An intelligent monitoring and analysis system for deep stress of geological disasters based on AI big data

Abstract

The present application relates to the field of geological disaster monitoring, and discloses an intelligent monitoring and analysis system for deep stress of geological disasters based on AI big data, comprising: obtaining a multi-source long-term big data set of a deep geological disaster monitoring area; calculating a viscoelastic constitutive parameter matrix based on core rheological test parameters to generate a rheological stress evolution baseline sequence of each monitoring point; subtracting the rheological baseline from the deep stress time series data and performing STL decomposition and system drift correction to generate a net anomaly trend sequence; extracting multi-scale trend change features and rheological deviation degree features to generate a composite long-period feature vector; inputting the feature vector into a hierarchical time series deep learning model to output long-period precursor intelligent recognition results; and generating deep stress intelligent monitoring and analysis results in association with historical geological disaster event records. The present application effectively improves the accuracy of precursor recognition of geological disasters in deep soft rock areas by explicitly separating rheological background effects and using a multi-time scale hierarchical deep learning model.

Description

Technical Field

[0001] This invention relates to the field of geological disaster monitoring technology, and more specifically, to an intelligent monitoring and analysis system for deep stress in geological disasters based on AI big data. Background Technology

[0002] In the field of deep geological hazard monitoring, multi-source sensors such as distributed optical fibers, borehole stress gauges, and microseismic monitoring systems continuously collect deep stress data over long periods, forming large-scale monitoring big data covering multiple years. Existing technologies use AI models to extract features and recognize patterns from time-series deep stress data through sliding windows. Stress anomaly patterns are learned within hourly to dayly windows, and precursor evolution states are inferred using time-series state models such as Hidden Markov Models, thereby enabling intelligent analysis and early warning of geological hazard risks.

[0003] However, in deep soft rock areas (such as rheological rock masses like mudstone and shale), the rock mass undergoes creep under sustained high geostress. The long-term stress redistribution trend generated by rheological processes highly overlaps with the long-term abnormal evolution caused by precursors of geological disasters on a timescale, both exhibiting slow, monotonous changes on a monthly to yearly scale. Existing AI big data analysis methods input all deep stress monitoring data into the model for processing without separating the rheological background effect. This leads to the AI ​​model misinterpreting rheological trends as precursor signals of disasters, resulting in systematic false alarms, or the true precursor signals being masked by the rheological baseline and thus missed. Furthermore, existing sliding windows are designed as short time windows, unable to cover the complete cycle of precursor evolution on a monthly scale in low-permeability rock masses. The stress change amplitude within the short window is close to the monitoring noise level, causing a significant decrease in the accuracy of intelligent monitoring and analysis of deep stress in deep soft rock areas. Summary of the Invention

[0004] This invention provides an intelligent monitoring and analysis system for deep stress in geological disasters based on AI big data, which solves the technical problems in related technologies, such as the difficulty in distinguishing between normal rheological trends and disaster precursor anomalies in monitoring data in deep soft rock areas due to rock mass rheological effects, the interference of system drift with long-term monitoring accuracy, and the lack of intelligent recognition capabilities for multi-scale time-series features.

[0005] This invention discloses an intelligent monitoring and analysis method for deep stress in geological disasters based on AI big data, including: acquiring a multi-source long-term big data set of the monitoring area of ​​deep geological disasters, wherein the multi-source long-term big data set includes multi-year continuous deep stress time series data, distributed optical fiber strain time series data and its historical cumulative records, core rheological test parameters, monitoring system calibration records, long-term environmental temperature change data and historical geological disaster event records; Based on the core rheological test parameters, the viscoelastic constitutive parameter matrix of different lithological regions is calculated, and the historical strain accumulation data of each monitoring point is integrated over time according to the viscoelastic constitutive model to generate the rheological stress evolution baseline sequence of each monitoring point. The rheological residual time series is obtained by subtracting the rheological stress evolution baseline sequence of the corresponding monitoring point from the multi-year continuous deep stress time series data. The rheological residual time series is decomposed into three components: trend, seasonality and residual. System drift correction is performed by combining the calibration record of the monitoring system to generate a net anomaly trend sequence. Multi-scale wavelet transform is performed on the net anomaly trend sequence to extract multi-scale trend change features. The deviation rate between the rheological stress evolution baseline sequence and the measured deep stress value is calculated and a rheological deviation feature vector is generated. The multi-scale trend change features and the rheological deviation feature vector are concatenated into a composite long-period feature vector. The composite long-period feature vector is input into a hierarchical time-series deep learning model, which outputs the intelligent recognition result of long-period precursors and the current evolution stage location. Based on the intelligent identification results of long-cycle precursors and the current evolution stage location, the statistical characteristics of the evolution cycle of similar precursor patterns in the historical geological disaster event records are associated to generate intelligent monitoring and analysis results of deep stress.

[0006] Furthermore, the calculation of the viscoelastic constitutive parameter matrix for different lithological regions includes: based on the creep curve data under different confining pressures and temperatures in the core rheological test parameters, the creep equation of the Burgers body model is fitted using the least squares method to obtain the parameter set corresponding to each confining pressure-temperature condition. The parameter set includes instantaneous elastic modulus, delayed elastic modulus, long-term viscosity coefficient, and delayed viscosity coefficient, forming the viscoelastic constitutive parameter matrix; the viscoelastic constitutive parameter matrix is ​​constructed with confining pressure and temperature as index dimensions, with row indices corresponding to discrete confining pressure value sequences and column indices corresponding to discrete temperature value sequences, and each element corresponding to a set of four parameters under specific confining pressure-temperature conditions; for each monitoring point, the estimated confining pressure and long-term environmental temperature variation data at the location of the monitoring point are obtained, and the constitutive parameter set corresponding to the monitoring point is obtained from the viscoelastic constitutive parameter matrix using a linear interpolation algorithm; when the estimated confining pressure or long-term environmental temperature variation data at the location of the monitoring point changes, the updated constitutive parameter set is obtained again by interpolation from the viscoelastic constitutive parameter matrix.

[0007] Furthermore, the step of integrating the historical strain accumulation data of each monitoring point over time according to the viscoelastic constitutive model to generate the rheological stress evolution baseline sequence of each monitoring point includes: based on the Boltzmann superposition principle, transforming the stress-strain relationship of each monitoring point under complex loading history into the convolution integral form of the relaxation modulus function and the strain rate history, wherein the relaxation modulus function is determined by the constitutive parameter set corresponding to the Burgers volume model; discretizing the convolution integral of the historical strain accumulation data of each monitoring point into a recursive calculation form that is gradually accumulated according to the sampling interval; multiplying the relaxation modulus function value with the strain rate obtained by the difference between the strain values ​​at adjacent times at each sampling time and then accumulating the results to generate the rheological stress evolution baseline sequence of each monitoring point. The rheological stress evolution baseline sequence characterizes the long-term trend of deep stress generated by pure rheological action under conditions without disaster disturbance.

[0008] Furthermore, the step of performing trend-seasonal-residual three-component decomposition on the rheological residual time series and combining it with the monitoring system calibration records to perform system drift correction and generate a net abnormal trend sequence includes: decomposing the rheological residual time series using the STL decomposition algorithm, which iteratively performs seasonal smoothing and trend smoothing to separate the long-term trend component, the annual cycle seasonal component, and the random residual component; obtaining the sensor reference drift at each calibration time node in the monitoring system calibration records, which is obtained by comparing the deviation between the sensor output value at the calibration time and the standard reference value, and obtaining a continuous drift estimation sequence using piecewise linear interpolation; subtracting the continuous drift estimation sequence from the long-term trend component to generate a net abnormal trend sequence; and for segments where the interval between adjacent calibration nodes exceeds a preset time threshold, using the drift correction of adjacent sensors within the same monitoring section for weighted auxiliary estimation, with the weight inversely proportional to the distance between sensors.

[0009] Furthermore, the step of extracting multi-scale trend change features from the net anomaly trend sequence using multi-scale wavelet transform includes: performing a multi-level discrete wavelet transform on the net anomaly trend sequence to obtain wavelet approximation coefficient sequences and wavelet detail coefficient sequences corresponding to three decomposition levels: monthly, quarterly, and annual scales. The multi-level discrete wavelet transform uses the Daubechies wavelet basis function. For the wavelet approximation coefficient sequences at each scale, the standard deviation within the sliding window is calculated after first-order differencing to obtain the trend slope change rate, and curvature features are obtained through second-order differencing. Second-order differences are then used to detect the curve characteristics. The zero-crossing points of the difference sequence are used to obtain the inflection point positions. The above features at each scale are combined into multi-scale trend change features. The calculation of the deviation rate between the rheological stress evolution baseline sequence and the measured deep stress value and the generation of the rheological deviation feature vector include: calculating the ratio of the difference between the measured deep stress value and the rheological stress evolution baseline value to the rheological stress evolution baseline value as the deviation rate; calculating the mean and rate of change of the deviation rate in the time windows of the monthly, quarterly and annual scales respectively, and generating the rheological deviation feature vector. The rate of change is obtained by the linear regression slope of the deviation rate sequence within the window.

[0010] Furthermore, the hierarchical temporal deep learning model includes a bottom-level weekly encoder, a middle-level monthly aggregator, and a top-level quarterly classifier. The bottom-level weekly encoder adopts a gated recurrent unit structure, receives time segments divided by week from the composite long-period feature vector as input sequences, encodes each weekly time segment, and obtains a weekly-level hidden state representation sequence. The middle-level monthly aggregator receives the weekly-level hidden state representation sequence as input, groups the weekly-level hidden state representation sequences according to the number of weeks contained in each month, aggregates each group through a cross-level attention mechanism, and outputs a monthly-level aggregated representation sequence. The top-level quarterly classifier receives the monthly-level aggregated representation sequence as input, groups each group according to the number of months contained in each quarter, performs average pooling operation on each group along the time dimension to obtain a fused representation vector, and outputs the long-period precursor intelligent identification result and the current evolution stage location through a fully connected layer. The long-period precursor intelligent identification result is used to distinguish whether the current signal belongs to the system deviation of the rheological model or the abnormal evolution of the real precursor of geological disaster. The current evolution stage location includes four discrete stage identifiers: initial stage, development stage, acceleration stage, and critical stage.

[0011] Furthermore, the cross-level attention mechanism uses the weekly-scale hidden state group corresponding to the current month as the query and the weekly-scale hidden state groups corresponding to adjacent months as the key and value. It employs a scaled dot product attention algorithm to calculate attention weights and perform weighted aggregation. The calculation process of the scaled dot product attention algorithm is as follows: calculate the matrix product between the transpose of the query matrix and the key matrix, divide by the arithmetic square root of the key vector dimension as a scaling factor, apply the softmax function to the result to obtain the attention weight matrix, and multiply the attention weight matrix by the value matrix to obtain the weighted representation vector. A temporal smoothing layer is set before the fully connected layer of the top-level quarterly-scale classifier. This temporal smoothing layer performs one-dimensional convolutional filtering on the quarterly-scale fused representation sequence along the time dimension. The filtering window length covers the typical transition duration between two adjacent evolutionary stages. The convolution kernel parameters are adaptively learned through the training process of the hierarchical temporal deep learning model.

[0012] Furthermore, the training process of the hierarchical temporal deep learning model includes: based on the marked precursor segments and confirmed non-precursor segments in the historical geological disaster event records, forming training sample pairs with the composite long-cycle feature vectors of the corresponding time periods and the labeled labels; jointly training the two output branches, the long-cycle precursor intelligent recognition result and the current evolution stage location, using a classification cross-entropy loss function, where the classification cross-entropy loss function multiplies the loss terms of the two output branches by their corresponding weight coefficients and then sums them, with the weight coefficient corresponding to the long-cycle precursor intelligent recognition result being greater than the weight coefficient corresponding to the current evolution stage location; introducing a category weight factor into the classification cross-entropy loss function, where the loss weight corresponding to the positive precursor sample is determined according to the reciprocal of the ratio of positive to negative samples, and the loss weight corresponding to the negative sample is 1; and optimizing using the Adam optimization algorithm.

[0013] Furthermore, the generation of deep stress intelligent monitoring and analysis results includes: when the long-period precursor intelligent identification result determines that the current signal is a real precursor abnormal evolution of geological disaster, historical events with the same precursor classification label are retrieved from the historical geological disaster event records to obtain the average evolution cycle length of the same precursor pattern; based on the ratio of the difference between the current time and the start time of the precursor signal to the average evolution cycle length, the relative position ratio of the current precursor evolution stage in the complete evolution cycle is calculated; the geological disaster precursor discrimination level is divided according to the classification confidence level output by the hierarchical time-series deep learning model, and the risk warning level is divided according to the relative position ratio, the closer the relative position ratio is to 1, the higher the corresponding risk warning level; when multiple monitoring points in the same monitoring section generate long-period precursor intelligent identification results at the same time, the spatial consistency of the classification confidence level of each monitoring point is checked, and only when the geological disaster precursor discrimination level of spatially adjacent monitoring points reaches the preset spatial consistency condition is the corresponding risk warning level upgraded to the corresponding level.

[0014] This invention provides an intelligent monitoring and analysis system for deep stress in geological disasters based on AI big data, comprising: The multi-source big data acquisition module is used to acquire multi-source long-term big data sets of the deep geological disaster monitoring area. The multi-source long-term big data sets include multi-year continuous deep stress time series data, distributed optical fiber strain time series data and its historical cumulative records, core rheological test parameters, monitoring system calibration records, long-term environmental temperature change data and historical geological disaster event records. The rheological baseline generation module is used to calculate the viscoelastic constitutive parameter matrix of different lithological regions based on the rheological test parameters of the rock core, and to perform time integration on the historical strain accumulation data of each monitoring point according to the viscoelastic constitutive model to generate the rheological stress evolution baseline sequence of each monitoring point. The net anomaly trend generation module is used to obtain the rheological residual time series by subtracting the rheological stress evolution baseline sequence of the corresponding monitoring point from the multi-year continuous deep stress time series data point by point, decompose the rheological residual time series into trend-seasonal-residual three components and combine it with the monitoring system calibration record to perform system drift correction, and generate the net anomaly trend sequence. The composite feature extraction module is used to extract multi-scale trend change features by performing multi-scale wavelet transform on the net anomaly trend sequence, calculate the deviation rate between the rheological stress evolution baseline sequence and the measured deep stress value and generate a rheological deviation feature vector, and concatenate the multi-scale trend change features and the rheological deviation feature vector into a composite long-period feature vector. The precursor intelligent recognition module is used to input the composite long-period feature vector into the hierarchical time-series deep learning model and output the long-period precursor intelligent recognition result and the current evolution stage location. The monitoring and analysis result generation module is used to generate deep stress intelligent monitoring and analysis results based on the long-cycle precursor intelligent identification results and the current evolution stage location, and to associate the evolution cycle statistical characteristics of similar precursor patterns in the historical geological disaster event records.

[0015] This invention calculates the viscoelastic constitutive parameter matrix based on core rheological test parameters and generates a baseline sequence of rheological stress evolution using the Boltzmann superposition principle. This explicitly separates the rheological background trend before the data enters the AI ​​model for processing, eliminating a systematic source of misjudging rheological trends as precursor signals. By extracting the rheological deviation feature vector and concatenating it with multi-scale trend change features to form a composite long-period feature vector, it provides a quantitative criterion for distinguishing rheological model residuals from real precursor anomalies in a hierarchical time-series deep learning model. By employing a hierarchical time-series deep learning model consisting of a bottom-level weekly encoder, a middle-level monthly aggregator, and a top-level quarterly classifier, it covers multi-timescale information ranging from short-term fluctuations to slow evolution over several months. This overcomes the limitation of existing short-time-window sliding windows in capturing the complete cycle of precursor evolution over several months in low-permeability rock masses. It solves the technical problem of decreased identification accuracy caused by the high overlap of rheological background and geological disaster precursor signals in deep soft rock areas over long-period timescales, achieving the technical effect of improving the accuracy of disaster precursor identification in intelligent monitoring and analysis of deep stress in deep soft rock areas. Attached Figure Description

[0016] Figure 1 This is a flowchart of the intelligent monitoring and analysis method for deep stress in geological disasters based on AI big data provided in this embodiment of the invention; Figure 2 This is a schematic diagram comparing the baseline of the rheological stress at monitoring point M06 with the measured stress provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the STL decomposition results of the rheological residual at monitoring point M06 provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the net anomaly trend sequence (comparison before and after drift correction) of monitoring point M06 provided in an embodiment of the present invention; Figure 5 This is a schematic diagram illustrating the relationship between viscoelastic constitutive parameters and confining pressure provided in an embodiment of the present invention; Figure 6 This is a schematic diagram comparing the trend features of multi-scale wavelet decomposition provided in an embodiment of the present invention; Figure 7 This is a schematic diagram of the probability distribution for precursor identification and classification provided in an embodiment of the present invention; Figure 8 This is a schematic diagram of the probability distribution of evolutionary stages provided in an embodiment of the present invention. Detailed Implementation

[0017] In the field of deep geological hazard monitoring, multi-source sensors such as distributed optical fibers, borehole stress gauges, and microseismic monitoring systems continuously collect deep stress data over long periods, forming large-scale monitoring big data covering multiple years. Existing technologies use AI models to extract features and recognize patterns from time-series deep stress data through sliding windows. Stress anomaly patterns are learned within hourly to dayly windows, and precursor evolution states are inferred using time-series state models such as Hidden Markov Models, thereby enabling intelligent analysis and early warning of geological hazard risks.

[0018] However, in deep soft rock areas (such as rheological rock masses like mudstone and shale), the rock mass undergoes creep under sustained high geostress. The long-term stress redistribution trend generated by rheological processes highly overlaps with the long-term abnormal evolution caused by precursors of geological disasters on a timescale, both exhibiting slow, monotonous changes on a monthly to yearly scale. Existing AI big data analysis methods input all deep stress monitoring data into the model for processing without separating the rheological background effect, leading to the following problems: First, rheological trends are misjudged by the AI ​​model as precursor signals of disasters, resulting in systematic false alarms; second, true precursor signals are masked by the rheological baseline and thus missed. Furthermore, the existing sliding window is a short-time window design, unable to cover the complete cycle of precursor evolution on a monthly scale in low-permeability rock masses. The stress change amplitude within the short window is close to the monitoring noise level, causing a significant decrease in the accuracy of intelligent monitoring and analysis of deep stress in such areas.

[0019] Therefore, there is a need for a deep stress intelligent monitoring and analysis method that can explicitly separate rheological background effects in the deep stress intelligent monitoring and analysis process and cover the complete cycle of long-period precursor evolution.

[0020] According to an embodiment of this invention, this invention provides an intelligent monitoring and analysis method for deep stress in geological disasters based on AI big data. It should be understood that the computer system executing this intelligent monitoring and analysis method for deep stress is communicatively connected to multi-source sensing devices such as distributed fiber optic sensors, borehole stress gauges, and microseismic monitoring systems, and is capable of receiving and storing the deep stress monitoring big data continuously collected by these sensing devices.

[0021] At least one embodiment of the present invention discloses an intelligent monitoring and analysis method for deep stress in geological disasters based on AI big data, such as... Figure 1 As shown, it includes the following steps: Step 1: Obtain a multi-source long-term large dataset of the deep geological hazard monitoring area; To acquire a multi-source long-term big data set of the deep geological hazard monitoring area, the multi-source long-term big data set includes multi-year continuous deep stress time series data, distributed optical fiber strain time series data and its historical cumulative records, core rheological test parameters, monitoring system calibration records, long-term environmental temperature variation data and historical geological hazard event records.

[0022] It should be noted that the aforementioned multi-year continuous deep stress time series data refers to the stress value sequence continuously collected by borehole stress gauges deployed at various monitoring points in deep soft rock areas at preset sampling intervals, with a time span covering at least two complete annual cycles. The aforementioned distributed fiber optic strain time series data and its historical cumulative records refer to strain measurements continuously distributed along the fiber optic survey line, as well as the cumulative strain values ​​at each measuring point from deployment to the current moment. The aforementioned core rheological test parameters refer to the mechanical parameters obtained from indoor creep tests conducted on core samples taken from the monitoring area under different confining pressures and temperatures.

[0023] It should be noted that the above-mentioned monitoring system calibration records refer to the calibration reference values ​​and calibration times of each sensor during deployment and periodic maintenance, which are used to quantify the long-term system drift of the sensors.

[0024] Step 2: Calculate the viscoelastic constitutive parameter matrix and generate the rheological stress evolution baseline sequence for each monitoring point; Based on core rheological test parameters, the viscoelastic constitutive parameter matrix for different lithological regions was calculated. This viscoelastic constitutive parameter matrix includes the instantaneous elastic modulus. Delayed elastic modulus and viscosity coefficient The relationship between strain and confining pressure and temperature was investigated; historical strain accumulation data at each monitoring point were integrated over time using a viscoelastic constitutive model to calculate the memory contribution of strain history to the current stress state, generating a baseline sequence of rheological stress evolution for each monitoring point. This baseline sequence of rheological stress evolution characterizes the long-term trend of deep stress generated by pure rheological action under conditions of no hazard disturbance.

[0025] The calculation process of the above viscoelastic constitutive parameter matrix includes the following sub-steps: Step 201: Based on different confining pressures in the core rheological test parameters and temperature The creep curve data under the given conditions were used to fit the creep equation of the Burgers body model using the least squares method.

[0026] in, for Creep strain at any moment To apply constant stress, For instantaneous elastic modulus, For the delayed elastic modulus, The long-term viscosity coefficient, The delayed viscosity coefficient, For time, It is a natural constant. Obtain the confining pressure-temperature conditions for each group. The corresponding parameter group The constitutive parameter matrix is ​​composed of the following: For enclosing pressure index, For temperature indexing.

[0027] Furthermore, the aforementioned viscoelastic constitutive parameter matrix is ​​constructed using confining pressure and temperature as index dimensions. Each element in the viscoelastic constitutive parameter matrix corresponds to a set of four parameters under specific confining pressure-temperature conditions. The row index of the viscoelastic constitutive parameter matrix corresponds to a discrete confining pressure value sequence, and the column index corresponds to a discrete temperature value sequence, thus forming a two-dimensional parameter lookup table.

[0028] Step 202: For each monitoring point, obtain the estimated confining pressure and long-term environmental temperature variation data at the location of the monitoring point, and obtain the constitutive parameter set corresponding to the monitoring point from the viscoelastic constitutive parameter matrix through a linear interpolation algorithm.

[0029] Furthermore, the above-mentioned confining pressure estimates are obtained based on the burial depth of the monitoring points and the density of the overlying strata. The calculations are performed using the hydrostatic pressure assumption or by estimating the lateral pressure coefficient after correction based on the regional tectonic stress field.

[0030] Step 203: Based on the Boltzmann superposition principle, perform time integration on the historical strain accumulation data of each monitoring point to calculate the rheological stress evolution baseline sequence:

[0031] in, for The baseline value of rheological stress at time t, The relaxation modulus function is determined based on the viscoelastic constitutive parameter set. for The measured cumulative strain value at time t. For strain rate, For the current moment, Let be the integral variable representing a historical moment.

[0032] Furthermore, the aforementioned Boltzmann superposition principle states that the stress response of a viscoelastic material under arbitrary strain history can be expressed as the superposition of the contribution of historical strain increments to the current stress. This Boltzmann superposition principle transforms the stress-strain relationship of a material under complex loading history into the convolution integral form of the relaxation modulus function and the strain rate history.

[0033] It should be noted that the above relaxation modulus function The relaxation modulus corresponding to the Burgers volume model is specifically derived from the constitutive parameter set obtained in step 201. Confirmed. In numerical calculations, the aforementioned time integration is performed using a discretized recursive method, that is, the continuous integration is transformed into a step-by-step accumulation calculation according to the sampling interval.

[0034] Furthermore, in the above discretization recursive calculation process, the continuous-time integral is discretized into time steps of... The summation form:

[0035] in, For the first Each sampling time, The sampling interval is... This is the index of the current sampling time. To sum the indices and take values ​​from... arrive The integer value is obtained by approximating the strain rate through the difference between strain values ​​at adjacent time points.

[0036] In this embodiment of the application, in order to adapt the viscoelastic constitutive parameters to the long-term changes in confining pressure and temperature in the monitoring area, based on step 202, the constitutive parameter set of each monitoring point is also dynamically updated over time: when the estimated value of confining pressure or the long-term change data of ambient temperature at the location of the monitoring point changes, the updated constitutive parameter set is re-interpolated from the viscoelastic constitutive parameter matrix, and the updated constitutive parameter set is used to continue the calculation in the subsequent time integration step, so that the rheological stress evolution baseline sequence can reflect the influence of long-term changes in confining pressure and temperature on rheological behavior.

[0037] Step 3: Separate the rheological background and system drift from the deep stress time series data to generate a net anomaly trend sequence; The rheological residual time series is obtained by subtracting the rheological stress evolution baseline sequence of the corresponding monitoring point from the continuous deep stress time series data over many years. :

[0038] in, for The measured deep stress value at time [time] For the corresponding monitoring points The baseline value of rheological stress evolution at time t.

[0039] rheotropic residual timing The STL decomposition algorithm is used to decompose the trend, seasonality, and residual components, separating the long-term trend component. Annual cycle seasonal components and random residual components :

[0040] The input to this STL decomposition algorithm is the rheological residual time series. The output consists of three component sequences: an annual cycle and a seasonal component. It reflects the periodic changes in stress related to environmental factors such as seasonal temperature changes and annual fluctuations in groundwater levels.

[0041] Furthermore, the aforementioned STL decomposition algorithm is a time series decomposition method based on local weighted regression. This STL decomposition algorithm extracts seasonal components and trend components sequentially through an iterative process. Specifically, the algorithm first performs seasonal smoothing on the original sequence to extract the initial seasonal components, then subtracts the seasonal components from the original sequence and performs trend smoothing to extract the trend components, and finally optimizes the decomposition results through multiple iterations to achieve convergence. The random residual components are obtained by subtracting the seasonal and trend components from the original sequence.

[0042] Combining monitoring system calibration records with long-term trend components System drift correction is performed, the sensor reference drift at each calibration time point is obtained, and a continuous drift estimation sequence is obtained by piecewise linear interpolation. Calculate the corrected net abnormal trend sequence :

[0043] Furthermore, the aforementioned sensor reference drift is obtained by comparing the deviation between the sensor output value at the calibration time and the standard reference value, and the drift estimation sequence between each calibration node is also described. It is obtained by linear interpolation calculation of the sensor reference drift of adjacent calibration nodes.

[0044] In this embodiment of the application, in order to avoid insufficient drift estimation accuracy of piecewise linear interpolation for sections with excessively long calibration intervals, the following processing is also included: for sections where the interval between adjacent calibration nodes exceeds a preset time threshold, the drift correction amount of adjacent sensors within the same monitoring section is used for weighted auxiliary estimation, and the weight is inversely proportional to the distance between the sensors, thereby improving the drift correction accuracy of long calibration interval sections.

[0045] Furthermore, the aforementioned preset time threshold is determined based on the sensor type and monitoring environment conditions. For borehole stress gauges, the preset time threshold ranges from 6 months to 12 months.

[0046] Step 4: Extract multi-scale trend change features and rheological deviation features to generate a composite long-period feature vector; Net abnormal trend sequence Perform multi-scale wavelet transform to extract trend variation features from monthly to annual scales. This sub-step includes: Step 401: For Perform multi-level discrete wavelet transform to obtain wavelet approximation coefficient sequences and wavelet detail coefficient sequences corresponding to three decomposition levels: monthly, quarterly, and annual scales.

[0047] Furthermore, the above-mentioned multi-layer discrete wavelet transform uses the Daubechies wavelet basis function, and the number of decomposition layers is determined according to the sampling frequency and the target time scale, so that the first layer corresponds to the monthly scale, the second layer corresponds to the quarterly scale, and the third layer corresponds to the annual scale.

[0048] Furthermore, the aforementioned multi-level discrete wavelet transform is a multi-resolution analysis method that decomposes a signal into different frequency components. This multi-level discrete wavelet transform obtains low-frequency approximation coefficients and high-frequency detail coefficients layer by layer by convolving the signal with wavelet basis functions and then performing downsampling operations. The wavelet approximation coefficient sequence reflects the trend characteristics of the signal, while the wavelet detail coefficient sequence reflects the fluctuation characteristics of the signal.

[0049] Step 402: Calculate the rate of change of trend slope for the wavelet approximation coefficient sequence at each scale. Curvature characteristics and inflection point position ,in This indicates a decomposition hierarchical index. Trend slope rate of change. Curvature characteristics are obtained by calculating the variation amplitude of wavelet approximation coefficient sequences at various scales after performing first-order differencing; The inflection point position is obtained by performing second-order difference on the wavelet approximation coefficient sequences at various scales. This is obtained by detecting the zero-crossing points of the second-order difference sequence. The above features at each scale are then combined to form multi-scale trend change features. .

[0050] Furthermore, the rate of change of the slope of the aforementioned trend The calculation method is as follows: for the first The wavelet approximation coefficient sequence is subjected to first-order difference, and then the standard deviation of the difference sequence within the sliding window is calculated as a quantitative indicator of the rate of change of the trend slope.

[0051] Step 403: Calculate the baseline sequence of rheological stress evolution. Compared with the measured deep stress value Deviation rate between :

[0052] For deviation rate The mean and rate of change are calculated within monthly, quarterly, and annual time windows, respectively, to generate rheological deviation feature vectors. .

[0053] Furthermore, the monthly time window is 30 days long, the quarterly time window is 90 days long, and the annual time window is 365 days long. The mean within each time window is obtained by arithmetic average, and the rate of change is obtained by linear regression slope of the deviation rate sequence within the window.

[0054] Step 404: Analyze the multi-scale trend change characteristics rheological deviation eigenvector Concatenate into a composite long-period feature vector : in, This indicates a vector concatenation operation.

[0055] It should be noted that the monthly, quarterly, and annual scales in the above multi-scale wavelet transform refer to the time frequency ranges corresponding to different levels of wavelet decomposition matching the monthly, quarterly, and annual time scales, respectively. Specifically, based on the relationship between the sampling interval and the number of decomposition levels, the number of decomposition levels is selected to ensure that the frequency passband range of each level covers the above three scales.

[0056] In this embodiment of the application, in order to reduce the interference of redundant features in the composite long-period feature vector on subsequent recognition steps, based on step 404, further steps are taken to... The variance contribution rate of each dimension feature is calculated, and the feature dimensions with variance contribution rates lower than the preset contribution threshold are filtered out to obtain the dimension-reduced composite long-period feature vector. This dimension-reduced composite long-period feature vector is used as the input for subsequent steps.

[0057] Furthermore, the aforementioned preset contribution threshold is set to 5% of the total variance, which means that feature dimensions with a variance contribution rate of less than 5% are filtered out.

[0058] In this embodiment of the application, in order to eliminate composite long-period eigenvectors The impact of differences in the dimensions of different features on the subsequent training of deep learning models is discussed after obtaining the composite long-period feature vector in step 404. The features in each dimension are Z-score standardized to make the mean of each feature zero and the standard deviation one, thereby ensuring that features of different dimensions have the same numerical scale in the model input layer.

[0059] Step 5: Input the composite long-period feature vector into the hierarchical time-series deep learning model and output the intelligent recognition result of long-period precursors; Composite long-period eigenvectors The input is fed into a hierarchical time-series deep learning model, which outputs long-cycle precursor intelligent recognition results and the current evolution stage location.

[0060] It should be noted that the above-described hierarchical temporal deep learning model is a multi-timescale hierarchical deep network structure. This hierarchical temporal deep learning model consists of three components: a bottom-level weekly encoder, a middle-level monthly aggregator, and a top-level quarterly classifier. The construction process of this hierarchical temporal deep learning model is as follows: Step 501: Construct the bottom-level periodic encoder. The bottom-level periodic encoder adopts a gated cyclic unit structure and receives... The input sequence is a time segment divided into weeks. Each week-scale time segment is encoded to obtain a week-scale hidden state representation sequence. It is used to capture short-term fluctuation patterns of rheological deviations.

[0061] Furthermore, the aforementioned time segments divided by week refer to the composite long-period feature vectors. The system is divided into 7-day time segments, with each segment containing feature vectors corresponding to each sampling moment within that week. Furthermore, the aforementioned gated recurrent unit is a recurrent neural network structure. This unit includes two gating mechanisms: an update gate and a reset gate. The update gate controls the proportion of hidden state information from the previous moment retained in the current moment, while the reset gate controls the proportion of the previous hidden state used in calculating the current candidate hidden state. This gating mechanism enables effective learning of long-term dependencies.

[0062] Step 502: Construct a mid-level monthly-scale aggregator. The mid-level monthly-scale aggregator receives the weekly-scale hidden state representation sequence. As input, sort by the number of weeks contained in each month. The data is grouped and aggregated within each group. The mid-level monthly-scale aggregator incorporates a cross-level attention mechanism. This mechanism uses the weekly-scale hidden state group corresponding to the current month as the query and the weekly-scale hidden state groups corresponding to adjacent months as the key and value. It employs a scaled dot product attention algorithm to calculate attention weights and performs weighted aggregation, outputting a monthly-scale aggregated representation sequence. It is used to identify the trend characteristics of gradual deviation accumulation between weeks.

[0063] Furthermore, the calculation process of the above-mentioned scaled dot product attention algorithm is as follows: First, calculate the query matrix. AND key matrix The product of the transposes of the matrices, then divided by the scaling factor. (in (where the dimension is the key vector), then apply the softmax function to the result to obtain the attention weight matrix, and finally combine the attention weight matrix with the value matrix. Multiplying yields the weighted representation vector. Wherein, The transpose is denoted as Furthermore, the above is based on the number of weeks included in each month. When grouping, each monthly group contains 4 to 5 weekly-scale hidden state vectors, with the specific number determined based on the actual number of days in that month.

[0064] Step 503: Construct the top-level quarterly-scale classifier. The top-level quarterly-scale classifier receives the monthly-scale aggregated representation sequence. As input, calculate based on the number of months contained in each quarter. The data is grouped, fused, and then output as a classification result via a fully connected layer. The output of the top-level quarterly-scale classifier consists of two parts: the intelligent identification result of long-term precursors and the current evolution stage location. The intelligent identification result of long-term precursors serves as a classification label, used to distinguish whether the current signal belongs to a deviation of the rheological model system or a true precursory anomaly evolution of a geological disaster; the current evolution stage location serves as a stage identifier, characterizing the stage of the precursor signal within the complete evolution cycle.

[0065] It should be noted that the specific method for fusion processing of each group is as follows: average pooling is performed on the monthly aggregated representation sequence within each quarterly group along the time dimension to obtain the fusion representation vector for that quarter.

[0066] It should be noted that the above-mentioned long-cycle precursor intelligent recognition results and current evolution stage positioning are discrete outputs in the form of classification labels. After the output of the fully connected layer of the top-level quarterly scale classifier, they are converted into probability distributions through the softmax activation function. The category with the highest probability is selected as the final long-cycle precursor intelligent recognition result and current evolution stage positioning, thereby realizing the decoding of the neural network output into specific executable classification results.

[0067] Furthermore, the calculation formula for the softmax activation function is as follows:

[0068] in, The first output of the fully connected layer The original scores for each category, The total number of categories, To sum the indices and take values ​​from... arrive integers, This represents an exponential function.

[0069] Furthermore, the input layer of the aforementioned hierarchical temporal deep learning model receives composite long-period feature vectors. The composite long-period feature vector is divided into weekly time segments and then input into the bottom weekly-scale encoder; the output layer is the fully connected layer of the top quarterly-scale classifier, which includes two output branches: the first branch outputs the classification probability distribution of the long-period precursor intelligent recognition result, and the second branch outputs the stage identifier probability distribution of the current evolution stage.

[0070] Furthermore, the current evolutionary stage positioning mentioned above includes four discrete stage identifiers: initial stage, development stage, acceleration stage, and critical stage. Each stage identifier corresponds to different dynamic characteristics in the evolution of precursor signals.

[0071] Furthermore, the training process of the above-mentioned hierarchical time-series deep learning model is as follows: the training data is based on the marked precursor segments and confirmed non-precursor segments in the historical geological disaster event records. The composite long-cycle feature vectors of the corresponding time periods are combined with the labeled labels to form training sample pairs. The training process adopts the classification cross-entropy loss function to jointly train the two output branches of long-cycle precursor intelligent recognition results and current evolution stage positioning. The optimization process adopts the Adam optimization algorithm.

[0072] Furthermore, the formula for calculating the classification cross-entropy loss function is as follows:

[0073] in, For the total loss, The number of training samples. The sample index is from [index] and the value range is [range]. arrive integers, For the first Identify the true label from the precursor of each sample. For the first Probability of predicting the precursors of a sample For the first True labels for the evolutionary stages of each sample. For the first The probability of predicting the evolutionary stage of a sample. and These are the weighting coefficients for the two output branches.

[0074] Furthermore, the aforementioned weighting coefficients and The value of is determined based on the importance of the two classification tasks. In this embodiment, _ _ is... The value is 0.6. A value of 0.4 is used to give the precursor recognition task a higher optimization weight.

[0075] Furthermore, the time dimension in the aforementioned classification cross-entropy loss function is reflected as follows: the classification cross-entropy loss function sums the samples across all time periods in the training sample set, with each sample... By using feature vectors and labels corresponding to specific time periods in the monitoring time series, and through cumulative optimization of sample loss in different time periods, the hierarchical time series deep learning model learns the dynamic feature patterns of precursor signals during the time evolution process.

[0076] In this embodiment of the application, in order to mitigate the impact of class imbalance caused by the number of positive precursor samples being far less than the number of negative non-precursor samples on the training of the hierarchical temporal deep learning model, a class weight factor is introduced into the classification cross-entropy loss function during the training process. The loss weight corresponding to the positive precursor samples is determined according to the reciprocal of the ratio of the number of positive to negative samples, thereby enabling the hierarchical temporal deep learning model to give more attention to the scarce precursor samples during the training process.

[0077] Furthermore, the above-mentioned category weighting factors are calculated as follows: The number of positive precursor samples in the training set is counted. Number of negative samples without precursors Calculate the positive sample weight factor as follows: The negative sample weight factor is 1, and the class weight factor is multiplied into the loss term of the corresponding sample.

[0078] In this embodiment of the application, in order to enhance the ability of the hierarchical temporal deep learning model to identify the transition intervals of different evolutionary stages, a temporal smoothing layer is added before the fully connected layer of the top-level quarterly scale classifier. This temporal smoothing layer performs one-dimensional convolutional filtering on the quarterly scale fused representation sequence along the time dimension. The length of the filtering window covers the typical transition duration of two adjacent evolutionary stages. The filtered representation is then input into the fully connected layer for classification, thereby enabling the hierarchical temporal deep learning model to perceive the gradual features within the stage transition intervals.

[0079] Furthermore, the window length of the aforementioned one-dimensional convolutional filter is set to 3 quarters, and the convolutional kernel parameters are obtained through adaptive learning during the training process of the hierarchical temporal deep learning model.

[0080] Step 6: Link historical geological disaster event records to generate intelligent monitoring and analysis results of deep stress; Based on the intelligent identification results of long-period precursors and the current evolutionary stage location, the statistical characteristics of the evolutionary cycles of similar precursor patterns in historical geological disaster event records are correlated to generate intelligent monitoring and analysis results of deep stress. This step includes the following sub-steps: Step 601: When the long-period precursor intelligent identification result determines that the current signal is a real precursor abnormal evolution of geological disaster, retrieve historical events with the same precursor classification label from the historical geological disaster event record, and obtain statistical distribution data of the complete evolution cycle length of the same precursor pattern, including the average evolution cycle length and standard deviation.

[0081] Furthermore, the aforementioned identical precursor classification labels refer to event labels in historical geological disaster event records that have the same disaster type identifier and the same evolutionary pattern characteristics as the current long-term precursor intelligent identification results.

[0082] Step 602: Based on the current evolutionary stage location and the evolutionary cycle statistical distribution data of similar precursor patterns, calculate the relative position ratio of the current precursor evolutionary stage within the complete evolutionary cycle. :

[0083] in, For the current moment, This is the start time of the current precursor signal (determined by the initial stage start point in the current evolutionary stage positioning). This represents the average evolutionary cycle length of similar precursor patterns.

[0084] Furthermore, the starting time of the aforementioned precursor signals The initial stage identifier is determined by tracing back to the time point when it first appears in the current monitoring sequence.

[0085] Step 603: Classification confidence and relative position ratio based on the intelligent identification results of long-period precursors This generates geological hazard precursor classification levels and risk warning levels. Geological hazard precursor classification levels are divided into three levels based on classification confidence: suspected precursors, possible precursors, and confirmed precursors; risk warning levels are based on relative location ratios. Divided into three levels: Attention, Warning, and Emergency. The closer the value is to 1, the closer the evolutionary process is to the end of the complete evolutionary cycle, and the higher the corresponding risk warning level.

[0086] Furthermore, the classification criteria for the aforementioned geological disaster precursors are as follows: when the classification confidence level is less than 0.6, it is judged as a suspected precursor; when the classification confidence level is between 0.6 and 0.8, it is judged as a possible precursor; and when the classification confidence level is greater than 0.8, it is judged as a confirmed precursor.

[0087] Furthermore, the criteria for classifying the above-mentioned risk warning levels are as follows: when the relative position ratio A value less than 0.5 is considered a level of concern. A reading between 0.5 and 0.8 indicates a warning level. A value greater than 0.8 is considered an emergency level.

[0088] Step 604: Summarize the geological disaster precursor identification level, the relative position of the precursor evolution stage in the complete evolution cycle, and the risk warning level to generate the deep stress intelligent monitoring and analysis results.

[0089] It should be noted that the above classification confidence level refers to the probability value of the softmax probability value output by the top-level quarterly scale classifier that corresponds to the abnormal evolution category of the true precursor of geological disaster. When the long-period precursor intelligent identification result determines that the current signal is a systemic bias of the rheological model, the geological disaster precursor discrimination level in the deep stress intelligent monitoring and analysis result output in step 604 is marked as no precursor, and the risk warning level is marked as normal.

[0090] In this embodiment of the application, in order to avoid false alarms caused by occasional anomalies of a single monitoring point, the following processing is also included in step 603: when multiple monitoring points in the same monitoring section generate long-period precursor intelligent identification results at the same time, the classification confidence of each monitoring point is checked for spatial consistency. Only when the geological disaster precursor discrimination level of spatially adjacent monitoring points reaches the preset spatial consistency condition will the corresponding risk warning level be upgraded to the corresponding level; otherwise, the current risk warning level will be maintained or downgraded.

[0091] Furthermore, the aforementioned pre-set spatial consistency condition is as follows: within the same monitoring section, at least three monitoring points with a spatial distance of less than 50 meters simultaneously output the same geological disaster precursor judgment level, and the classification confidence of each monitoring point is greater than 0.5.

[0092] This embodiment addresses the problem of high overlap between rheological trends and geological disaster precursor signals in deep soft rock areas on time scales ranging from monthly to yearly, by providing a method for intelligent monitoring and analysis of deep stress.

[0093] Because this implementation method calculates the viscoelastic constitutive parameter matrix based on core rheological test parameters and performs time integration on the historical strain accumulation data of each monitoring point using the Boltzmann superposition principle, a rheological stress evolution baseline sequence characterizing the long-term trend of deep stress generated by pure rheological action is generated. Then, this rheological stress evolution baseline sequence is explicitly subtracted from the measured deep stress time series data. Therefore, the rheological background trend is separated before the data enters the AI ​​model for processing, eliminating the systematic source of rheological trends being misjudged as precursor signals and reducing false alarms caused by rheological effects.

[0094] Furthermore, because this implementation further calculates the rheological deviation feature vector after separating the rheological stress evolution baseline sequence, and concatenates the rheological deviation feature vector with the multi-scale trend change feature vector into a composite long-period feature vector, it provides a quantitative criterion for the subsequent hierarchical time-series deep learning model to distinguish between the residuals of the rheological model and the real precursor anomalies. Therefore, the hierarchical time-series deep learning model can further distinguish the difference between the residual system deviation and the precursor signal under the premise that the rheological background has been weakened.

[0095] In addition, because this implementation adopts a hierarchical time-series deep learning model consisting of a bottom-level weekly encoder, a middle-level monthly aggregator, and a top-level quarterly classifier, it covers multi-time-scale information from short-term fluctuations to slow evolution over several months through a layer-by-layer aggregation structure from weekly to monthly to quarterly. Therefore, it overcomes the limitation of existing short-term sliding windows that cannot capture the complete cycle of precursory evolution over several months in low-permeability rock masses, and enables the long-term slow state migration features to be extracted and identified within an appropriate time scale.

[0096] It is evident that this implementation method improves the accuracy of disaster precursor identification in intelligent monitoring and analysis of deep stress in deep soft rock areas by combining explicit separation of rheological background with hierarchical multi-timescale long-period precursor signal extraction.

[0097] A deep coal mine located in North China has a main coal seam buried at a depth of 850 meters. The roof of the coal seam consists of thick mudstone layers, while the floor consists of shale interbedded with thin sandstone layers. In 20XX, the mine initiated the construction of a deep stress intelligent monitoring system, deploying 12 borehole stress monitoring points, 3 distributed fiber optic survey lines, and 1 microseismic monitoring network in the working face return airway, transport roadway, and roof strata. The monitoring system began continuously collecting data in January 20XX, accumulating three years of complete monitoring records by December 20XX. The mudstone in this area exhibits significant creep characteristics, and under continuous high ground stress, the stress shows a clear long-term redistribution trend. In September 20XX, the stress data at monitoring point M06 showed a slow increase over three consecutive months. It is necessary to determine whether this trend is a rheological background effect or a precursor signal of rockburst.

[0098] For monitoring point M06, the system extracted continuous deep stress time-series data for 36 months from January 20XX to December 20XX from the database, with a sampling interval of 2 hours, totaling 15,768 data points. Simultaneously, distributed fiber optic strain data for section F2, where the monitoring point is located, was extracted. The fiber optic measuring line length is 120 meters, with a spatial resolution of 1 meter and a temporal resolution of 12 hours. Furthermore, indoor creep test data of mudstone cores from the mining area under confining pressures ranging from 15 MPa to 30 MPa and temperatures ranging from 20℃ to 35℃ were retrieved, including six sets of creep curves for different confining pressure-temperature combinations. The monitoring system calibration record shows that monitoring point M06 underwent three calibration maintenances in March 20XX, September 20XX, and March 20XX. Environmental temperature monitoring data shows that the temperature of the rock strata at this depth fluctuates seasonally between 28℃ and 32℃. The historical geological disaster event record database contains detailed records of 17 rockburst events that occurred in the mining area and adjacent mining areas within the past 10 years.

[0099] Table 1. Basic Information and Data Overview of Monitoring Point M06:

[0100] First, Burgers volume model parameters were fitted to six sets of core creep test data. Taking the test curves under confining pressure of 20 MPa and temperature of 30℃ as an example, under constant stress... Under the action, the creep strain evolution data over time was obtained by least squares fitting: instantaneous elastic modulus Delayed elastic modulus Long-term viscosity coefficient Delayed viscosity coefficient The remaining five sets of experimental data were fitted sequentially to construct the viscoelastic constitutive parameter matrix.

[0101] Table 2 Viscoelastic constitutive parameter matrix:

[0102] Monitoring point M06 is buried at a depth of 850 meters, with an average density of 2.65 in the overlying rock strata. The confining pressure, estimated based on hydrostatic pressure, is 22.1 MPa. Environmental temperature monitoring at this depth shows an average temperature of 30.5℃ in September 20XX. By performing bilinear interpolation on the viscoelastic constitutive parameter matrix, the constitutive parameter set corresponding to this monitoring point was obtained: , , , .

[0103] Based on this constitutive parameter set, Boltzmann superposition integral calculations were performed on the distributed fiber optic cumulative strain data of monitoring point M06 from January 20XX to December 20XX. A discretized recursive method was adopted, with a time step of [missing information]. (Corresponding to a 12-hour sampling interval). The relaxation modulus function is determined based on the Burgers volume model, within the time interval. The relaxation modulus at that point is:

[0104] Taking September 15, 20XX as an example, this moment represents 987 time steps since the monitoring started on January 1, 20XX. By progressively accumulating the historical strain increments, the baseline value of the rheological stress at this moment is obtained. .

[0105] Table 3. Baseline sequence of rheological stress evolution at monitoring point M06 (partial data from July 20XX to October 20XX):

[0106] The rheological residual time series is obtained by subtracting the corresponding rheological stress evolution baseline sequence point by point from the measured deep stress time series data at monitoring point M06. Taking 14:00 on September 15, 20XX as an example, the measured stress value... Corresponding rheological baseline Calculate the rheological residuals:

[0107] The STL decomposition algorithm was used to decompose the rheological residual time series over the entire three-year time span. In the STL decomposition, the seasonality window was set to 365 days (corresponding to an annual cycle), and the trend window was set to 730 days (corresponding to a two-year smoothing cycle). The decomposition results show that the annual cycle seasonal component... The amplitude range is -0.8MPa to +0.9MPa, reflecting the impact of annual fluctuations in groundwater level on stress; the long-term trend component It showed a slow upward trend from January 20XX to June 20XX, remained basically stable from July 20XX to December 20XX, and began to accelerate upward in September 20XX.

[0108] Table 4. STL decomposition results of rheological residuals at monitoring point M06 (partial data from September 20XX):

[0109] System drift correction was performed on the long-term trend components based on the calibration records of the monitoring system. The sensor baseline drift measured at monitoring point M06 during the three calibrations were: +0.15 MPa in March 20XX, +0.28 MPa in September 20XX, and +0.42 MPa in March 20XX. A continuous drift estimation sequence was generated using piecewise linear interpolation. Taking September 15, 20XX as an example, this moment falls between the September 20XX calibration node (drift +0.28MPa) and the March 20XX calibration node (drift +0.42MPa), a time interval of 18 months. Twelve months have passed, and the interpolation calculation yields:

[0110] Calculate the corrected net abnormal trend sequence:

[0111] Table 5. Drift Correction and Net Anomaly Trends of Monitoring Point M06 (Partial Data from September 20XX):

[0112] Net abnormal trend sequence A three-level discrete wavelet transform is performed using the Daubechiesdb4 wavelet basis. The first level of decomposition corresponds to a monthly scale (approximately 30 days), the second level to a quarterly scale (approximately 90 days), and the third level to an annual scale (approximately 365 days). Taking September 15, 20XX, as an example, the wavelet approximation coefficient sequences at each scale are extracted.

[0113] The first-order difference of the wavelet approximation coefficient sequence at the monthly scale (first level) is performed, and the standard deviation is calculated within a 30-day sliding window to obtain the rate of change of the trend slope. The curvature features are obtained by performing a second-order difference on the sequence. The second-order difference zero-crossing point was detected on September 8, 20XX, and is denoted as inflection point 1. The same calculation is performed on the wavelet approximation coefficient sequences at the quarterly (second layer) and annual (third layer) scales to obtain... , The inflection point 2 is located on August 20, 20XX; , The inflection point 3 is located on July 15, 20XX.

[0114] Calculate the deviation rate between the baseline sequence of rheological stress evolution and the measured stress values. Taking September 15, 20XX as an example:

[0115] The mean deviation rate and rate of change were calculated for monthly (30-day window), quarterly (90-day window), and annual (365-day window) timeframes, respectively. The mean deviation rate for the monthly window (August 16, 20XX to September 15, 20XX) was 0.128, and the rate of change was... The average deviation rate within the quarterly window (June 17, 20XX to September 15, 20XX) was 0.115, and the rate of change was... The average deviation rate within the annual window (September 16, 20XX to September 15, 20XX) was 0.098, and the rate of change was... .

[0116] Multi-scale trend change features and rheological deviation features are concatenated into a composite long-period feature vector, which contains 15 feature dimensions (inflection point positions are encoded as timestamps and then converted into numerical features).

[0117] Table 6. Composition of the composite long-period feature vector of monitoring point M06:

[0118] The composite long-period feature vector was divided into weekly time segments (7 days per unit). For the period from September 1st to September 30th, 20XX, it was divided into 4 weekly time segments (Week 1: September 1st-7th, Week 2: September 8th-14th, Week 3: September 15th-21st, Week 4: September 22nd-30th). Each time segment corresponds to the composite long-period feature vector sequence at each sampling time within that week.

[0119] The underlying weekly encoder receives these four weekly time segments and encodes each segment using a gated recurrent unit (ROU) structure. Taking week 3 (September 15-21) as an example, this week contains 84 sampling times (2-hour intervals), corresponding to 84 feature vector input sequences. The hidden state dimension of the gated ROU is set to 64 dimensions, and after encoding, it outputs the weekly-scale hidden state representation vector (64-dimensional vector). By encoding each of the four weekly segments, four weekly-scale hidden state representation vectors are obtained.

[0120] The mid-level monthly-scale aggregator groups these four weekly-scale hidden states by month (September contains four weeks) and aggregates them through a cross-level attention mechanism. Taking September as the query, the weekly-scale hidden states of adjacent months (August and October) are the keys and values. Scaled dot-product attention weights are calculated, with the key vector having a dimension of 64 and a scaling factor of 8. After calculating the attention weights, the four weekly-scale hidden states of September are weighted and aggregated, outputting a monthly-scale aggregated representation vector (128-dimensional vector).

[0121] The top-level quarterly-scale classifier groups the monthly-scale aggregated representations by quarter. The third quarter of 20XX (July-September) contains three monthly-scale aggregated representation vectors, which are averaged along the time dimension to obtain the fused representation vector (128 dimensions) for that quarter. This fused representation vector is processed by a fully connected layer, outputting two branches: the first branch outputs the probability distribution for precursor identification and classification, and the second branch outputs the probability distribution for evolutionary stage localization.

[0122] After transformation using the softmax activation function, the precursor identification and classification results show: the probability of systematic bias in the rheological model is 0.18, and the probability of anomalous evolution of true precursors of geological disasters is 0.82. Selecting the category with the highest probability, the long-term precursor intelligent identification result is determined to be "anomalous evolution of true precursors of geological disasters." The evolution stage location results show: initial stage probability 0.05, development stage probability 0.68, acceleration stage probability 0.25, and critical stage probability 0.02. Selecting the category with the highest probability, the current evolution stage is located as "development stage."

[0123] Table 7 Output results of the hierarchical temporal deep learning model for monitoring point M06:

[0124] Since the long-period precursor intelligent identification results determined that the current signal was a true precursor to a geological disaster, the system retrieved historical events tagged with "rockburst - mudstone roof" from the historical geological disaster event record database. A total of 8 historical events with similar precursor patterns were retrieved. The complete evolution cycle length from the initial stage to the occurrence of the disaster for these 8 events was calculated, and the average evolution cycle length was 156 days with a standard deviation of 23 days.

[0125] The precursor signal at monitoring point M06, through backtracking analysis, has had its initial occurrence dated July 12, 20XX (the time point when the initial stage marker first appeared). The current time is September 15, 20XX. The relative position ratio is calculated as follows:

[0126] The current time is 65 days from the start time, and the average evolution cycle length is 156 days.

[0127] Based on a classification confidence level of 0.82 (greater than 0.8), the geological hazard precursor classification level is determined to be "confirmed precursor". Based on the relative location ratio... (less than 0.5), the risk warning level is determined to be "attention".

[0128] The results of the intelligent monitoring and analysis of deep stress are summarized as follows: Monitoring point M06 is currently in the confirmed precursor state, the precursor evolution stage is the development stage, its relative position in the complete evolution cycle is 41.7%, and the risk warning level is attention.

[0129] Table 8. Analysis results of intelligent monitoring of deep stress at monitoring point M06:

[0130] The entire intelligent monitoring and analysis process for deep stress achieves a complete transformation from raw monitoring big data to risk warning results: First, a multi-source long-term dataset spanning three years was extracted from monitoring point M06, including 15,768 deep stress data points, distributed fiber optic strain data, and core rheological test parameters; then, through viscoelastic constitutive model calculations, the core test parameters were transformed into a constitutive parameter set under specific conditions at the monitoring point, thereby generating a rheological stress evolution baseline sequence covering the entire time span. This baseline sequence quantifies the long-term stress trend generated by pure rheological action; based on this, by subtracting the rheological baseline from the measured stress time series and performing STL decomposition and system drift correction, the original stress data was converted into a net anomaly trend sequence that eliminates the rheological background and sensor drift. The net anomaly trend sequence is then transformed into a composite long-cycle feature vector with 15 dimensions through multi-scale wavelet transform and rheological deviation calculation. This feature vector integrates trend change features and rheological deviation features from monthly to annual multi-timescales. The hierarchical time series deep learning model receives this feature vector and converts it into precursor identification classification probability (0.82) and evolution stage location (development stage) through three-level hierarchical processing of week, month, and quarter. Finally, combined with the average evolution cycle (156 days) of historical event statistics and the current relative evolution position (41.7%), a complete monitoring and analysis result including confirmed precursor judgment, development stage location, and attention-level early warning is generated, providing quantitative intelligent decision support for the prevention and control of rockbursts in mining areas.

[0131] The embodiments of the present invention have been described above. However, the embodiments are 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 more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for intelligent monitoring and analysis of deep stress in geological hazards based on AI big data, characterized in that, Includes the following steps: To acquire a multi-source long-term large dataset of deep geological disaster monitoring areas, the multi-source long-term large dataset includes multi-year continuous deep stress time series data, distributed optical fiber strain time series data and its historical cumulative records, core rheological test parameters, monitoring system calibration records, long-term environmental temperature variation data and historical geological disaster event records. Based on the core rheological test parameters, the viscoelastic constitutive parameter matrix of different lithological regions is calculated, and the historical strain accumulation data of each monitoring point is integrated over time according to the viscoelastic constitutive model to generate the rheological stress evolution baseline sequence of each monitoring point. The rheological residual time series is obtained by subtracting the rheological stress evolution baseline sequence of the corresponding monitoring point from the multi-year continuous deep stress time series data. The rheological residual time series is decomposed into three components: trend, seasonality and residual. System drift correction is performed by combining the calibration record of the monitoring system to generate a net anomaly trend sequence. Multi-scale wavelet transform is performed on the net anomaly trend sequence to extract multi-scale trend change features. The deviation rate between the rheological stress evolution baseline sequence and the measured deep stress value is calculated and a rheological deviation feature vector is generated. The multi-scale trend change features and the rheological deviation feature vector are concatenated into a composite long-period feature vector. The composite long-period feature vector is input into a hierarchical time-series deep learning model, which outputs the intelligent recognition result of long-period precursors and the current evolution stage location. Based on the intelligent identification results of long-cycle precursors and the current evolution stage location, the statistical characteristics of the evolution cycle of similar precursor patterns in the historical geological disaster event records are associated to generate intelligent monitoring and analysis results of deep stress.

2. The method for intelligent monitoring and analysis of deep stress in geological disasters based on AI big data as described in claim 1, characterized in that, The calculation of the viscoelastic constitutive parameter matrix for different lithological regions includes: Based on the creep curve data under different confining pressure and temperature conditions in the core rheological test parameters, the creep equation of the Burgers body model is fitted by the least squares method to obtain the parameter set corresponding to each confining pressure-temperature condition. The parameter set includes instantaneous elastic modulus, delayed elastic modulus, long-term viscosity coefficient and delayed viscosity coefficient, which constitute the viscoelastic constitutive parameter matrix. The viscoelastic constitutive parameter matrix is ​​constructed with confining pressure and temperature as index dimensions. The row index corresponds to the discrete confining pressure value sequence, the column index corresponds to the discrete temperature value sequence, and each element corresponds to a set of four parameters under a specific confining pressure-temperature condition. For each monitoring point, the estimated confining pressure and long-term environmental temperature variation data at the location of the monitoring point are obtained, and the constitutive parameter set corresponding to the monitoring point is obtained from the viscoelastic constitutive parameter matrix through a linear interpolation algorithm; when the estimated confining pressure or long-term environmental temperature variation data at the location of the monitoring point changes, the updated constitutive parameter set is obtained again by interpolation from the viscoelastic constitutive parameter matrix.

3. The intelligent monitoring and analysis method for deep stress in geological disasters based on AI big data as described in claim 2, characterized in that, The process of integrating the historical strain accumulation data of each monitoring point over time using a viscoelastic constitutive model to generate a baseline sequence of rheological stress evolution for each monitoring point includes: Based on the Boltzmann superposition principle, the stress-strain relationship of each monitoring point under complex loading history is transformed into the convolution integral form of the relaxation modulus function and the strain rate history. The relaxation modulus function is determined by the constitutive parameter set corresponding to the Burgers volume model. For the historical strain accumulation data of each monitoring point, the convolution integral is discretized into a recursive calculation form that is gradually accumulated according to the sampling interval. At each sampling time, the strain rate obtained by approximating the difference between the relaxation modulus function value and the strain value at the adjacent time is multiplied and accumulated to generate the rheological stress evolution baseline sequence of each monitoring point. The rheological stress evolution baseline sequence characterizes the long-term trend of deep stress generated by pure rheological action under no disaster disturbance conditions.

4. The intelligent monitoring and analysis method for deep stress in geological disasters based on AI big data as described in claim 1, characterized in that, The process of performing trend-seasonal-residual three-component decomposition on the rheological residual time series and combining it with the system drift correction based on the monitoring system calibration records to generate a net anomaly trend sequence includes: The rheological residual time series is decomposed using the STL decomposition algorithm. The STL decomposition algorithm performs seasonal smoothing and trend smoothing in an iterative manner to separate the long-term trend component, the annual cycle seasonal component and the random residual component. The sensor reference drift at each calibration time node in the calibration record of the monitoring system is obtained. The sensor reference drift is obtained by comparing the deviation between the sensor output value at the calibration time and the standard reference value. A continuous drift estimation sequence is obtained by piecewise linear interpolation. Subtract the continuous drift estimation sequence from the long-term trend component to generate the net abnormal trend sequence; For sections where the interval between adjacent calibration nodes exceeds a preset time threshold, a weighted auxiliary estimation is performed using the drift correction values ​​of adjacent sensors within the same monitoring section. The weights are inversely proportional to the distance between the sensors.

5. The intelligent monitoring and analysis method for deep stress in geological disasters based on AI big data as described in claim 1, characterized in that, The step of extracting multi-scale trend change features from the net anomaly trend sequence by performing multi-scale wavelet transform includes: The net anomaly trend sequence is subjected to multi-level discrete wavelet transform to obtain wavelet approximation coefficient sequences and wavelet detail coefficient sequences corresponding to three decomposition levels: monthly, quarterly, and annual. The multi-level discrete wavelet transform uses the Daubechies wavelet basis function. For wavelet approximation coefficient sequences at various scales, the standard deviation within the sliding window is calculated after first-order difference to obtain the rate of change of trend slope, the curvature feature is obtained through second-order difference, and the inflection point position is obtained by detecting the zero-crossing point of the second-order difference sequence. The above features at various scales are combined into multi-scale trend change features. Calculating the deviation rate between the baseline sequence of rheological stress evolution and the measured deep stress value and generating a rheological deviation feature vector includes: calculating the ratio of the difference between the measured deep stress value and the baseline value of rheological stress evolution to the baseline value of rheological stress evolution as the deviation rate; calculating the mean and rate of change of the deviation rate within time windows on a monthly, quarterly, and annual scale respectively; generating a rheological deviation feature vector; and obtaining the rate of change through the linear regression slope of the deviation rate sequence within the window.

6. The intelligent monitoring and analysis method for deep stress in geological disasters based on AI big data as described in claim 1, characterized in that, The hierarchical temporal deep learning model includes a bottom-level weekly encoder, a middle-level monthly aggregator, and a top-level quarterly classifier. The underlying periodic encoder adopts a gated cyclic unit structure, receives time segments divided into weeks in the composite long-period feature vector as input sequences, encodes each periodic time segment, and obtains a periodic hidden state representation sequence. The mid-level monthly scale aggregator receives the weekly scale hidden state representation sequence as input, groups the weekly scale hidden state representation sequence according to the number of weeks contained in each month, aggregates each group through a cross-level attention mechanism, and outputs the monthly scale aggregated representation sequence. The top-level quarterly-scale classifier receives the monthly-scale aggregated representation sequence as input, groups it according to the number of months in each quarter, performs average pooling operation on each group along the time dimension to obtain a fused representation vector, and outputs the long-cycle precursor intelligent identification result and the current evolution stage location through a fully connected layer. The long-cycle precursor intelligent identification result is used to distinguish whether the current signal belongs to the system deviation of the rheological model or the abnormal evolution of the real precursor of geological disaster. The current evolution stage location includes four discrete stage identifiers: initial stage, development stage, acceleration stage and critical stage.

7. The intelligent monitoring and analysis method for deep stress in geological disasters based on AI big data as described in claim 6, characterized in that, The cross-level attention mechanism uses the weekly-scale hidden state group corresponding to the current month as the query and the weekly-scale hidden state groups corresponding to adjacent months as the key and value. It uses the scaled dot product attention algorithm to calculate the attention weights and perform weighted aggregation. The calculation process of the scaled dot product attention algorithm is as follows: calculate the matrix product between the transpose of the query matrix and the key matrix, divide by the arithmetic square root of the key vector dimension as the scaling factor, apply the softmax function to the result to obtain the attention weight matrix, and multiply the attention weight matrix with the value matrix to obtain the weighted representation vector. A temporal smoothing layer is set before the fully connected layer of the top-level quarterly-scale classifier. The temporal smoothing layer performs one-dimensional convolutional filtering on the quarterly-scale fused representation sequence along the time dimension. The filtering window length covers the typical transition duration of two adjacent evolutionary stages. The convolution kernel parameters are adaptively learned through the training process of the hierarchical temporal deep learning model.

8. The intelligent monitoring and analysis method for deep stress in geological disasters based on AI big data as described in claim 6, characterized in that, The training process of the hierarchical temporal deep learning model includes: Based on the marked precursor sections and confirmed non-precursor sections in the historical geological disaster event records, the composite long-period feature vectors of the corresponding time periods and the labeled labels are used to form training sample pairs. The classification cross-entropy loss function is used to jointly train the two output branches of the long-term precursor intelligent recognition result and the current evolution stage location. The classification cross-entropy loss function multiplies the loss terms of the two output branches by their corresponding weight coefficients and then sums them. The weight coefficient corresponding to the long-term precursor intelligent recognition result is greater than the weight coefficient corresponding to the current evolution stage location. The classification cross-entropy loss function introduces a category weight factor. The loss weight corresponding to the positive precursor sample is determined according to the reciprocal of the ratio of the number of positive to negative samples, and the loss weight corresponding to the negative sample is 1. The Adam optimization algorithm is used for optimization.

9. The intelligent monitoring and analysis method for deep stress in geological disasters based on AI big data according to claim 1, characterized in that, The generated intelligent monitoring and analysis results of deep stress include: When the long-period precursor intelligent identification result determines that the current signal is an abnormal evolution of a real precursor of a geological disaster, historical events with the same precursor classification label are retrieved from the historical geological disaster event records to obtain the average evolution cycle length of the same precursor pattern. Based on the ratio of the difference between the current time and the start time of the precursor signal to the average evolution period length, the relative position ratio of the current precursor evolution stage in the complete evolution period is calculated. The geological disaster precursor discrimination level is determined based on the classification confidence output by the hierarchical time-series deep learning model, and the risk warning level is determined based on the relative position ratio. The closer the relative position ratio is to 1, the higher the corresponding risk warning level. When multiple monitoring points within the same monitoring section simultaneously generate long-term precursor intelligent identification results, the classification confidence of each monitoring point is spatially consistent. Only when the geological disaster precursor discrimination level of spatially adjacent monitoring points reaches the preset spatial consistency condition will the corresponding risk warning level be upgraded to the appropriate level.

10. A smart monitoring and analysis system for deep stress in geological hazards based on AI big data, used to execute the smart monitoring and analysis method for deep stress in geological hazards based on AI big data as described in any one of claims 1 to 9, characterized in that, include: The multi-source big data acquisition module is used to acquire multi-source long-term big data sets of the deep geological disaster monitoring area. The multi-source long-term big data sets include multi-year continuous deep stress time series data, distributed optical fiber strain time series data and its historical cumulative records, core rheological test parameters, monitoring system calibration records, long-term environmental temperature change data and historical geological disaster event records. The rheological baseline generation module is used to calculate the viscoelastic constitutive parameter matrix of different lithological regions based on the rheological test parameters of the rock core, and to perform time integration on the historical strain accumulation data of each monitoring point according to the viscoelastic constitutive model to generate the rheological stress evolution baseline sequence of each monitoring point. The net anomaly trend generation module is used to obtain the rheological residual time series by subtracting the rheological stress evolution baseline sequence of the corresponding monitoring point from the multi-year continuous deep stress time series data point by point, decompose the rheological residual time series into trend-seasonal-residual three components and combine it with the monitoring system calibration record to perform system drift correction, and generate the net anomaly trend sequence. The composite feature extraction module is used to extract multi-scale trend change features by performing multi-scale wavelet transform on the net anomaly trend sequence, calculate the deviation rate between the rheological stress evolution baseline sequence and the measured deep stress value and generate a rheological deviation feature vector, and concatenate the multi-scale trend change features and the rheological deviation feature vector into a composite long-period feature vector. The precursor intelligent recognition module is used to input the composite long-period feature vector into the hierarchical time-series deep learning model and output the long-period precursor intelligent recognition result and the current evolution stage location. The monitoring and analysis result generation module is used to generate deep stress intelligent monitoring and analysis results based on the long-cycle precursor intelligent identification results and the current evolution stage location, and to associate the evolution cycle statistical characteristics of similar precursor patterns in the historical geological disaster event records.

Images