainage data, working face advance data, water pump start and stop data, blasting operation data, hydrogeological connection data, and data acquisition timestamps.

3. The mine water hazard early warning method based on big data analysis according to claim 1, characterized in that, The preprocessing of multi-source correlation data of mine water hazards specifically includes time alignment, missing data completion, anomaly removal, spatial correlation and normalization.

4. The mine water hazard early warning method based on big data analysis according to claim 1, characterized in that, The construction of the dual-root-cause causal framework includes: Based on the data collection timestamp, the mine water hazard associated state sequence is extracted, and the state difference, rate of change and cumulative change at adjacent collection times are calculated to obtain the state change characteristics. The aquifer water pressure and the water level in the old workings water accumulation area are classified as water source root variables, the working face advancement, water pump start-up and shutdown and blasting operations are classified as production operation root variables, the borehole water pressure, micro-vibration energy and apparent resistivity are classified as channel state variables, and the working face water inflow, water tank water level and water pump drainage are classified as water receiving state variables. The causal analysis sample is composed according to the data acquisition timestamp. The LiNGAM causal discovery algorithm is executed on the causal analysis samples to determine the causal order of variables and calculate the connection strength between variables, thereby obtaining candidate causal edges. Perform self-service resampling on candidate causal edges, and retain causal edges that have an occurrence rate of 80% and an absolute value of connection strength of 0.20, pointing from water source root cause or production operation root cause to channel state and from channel state to water receiving state. By connecting the retained causal edges according to the causal direction, associating hydrogeological data with data acquisition timestamps, a dual-root cause causal framework is constructed with water source root cause and production operation root cause as the starting node, channel status as the intermediate node, and water reception status as the ending node.

5. The mine water hazard early warning method based on big data analysis according to claim 1, characterized in that, The extrapolated water-receiving activation trajectory forms a water-receiving state prediction trajectory, including: An improved N-HiTS model is constructed, which includes a causal path erasure counterfactual layer, a minimum sufficient path inversion layer, a path delay interpolation layer, and a conjugate back-injection closed-loop layer. The causal path erasure counterfactual layer extracts candidate causal paths and corresponding variable sequences from the mine water hazard associated state sequence, forming a path retention sequence and a path erasure sequence that sets the current path connection strength to 0; Hierarchical sampling, residual fitting, and timestamp back-interpolation were performed on the path-preserved sequence, path-erased sequence, and water-receiving state variable sequence at 1, 2, and 4 sampling intervals to obtain the path-preserved water-receiving layer component, the path-erased water-receiving layer component, and the actual water-receiving layer component. The path-specific activation amount was obtained by subtracting the path-preserved water-receiving layer component and the path-erased water-receiving layer component. The actual water-receiving layer components are reconstructed using the minimum sufficient path inversion layer superimposed with path-specific activation. The path combinations with the smallest reconstruction deviation ratio (not exceeding 20%) and the fewest number of paths are selected. The path combinations with a reconstruction deviation ratio exceeding 20% ​​after deleting any path are determined as the minimum sufficient disaster-causing path set. The path delay interpolation layer calculates the activation start time difference between adjacent nodes in the minimum sufficient disaster-causing path set to obtain the path propagation delay. It then allocates path-specific activation amounts according to the connection strength and translates and interpolates along the causal direction to form the initial water source activation trajectory, initial channel activation trajectory, initial water receiving activation trajectory, and initial hierarchical activation amount. Finally, it extends and superimposes the initial water receiving activation trajectory to form the initial water receiving state prediction trajectory. The conjugate back-feeding closed-loop layer back-feeds the initial water-receiving state prediction trajectory to the channel state and water source root cause in the opposite direction of the path transmission time delay, forming the channel back-feeding trajectory and the water source back-feeding trajectory. Calculate the closed-loop deviation between the channel return trajectory and the water source return trajectory and the corresponding initial activation trajectory. Delete paths whose closed-loop deviation exceeds 20% of the cumulative absolute value of the corresponding initial activation trajectory. Re-execute the minimum sufficient path inversion and path time delay interpolation until the minimum sufficient disaster-causing path set remains unchanged. Determine the final iteration results as the water source activation trajectory, channel activation trajectory, water receiving activation trajectory, hierarchical activation amount, and water receiving state prediction trajectory. The improved N-HiTS model was trained by weighting the prediction error of water state, the reconstruction error of hierarchical components, and the deviation of conjugate back-projection. The parameters of the causal path erasure counterfactual layer, the minimum sufficient path inversion layer, the path delay interpolation layer, and the conjugate back-projection closed loop layer were optimized. The improved N-HiTS model was determined to have converged when the change in loss was less than 0.001 and the change rate of closed loop deviation was less than 1% after 10 consecutive rounds of verification.

6. The mine water hazard early warning method based on big data analysis according to claim 1, characterized in that, The calculation of the causal contribution of water damage and the operational causal contribution includes: Read the causal edges, connection strength and hierarchical activation amount in the dual-root cause causal skeleton, and associate the aquifer water pressure and the water level in the old working water accumulation area as water source root causes, and associate the working face advancement, water pump start and stop and blasting operation as production operation root causes, and associate the channel status and water receiving status. The connection strength pointing to the same node is normalized, and the hierarchical activation amount of the upstream node is allocated according to the normalized connection strength, and then passed to the channel state and the water-receiving state in sequence along the causal direction. The hierarchical activation amounts of water source root causes and production operation root causes transmitted along each causal edge to the channel state and water-receiving state are summarized to obtain the causal contribution of water damage and the causal contribution of operation.

7. The mine water hazard early warning method based on big data analysis according to claim 1, characterized in that, The formation of the causal contribution closed path includes: Based on the data collection timestamp and hierarchical alignment of the causal contribution of water hazards, the operational causal contribution, and the hierarchical activation amount, the proportions of the causal contribution of water hazards and the operational causal contribution are calculated to obtain the water hazard competition weight and the operational competition weight. The activation amount of each level is allocated according to the competition weight of water hazards and the competition weight of operations. The activation amount of water hazards and the activation amount of operations are obtained. The deviation ratio between the activation amount of water hazards and the causal contribution of water hazards and the deviation ratio between the activation amount of operations and the causal contribution of operations are calculated. The causal contribution of water hazards and the causal contribution of operations with a deviation ratio of no more than 20% are retained. Connect the root cause nodes, channel state nodes, and water-receiving state nodes corresponding to the retained water hazard causal contribution or operational causal contribution along the dual root cause causal framework, associate the water hazard matching activation amount or operational matching activation amount, and form a closed path of causal contribution.

8. The mine water hazard early warning method based on big data analysis according to claim 1, characterized in that, The formation of the active flood-causing chain and the determination of the abnormal transmission time delay include: Based on the data collection timestamp, the causal contribution closed path is aligned with the water source activation trajectory, channel activation trajectory and water receiving activation trajectory. The start time of water source activation, the start time of channel activation and the start time of water receiving activation are extracted. The causal contribution closed path with the activation start time increasing sequentially and satisfying the hydrogeological connection relationship is retained. From the preserved closed causal contribution paths, root cause nodes, channel state nodes, and water-receiving state nodes that contain the causal contribution of water hazards are selected, and connected sequentially according to hydrogeological connection relationships and causal directions to form an active water hazard disaster chain. The time difference between the activation start time of the water source and the activation start time of the channel, and between the activation start time of the channel and the activation start time of the water receiving are calculated along the active flood disaster chain, and the time differences are accumulated to obtain the abnormal transmission time delay of the active flood disaster chain.

9. A mine water hazard early warning method based on big data analysis according to claim 1, characterized in that, The process of determining the state of flood damage and outputting early warning results includes: Read the activation start time of the water source in the active water disaster chain, decompose the abnormal transmission time delay into the water source to channel time delay and the channel to water receiving time delay, and accumulate them segment by segment to obtain the arrival time of the water disaster causal contribution to the channel state and the water receiving state. According to the arrival time of the causal contribution of water damage to the water-receiving state, the predicted water inflow is read from the predicted trajectory of the water-receiving state, and the water inflow at the working face, the water pump discharge and the water level in the water tank are read according to the current data collection timestamp, and the water level change in the water tank at adjacent collection times is calculated. The predicted water inflow is compared with the water inflow at the working face and the drainage volume of the water pump. The water hazard status is determined by combining the water level change in the water tank. The results of the active water hazard disaster chain, the arrival time of the causal contribution of the water hazard, the predicted water inflow, and the early warning results corresponding to the water hazard status are output.

10. A mine water hazard early warning system based on big data analysis, comprising executing the mine water hazard early warning method based on big data analysis as described in any one of claims 1 to 9, characterized in that, Includes the following modules: The data acquisition module is used to collect and preprocess multi-source correlation data of mine water hazards and generate a mine water hazard correlation state sequence. The causal skeleton construction module is used to extract state change features and execute the LiNGAM causal discovery algorithm to construct a dual-root cause causal skeleton. The water hazard activation prediction module is used to extract the water source activation trajectory, channel activation trajectory, water receiving activation trajectory and hierarchical activation amount by improving the N-HiTS model, and form the water receiving state prediction trajectory. The causal contribution calculation module is used to pass hierarchical activation along the double root cause causal framework to obtain the causal contribution of water damage and the causal contribution of operation. The contribution competition matching module is used to compete and match the causal contribution of water damage, the causal contribution of operation, and the hierarchical activation amount to form a closed causal contribution path; The disaster chain identification module is used to combine the time sequence of activation trajectories and hydrogeological connections to screen closed paths of causal contribution, form active water hazard disaster chains and determine the time lag of anomaly transmission. The early warning output module is used to deduce the arrival time of the causal contribution of water damage, determine the state of water damage, and output early warning results.

Images