Monitoring and early warning method for coal rock fracture of argillaceous weakly cemented thick top coal roadway

Abstract

The invention discloses a monitoring and early warning method for coal rock fracture of an argillaceous weakly-cemented thick top coal roadway, and belongs to the field of top coal roadway coal rock fracture monitoring. The method comprises the following steps: constructing a prediction framework of long-term structural evolution, and judging a potential risk area of a long-term fracture trend; according to the potential risk area of the long-term fracture trend, integrating the feature data of multiple time scales, and obtaining a comprehensive feature vector of the change of the whole coal rock fracture process by applying a time sequence analysis method; dynamically adjusting parameters of an early warning model through an instant risk assessment result in combination with comprehensive data of medium-term stress change and long-term fracture trend, and determining risk grade division under multiple time scales; and according to risk grade division, generating graded early warning information, mapping the graded early warning information to a roadway spatial distribution model, and obtaining a risk visualization layer of a specific area.

Description

Technical Field

[0001] This invention belongs to the field of coal and rock fracture monitoring in top coal roadways, and particularly relates to a monitoring and early warning method for coal and rock fracture in thick top coal roadways with weak cementation in argillaceous soil. Background Technology

[0002] In the field of coal mining, the monitoring and early warning of coal and rock fracturing in roadways is directly related to safe production and efficient resource utilization, and is a crucial link in ensuring the stable operation of the mine. With the increase of mining depth, the special geological conditions of roadways with thin, weakly cemented, muddy roofs make the coal and rock fracturing problem increasingly complex, placing higher demands on the accuracy and timeliness of monitoring technology.

[0003] However, current mainstream monitoring methods have significant shortcomings when dealing with complex geological environments. These methods are often unable to adapt to the dynamic changes at different stages of coal and rock fracturing, especially when dealing with materials such as argillaceous weakly cemented coal and rock, which have heterogeneous mechanical properties and complex fracturing evolution patterns. Traditional technologies are inadequate in capturing multi-time-dimensional changes, which can easily lead to delayed or misjudgments in early warning.

[0004] A more serious challenge lies in the fact that the coal and rock fracturing process involves multi-timescale characteristics, ranging from the instantaneous expansion of microcracks to the long-term evolution of the overall structure, while existing technologies have significant shortcomings in comprehensive analysis along the time dimension. These shortcomings are first manifested in insufficient sensitivity to short-term instantaneous changes, making it difficult to capture the initiation and propagation signals of microcracks in real time; furthermore, the lack of effective integrated analysis of medium-term stress concentration and long-term structural evolution limits the predictive ability for the entire fracturing process. These two deficiencies are interconnected, preventing monitoring systems from achieving comprehensive early warning from local to overall, and from short-term to long-term, in complex environments.

[0005] Therefore, how to construct a monitoring technology that can integrate features of multiple time scales to accurately capture the changes in coal and rock fracturing in muddy, weakly cemented, thick-roofed coal roadways from instantaneous to long-term processes has become a key issue that urgently needs to be addressed. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a method for monitoring and early warning of coal and rock fracturing in argillaceous, weakly cemented thick-roofed coal roadways, comprising:

[0007] A high-precision sensor network is deployed based on the tunnel environment to acquire minute vibration signals inside the coal and rock mass.

[0008] Based on the minute vibration signal, short-time high-energy frequency components are extracted to obtain crack propagation characteristic signals;

[0009] Based on the crack propagation characteristic signals, a multi-scale risk early warning model is constructed, and the fracture risk level results are output.

[0010] Preferably, the step of acquiring minute vibration signals inside the coal and rock mass includes:

[0011] The sensor placement locations are determined based on the roadway lithology information to acquire multi-channel raw vibration signal data;

[0012] Based on the original vibration signal data, a time-frequency feature matrix is ​​obtained by applying moving average filtering and short-time Fourier transform.

[0013] Based on the time-frequency feature matrix, high-confidence anomalous components are filtered to obtain a microcrack signal set.

[0014] Preferably, the step of extracting short-time high-energy frequency components includes:

[0015] Based on the microcrack signal set, low-pass filtering and fast Fourier transform were applied to extract the key frequency components of 50-500Hz.

[0016] Based on the key frequency components, wavelet transform is used to analyze their energy standard deviation and identify dynamically changing frequency bands.

[0017] Based on the dynamically changing frequency bands, an LSTM neural network input sequence is constructed to obtain the dynamic features of short-term crack propagation.

[0018] Preferably, the step of constructing the multi-scale risk early warning model includes:

[0019] Based on the dynamic characteristics of short-term crack propagation and combined with the three-dimensional stress distribution model of the tunnel, the mid-term stress anomaly area is obtained;

[0020] Based on the stress anomaly region, geological structure and historical displacement data are integrated to form a structural evolution feature vector;

[0021] Based on the structural evolution feature vectors, a support vector machine model is trained to establish a fracture trend discriminant classifier.

[0022] Preferably, the step of forming the structural evolution feature vector includes:

[0023] Based on the node locations within the mid-term stress anomaly region, extract its 24-hour stress growth curve;

[0024] Based on the stress growth curve, the risk characteristic combination is constructed by integrating joint density and surrounding rock strength information;

[0025] Based on the combination of risk characteristics, principal component analysis is performed to obtain the dominant factors in structural evolution.

[0026] Preferably, the step of establishing the fracture trend discriminant classifier includes:

[0027] Based on the dominant factors of structural evolution, historical rupture sample data are labeled and a training dataset is constructed.

[0028] Based on the training dataset, radial basis function support vector machine is applied to train three types of rupture trends;

[0029] Based on the trained model, the current dominant factors are classified and labeled as stable, fluctuating, or trending upward.

[0030] Preferably, the step of outputting the rupture risk level result includes:

[0031] Based on trend labels, short-term dynamic features and medium-term stress results are integrated to construct a multi-scale fused feature vector;

[0032] Based on the fused feature vector, it is input into the entropy-weighted risk scoring model to calculate the comprehensive risk value;

[0033] Risk levels are determined based on the relationship between the overall risk value and a preset threshold.

[0034] Preferably, the step of constructing the multi-scale fused feature vector includes:

[0035] Frequency fluctuation values ​​are extracted based on short-term dynamic characteristics, and gradient change rates are extracted based on medium-term stress results.

[0036] Based on the fluctuation value and the rate of change, a combined feature matrix is ​​generated;

[0037] Based on the combined feature matrix, a standardization process is performed to form the final fusion vector that can be used in the scoring model.

[0038] On the other hand, the present invention also provides an electronic device including a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein the processor implements the method when executing the computing program.

[0039] On the other hand, the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method.

[0040] Compared with the prior art, the present invention has the following advantages and technical effects:

[0041] This invention discloses a multi-timescale coal and rock fracturing early warning method for roadways with weak cementation and thick roof coal seams. It acquires minute vibration signal data by deploying a high-precision sensor network, extracts short-term fracturing behavior features using signal processing techniques, analyzes medium-term stress changes using a stress distribution model, and predicts long-term fracturing trends by integrating geological parameters. This invention integrates multi-timescale feature data to construct a machine learning-based fracturing early warning model, enabling real-time short-term risk assessment, medium-term stress change analysis, and long-term fracturing trend prediction. It generates graded early warning information based on risk levels and maps it to a roadway spatial distribution model for visualization. This invention can dynamically adjust monitoring priorities and data acquisition strategies, effectively improving the accuracy and timeliness of coal and rock fracturing early warning, and providing important technical support for safe mine production. Attached Figure Description

[0042] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0043] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention. Detailed Implementation

[0044] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0045] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0046] Example 1

[0047] This embodiment provides a method for monitoring and early warning of coal and rock fracturing in a silty, weakly cemented thick-roofed coal roadway, including:

[0048] By deploying a high-precision sensor network, we can acquire minute vibration signal data within the coal and rock mass of a roadway with weak cementation in silty clay, covering short-term instantaneous change characteristics, and obtain a preliminary dataset of microcrack propagation signals.

[0049] Based on the preliminary microcrack propagation signal dataset, signal filtering and feature extraction methods were used to separate the key frequency components related to short-term instantaneous changes and determine the dynamic change feature set of short-term fracture behavior.

[0050] Based on the dynamic change characteristics of short-term rupture behavior and combined with the roadway stress distribution model, the local stress accumulation characteristics related to medium-term stress concentration are analyzed to obtain the distribution map of medium-term stress changes.

[0051] By integrating the distribution map of medium-term stress changes with tunnel geological parameters and historical deformation data, a prediction framework for long-term structural evolution is constructed to identify potential risk areas for long-term rupture trends.

[0052] Based on the potential risk areas of long-term fracturing trends, feature data from multiple time scales are integrated, and time series analysis methods are applied to obtain a comprehensive feature vector of changes in the entire process of coal and rock fracturing.

[0053] For the comprehensive feature vector, a rupture early warning model based on machine learning is constructed. If the short-term change component of the feature vector exceeds the preset threshold, a short-term early warning signal is triggered, and an immediate risk assessment result is obtained.

[0054] By combining real-time risk assessment results with comprehensive data on medium-term stress changes and long-term rupture trends, the parameters of the early warning model are dynamically adjusted to determine the risk level classification at multiple time scales.

[0055] Based on the risk level classification, hierarchical early warning information is generated and mapped onto the roadway spatial distribution model to obtain a risk visualization layer for a specific area;

[0056] For the risk visualization layer, the key monitoring areas of the sensor network are updated in real time. If an abnormal signal is detected, the data acquisition frequency is automatically adjusted to obtain more accurate dynamic change data of the rupture.

[0057] Optionally, by deploying a high-precision sensor network to acquire minute vibration signal data within the coal and rock mass of a silty, weakly cemented, thick-roofed coal roadway environment, covering short-term instantaneous change characteristics, a preliminary microcrack propagation signal dataset is obtained, including:

[0058] A vibration sensor array with a sampling rate of 100Hz was deployed in a silty, weakly cemented, thick-roofed coal roadway to collect vibration signals inside the coal and rock mass. High-frequency noise from the equipment was removed by moving average filtering to generate the original signal set.

[0059] Apply a Hanning window function to the original signal set and perform a short-time Fourier transform to extract the energy spectrum features of the 0-50Hz frequency band, forming a time-frequency feature matrix;

[0060] The time-frequency feature matrix is ​​input into a support vector machine with a preset radial basis kernel function. Cross-validation is used to divide the training set and the test set, and the output is a classification result set containing microcrack labels.

[0061] Abnormal signals with a confidence level greater than 90% are selected from the classification results set, and their peak intensity is calculated with Pearson correlation coefficient between roadway humidity and stress. When the absolute value of the correlation coefficient is greater than 0.7, it is determined to be a valid association.

[0062] The ARIMA model was used to predict the 72-hour trend of the effective correlation signal. The prediction results were matched with the daily average vibration energy in the historical database using a sliding window to identify data points that deviated from the mean by 3 times the standard deviation.

[0063] Based on the spatial distribution density of the deviation data points, the chi-square test is used to determine the optimal sampling interval of the sensor, and the sampling rate is increased to 200Hz in the active crack propagation region.

[0064] A Butterworth bandpass filter is used to filter the newly sampled data in the 5-45Hz frequency band, and the time-frequency feature matrix is ​​updated.

[0065] Optionally, based on the preliminary microcrack propagation signal dataset, signal filtering and feature extraction methods are used to separate key frequency components related to short-term instantaneous changes, and to determine the dynamic change feature set of short-term fracture behavior, including:

[0066] The original signal dataset of microcrack propagation was obtained from the initial data source, and high-frequency noise above 20kHz was filtered out using a Butterworth low-pass filter to obtain a set of denoised signals.

[0067] Perform a Fast Fourier Transform on the denoised signal set and extract the frequency components with amplitudes 3dB higher than the baseline in the 50-500Hz frequency band as key frequency components.

[0068] When the peak fluctuation of the key frequency component exceeds the preset threshold of 0.5V, the mixed signal is decomposed into 5 refined frequency sub-bands using the independent component analysis algorithm.

[0069] The Morlet wavelet basis function is applied to the refined frequency subbands for time-frequency transformation, and the energy standard deviation of each subband within a 0.1s time window is calculated.

[0070] Based on the linear relationship between energy standard deviation and crack propagation rate, the frequency band exceeding the critical value of 0.8 is determined to be a significant mode;

[0071] The center frequency and energy gradient corresponding to the saliency pattern are used as input features of the LSTM neural network to generate a dynamically changing feature set.

[0072] Optionally, the set of dynamic change characteristics for short-term rupture behavior, combined with a roadway stress distribution model, analyzes the local stress accumulation characteristics related to mid-term stress concentration to obtain a distribution map of mid-term stress changes, including:

[0073] The raw fracture signal collected by the roadway monitoring sensor is obtained, and then decomposed into 5 layers using the db4 wavelet basis function. The low-frequency coefficients of 3-5 layers are retained to reconstruct the signal and obtain the denoised fracture signal.

[0074] The denoised fracture signal is subtracted from the mean for each channel and then divided by the standard deviation to output a standardized fracture feature vector.

[0075] Import the tunnel CAD model into ABAQUS, set the parameters of Q235 steel, apply the measured in-situ stress boundary conditions, perform static analysis and extract the VonMises stress of the nodes to form a stress distribution array;

[0076] Using the material yield strength as a threshold, NumPy is used to filter out data points exceeding the limit in the stress array, and their grid numbers and coordinates are marked.

[0077] The griddata function of scipy is used to perform linear interpolation on discrete stress points to generate a stress matrix with a 100×100 grid.

[0078] Input the stress matrix into Matplotlib’s contourf function, set 20 color levels to display the stress range of 0-300MPa, and output the stress contour plot.

[0079] For regions in the graph where the gradient is greater than 5 MPa / m, the stress monitoring data of the nodes in that region over the past 24 hours is retrieved, and the slope parameter is fitted using the least squares method to output the stress growth rate of the abnormal region.

[0080] Optionally, the process of constructing a predictive framework for long-term structural evolution by integrating medium-term stress change distribution maps with tunnel geological parameters and historical deformation data, and identifying potential risk areas for long-term rupture trends, includes:

[0081] The first three principal components after principal component analysis were extracted from the mid-term stress distribution map and used as key feature values. Combined with the rock mass strength and joint density parameters of the tunnel environment, a standardized initial comprehensive dataset was formed.

[0082] Based on the standardized dataset, the historical data of roof subsidence stored in the database were integrated. The radial basis function in the support vector machine algorithm was used, and the penalty coefficient C was determined to be 1.0 through 5-fold cross-validation. The structural evolution trend was divided into three categories: stable, fluctuating, and rising.

[0083] After obtaining the classification results, SPSS software is used to calculate the Pearson correlation coefficient for the rock mass strength parameters and the change pattern category. When the absolute value of the correlation coefficient is greater than 0.7, it is determined to be a strong correlation direction.

[0084] If the evolution direction is upward and the correlation coefficient is greater than 0.7, then a threshold of 1.5 is set based on the ratio of the maximum stress value to the safety factor in historical cases, and the risk level is divided into three levels: red, orange and yellow.

[0085] Based on the risk level, new stress data are acquired from the microseismic monitoring system every 24 hours. The DBSCAN clustering algorithm is used to merge the measuring points with an Euclidean distance of less than 2MPa to determine the spatial boundary of the high-risk area.

[0086] For the boundary coordinates of high-risk areas, a network of displacement sensors is deployed to collect data every 8 hours. The daily displacement increment is fitted using the ARIMA model in MATLAB. When the predicted value exceeds the threshold of 1.5, the training data update cycle of the support vector machine is automatically shortened from 7 days to 1 day.

[0087] Optionally, based on the potential risk areas of long-term fracturing trends, the feature data of multiple time scales are integrated, and time series analysis methods are applied to obtain a comprehensive feature vector of changes throughout the entire process of coal and rock fracturing, including:

[0088] The raw signals at multiple time scales were obtained from coal and rock fracture monitoring data, and median filtering was used to remove impulse noise. The first dataset was then obtained by Z-score standardization.

[0089] Input the first dataset into the STL time series decomposition tool to separate the trend term, periodic term and residual term, mark points where the residual exceeds 3 times the standard deviation as outliers, and output the second dataset.

[0090] Based on the magnitude of the trend and periodic terms in the second dataset, wavelet transform is used to extract energy features at three scales: 24 hours, 72 hours, and 168 hours, which are then combined into a third feature matrix.

[0091] Calculate the Euclidean distance of each sampling point in the third feature matrix. If the distance exceeds the preset threshold of 0.85, it is marked as a high-risk area and a risk level label is generated.

[0092] In the time series corresponding to the risk area label, the Bayesian change point detection algorithm is used to locate the trend change point and output the change point timestamp and change direction.

[0093] Using data from 30 minutes before and after the mutation point as positive and negative samples, and the third feature matrix as input, a support vector machine classifier is trained to predict the risk probability of new sampling points in the next 6 hours.

[0094] When the predicted probability of a new data point exceeds 0.7, the monitoring device is triggered to increase the sampling frequency from 10 minutes to 1 minute and generate a dynamic monitoring instruction table.

[0095] Optionally, for the comprehensive feature vector, a machine learning-based rupture early warning model is constructed. If the short-term change component of the feature vector exceeds a preset threshold, a short-term early warning signal is triggered to obtain an immediate risk assessment result, including:

[0096] Three dimensions of features—time-domain mean, frequency-domain wavelet energy, and nonlinear entropy—are extracted from the raw data to form an initial set of feature vectors.

[0097] After sorting the feature vectors by timestamp, the STL decomposition algorithm is used to separate the trend term, seasonal term and residual term, and the residual component is obtained as the short-term fluctuation feature set.

[0098] Calculate the standard deviation of each residual component. If the deviation exceeds the preset threshold 2σ, mark the component as an abnormal state.

[0099] After converting the anomaly labels into numerical vectors, they are input into a pre-trained RBF kernel support vector machine. This model is trained using historical rupture event data and outputs a binary classification result of high risk or low risk.

[0100] The weights of each feature are determined by the entropy weight method. The weighted summation of the risk classification results and the deviation value are used to obtain a comprehensive risk score. The score range of 0-100 is divided into three risk levels: red, yellow and blue.

[0101] Based on the risk level, a corresponding early warning signal is triggered. At the same time, a dynamic time warping algorithm is used to match similar risk patterns in the historical database and update the risk label of the current feature vector.

[0102] The latest feature vectors and risk labels are stored in the database to complete the closed-loop update.

[0103] Optionally, the step of dynamically adjusting the parameters of the early warning model based on real-time risk assessment results, combined with comprehensive data on medium-term stress changes and long-term rupture trends, to determine the risk level classification across multiple time scales includes:

[0104] Real-time data is acquired from multiple sources of sensors and merged with historical stress change records to generate a raw dataset.

[0105] Missing values ​​are imputed and units are standardized in the original dataset to form a preprocessed dataset;

[0106] The Z-score method is used to normalize the preprocessed dataset to generate a standardized data matrix;

[0107] Calculate the sliding variance of the standardized data matrix in each time window. If the variance exceeds 1.5 times the historical variance mean, it is marked as an abnormal time scale.

[0108] Extract the mean, range, and frequency domain energy within the abnormal time scale as feature vectors;

[0109] Align the feature vectors with the historical database by timestamp and calculate the cosine similarity to match high-risk patterns.

[0110] Matches with a result greater than 0.7 are marked as high-risk.

[0111] Adjust the weight parameters of the early warning model according to the risk level;

[0112] At high-risk levels, the frequency domain energy weight is increased by 20%, and the model parameters are updated using gradient descent.

[0113] When the Euclidean distance between the new parameters and the original parameters is greater than 0.3, refit the least squares solution of the model;

[0114] The calibrated model is used to scan the fused data matrix and output the risk probability at each time scale.

[0115] When the probability of risk is greater than 80%, it is determined to be a high-risk level.

[0116] Optionally, the step of generating graded early warning information based on risk level classification, mapping it to the roadway spatial distribution model, and obtaining a risk visualization layer for a specific area includes:

[0117] Risk level data is obtained from the data source, and the random forest algorithm is used to classify the data, outputting preliminary risk level classification results;

[0118] Based on the classification results, hierarchical early warning information is generated, and the early warning level and content description fields are stored in a structured manner.

[0119] When the risk value in the early warning information exceeds the threshold of 0.7, the roadway number field in the early warning information is extracted and its attributes are linked with the GIS layer of the roadway spatial data.

[0120] The spatial point density is analyzed using the DBSCAN clustering algorithm to construct a risk hotspot distribution model;

[0121] High-density areas with a cluster radius of less than 50 meters are extracted from the distribution model, and the early warning information is matched to the center point of the corresponding lane polygon using ArcGIS's feature conversion tool.

[0122] A risk intensity distribution layer is generated using the MapboxGLJS heatmap rendering module;

[0123] If data gaps appear in the heatmap, the Kriging interpolation algorithm is used to fill in the missing grids and output the final risk visualization layer.

[0124] Write the latitude and longitude coordinates, risk values, and timestamp fields from the layer into a PostgreSQL database to complete the storage of the spatial distribution report.

[0125] Optionally, the risk visualization layer is updated in real time with the key monitoring areas of the sensor network. If an abnormal signal is detected, the data acquisition frequency is automatically adjusted to obtain more accurate dynamic change data of the rupture, including:

[0126] Temperature, vibration, and deformation data of the monitored area are acquired through a sensor network to form an initial environmental state matrix;

[0127] When any sensor data deviates from the baseline value by more than twice the standard deviation, the sampling frequency of that node is increased from 1Hz to 10Hz.

[0128] Perform a sliding window Fourier transform on the high-frequency acquired data to extract the mean amplitude within the 0-100Hz frequency band as a feature vector;

[0129] When the cosine similarity between the feature vector and the baseline spectrum is less than 0.7, it is marked as an anomalous time segment;

[0130] For areas where the cumulative abnormal time segment exceeds 30 minutes, multi-sensor collaborative monitoring will be initiated.

[0131] Redundant nodes are deployed within a 50-meter radius of the anomaly point to generate three-dimensional deformation field data with millimeter-level accuracy;

[0132] Based on the deformation field data, the daily displacement increment is calculated. When the vertical displacement rate exceeds 5 mm / day and the horizontal displacement vector direction is consistent, a level three risk zone is defined.

[0133] The risk zone boundary coordinates are projected onto the GIS map in real time, and the risk level is marked with red, yellow and blue colors.

[0134] The GIS engine automatically calculates the sensor blind spots covered by the risk area and schedules mobile monitoring nodes to fill the blank grid.

[0135] The newly added node acquires multi-axis acceleration data at a frequency of 20Hz, and Kalman filtering is used to eliminate measurement noise.

[0136] For data from a Level 3 risk zone lasting 72 hours, 12-dimensional features, including time-domain peak value, frequency-domain energy, and displacement covariance, were extracted.

[0137] The LibSVM tool is used to perform binary classification and output the results of rock mass fracture or equipment failure.

[0138] Example 2

[0139] like Figure 1 As shown, this embodiment provides a method for monitoring and early warning of coal and rock fracturing in a silty, weakly cemented thick-roofed coal roadway, including:

[0140] A high-precision sensor network is deployed based on the tunnel environment to acquire minute vibration signals inside the coal and rock mass.

[0141] Based on the minute vibration signal, short-time high-energy frequency components are extracted to obtain crack propagation characteristic signals;

[0142] Based on the crack propagation characteristic signals, a multi-scale risk early warning model is constructed, and the fracture risk level results are output.

[0143] Specifically:

[0144] Step S101: By deploying a high-precision sensor network, for the environment of a silty, weakly cemented, thick-roofed coal roadway, small vibration signal data in the coal and rock mass are acquired, covering short-term instantaneous change characteristics, to obtain a preliminary microcrack propagation signal dataset.

[0145] A vibration sensor array with a sampling rate of 100Hz was deployed in a thick-top coal roadway with weak cementation in silty clay to collect vibration signals from within the coal and rock mass. High-frequency noise from the equipment was removed using a moving average filter to generate an original signal set. A Hanning window function was applied to the original signal set for short-time Fourier transform to extract the energy spectrum features in the 0-50Hz frequency band, forming a time-frequency feature matrix. This matrix was input into a support vector machine with a preset radial basis function kernel, and cross-validation was used to divide the training and testing sets, outputting a classification result set containing microcrack labels. Abnormal signals with a confidence level greater than 90% were selected from the classification result set, and their peak intensity was calculated using the Pearson correlation coefficient between the signal and roadway humidity and stress. A correlation coefficient with an absolute value greater than 0.7 was considered a valid association. An ARIMA model was used to predict the 72-hour trend of the validly associated signals. The prediction results were matched with the daily average vibration energy in the historical database using a sliding window to identify data points deviating from the mean by three standard deviations. Based on the spatial distribution density of the deviating data points, a chi-square test was used to determine the optimal sampling interval for the sensors, increasing the sampling rate to 200Hz in areas with active crack propagation. A Butterworth bandpass filter is used to filter the newly sampled data in the 5-45Hz frequency band, and the time-frequency feature matrix is ​​updated.

[0146] Specifically, a high-precision sensor network was deployed in a 500-meter-long and 50-meter-wide monitoring area within a silty, weakly cemented, thick-roofed coal roadway. A vibration sensor was placed every 5 meters, totaling 1000 sensors, each with a sensitivity of 0.001 mm / s, to capture minute vibration signals within the coal and rock mass. The data acquisition frequency was set to 1000 times per second to cover short-term transient changes and ensure the capture of microsecond-level vibration fluctuations. Next, the acquired vibration signals were analyzed in the frequency domain using a Fast Fourier Transform (FFT) algorithm, decomposing the signals into different frequency components. Feature signals with frequencies ranging from 0.5 Hz to 50 Hz were extracted, and background noise was filtered out using wavelet denoising. The wavelet decomposition was set to 5 levels with a threshold of 0.02 to improve the signal-to-noise ratio. The analysis results showed that the dominant frequency of the vibration signal was concentrated around 10 Hz, indicating potential precursory features of microcrack propagation. Subsequently, a microcrack propagation signal dataset was constructed. The processed vibration signal data was stored in time series format and classified using the Support Vector Machine (SVM) algorithm from machine learning. The kernel function was set as a radial basis function, the penalty parameter C was set to 1.0, and the ratio of training set to test set was 7:3. The classification accuracy reached over 85%, and the characteristic patterns of microcrack propagation signals were preliminarily identified. For example, signals with an amplitude exceeding 0.005 mm / s and a duration exceeding 0.1 seconds were marked as potential crack propagation events. Finally, through time-space correlation analysis, the propagation speed of the signal between different sensors was calculated. A speed threshold of 200 m / s was set, and the initial location of crack propagation was determined using a triangulation algorithm, with the positioning error controlled within 0.5 meters.

[0147] Step S102: Based on the preliminary microcrack propagation signal dataset, signal filtering and feature extraction methods are used to separate the key frequency components related to short-term instantaneous changes and determine the dynamic change feature set of short-term fracture behavior.

[0148] The original signal dataset of microcrack propagation was obtained from the initial data source. A Butterworth low-pass filter was used to remove high-frequency noise above 20kHz, resulting in a denoised signal set. A Fast Fourier Transform (FFT) was performed on the denoised signal set to extract frequency components with amplitudes exceeding the baseline by 3dB within the 50-500Hz band as key frequency components. When the peak fluctuation of the key frequency components exceeded a preset threshold of 0.5V, an independent component analysis (ICA) algorithm was used to decompose the mixed signal into five refined frequency sub-bands. A Morlet wavelet basis function was applied to the refined frequency sub-bands for time-frequency transformation, and the energy standard deviation of each sub-band within a 0.1s time window was calculated. Based on the linear relationship between the energy standard deviation and the crack propagation rate, frequency bands exceeding the critical value of 0.8 were identified as significant patterns. The center frequency and energy gradient corresponding to the significant patterns were used as input features of an LSTM neural network to generate a dynamically changing feature set.

[0149] Specifically, for the processing of the microcrack propagation signal dataset, the original data was first preprocessed using signal filtering techniques. A Butterworth low-pass filter was used with a cutoff frequency of 50 Hz to remove high-frequency noise and retain signal components related to short-term transient changes. Assuming the original signal sampling rate was 1000 Hz, spectral analysis of the filtered signal showed that the frequency components below 50 Hz accounted for 85% of the energy, indicating that the filtering effectively preserved the main information. Next, feature extraction was performed using wavelet transform with the Daubechies wavelet basis function (db4) and a decomposition level of 5. The wavelet coefficients of each level were extracted, and their energy distribution was calculated. It was found that the energy of the coefficients in the 3rd and 4th levels was concentrated in the 10 to 30 Hz range, accounting for 70% of the total energy, indicating that this frequency band is highly correlated with short-term transient changes. Subsequently, for the separation of key frequency components, the Fast Fourier Transform (FFT) algorithm was used to transform the signal into the frequency domain, extracting the dominant frequency components in the 10-30 Hz range. Combined with power spectral density analysis, significant peaks were found at 15 Hz and 25 Hz, with power densities reaching 0.45 and 0.38, respectively, indicating that these two frequencies are the main driving factors of short-term rupture behavior. Finally, the dynamic change feature set of short-term rupture behavior was determined. Principal component analysis (PCA) was used to reduce the dimensionality of the extracted frequency components and energy features, retaining the first two principal components, explaining 90% of the variance. Combined with the detection of abrupt change points in the time-domain signal (using the CUSUM algorithm with a threshold set to 3 times the standard deviation), the abrupt change times of short-term rupture behavior in the signal were identified. For example, significant changes were detected at 200 ms and 450 ms, with amplitude change rates reaching 2.5 and 3.1, respectively. A dynamic change set containing frequency, energy, and time-domain abrupt change features was formed.

[0150] Step S103: Based on the dynamic change characteristics set of short-term rupture behavior, and combined with the roadway stress distribution model, analyze the local stress accumulation characteristics related to mid-term stress concentration, and obtain the distribution map of mid-term stress change.

[0151] The raw fracture signals collected by the tunnel monitoring sensors were acquired and decomposed into 5 levels using the db4 wavelet basis function. The low-frequency coefficients of levels 3-5 were retained to reconstruct the signals, resulting in denoised fracture signals. The denoised fracture signals were then subtracted from their mean values ​​for each channel and divided by the standard deviation to output standardized fracture feature vectors. The tunnel CAD model was imported into ABAQUS, Q235 steel parameters were set, and measured in-situ stress boundary conditions were applied. Static analysis was performed to extract VonMises stresses at the nodes, forming a stress distribution array. Using the material yield strength as a threshold, NumPy was used to filter out data points exceeding the limit in the stress array, marking their grid numbers and coordinates. The SciPy griddata function was used to perform linear interpolation on the discrete stress points, generating a 100×100 grid stress matrix. The stress matrix was then input into Matplotlib's contourf function, setting 20 color levels to display the 0-300MPa stress range and outputting a stress contour map. For regions in the graph where the gradient is greater than 5 MPa / m, the stress monitoring data of the nodes in that region over the past 24 hours is retrieved, and the slope parameter is fitted using the least squares method to output the stress growth rate of the abnormal region.

[0152] Specifically, regarding the dynamic change characteristics of short-term fracturing behavior, combined with the roadway stress distribution model, we analyze the local stress accumulation characteristics related to medium-term stress concentration and obtain the distribution map of medium-term stress changes. This can be achieved through the following specific implementation method: First, we use numerical simulation software (such as FLAC3D) to dynamically model the short-term fracturing behavior of the rock mass surrounding the roadway. Assuming a roadway depth of 800 meters, a rock mass density of 2500 kg / m³, and an initial ground stress of 20 MPa, we calculate the fracturing characteristics using the built-in Mohr-Coulomb criterion. This yields stress abrupt change data within a 2-meter radius of the roadway fracturing zone. The time step is set to 0.1 seconds, and the total simulation duration is 10 seconds, generating a dynamic change characteristic set data table. Then, these data were combined with the roadway stress distribution model, and the finite element method was used. The roadway width was set to 5 meters, the height to 3 meters, the elastic modulus of the surrounding rock mass to 30 gigapascals, and the Poisson's ratio to 0.25. Three-dimensional mesh data of the stress distribution around the roadway was obtained through iterative calculation, with mesh cell sizes of 0.5 meters × 0.5 meters × 0.5 meters. The calculation results showed that the stress concentration factor at the top of the roadway was 1.5, and at the bottom it was 1.2. Next, regarding the local stress accumulation characteristics related to mid-term stress concentration, based on the above mesh data, a time series analysis algorithm was used, setting the time window to 24 hours, to calculate the local stress accumulation rate. Assuming an accumulation rate of 0.02 megapascals per hour, combined with historical monitoring data, it was found that the stress accumulation value within 5 meters of the top of the roadway reached 30 megapascals, indicating a significant concentration phenomenon. Finally, a medium-term stress variation distribution map was plotted using visualization tools (such as MATLAB). The stress value range was set to 10 to 50 MPa, and the map was displayed in the form of a heat map. The intensity of the stress was represented by the color, with the top concentrated area in red and the value marked as 30 MPa, and the bottom area in yellow and the value marked as 15 MPa, forming an intuitive stress distribution characteristic map.

[0153] Step S104: By integrating the distribution map of medium-term stress changes with tunnel geological parameters and historical deformation data, a prediction framework for long-term structural evolution is constructed to determine potential risk areas for long-term rupture trends.

[0154] The first three principal components extracted from the mid-term stress distribution map after principal component analysis were used as key feature values. Combined with the rock mass strength and joint density parameters of the tunnel environment, a standardized initial comprehensive dataset was formed. Based on the standardized dataset, historical roof subsidence data stored in the database were integrated. Using the radial basis function in the support vector machine algorithm, a penalty coefficient C of 1.0 was determined through 5-fold cross-validation. The structural evolution trend was classified into three categories: stable, fluctuating, and rising. After obtaining the classification results, Pearson correlation coefficients were calculated using SPSS software for the rock mass strength parameters and change pattern categories. A strong correlation direction was determined when the absolute value of the correlation coefficient was greater than 0.7. If the evolution direction was rising and the correlation coefficient was greater than 0.7, a threshold of 1.5 was set based on the ratio of the maximum stress value to the safety factor in historical cases, and the risk level was divided into three levels: red, orange, and yellow. According to the risk level ranking, new stress data were obtained from the microseismic monitoring system every 24 hours. The DBSCAN clustering algorithm was used to merge measuring points with an Euclidean distance of less than 2 MPa to determine the spatial boundary of high-risk areas. For the boundary coordinates of high-risk areas, a network of displacement sensors is deployed to collect data every 8 hours. The daily displacement increment is fitted using the ARIMA model in MATLAB. When the predicted value exceeds the threshold of 1.5, the training data update cycle of the support vector machine is automatically shortened from 7 days to 1 day.

[0155] Specifically, by constructing a comprehensive predictive framework to identify potential risk areas for long-term fracturing trends, the following steps are taken. First, a mid-term stress distribution map is used, and stress field simulation is performed using finite element analysis software such as ANSYS. Assuming the rock mass elastic modulus of the tunnel area is 25 GPa, Poisson's ratio is 0.3, and the applied boundary stress is 10 MPa, the maximum principal stress value of the stress concentration area is calculated to be 15 MPa. A stress distribution map is then plotted, revealing that the stress concentration area is mainly located within 5 meters of the top of the tunnel. Next, tunnel geological parameters are integrated. Geological exploration data, such as a rock layer thickness of 10 meters, a fault dip angle of 30 degrees, and a ore body burial depth of 500 meters, are input into the geological modeling software GeoModeller to generate a three-dimensional geological model. This model is then overlaid with the stress distribution map for analysis, identifying anomaly zones near the faults, where stress values ​​exceed the average by 20%, indicating a potential fracturing risk. Subsequently, combining historical deformation data, assuming the tunnel roof subsidence was 0.5 meters per year over the past 5 years, the ARIMA time-series analysis model was used to predict the deformation trend over the next 3 years. The predicted subsidence rate could increase to 0.6 meters per year. Combined with the location of stress concentration zones, the risk of rupture in the roof area was determined to be high. Finally, a long-term structural evolution prediction framework was constructed. Using machine learning algorithms such as random forests, stress values, geological parameters, and deformation rates were used as input features to train the model to predict rupture probabilities. It was assumed that the predicted rupture probability was 75% near the fault, while it was only 30% in other areas. This identified the area near the fault as a high-risk zone and generated a risk distribution map, providing a basis for subsequent reinforcement measures.

[0156] Step S105: Based on the potential risk areas of long-term fracturing trends, integrate feature data from multiple time scales and apply time series analysis methods to obtain a comprehensive feature vector of changes throughout the entire coal and rock fracturing process.

[0157] Raw signals at multiple time scales were obtained from coal and rock fracture monitoring data. Median filtering was used to remove impulse noise, and Z-score standardization was applied to obtain the first dataset. The first dataset was input into an STL time series decomposition tool to separate the trend, periodic, and residual terms. Points with residuals exceeding three times the standard deviation were marked as outliers, resulting in the second dataset. Based on the amplitudes of the trend and periodic terms in the second dataset, wavelet transform was used to extract energy features at three scales: 24 hours, 72 hours, and 168 hours, which were combined to form a third feature matrix. The Euclidean distance of each sampling point in the third feature matrix was calculated. If the distance exceeded a preset threshold of 0.85, it was marked as a high-risk area, generating a risk level label. On the time segment corresponding to the risk area label, a Bayesian change point detection algorithm was used to locate trend abrupt changes, outputting the timestamp and direction of the abrupt change. Data from 30 minutes before and after the abrupt change point were used as positive and negative samples, and the third feature matrix was used as input to train a support vector machine classifier to predict the risk probability of new sampling points within the next 6 hours. When the predicted probability of a new data point exceeds 0.7, the monitoring device is triggered to increase the sampling frequency from 10 minutes to 1 minute and generate a dynamic monitoring instruction table.

[0158] Specifically, to obtain the comprehensive feature vector of changes throughout the entire process of coal and rock fracturing, the process first involves collecting data on potential risk areas with long-term fracturing trends using multi-source sensors. For example, stress sensors, acoustic sensors, and displacement sensors are set up in a mining area, with collection frequencies of once per minute, 10 times per second, and once per hour, respectively. The data covers a 1000m × 1000m area within the mining area, and continuous monitoring is performed for 30 days, generating a time-series dataset containing stress values ​​(e.g., a maximum value of 50MPa), acoustic frequencies (e.g., a peak frequency of 200Hz), and displacement (e.g., a maximum displacement of 5.2mm). Next, the feature data from multiple time scales are integrated. The data from the different frequencies are unified to a time scale of once per minute using a time alignment algorithm (e.g., linear interpolation). The mean, variance, and peak characteristics of each sensor data are calculated. For example, the 30-day mean of stress data is 30.5MPa with a variance of 8.2, the mean of acoustic frequency is 150Hz, and the mean of displacement is 2.1mm, forming a multi-dimensional feature matrix. Then, time series analysis methods are applied, such as using an autoregressive moving average (ARIMA) model to model the feature matrix. Model parameters p=2, d=1, and q=1 are set. By fitting historical data, the trend of rupture risk in the next 24 hours is predicted, with a mean prediction error of 0.8 MPa. Wavelet transform is used to extract high-frequency anomalous signals from the data (such as signals with frequency abrupt changes exceeding 50 Hz) to identify potential rupture precursors. Finally, based on the above analysis results, a feature vector for the entire coal and rock rupture process is constructed. The feature values ​​of stress, acoustic waves, and displacement are standardized and weighted, with weights of 0.4, 0.3, and 0.3 respectively, to obtain a comprehensive feature vector. For example, the vector value at a certain moment is [0.75, 0.62, 0.58]. The support vector machine (SVM) algorithm is then used to classify the vector and determine the rupture risk level, achieving a classification accuracy of 85%, thus providing data support for subsequent early warning.

[0159] Step S106: For the comprehensive feature vector, a machine learning-based rupture early warning model is constructed. If the short-term change component of the feature vector exceeds a preset threshold, a short-term early warning signal is triggered, and an immediate risk assessment result is obtained.

[0160] Three dimensions of features—time-domain mean, frequency-domain wavelet energy, and nonlinear entropy—are extracted from the raw data to form an initial feature vector set. After sorting the feature vectors by timestamp, the STL decomposition algorithm is used to separate the trend term, seasonal term, and residual term, obtaining the residual components as the short-term fluctuation feature set. The standard deviation of each residual component is calculated; if the deviation exceeds a preset threshold of 2σ, the component is marked as an anomalous state. The anomalous labels are converted into numerical vectors and input into a pre-trained RBF kernel support vector machine. This model is trained using historical rupture event data and outputs a binary classification result of high or low risk. The weights of each feature are determined based on the entropy weight method, and the weighted sum of the risk classification result and the deviation value yields a comprehensive risk score. The score range of 0-100 is divided into three risk levels: red, yellow, and blue. Corresponding warning signals are triggered based on the risk level, and a dynamic time warping algorithm is used to match similar risk patterns in the historical database, updating the risk label of the current feature vector. The latest feature vector and risk label are stored in the database, completing the closed-loop update.

[0161] Specifically, a machine learning-based rupture early warning model is constructed based on the comprehensive feature vector. First, data preprocessing is performed to integrate vibration, stress, and temperature data collected from multiple sensors into a high-dimensional feature vector, assuming it contains 10 dimensions, such as a vibration frequency of 50.2 Hz, a stress value of 120.5 MPa, and a temperature of 35.8 degrees Celsius. This data undergoes standardization (e.g., Z-score normalization, using the formula x' = (x - μ) / σ, where μ is the mean and σ is the standard deviation; the standardized vibration frequency value is 0.85 after calculation) to eliminate the influence of dimensions and form a unified feature vector. Next, a machine learning model is constructed using the Support Vector Machine (SVM) algorithm, with the kernel function set to Radial Basis Function (RBF), parameter C of 1.0, and gamma of 0.1. The model is trained using historical data. Assuming the training set contains 1000 samples, with rupture samples accounting for 10%, the model achieves an accuracy of 92.3% on the validation set, indicating strong classification ability. Then, for the short-term change components of the feature vector, a sliding window analysis method is designed with a window size of 5 minutes. The rate of change of each dimension feature is calculated. For example, if the vibration frequency changes from 50.2 Hz to 52.1 Hz within 5 minutes, the rate of change is 3.78%. If the preset threshold is 3.5%, then this dimension triggers an early warning condition. Furthermore, the rate of change of all dimensions is weighted and summed (weights are based on feature importance analysis, with vibration frequency having a weight of 0.4, stress 0.3, and temperature 0.3), resulting in a comprehensive change index of 3.9%, which exceeds the preset comprehensive threshold of 3.0%, triggering a short-term early warning signal. Finally, the system predicts the real-time risk assessment results through the model. Assuming that the current feature vector, after being input into the SVM model, outputs a rupture probability of 78.6%, combined with the early warning signal, the system automatically generates a high-risk warning and records it in the log database.

[0162] Step S107: Based on the real-time risk assessment results and combined with comprehensive data on medium-term stress changes and long-term rupture trends, dynamically adjust the parameters of the early warning model to determine the risk level classification under multiple time scales.

[0163] Real-time data is acquired from multiple sensors and merged with historical stress change records to generate a raw dataset. Missing values ​​are imputed and units are standardized in the raw dataset to form a preprocessed dataset. The Z-score method is used to normalize the preprocessed dataset, generating a standardized data matrix. The sliding variance of the standardized data matrix is ​​calculated for each time window; if the variance exceeds 1.5 times the historical variance mean, it is marked as an abnormal time scale. The mean, range, and frequency domain energy within the abnormal time scale are extracted as feature vectors. The feature vectors are aligned with the historical database by timestamp, and cosine similarity is calculated to match high-risk patterns. A matching result greater than 0.7 is marked as a high-risk level. The weight parameters of the early warning model are adjusted according to the risk level. Under the high-risk level, the frequency domain energy weight is increased by 20%, and the model parameters are updated using gradient descent. When the Euclidean distance between the new parameters and the original parameters is greater than 0.3, the least squares solution of the model is refitted. The calibrated model is used to scan the fused data matrix and output the risk probability at each time scale. A risk probability greater than 80% is determined as the final high-risk level.

[0164] Specifically, in constructing the dynamic early warning model, the first step is to obtain risk data under the current environment through real-time risk assessment results. For example, if the crustal stress value monitored by seismic sensors in a certain area is 5.3 MPa, exceeding the normal range of 4.0 MPa, the risk index is calculated to be 1.32 using an algorithm (calculated by dividing the current stress value by the normal value). This index is then compared with historical data, indicating that the current risk is at a moderately high level. Next, combined with medium-term stress change data, assuming that the stress value has gradually increased from 4.5 MPa to 5.3 MPa over the past 30 days, a linear regression algorithm is used to predict the stress change trend for the next 7 days, yielding a predicted value of 5.5 MPa, indicating a continued increase in risk. Then, long-term rupture trend analysis is incorporated. Based on geological data from the past 10 years, a rupture probability of 0.15 is calculated (using a Poisson distribution model, combined with historical rupture frequency and time intervals). A weighted algorithm (40% for the immediate risk index, 35% for the medium-term trend, and 25% for the long-term probability) is used to synthesize the total risk value, resulting in a total risk value of 1.18. Next, the parameters of the early warning model were dynamically adjusted, raising the risk threshold from 1.0 to 1.2 to adapt to the current high-risk environment. This adjustment was based on a deviation analysis of the total risk value from historical thresholds to ensure model sensitivity. Finally, based on comprehensive data across multiple time scales, the risk level was divided into three levels: immediate risk (medium, 1.32), medium-term risk (high, predicted value 5.5 MPa), and long-term risk (low, probability 0.15). The classification results were automatically generated using a logistic regression algorithm and mapped to the early warning system.

[0165] Step S108: Based on the risk level classification, generate graded early warning information, map it onto the roadway spatial distribution model, and obtain a risk visualization layer for a specific area.

[0166] Risk level data is obtained from the data source, and the random forest algorithm is used to classify the data, outputting preliminary risk level classification results. Based on the classification results, tiered early warning information is generated, and the early warning level and content description fields are stored in a structured manner. When the risk value in the early warning information exceeds the threshold of 0.7, the roadway number field is extracted from the early warning information and linked to the GIS layer of the roadway spatial data. The spatial point density is analyzed using the DBSCAN clustering algorithm to construct a risk hotspot distribution model. High-density areas with a cluster radius of less than 50 meters are extracted from the distribution model, and the early warning information is matched to the corresponding roadway polygon center point using ArcGIS's feature conversion tool. A risk intensity distribution layer is generated using the MapboxGLJS heatmap rendering module. If data holes appear in the heatmap, the Kriging interpolation algorithm is used to complete the missing grid, outputting the final risk visualization layer. The latitude and longitude coordinates, risk values, and timestamp fields from the layer are written to a PostgreSQL database to complete the spatial distribution report storage.

[0167] Specifically, in the technical process of constructing risk level classification and mapping graded early warning information to the roadway spatial distribution model and generating a risk visualization layer for a specific area, the environmental parameters in the roadway are first analyzed through data collection and risk assessment algorithms. Assuming that the gas concentration data of a certain roadway area is 2.5% and the temperature is 35 degrees Celsius, combined with the risk assessment model, the area with a gas concentration exceeding 2.0% and a temperature exceeding 30 degrees Celsius is defined as a high-risk level, and the risk index is calculated to be 0.85 (out of 1.0). Through a weighted algorithm (gas concentration weight 0.6, temperature weight 0.4), the comprehensive risk value is obtained as 0.85*0.6+0.85*0.4=0.85, which belongs to the high-risk range (0.8-1.0). Subsequently, based on the risk index, a tiered early warning system is generated. The system automatically maps risk values ​​to red high-risk warning signals and associates the warning information with the roadway spatial distribution model through a data interface. Assuming the roadway is 1000 meters long and the risk area is located between 300 and 500 meters, the system uses GIS technology to mark this section as a high-risk area, generating a spatial distribution data layer. Next, for the generation of a risk visualization layer for a specific area, the system calls a 3D modeling algorithm to overlay the high-risk area between 300 and 500 meters onto the roadway's 3D model with a red transparent layer, while simultaneously overlaying a risk value label of 0.85. Combined with a heatmap analysis algorithm, the risk diffusion trend within a 200-meter radius is calculated, resulting in a diffusion risk value of 0.65, which is marked as a yellow medium-risk area, forming a complete visualization layer.

[0168] Step S109: For the risk visualization layer, the monitoring key areas of the sensor network are updated in real time. If an abnormal signal is detected, the data acquisition frequency is automatically adjusted to obtain more accurate dynamic change data of the rupture.

[0169] Temperature, vibration, and deformation data of the monitored area are acquired through a sensor network to form an initial environmental state matrix. When any sensor data deviates from the baseline value by more than two standard deviations, the acquisition frequency of that node is increased from 1Hz to 10Hz. A sliding window Fourier transform is performed on the high-frequency acquired data to extract the average amplitude within the 0-100Hz frequency band as a feature vector. When the cosine similarity between the feature vector and the baseline spectrum is less than 0.7, it is marked as an abnormal time segment. For areas where the cumulative abnormal time segments exceed 30 minutes, multi-sensor collaborative monitoring is initiated. Redundant nodes are deployed within a 50-meter radius of the abnormal point to generate three-dimensional deformation field data with millimeter-level accuracy. The daily displacement increment is calculated based on the deformation field data. When the vertical displacement rate exceeds 5mm / day and the horizontal displacement vector direction is consistent, a three-level risk zone is delineated. The boundary coordinates of the risk zone are projected onto the GIS map in real time, and the risk level is marked with red, yellow, and blue colors. The GIS engine automatically calculates the sensor blind spots covered by the risk zone and schedules mobile monitoring nodes to fill the blank grid. Newly added nodes acquire multi-axis acceleration data at a frequency of 20Hz, and measurement noise is eliminated through Kalman filtering. For 72 hours of data from a level-3 risk zone, 12-dimensional features, including time-domain peak value, frequency-domain energy, and displacement covariance, were extracted. The LibSVM tool was used to perform binary classification, outputting the determination results of rock mass fracture or equipment failure.

[0170] On the other hand, this embodiment also provides an electronic device, including a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein the processor implements the method when executing the computing program.

[0171] On the other hand, this embodiment also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method.

[0172] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for monitoring and early warning of coal and rock fracturing in a silty, weakly cemented thick-roofed coal roadway, characterized in that, include: A high-precision sensor network is deployed based on the tunnel environment to acquire minute vibration signals inside the coal and rock mass. Based on the minute vibration signal, short-time high-energy frequency components are extracted to obtain crack propagation characteristic signals; Based on the crack propagation characteristic signals, a multi-scale risk early warning model is constructed, and the fracture risk level results are output.

2. The method according to claim 1, characterized in that, The steps for obtaining minute vibration signals inside the coal and rock mass include: The sensor placement locations are determined based on the roadway lithology information to acquire multi-channel raw vibration signal data; Based on the original vibration signal data, a time-frequency feature matrix is ​​obtained by applying moving average filtering and short-time Fourier transform. Based on the time-frequency feature matrix, high-confidence anomalous components are filtered to obtain a microcrack signal set.

3. The method according to claim 1, characterized in that, The step of extracting short-time high-energy frequency components includes: Based on the microcrack signal set, low-pass filtering and fast Fourier transform were applied to extract the key frequency components of 50-500Hz. Based on the key frequency components, wavelet transform is used to analyze their energy standard deviation and identify dynamically changing frequency bands. Based on the dynamically changing frequency bands, an LSTM neural network input sequence is constructed to obtain the dynamic features of short-term crack propagation.

4. The method according to claim 1, characterized in that, The steps for constructing a multi-scale risk early warning model include: Based on the dynamic characteristics of short-term crack propagation and combined with the three-dimensional stress distribution model of the tunnel, the mid-term stress anomaly area is obtained; Based on the stress anomaly region, geological structure and historical displacement data are integrated to form a structural evolution feature vector; Based on the structural evolution feature vectors, a support vector machine model is trained to establish a fracture trend discriminant classifier.

5. The method according to claim 4, characterized in that, The steps for forming the structural evolution feature vector include: Based on the node locations within the mid-term stress anomaly region, extract the 24-hour stress growth curve; Based on the stress growth curve, the risk characteristic combination is constructed by integrating joint density and surrounding rock strength information; Based on the combination of risk characteristics, principal component analysis is performed to obtain the dominant factors in structural evolution.

6. The method according to claim 4, characterized in that, The steps for establishing the fracture trend discriminant classifier include: Based on the dominant factors of structural evolution, historical rupture sample data are labeled and a training dataset is constructed. Based on the training dataset, radial basis function support vector machine is applied to train three types of rupture trends; Based on the trained model, the current dominant factors are classified and labeled as stable, fluctuating, or trending upward.

7. The method according to claim 1, characterized in that, The steps for outputting the rupture risk level result include: Based on trend labels, short-term dynamic features and medium-term stress results are integrated to construct a multi-scale fused feature vector; Based on the fused feature vector, it is input into the entropy-weighted risk scoring model to calculate the comprehensive risk value; Risk levels are determined based on the relationship between the overall risk value and a preset threshold.

8. The method according to claim 7, characterized in that, The steps for constructing the multi-scale fused feature vector include: Frequency fluctuation values ​​are extracted based on short-term dynamic characteristics, and gradient change rates are extracted based on medium-term stress results. Based on the fluctuation value and the rate of change, a combined feature matrix is ​​generated; Based on the combined feature matrix, a standardization process is performed to form the final fusion vector that can be used in the scoring model.

9. An electronic device comprising a memory, a processor, and a computing program stored in the memory and executable on the processor, characterized in that, When the processor executes the computing program, it implements the method of any one of claims 1-8.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1-8.

Images