A mine environment risk multi-modal analysis and early warning decision method

Abstract

The present application relates to the technical field of mine environment risk, and discloses a mine environment risk multi-modal analysis and early warning decision method, comprising: constructing a unified space-time data set by time and space benchmark alignment of multi-source observation data, and establishing a troposphere disturbance model to realize observation disturbance coupling. Based on the seepage mechanics relationship, a physical evolution operator and an external input operator driven by rainfall are constructed to generate a state evolution mechanism. By mapping semantic query information to the target attention dimension, a semantic direction parameter is generated, and the end time of the rainstorm is set as the starting point to construct a lag propagation operator. Through the non-regularity measurement and dynamic change rate of each lag propagation operator, the non-regular spectrum transition index and peak lag time are determined, and the reference threshold is set based on the spectrum transition index of the historical non-risk period. Finally, when the warning intensity meets the condition, the warning is output, and the query result is generated combining the lag propagation operator and the semantic direction parameter.

Description

Technical Field

[0001] This invention relates to the field of mine environmental risk technology, and more specifically, to a multimodal analysis and early warning decision-making method for mine environmental risks. Background Technology

[0002] Tailings dams are among the most risky man-made structures in mining production, their safety status influenced by factors such as rainfall, seepage, dam structure, and groundwater level changes. With the development of remote sensing and IoT technologies, mine monitoring systems can now simultaneously collect multi-source information, including radar interferometric images, rainfall data, water level, and displacement. This data, after spatiotemporal registration, forms a high-dimensional monitoring matrix, providing rich evidence for predicting potential instability. However, existing monitoring systems primarily rely on independent numerical threshold judgments or statistical feature identification from single data channels. When faced with nonlinear, time-delayed rainstorm-induced processes, they often exhibit problems such as delayed monitoring response and scattered judgment results. While traditional early warning algorithms can capture deformation acceleration, they struggle to correspond to the actual rainfall triggering stage and dynamic changes in groundwater.

[0003] Extreme rainfall events are characterized by their sudden onset, while the seepage diffusion and pore pressure response within tailings bodies exhibit significant time lags. The asynchronicity of monitoring data leads to asynchronous peak values ​​and phase differences in the same natural event recorded by different sensors. Furthermore, InSAR data is affected by tropospheric delays and orbital sampling intervals, making it prone to misjudgments during critical periods. When processing this data, large-scale question-and-answer models must map natural language questions to specific physical variables and time periods. However, due to time lags, sampling differences, and atmospheric noise, the indexes generated by the models often deviate from the actual physical response range. This deviation directly results in unstable risk response outcomes, preventing the system from reliably outputting a quantitative description of post-rainstorm risk changes.

[0004] Fundamentally, the problem stems from the non-simultaneous natural processes within the high-dimensional monitoring matrix. Rainfall input gradually affects pore pressure and stress distribution through infiltration and diffusion, which then propagates to surface deformation signals. The response rate differs at each physical level, forming a complex time-delay network. When a question-answering model performs natural language indexing within this high-dimensional matrix, without modeling the time-delay diffusion and anisotropic propagation patterns, it cannot determine the correct time window corresponding to the current semantic question. Furthermore, the inherent meteorological interference and non-equidistant sampling in remote sensing amplify this bias, causing the model to misjudge data sources during critical periods. Ultimately, while the early warning system can capture a large amount of monitoring data, it cannot immediately identify the escalating risk phase after rainfall, thus losing the timeliness of early intervention. Summary of the Invention

[0005] This invention provides a multimodal analysis and early warning decision-making method for mine environmental risks, which solves the technical problems mentioned in the background art.

[0006] This invention provides a multimodal analysis and early warning decision-making method for mine environmental risks, including:

[0007] Align the multi-source observation data with temporal and spatial references to construct an observation dataset with a unified spatiotemporal reference, and establish a tropospheric perturbation model to obtain the observation perturbation coupling;

[0008] A physical evolution operator is constructed based on seepage mechanics, and an external input operator is established to characterize the driving effect of rainfall. A state evolution mechanism is generated through the physical evolution operator and the external input operator.

[0009] Semantic query information is mapped to the target focus dimension to generate semantic direction parameters;

[0010] Starting from the end of the rainstorm, a lag time range is set. For each lag time within the lag time range, a lag propagation operator is generated through the sequential action of the physical evolution operator, the external input operator, the semantic direction parameter, and the coupling of the observation perturbation.

[0011] Based on the calculation of the non-regularity measure and its dynamic rate of change of each hysteresis propagation operator, the non-regular spectral transition index, peak hysteresis time and spectral hot zone time window are determined.

[0012] A baseline threshold is determined based on the non-regular spectral transition index during historical non-risk periods, and the current non-regular spectral transition index is mapped to the warning intensity.

[0013] When the warning intensity meets the preset triggering conditions, the warning is output, and the query results are generated only within the spectral hot zone time window by combining the hysteresis propagation operator and semantic direction parameters.

[0014] The beneficial effects of this invention include: by aligning spatiotemporal benchmarks and establishing perturbation models based on multi-source data, combining seepage mechanics and rainfall-driven physical evolution analysis, and employing intelligent applications of semantic query mapping and hysteresis propagation operators, multimodal analysis and accurate early warning of mine environmental risks are achieved. This method can generate real-time early warnings based on historical data and dynamic changes, improving the accuracy and response efficiency of mine risk management and effectively reducing the risk of environmental disasters. Attached Figure Description

[0015] Figure 1 This is the first flowchart of the present invention;

[0016] Figure 2 This is the second flowchart of the present invention. Detailed Implementation

[0017] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0018] like Figures 1 to 2 As shown, a multimodal analysis and early warning decision-making method for mine environmental risks includes:

[0019] Align the multi-source observation data with temporal and spatial references to construct an observation dataset with a unified spatiotemporal reference, and establish a tropospheric perturbation model to obtain the observation perturbation coupling;

[0020] A physical evolution operator is constructed based on seepage mechanics, and an external input operator is established to characterize the driving effect of rainfall. A state evolution mechanism is generated through the physical evolution operator and the external input operator.

[0021] Semantic query information is mapped to the target focus dimension to generate semantic direction parameters;

[0022] Starting from the end of the rainstorm, a lag time range is set. For each lag time within the lag time range, a lag propagation operator is generated through the sequential action of the physical evolution operator, the external input operator, the semantic direction parameter, and the coupling of the observation perturbation.

[0023] Based on the calculation of the non-regularity measure and its dynamic rate of change of each hysteresis propagation operator, the non-regular spectral transition index, peak hysteresis time and spectral hot zone time window are determined.

[0024] A baseline threshold is determined based on the non-regular spectral transition index during historical non-risk periods, and the current non-regular spectral transition index is mapped to the warning intensity.

[0025] When the warning intensity meets the preset triggering conditions, the warning is output, and the query results are generated only within the spectral hot zone time window by combining the hysteresis propagation operator and semantic direction parameters.

[0026] In one embodiment of the present invention, multi-source observation data are aligned with temporal and spatial references to construct an observation dataset with a unified spatiotemporal reference, and a tropospheric perturbation model is established to obtain observation perturbation coupling, including:

[0027] Set a unified timeline as ;in, For the unified first A point in time;

[0028] Set a unified spatial grid as ;in, For the unified first A spatial point;

[0029] The multi-source observation data includes rainfall data, pore pressure data, water level data, displacement data, and... Time series data, the original sequence of each data point is denoted as ;in, Indicates the data source type. For data source The An original time point, For data source At the original time point Observed values;

[0030] Time alignment of the original sequences yields the following results: ;in, For data source At the same time point Alignment value, For data source At the original time point With a unified time point Time resampling weights between;

[0031] Spatial alignment is performed based on the temporal alignment results to obtain ;in, For data source At a unified spatial point and unified time point The observed values, For data source The original spatial location, For data source In the original spatial location With unified spatial point Spatial interpolation weights between them;

[0032] Establish a tropospheric disturbance model :

[0033] ;in, To unify spatial points and unified time point The amount of tropospheric disturbance, To unify the time point The time-dependent average disturbance term, To unify the time point The altitude linearity coefficient, To unify spatial points altitude;

[0034] Calculate the coupling of observational disturbances :

[0035] ;in To observe the perturbation coupling, for Time series data at a unified spatial point and unified time point The observed values.

[0036] Because the original time points (e.g., rainfall data recorded hourly, InSAR data recorded every 6 / 12 days) and spatial monitoring locations (e.g., pore pressure sensors at different depths within the dam, InSAR monitoring points being grid points covered by satellite radar) of multi-source data such as rainfall, pore pressure, water level, displacement, and InSAR are not consistent, direct correlation analysis is not possible. Therefore, a unified spatiotemporal coordinate system is first established:

[0037] Unified Timeline: Set a timeline containing multiple time points. Each point on the axis is a unified time point. All data sources must subsequently map their observations to the time points on this axis.

[0038] Unified spatial grid: Define a grid containing multiple spatial points. Each point in the grid is a unified spatial point. All data sources must subsequently map their observations to the spatial points of this grid.

[0039] Data types include rainfall data that records rainfall intensity, pore pressure data that records pore water pressure in soil and rock, water level data that records groundwater / tailings dam water level, displacement data that records dam / accumulation deformation, and InSAR time-series data of surface deformation acquired through satellite radar.

[0040] Original sequence: Each data source has its own original observation record sequence, which contains the observation values ​​corresponding to the original time point (i.e. the time point when the data source itself records the observation values). This record form is the original sequence.

[0041] The original observations from different time points across various data sources are uniformly adjusted to a unified timeline to ensure synchronization across the time dimension.

[0042] Weighting: For each original time point from each data source, a time resampling weight is assigned. This weight reflects the contribution of the observation at this original time point to the observation at the unified time point (e.g., the closer the original time point is to the unified time point, the greater the weight).

[0043] Alignment calculation: By weighted summation (multiplying the observation value of each original time point by the corresponding weight and then summing them), the observation value of each data source at the same time point is obtained, thereby achieving time synchronization of all data sources.

[0044] The original observations from different time points across various data sources are uniformly adjusted to a unified timeline to ensure synchronization across the time dimension.

[0045] Weighting: For each original time point from each data source, a time resampling weight is assigned. This weight reflects the contribution of the observation at this original time point to the observation at the unified time point (e.g., the closer the original time point is to the unified time point, the greater the weight).

[0046] Alignment calculation: By weighted summation (multiplying the observation value of each original time point by the corresponding weight and then summing them), the observation value of each data source at the same time point is obtained, thereby achieving time synchronization of all data sources.

[0047] Based on time alignment, observations from different spatial locations across various data sources are uniformly adjusted to spatial points on a unified spatial grid to ensure spatial dimensionality matching.

[0048] Weighting: For each original spatial location from each data source, a spatial interpolation weight is assigned. This weight reflects the contribution of the observation at this original spatial location to the observation at the unified spatial point (e.g., the closer the original spatial location is to the unified spatial point, the greater the weight).

[0049] Alignment calculation: By weighted summation (multiplying the observation value of each original spatial location by the corresponding weight and then summing them), the observation values ​​of each data source at a unified spatial point and the corresponding unified time point are obtained, realizing a complete match of all data sources in the spatiotemporal dimension.

[0050] Because InSAR data is acquired via satellite radar signals, the signals are affected by humidity and temperature changes as they propagate through the troposphere, resulting in delays that cause tropospheric interference errors in the observations. A model needs to be built to calculate these errors.

[0051] Model composition: The tropospheric disturbance at a unified spatial point and a unified time point is determined by two parts:

[0052] 1. Time-dependent average disturbance term at a unified time point: Reflects the average tropospheric disturbance shared by different spatial points at a unified time point (such as the disturbance caused by the atmospheric state of the entire region at the same time).

[0053] II. The product of the altitude linearity coefficient at a unified time point and the altitude at a unified spatial point. Since the higher the altitude, the thinner the troposphere and the smaller the interference, the altitude linearity coefficient is used to quantify the degree of altitude's influence on interference. Then, combined with the altitude of the specific spatial point, the altitude-related interference is calculated.

[0054] This model can be used to calculate the specific values ​​of tropospheric interference to InSAR data at each unified spatiotemporal point (unified spatial point and unified time point).

[0055] The purpose of observational perturbation coupling is to eliminate tropospheric interference in InSAR data and obtain observational values ​​that reflect the true surface deformation. This is achieved by subtracting the amount of tropospheric interference calculated using a tropospheric perturbation model from the original observational values ​​at a unified spatial and temporal point in InSAR time-series data. This result, having eliminated tropospheric interference, accurately reflects surface deformation.

[0056] In one embodiment of the present invention, a physical evolution operator is constructed based on seepage mechanics, and an external input operator is established to characterize the driving effect of rainfall. A state evolution mechanism is generated through the physical evolution operator and the external input operator, including:

[0057] The state variable vector is determined as follows , To unify spatial grids and unified timeline Below, a discrete state vector composed of pore pressure, water level, and displacement;

[0058] The hydraulic conductivity tensor matrix is ​​determined as follows: , For anisotropic hydraulic conductivity parameters in a unified spatial grid Discrete matrix on;

[0059] The discrete gradient matrix is ​​determined as follows: , The spatial discrete operator matrix from the scalar field to the gradient field;

[0060] The specific water storage coefficient matrix is ​​determined as follows: , The coefficient of water content change caused by unit pressure change in a unified spatial grid. Discrete matrix on;

[0061] The gravity term matrix is ​​determined as follows , For the directional flux term caused by gravity in a unified spatial grid Discrete matrix on;

[0062] Computational physical evolution operators ;in, Discrete gradient matrix The transpose of the matrix, For the water storage coefficient matrix The inverse matrix;

[0063] Determine the external input operator as , To incorporate rainfall intensity into the state variable vector The action matrix corresponds to the unified spatial grid. ;

[0064] The rainfall intensity was determined to be , To unify the timeline Rainfall sequence above;

[0065] The time interval is determined as , To unify the timeline The difference between two adjacent time points;

[0066] The time advancement matrix is ​​calculated as follows: ;in, Used in time interval The internal evolution of the internal propulsion state;

[0067] The time-progression kernel matrix is ​​calculated as follows: ;in, For integration variables, Used for aggregation time interval The cumulative impact of internal and external drivers on the state;

[0068] Generative state evolution mechanism ;in, For time points The state variable vector.

[0069] State variable vector: This is a discrete vector under a unified spatial grid and a unified time axis, containing pore pressure (reflecting pore water pressure, affecting the effective stress of the soil), water level (reflecting the groundwater / tailings dam water level), and displacement (reflecting the deformation of the deposit). These three quantities directly determine the seepage stability and mechanical safety of the mine deposit, and are the core indicators for characterizing the physical state.

[0070] Hydraulic conductivity tensor matrix: Discretizes the anisotropic hydraulic conductivity of the mine deposit (due to bedding, grain arrangement, etc., the seepage capacity varies in different directions) onto a unified spatial grid, quantifies the seepage capacity at different spatial points and in different directions, and is the key to calculating the seepage rate.

[0071] Discrete gradient matrix: Its function is to transform a scalar field (such as the spatial distribution of pore pressure) into a gradient field (the spatial rate of change of pore pressure). Since the seepage direction is determined by the pore pressure gradient (seepage from high pore pressure to low pore pressure), this matrix provides the basis for calculating the seepage direction and rate.

[0072] Specific water storage coefficient matrix: Discretizes the coefficient of water content change caused by unit pressure change to a spatial grid, quantifying the water-holding capacity of pores at different spatial points. The larger the coefficient, the more obvious the water content change under unit pore pressure change, which directly affects the rate of pore pressure change with seepage.

[0073] Gravity term matrix: Discretizes the directional seepage flux generated by gravity (such as gravity causing water to seep downwards) onto a spatial grid, reflecting the directional influence of gravity on the seepage direction and rate, which is consistent with the law that gravity is an important driving force for seepage in seepage mechanics.

[0074] The physical evolution operator calculates the spatial distribution and rate of seepage (reflecting the influence of seepage on the state) using the hydraulic conductivity tensor matrix, discrete gradient matrix, and gravity term matrix. Then, by combining the inverse of the specific storage coefficient matrix, it transforms the seepage effect into the rate of change of physical states (pore pressure, water level) over time. The physical evolution operator quantifies the natural evolution of physical states (e.g., pore pressure naturally dissipates due to seepage, water level changes naturally) when there is no external driving force (e.g., no rainfall).

[0075] External input operators: These act as a bridge connecting rainfall and state variables, taking the form of a matrix and corresponding to a unified spatial grid. Their function is to transform rainfall intensity (external driver) into an impact on the physical state (e.g., rainfall infiltration increases the water level at a spatial point or raises pore pressure), clarifying how rainfall affects the physical state of the mine deposit.

[0076] The generation of the state evolution mechanism requires combining the natural evolution of the physical state with the external driving force of rainfall to quantify the physical state at the next moment. The steps are as follows:

[0077] Determine the time-related parameters: rainfall intensity is the areal rainfall sequence on a unified time axis (to ensure time synchronization); time interval is the difference between adjacent time points on the unified time axis (as the time step of state evolution).

[0078] The computational time progression matrix quantifies the internal evolution of the physical state within a time interval when there is no rainfall, based on physical evolution operators and time intervals. For example, this matrix can be used to calculate how the pore pressure and water level at the current moment will naturally change to the next moment, solely due to seepage.

[0079] Calculate the time-progression kernel matrix: Quantify the cumulative impact of rainfall (an external driver) on the physical state within a time interval by integrating (accumulating the effects over the time interval). Since the impact of rainfall on the state is continuous, the total contribution over this period needs to be aggregated through integration.

[0080] The generation state evolution mechanism: The physical state at the next moment is equal to the natural state evolved from the current state by the time-progression matrix, plus the external driving state resulting from the combined effects of the time-progression kernel matrix, external input operators, and rainfall intensity. This fully describes the dynamic change of the physical state of the mine accumulation body over time under the influence of natural seepage evolution and rainfall.

[0081] In one embodiment of the present invention, mapping semantic query information to a target attention dimension to generate semantic direction parameters includes:

[0082] Determine the semantic query information as , For text queries related to mine environmental risks entered by the user;

[0083] The semantic encoder operator of the large question-answering model is determined as follows: , This refers to an algorithmic operator used to transform semantic query information in text form into a high-dimensional vector representation;

[0084] Computing semantic representation vector prototypes ;in, For semantic query information via encoder operator The high-dimensional vector obtained after processing;

[0085] semantic representation vector prototype After normalization, we obtain the normalized semantic representation vector. :

[0086] ,in, semantic representation vector prototype The Euclidean model is long;

[0087] The target focus selection matrix is ​​determined as follows: , This is a Boolean matrix used to filter dimensions related to mining environment data sources. 1 corresponds to the target focus dimension, and 0 corresponds to the non-target focus dimension. Each dimension corresponds to a data source.

[0088] Calculate semantic orientation parameters ;in, This is the vector mapped to the target's focus dimension.

[0089] Semantic query information refers to text content entered by the user that is related to mine environmental risks, such as changes in pore pressure on the west side of the tailings dam after a rainstorm and areas of excessive displacement in the mining area over the past week. This is the raw input for the entire mapping process.

[0090] The semantic encoder operator in the question-answering large model is a specialized algorithmic tool that transforms semantic query information in text form into word vector representations. Because text cannot be directly associated with mine observation data (such as pore pressure and displacement matrices), vectorization is needed to establish a mathematical connection between semantics and data.

[0091] The semantic encoder operator processes the user-input text query, and the resulting high-dimensional vector is the semantic representation vector prototype. This semantic representation vector prototype carries the semantic information of the text query, but the vector length may vary depending on the text length and the way it is expressed.

[0092] The Euclidean magnitude of the prototype semantic representation vector (i.e., the length of the vector) is calculated, and then the prototype vector is divided by this magnitude to obtain the normalized semantic representation vector. This eliminates the influence of vector length and retains only the directional information of the semantics (i.e., the dimension that the user focuses on).

[0093] The target focus selection matrix is ​​a Boolean type (containing only 0 and 1), and each dimension corresponds to a mining environment data source (such as pore pressure data dimension, displacement data dimension, rainfall data dimension, etc.): where 1 represents that the dimension is relevant to the user's query requirements (and should be retained), and 0 represents that the dimension is irrelevant to the query requirements (and should be removed). For example, if a user queries pore pressure changes, the pore pressure dimension is marked with 1, and the displacement dimension is marked with 0, thus achieving dimensional filtering of the semantic vector.

[0094] Multiplying the normalized semantic representation vector by the target attention selection matrix yields the semantic direction parameter. This semantic direction parameter has had dimensions irrelevant to the query removed, retaining only the semantic directions relevant to the user's interest in the mining environment data.

[0095] In one embodiment of the present invention, a lag time range is set starting from the end time of the rainstorm. For each lag time within the lag time range, a lag propagation operator is generated through the sequential action of the physical evolution operator, the external input operator, the semantic direction parameter, and the coupling of the observation perturbation. This includes:

[0096] Define the end time of the rainstorm as , This refers to the actual time point at which the monitored rainstorm process ended;

[0097] Set the lag time range as ;in, The minimum lag time is set to 0. The maximum hysteresis time is determined based on the permeability depth and equivalent diffusion rate of the mine deposit.

[0098] Within the lag time range Within a fixed time interval Selecting the lag time ;

[0099] For each lag time Calculate the total amplitude of the observed disturbance. ;in, To unify the total number of nodes in the spatial grid, Lag time The corresponding actual time;

[0100] Define the regularization constant as , A preset constant that is greater than 0;

[0101] Calculate each component of the observation perturbation weight vector ;in, For spatial points In the lag time Weighting coefficients;

[0102] Construct the observation perturbation weight matrix based on the observation perturbation weight vector. :

[0103] ;in, Construct a function for a diagonal matrix. for a diagonal matrix;

[0104] Define the time progression matrix as ;

[0105] The hysteresis operator is generated by multiplying matrices sequentially. .

[0106] The end time of the rainstorm is defined as the actual time point at which the rainstorm process ceases, serving as the starting point for lag analysis. Since rainstorms are a trigger for mine risks, subsequent analysis of state changes at different times (lag times) after the rainstorm's end is necessary; therefore, a clear starting point is required to standardize the lag time calculation.

[0107] The lag time range includes both minimum and maximum lag times:

[0108] The minimum lag time is set to 0 (i.e., the moment the rainstorm just ended).

[0109] The maximum lag time needs to be determined by combining the infiltration depth of the mine deposit (the maximum depth into which rainwater can penetrate) and the equivalent diffusion rate (the speed at which seepage propagates within the deposit).

[0110] The minimum and maximum lag times determine the longest time that rainfall will affect physical states (pore pressure, displacement), ensuring that the lag range covers the entire cycle of rainfall impact from occurrence to end.

[0111] Within a defined lag time range, multiple discrete lag moments are selected at fixed time intervals, thereby discretizing the continuous lag time. This facilitates subsequent calculation of operators for each moment, enabling time-segmented analysis of the lag response. For example, a lag moment can be set every 6 hours, covering a range of 0-20 days.

[0112] For each lag time, first determine the actual time corresponding to that time (the end time of the rainstorm plus the lag time), then summarize the absolute value of the observation disturbance coupling of all nodes in the unified spatial grid (the observation disturbance coupling is the InSAR data after removing tropospheric interference, and its absolute value reflects the residual amount of observation interference at that spatial point), and obtain the total amplitude of the observation disturbance at that lag time. The total amplitude of the observation disturbance is used to quantify the overall observation interference level at that time.

[0113] The regularization constant is a preset value greater than 0. The regularization constant is used to avoid the denominator being 0 or too small when calculating the weights in the subsequent calculation, so as to ensure the numerical stability of the weight calculation.

[0114] For each spatial point in the unified spatial grid, the weight coefficient of that spatial point is obtained by dividing the absolute value of the observation perturbation coupling at the current lag time by the total amplitude of the observation perturbation plus a regularization constant. The weight coefficient is inversely proportional to the absolute value of the observation perturbation coupling: the smaller the perturbation (the more reliable the observation), the greater the weight of the spatial point.

[0115] By constructing a diagonal matrix, each component of the observation perturbation weight vector is used as a diagonal element, and the off-diagonal elements are set to 0, thus generating an observation perturbation weight matrix. The observation perturbation weight matrix can weight and filter the observation information of different spatial points according to the weights, highlighting the contribution of reliable observations.

[0116] First, a time-progression matrix is ​​defined (based on the physical evolution operator and the current lag time, quantifying the natural evolution of the physical state at that lag time). Then, the observation perturbation weight matrix, the time-progression matrix, the external input operator (rainfall-driven effect), and the semantic direction parameter (user-focused dimension) are sequentially multiplied by matrix multiplication to obtain the lag propagation operator. The lag propagation operator characterizes the correlation between the user-focused semantic dimension → the rainfall-driven physical state → the weighted and filtered reliable observations at the current lag time.

[0117] In one embodiment of the present invention, the nonregularity metric and its dynamic rate of change are calculated based on each hysteresis propagation operator to determine the nonregular spectral transition index, peak hysteresis time, and spectral hot zone time window, including:

[0118] For each lag propagation operator Calculate the irregularity measure ;in, For the delay propagation operator The conjugate transpose of the matrix. Let Frobenius norm be the matrix;

[0119] Calculate the dynamic rate of change of the nonregularity measure ;in, and Lag time The discrete lag times that are adjacent to each other;

[0120] Determine the non-canonical spectral transition index ;

[0121] Determine the peak lag time ;

[0122] Determining the time window of the spectral thermal region ;in, As a preset constant, , This is the critical value of the dynamic change rate of the spectral thermal region.

[0123] For the lag propagation operator at each lag time, first perform the conjugate transpose (transposing the operator matrix and then taking the conjugate of each element is a common operation in matrix analysis to characterize asymmetry); then calculate the Frobenius norm (the square root of the sum of the squares of all elements of the matrix, used to measure the size of the matrix) of the difference between the product of the operator and its conjugate transpose, and the Frobenius norm of the operator itself; finally, divide the former by the square of the latter to obtain the irregularity measure. The role of the irregularity measure is to quantify the asymmetry of the lag propagation operator. Due to the anisotropy of the mine deposits (non-uniform hydraulic conductivity) and the spatial differences in the coupling of observational perturbations, the operator exhibits asymmetric characteristics. The strength of this asymmetry is positively correlated with the degree of abrupt change in physical states (pore pressure, displacement); the larger the irregularity measure, the higher the probability of abrupt state changes.

[0124] For each lag time, select two adjacent discrete lag times (e.g., the current time is lag 3 days, and the adjacent times are 2.5 days and 3.5 days). First, calculate the logarithm of the irregularity measure for each of these three times (compressing the numerical range to highlight the trend). Then, divide the difference between the logarithm of the next time step and the logarithm of the previous time step by the time interval between the two adjacent times to obtain the dynamic rate of change. The dynamic rate of change measures how quickly the irregularity measure changes over time. That is, the larger the absolute value of the dynamic rate of change, the more drastic the change in irregularity (i.e., the trend of abrupt changes in physical state), which is a signal for identifying risk abrupt changes.

[0125] The dynamic change rate at all lag times is iterated, and the maximum absolute value is taken as the non-regular spectral transition index. The non-regular spectral transition index is a quantitative indicator of the intensity of sudden changes in mine risk after a rainstorm. That is, the larger the index, the stronger the sudden change trend of physical state (such as sudden increase in pore pressure, acceleration of displacement) and the higher the risk level.

[0126] The peak lag time is the time when the absolute value of the dynamic rate of change is at its maximum, corresponding to the non-canonical spectral transition index. This time is the critical point at which the mine risk is most likely to change abruptly after a rainstorm. For example, if the peak lag time is 4 days after the rainstorm, it means that the risk of abrupt change in the physical state of the accumulation body is highest at this time, and it needs to be closely monitored.

[0127] A preset constant between 0 and 1 (e.g., 0.7, used to filter periods of significant change) is set. The product of this preset constant and the non-regular spectral transition index is calculated as the critical value of the dynamic change rate. Then, all lag moments with an absolute value of dynamic change rate not less than this critical value are selected. The set of these moments constitutes the spectral hot zone time window. The purpose of the spectral hot zone time window is to delineate the critical period of mine risk after heavy rain. Within the spectral hot zone time window, non-regular changes are significant, and the physical state is in a high-risk fluctuation period. Subsequent semantic query results are only generated within this time window, ensuring focus on high-risk periods.

[0128] In one embodiment of the present invention, a baseline threshold is determined based on the non-regular spectral transition index during historical non-risk periods, and the current non-regular spectral transition index is mapped to an early warning intensity, including:

[0129] Define the set of historical non-risk periods as , It is a collection of historical periods in which no mining environmental risk events have occurred and meteorological and geological conditions have been stable;

[0130] For each historical non-risk period Calculate the corresponding non-regular spectral transition index to form a sample set of historical non-risk non-regular spectral transition indices. ;in, Historically non-risk period The corresponding non-canonical spectral transition index;

[0131] Define the quantile function as , For calculating the probability of the dataset A function that deals with digit values, takes a dataset as input, and outputs the probability corresponding to that dataset. quantiles;

[0132] Set the starting quantile parameter as Set the upper limit quantile parameter as ;in, and All are satisfied The preset probability values ​​are used to determine the initial critical benchmark and the upper limit critical benchmark for early warning, respectively;

[0133] Calculate the starting threshold and upper limit threshold :

[0134] ;

[0135] ;

[0136] in, The sample set of non-regular spectral transition indices during historical non-risk periods is in probability The quantile at that point is used as the critical value for a warning intensity of 0; The sample set of non-regular spectral transition indices during historical non-risk periods is in probability The quantile at that point is used as the critical value for an early warning intensity of 1;

[0137] The current non-regular spectral transition exponent is obtained by using a piecewise function. Mapped to warning intensity :

[0138] ;in, The value range is from 0 to 1.

[0139] The set of historical non-risk periods includes all historical periods in which no mining environmental risk events have occurred and where meteorological (e.g., no extreme rainfall) and geological (e.g., no fault activity) conditions are stable. These periods were chosen because their corresponding non-regular spectral transition indices reflect the background level under normal mining conditions and can serve as a benchmark for judging whether there is a risk at present.

[0140] For each historical non-risk period in the historical non-risk non-regular spectral transition index sample set, calculate the non-regular spectral transition index corresponding to the historical non-risk period in the aforementioned manner, and summarize all non-regular spectral transition indices to form a sample set.

[0141] Quantile functions are tools for calculating quantile values ​​in a dataset: the input is a dataset (such as the historical index sample set mentioned above) and a probability value, and the output is the quantile value for the corresponding probability in the dataset (i.e., samples with that probability are below this value, and samples with the remaining probabilities are above this value). For example, with an input probability of 0.95, the output quantile represents 95% of historical non-risk indices being below this value, which can be used to define the boundary between normal and abnormal.

[0142] Both parameters are preset probability values ​​between 0 and 1, and the initial quantile parameter is less than the upper quantile parameter:

[0143] Starting quantile parameter: used to determine the critical benchmark for a warning intensity of 0 (i.e., below the index corresponding to this quantile, it is considered a normal state);

[0144] Upper quantile parameter: used to determine the critical benchmark for a warning intensity of 1 (i.e., an index higher than the corresponding quantile is considered an extremely high-risk state).

[0145] The gradient range from normal to extremely high risk is formed by combining the initial quantile parameter and the upper quantile parameter.

[0146] Initial threshold: The quantile is calculated using the quantile function on the historical non-risk index sample set according to the initial quantile parameter. This value is the critical value for the warning intensity to be 0. That is, when the current index is lower than this value, it means that it is consistent with the historical normal state and there is no risk.

[0147] Upper limit threshold: The quantile is calculated using the quantile function on the historical non-risk index sample set according to the upper limit quantile parameter. This value is the critical value for the warning intensity of 1. That is, when the current index is higher than this value, it means that it is far beyond the historical normal state and the risk is extremely high.

[0148] Mapping the current non-regular spectral transition index to the warning intensity requires a piecewise function, with the result ranging from 0 to 1.

[0149] If the current index is less than or equal to the initial threshold: the warning intensity is 0, indicating no risk;

[0150] If the current index is greater than or equal to the upper limit threshold: the warning intensity is 1, representing extremely high risk;

[0151] If the current index is between the two: the warning intensity is calculated linearly as (current index - starting threshold) / (upper limit threshold - starting threshold), representing medium risk (the closer the value is to 1, the higher the risk).

[0152] In one embodiment of the present invention, an early warning is output when the warning intensity meets a preset triggering condition, and query results are generated only within a spectral hot zone time window by combining the hysteresis propagation operator and semantic direction parameters, including:

[0153] Increase the warning intensity Compared with the preset warning intensity threshold:

[0154] If the warning intensity If the risk level is greater than or equal to the preset warning intensity threshold, a mine environmental risk warning will be output; otherwise, no warning will be output.

[0155] Define the query result matrix as Only in the spectral thermal region time window Generate a query result matrix: ;in, For the time window of the spectral hot zone Any time delay within the range, Used to characterize the physical state response corresponding to the semantic query at this lag time;

[0156] Define the query result set as Spectral hot zone time window All query result matrices are summarized to form a query result set: , Used to output the final result corresponding to the semantic query information.

[0157] The preset warning intensity threshold is a risk trigger standard set in advance based on the mine's risk tolerance and historical accident data. For example, 0.8 means that when the warning intensity reaches or exceeds this value, the risk level is sufficient to require intervention.

[0158] Compare the currently calculated warning intensity with the preset threshold:

[0159] If the current warning intensity is greater than or equal to the preset threshold, it indicates that the mine environmental risk has reached a level that requires vigilance, and a mine environmental risk warning will be issued immediately (such as pushing alarm information to the control system and prompting manual verification); if the current warning intensity is less than the preset threshold, it indicates that the risk is within an acceptable normal range, and no warning will be issued.

[0160] The query result matrix is ​​generated only within the spectral hot zone time window. Because the spectral hot zone time window represents a period of significant non-regular changes and high-risk fluctuations in physical conditions, generating results only within this interval avoids outputting invalid information from non-risk periods, ensuring that the results are correlated with risk. The query result matrix corresponds to any lag time within the spectral hot zone time window. Its function is to characterize the mine's physical state response to the user's semantic query (e.g., the change in pore pressure in a tailings dam 3 days after a rainstorm) at that lag time. For example, if a user queries pore pressure, the matrix will reflect the pore pressure response data at various spatial points at that moment, achieving a match between semantic requirements and physical conditions.

[0161] The query result set is a collection of query result matrices summarizing all lag times within the spectral hot zone time window. The purpose of the query result set is to integrate all physical state response data relevant to the user's semantic query throughout the entire high-risk period. Compared to matrices at a single time point, the query result set can present the state change trend within the risk period (such as the process of pore pressure rising to a peak and then falling back), providing users with comprehensive and continuous query results.

[0162] The embodiments of this example have been described above. However, this example is not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of this example, and all of them are within the protection scope of this example.

Claims

1. A mine environment risk multi-modal analysis and early warning decision method, characterized in that, The method comprises the following steps: time and space reference alignment of multi-source observation data, construction of an observation data set with unified time and space reference, establishment of a tropospheric disturbance model, and obtaining of observation disturbance coupling; construction of a physical evolution operator based on seepage mechanics relationship, establishment of an external input operator for representing rainfall driving effect, and generation of state evolution mechanism through the physical evolution operator and the external input operator; mapping of semantic query information to target concerned dimensions to generate semantic direction parameters; setting of a lag time range from the end time of a rainstorm, generation of a lag propagation operator at each lag time in the lag time range through sequential action of the physical evolution operator, the external input operator, the semantic direction parameters, and the observation disturbance coupling; calculation of non-regularity measure and its dynamic change rate based on each lag propagation operator, determination of non-regular spectrum transition index, peak lag time, and spectrum hot area time window; determination of a reference threshold based on the non-regular spectrum transition index of a historical non-risk period, and mapping of the current non-regular spectrum transition index to early warning intensity; output of early warning when the early warning intensity meets a preset triggering condition, and generation of a query result within the spectrum hot area time window in combination with the lag propagation operator and the semantic direction parameters.

2. The mine environment risk multi-modal analysis and early warning decision method according to claim 1, characterized in that, time and space reference alignment of multi-source observation data, construction of an observation data set with unified time and space reference, and establishment of a tropospheric disturbance model to obtain observation disturbance coupling, comprising: Setting a uniform time axis as ; wherein, is the uniformed first time point; A uniform spatial grid is set as ; wherein, is the uniformed first spatial point; The multi-source observation data includes rainfall data, pore pressure data, water level data, displacement data, and... Time series data, the original sequence of each data point is denoted as ;in, Indicates the data source type. For data source The An original time point, For data source At the original time point Observed values; The original sequences are time-aligned to obtain ; wherein, is a data source at a uniform time point ; and is a data source at an original time point and a time resampling weight between the original time point and the uniform time point . Based on the time alignment result, spatial alignment is performed to obtain ; wherein, is a data source At a unified spatial point and a unified time point , an observation value, is an original spatial position of a data source , is a spatial interpolation weight between the original spatial position of the data source and the unified spatial point ; Establishing a model of tropospheric disturbances : ; where is the tropospheric disturbance quantity at a uniform spatial point and a uniform time point , is the time-dependent average disturbance term at a uniform time point , is the altitude linear coefficient at a uniform time point , is the altitude at a uniform spatial point ; Computing observed disturbance coupling : ; wherein is the observed disturbance coupling, is the time series data at a uniform spatial point and a uniform time point of observation.

3. The mine environment risk multi-modal analysis and early warning decision method according to claim 2, characterized in that, construction of a physical evolution operator based on seepage mechanics relationship, establishment of an external input operator for representing rainfall driving effect, and generation of state evolution mechanism through the physical evolution operator and the external input operator, comprising: The state variable vector is determined as , is a uniform spatial grid and a uniform time axis The discrete state vector is composed of pore pressure, water level and displacement. The hydraulic conductivity tensor matrix is determined as , is a discrete matrix of anisotropic hydraulic conductivity parameters on a uniform spatial grid ; The discrete gradient matrix is determined as , is a matrix of spatial discrete operators from the scalar field to the gradient field. The specific water storage coefficient matrix is ​​determined as follows: , The coefficient of water content change caused by unit pressure change in a unified spatial grid. Discrete matrix on; The gravity term matrix is determined as , is the discrete matrix of the directional flux term due to gravity on the unified spatial grid ; Computing a physical evolution operator ; wherein is a discrete gradient matrix is a transpose matrix of is a specific storage coefficient matrix is an inverse matrix of determining the external input operator as , is the action matrix for introducing the rainfall intensity into the state variable vector corresponding to the uniform spatial grid ; determining the rainfall intensity as , is a sequence of areal rainfall on a unified time axis ; determining a time interval as , for a unified timeline the difference between the two adjacent time points The time advancement matrix is computed as ; where for the internal evolution of the state within the time interval ; The time advancement kernel matrix is computed as ; where is the integration variable, is the cumulative effect of the internal and external drivers on the state over the aggregation time interval ; Generating state evolution mechanisms ; wherein, is a state variable vector at time point .

4. The mine environment risk multi-modal analysis and early warning decision method according to claim 3, characterized in that, mapping of semantic query information to target concerned dimensions to generate semantic direction parameters, comprising: The semantic query information is determined to be , a text query content input by the user and related to the mine environment risk; The determining question and answer large model semantic encoder operator is , is an algorithm operator for converting semantic query information in text form into a high-dimensional vector representation; Computing semantic representation vectors prototypes ; wherein, are semantic query information encoded by an encoder operator a high-dimensional vector resulting from the processing Prototype of semantic representation vector : Unitization processing is performed to obtain a unitized semantic representation vector : wherein, is a semantic representation vector prototype Euclidean norm of The target attention selection matrix is determined as , is a Boolean matrix for screening dimensions related to the mine environment data source, 1 corresponds to a target attention dimension, 0 corresponds to a non-target attention dimension, and each dimension corresponds to a data source; Computing semantic direction parameters ; wherein, is the vector mapped to the target dimension of interest.

5. The mine environment risk multi-modal analysis and early warning decision method according to claim 4, characterized in that, setting of a lag time range from the end time of a rainstorm, generation of a lag propagation operator at each lag time in the lag time range through sequential action of the physical evolution operator, the external input operator, the semantic direction parameters, and the observation disturbance coupling, comprising: The end of the storm is defined as , the time point at which the storm actually ends is monitored; The hysteresis time range is set as ; wherein, is the minimum hysteresis time, and is 0, is the maximum hysteresis time determined based on the penetration depth of the mine accumulation body and the equivalent diffusion rate; In the hysteresis time range fixed time intervals hysteresis time points ; For each lag time Calculate the total amplitude of the observed disturbance. ;in, To unify the total number of nodes in the spatial grid, Lag time The corresponding actual time; The regularization constant is defined as , is a preset constant greater than 0; computing each component of the observed disturbance weight vector ; wherein for a spatial point at a lag time weighting coefficient Constructing an observation disturbance weight matrix based on an observation disturbance weight vector : ; wherein is a diagonal matrix constructor, is a diagonal matrix; The time advancing matrix is defined as ; Generating a lag propagator by sequentially multiplying matrices .

6. The mine environment risk multi-modal analysis and early warning decision method according to claim 5, characterized in that, calculation of non-regularity measure and its dynamic change rate based on each lag propagation operator, determination of non-regular spectrum transition index, peak lag time, and spectrum hot area time window, comprising: for each lagged propagator operator computing the non-regularity measure ; wherein, is the conjugate transpose matrix of the lagged propagator operator , is the Frobenius norm of the matrix Calculate the dynamic rate of change of the nonregularity measure ;in, and Lag time The discrete lag times that are adjacent to each other; Determining non-regular spectral transition index ; Determining peak lag time ; Determining a time window of a spectral hot area ; wherein, is a preset constant, , is a critical value of a dynamic change rate of the spectral hot area.

7. The mine environment risk multi-modal analysis and early warning decision method according to claim 6, characterized in that, determination of a reference threshold based on the non-regular spectrum transition index of a historical non-risk period, and mapping of the current non-regular spectrum transition index to early warning intensity, comprising: The set of historical non-risk periods is defined as , is a set of historical periods in which no mine environmental risk events occurred and the meteorological and geological conditions were stable; for each historical non-risk period , a corresponding irregular spectral transition index is calculated, forming a historical non-risk irregular spectral transition index sample set ; wherein, is the corresponding irregular spectral transition index of the historical non-risk period ​ Define quantile function as , is a function for computing quantile value of a dataset at probability , input is a dataset, output is the quantile of the dataset corresponding to probability ; The starting quantile parameter is set as The upper limit quantile parameter is set as ; wherein, and are preset probability values satisfying , and are respectively used for determining the starting critical reference and the upper limit critical reference of the early warning. calculating a start threshold and an upper threshold : ； ； wherein, is the quantile of the sample set of non-regular spectral transition indices at the probability place, as a threshold value for an early warning strength of 0; is the quantile of the sample set of non-regular spectral transition indices at the probability place, as a threshold value for an early warning strength of 1; The current non-routine spectral transition index is mapped to an early warning intensity by a piecewise function :​ ; wherein, the value of a ranges from 0 to 1.

8. The mine environment risk multi-modal analysis and early warning decision method according to claim 7, characterized in that, output of early warning when the early warning intensity meets a preset triggering condition, and generation of a query result within the spectrum hot area time window in combination with the lag propagation operator and the semantic direction parameters, comprising: comparing the alert intensity with a preset alert intensity threshold value: If the early warning intensity is greater than or equal to a preset early warning intensity threshold, a mine environment risk early warning is output; otherwise, no early warning is output. Define the query result matrix as Only in the spectral thermal region time window Generate a query result matrix: ;in, For the time window of the spectral hot zone Any time delay within, Used to characterize the physical state response corresponding to the semantic query at this lag time; The query result set is defined as All query result matrices in the spectrum hot area time window are summarized to form a query result set: , for outputting the final result corresponding to the semantic query information.

Images