A machine learning-based coal mine rock burst early warning method

Abstract

The application discloses a coal mine rock burst early warning method based on machine learning, relates to the technical field of coal mine safety monitoring and computer data processing, and acquires microseismic event data, coal rock mass stress data and acoustic emission signal data in a spatial monitoring unit, configures a unified time reference and forms multi-source monitoring data according to a continuous monitoring window; after preprocessing and feature extraction, multi-source time sequence feature sequences and microseismic-acoustic emission waveform data tensors are formed. The time sequence feature sequences are input into a long short-term memory network time sequence prediction model to obtain time sequence risk prediction values, and the waveform tensors are input into a one-dimensional convolutional neural network pattern recognition model to obtain precursor pattern probability values. The application captures cross-period evolution trends and local waveform precursors through a parallel double-model structure, dynamically adjusts a fusion weight according to historical performance, and improves early warning accuracy and adaptability.

Description

Technical Field

[0001] This invention relates to the field of coal mine safety monitoring and computer data processing technology, specifically to a coal mine rockburst early warning method based on machine learning. Background Technology

[0002] Rockbursts are dynamic disasters caused by the rapid release of energy accumulated in coal and rock during coal mining. Under deep mining conditions, the combined effects of ground stress levels, geological structures, working face disturbances, and roadway support conditions typically manifest rockburst precursors as microseismic activity, changes in coal and rock mass stress, and acoustic emission activity. Different monitoring methods reflect different physical processes. Microseismic event data can characterize large-scale coal and rock fracturing and energy release, coal and rock mass stress data can characterize load accumulation and stress concentration, and acoustic emission signal data can characterize small-scale crack initiation and propagation. When a single monitoring quantity is abnormal, equipment vibration, changes in sensor coupling, local construction disturbances, and communication packet loss can all generate signal changes similar to rockburst precursors.

[0003] Existing rockburst early warning methods already employ data-driven models to process monitoring indicators. Public document CN117312919A discloses a rockburst early warning method based on a BO-CNN-LSTM model. After preprocessing, the monitoring indicators are input into a prediction model formed by combining a convolutional neural network and a long short-term memory network, and the model parameters are determined using Bayesian optimization. In this scheme, the convolutional neural network and the long short-term memory network form a cascade relationship for feature extraction and temporal prediction. This cascaded structure typically requires the output of the preceding network to fully carry the information needed by the subsequent network; local transient waveform features and stress evolution features across monitoring periods are easily compressed into the same intermediate representation.

[0004] Public document CN120493068A discloses a method for early warning and prevention of mine rockbursts. The method acquires multi-source monitoring data, constructs a network prediction model, and builds a three-dimensional cellular automata model of the coal and rock mass fracturing process for prevention and control strategy analysis. The three-dimensional cellular automata model requires coal and rock mass element parameters and fracturing evolution parameters, resulting in a significant workload for parameter calibration. For monitoring systems whose primary objective is online risk output, the physical simulation process also includes data preparation and computational steps.

[0005] Different monitoring devices also have different sampling rates and data formats. Coal and rock mass stress data are often recorded continuously at intervals of seconds or minutes; microseismic systems mainly record event catalogs and event waveforms; acoustic emission systems may save trigger waveforms at a higher sampling rate. Relying solely on network clock synchronization cannot directly form data inputs that can be used by machine learning models. Monitoring systems still need to complete timestamp correction, event merging, window segmentation, missing data marking, waveform segment extraction, and data dimension unification. If the training labels use events that have already occurred within the input window, the model training results will contain future information leakage, making it difficult to form accurate early warnings during the operational phase.

[0006] Fixed fusion weights are also difficult to adapt to changes in the operating environment. Coal and rock stress sensors may drift, acoustic emission sensors may experience reduced signal quality due to changes in coupling state, and microseismic stations may experience waveform loss due to construction or communication failures. When the contribution ratio of the two models is fixed based on offline training performance, the model with degraded input data quality may still maintain a high fusion weight. It is necessary to establish a dual-model structure that processes continuous evolution information and local waveform information in parallel, and to update the model reliability using historical early warning cycles with confirmed results, so that the comprehensive early warning results can be adjusted according to changes in model performance. Summary of the Invention

[0007] The purpose of this invention is to provide a machine learning-based method for early warning of coal mine rockbursts. The technical problem to be solved by this invention is: under the conditions that the sampling rates, data forms, and data quality of microseismic event data, coal and rock mass stress data, and acoustic emission signal data are different and change with the operation process, a dual-path model with a clear time relationship is formed as input, the risk evolution trend across the monitoring cycle and the local waveform precursor mode within the monitoring window are extracted respectively, and the fusion weight is dynamically determined based on the historical prediction performance of the two models during the operation phase.

[0008] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A machine learning-based method for early warning of coal mine rockburst includes: acquiring microseismic event data, coal and rock mass stress data, and acoustic emission signal data within a spatial monitoring unit; configuring a unified time reference for the microseismic event data, the coal and rock mass stress data, and the acoustic emission signal data; and processing multi-source monitoring data according to the formation time of a continuous monitoring window; preprocessing and feature extraction of the multi-source monitoring data to form a multi-source time-series feature sequence and a microseismic-acoustic emission waveform data tensor, wherein the multi-source time-series feature sequence contains microseismic statistical features, coal and rock mass stress features, and acoustic emission statistical features at multiple consecutive time steps, and the microseismic-acoustic emission waveform data tensor contains microseismic waveform segments and acoustic emission waveform segments corresponding to the continuous monitoring window; and then processing the data... The multi-source temporal feature sequence is input into a long short-term memory network (LSTM) temporal prediction model to obtain the temporal risk prediction value within the target prediction time range after the current continuous monitoring window ends. The microseismic-acoustic emission waveform data tensor is input into a one-dimensional convolutional neural network (CNN) pattern recognition model to obtain the precursor mode probability value within the target prediction time range. From historical early warning periods with confirmed results, the model reliability of the LSTM temporal prediction model and the one-dimensional CNN pattern recognition model are determined respectively, and the fusion weights of the two models are updated according to the model reliability. A comprehensive risk prediction value is calculated based on the temporal risk prediction value, the precursor mode probability value, and the fusion weights, and the rockburst risk level is determined based on the comprehensive risk prediction value.

[0009] As a preferred technical solution of the present invention, a unified time reference is configured for the microseismic event data, the coal and rock mass stress data, and the acoustic emission signal data, including: clock correction of the microseismic acquisition nodes, stress acquisition nodes, and acoustic emission acquisition nodes using a precise time protocol; writing the corrected acquisition time into a data frame; dividing the continuous monitoring window according to a preset monitoring cycle; maintaining the coal and rock mass stress data to the corresponding time step according to the timestamp; merging the microseismic event data and acoustic emission signal data to the corresponding time step according to the event occurrence time; and writing a missing flag for missing time steps.

[0010] As a preferred technical solution of the present invention, the multi-source monitoring data is preprocessed and feature extracted, including: performing wavelet decomposition on the signal within the effective response frequency band obtained by sensor calibration, performing soft threshold shrinkage on the wavelet detail coefficients and completing signal reconstruction; extracting the mean, variance, peak value, root mean square value, dominant frequency, frequency band energy and spectral entropy from the reconstructed signal; standardizing the extraction results; projecting the principal component projection matrix obtained from the training data, selecting the principal components whose cumulative contribution rate reaches the preset contribution rate threshold, and forming the multi-source time series feature sequence.

[0011] As a preferred embodiment of the present invention, the training samples of the Long Short-Term Memory Network (LSTM) time-series prediction model consist of multi-source time-series feature sequences corresponding to multiple continuous monitoring windows, and the training labels are determined by the confirmed risk results within the target prediction time range after the input sequence ends; the training set, validation set, and test set are divided according to the monitoring time or mining stage, so that samples with overlapping time ranges are located in the same data subset; the LTM time-series prediction model outputs the time-series risk prediction value with a value between zero and one.

[0012] As a preferred technical solution of the present invention, the formation of the microseismic-acoustic emission waveform data tensor includes: extracting microseismic waveform segments and acoustic emission waveform segments according to the event trigger time; performing bandpass filtering, amplitude normalization, resampling, and fixed-length processing on the waveform segments; stacking the processed waveform segments into channels according to the sensor number and signal type; the positive class label determination rule of the one-dimensional convolutional neural network pattern recognition model is: when a confirmed rockburst event exists within the target prediction time range after the end of the continuous monitoring window to which the waveform segment belongs, it is marked as positive.

[0013] As a preferred embodiment of the present invention, for the model identifier m, where model identifier m is either a long short-term memory network time-series prediction model identifier L or a one-dimensional convolutional neural network pattern recognition model identifier C, the model reliability is calculated based on the historical early warning period for which confirmation results have been obtained:

[0014] in, This represents the model reliability corresponding to the continuous monitoring window t. This represents the correct prediction prior count corresponding to model identifier m. This represents the prior count of incorrect predictions corresponding to model identifier m. This indicates the number of correct predictions corresponding to model identifier m within a preset historical range prior to the continuous monitoring window t. This represents the number of incorrect predictions corresponding to model identifier m within a preset historical range prior to the continuous monitoring window t; candidate weights are calculated based on model reliability, and exponential smoothing and normalization are performed on the candidate weights to make the sum of the fusion weights of the long short-term memory network time-series prediction model and the fusion weights of the one-dimensional convolutional neural network pattern recognition model equal to one.

[0015] As a preferred technical solution of the present invention, a data quality coefficient is calculated based on the input data corresponding to the model identifier m within the current continuous monitoring window. The data quality coefficient is determined by data integrity, time synchronization quality, signal-to-noise ratio, and outlier ratio. The model reliability is corrected using the data quality coefficient, and the candidate weights are calculated based on the corrected model reliability.

[0016] As a preferred embodiment of the present invention, after clock calibration is performed using a precise time protocol, the propagation delay is calculated based on the microseismic event location coordinates, the sensor's three-dimensional coordinates, and the propagation velocity of the coal and rock medium:

[0017]

[0018] in, This indicates that the wave corresponding to event number e propagates to sensor number e. The propagation delay Indicates sensor number The three-dimensional coordinate vector, This represents the location coordinate vector of event number e. This indicates the calibration propagation speed in the coal and rock region where event number e is located. Indicates sensor number The recorded waveform arrival time, Indicates based on sensor number The event occurrence time is calculated inversely; the microseismic waveform segment and acoustic emission waveform segment are aligned according to the event occurrence time.

[0019] As a preferred technical solution of the present invention, the starting time of the stress continuous growth stage, the starting time of the acoustic emission activity enhancement stage, and the starting time of the microseismic activity enhancement stage are determined respectively. The consistency coefficient of the precursor stage is calculated based on the degree of matching between the three starting times and the preset precursor sequence. The comprehensive risk prediction value is incrementally corrected based on the consistency coefficient of the precursor stage.

[0020] 10. A method for early warning of coal mine rockburst based on machine learning according to claim 1, characterized in that: the model divergence between the time-series risk prediction value and the precursor mode probability value is calculated; when the model divergence exceeds the divergence trigger threshold within multiple consecutive monitoring windows, a sampling parameter adjustment instruction is sent to the monitoring and acquisition node, wherein the sampling parameter adjustment instruction is used to shorten the sampling interval of coal and rock mass stress data and increase the original data retention ratio of microseismic waveform segments and acoustic emission waveform segments.

[0021] Compared with the prior art, the present invention has the following beneficial effects: This invention uses a unified time reference and a continuous monitoring window to establish the time correspondence between three types of monitoring data. The continuous recording of coal and rock stress, the catalog of microseismic events and the acoustic emission trigger waveform do not need to be forcibly spliced ​​into the same original sampling sequence, which can reduce data duplication and data volume expansion caused by sampling rate differences.

[0022] This invention employs a parallel dual-model structure. The long short-term memory network time-series prediction model processes the statistical characteristics and stress changes across continuous monitoring windows, while the one-dimensional convolutional neural network pattern recognition model processes the local waveform morphology within the monitoring window. The two models respectively retain long-term evolutionary information and short-term local information.

[0023] This invention updates model reliability based on historical early warning periods with confirmed results, while early warning periods without confirmed results are not included in weight updates. This avoids using future results or incomplete labels to influence current weights. The data quality coefficient can reduce the fusion contribution of the model due to missing inputs, synchronization errors, or degraded signal quality.

[0024] This invention reduces the impact of cross-sensor waveform misalignment, single-channel transient anomalies, and repeated level switching near thresholds on early warning results through propagation delay compensation, precursor stage consistency verification, and partitioned hysteresis thresholding. This invention also uses model divergence-triggered sampling parameter adjustment to ensure that high-sampling-rate original waveforms are primarily preserved during stages of model conclusion conflict or increased risk. Historical case local residual correction and hierarchical memory playback can adapt to changes in data distribution across different spatial monitoring units and different mining stages. Attached Figure Description

[0025] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0026] Figure 1 This is the main flowchart of the overall method of the present invention.

[0027] Figure 2 This is a flowchart of the unified time base configuration sub-process of the present invention.

[0028] Figure 3 This is a flowchart of the multi-source monitoring data preprocessing and feature extraction sub-process of the present invention.

[0029] Figure 4 This is a flowchart of the model reliability update and dynamic fusion sub-process of the present invention. Detailed Implementation

[0030] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the embodiments of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0031] The following is in conjunction with the appendix Figures 1-4 The embodiments of the present invention will be described in detail below.

[0032] Example 1: This example divides the mine monitoring area into spatial monitoring units. A spatial monitoring unit can correspond to a coal mining face, a tunneling face, a roadway section, or a three-dimensional grid. Each monitoring record includes a spatial monitoring unit number, sensor number, acquisition timestamp, data type, measured value or waveform array, data quality flag, and equipment status flag. The spatial monitoring unit number is used to constrain the subsequent model training and reliability statistics range, avoiding direct mixing of monitoring results from areas with different geological conditions.

[0033] The micro-vibration sensor array uses a three-component accelerometer, and the effective response frequency band of the sensor is determined by the amplitude-frequency calibration results after installation.

[0034] This embodiment uses a microseismic signal processing frequency band of 0.5Hz to 500Hz, with a sampling frequency of 2000Hz. Coal and rock mass stress data are acquired by anchor bolt stress sensors and borehole stress sensors, with a measurement range of 0MPa to 60MPa and a sampling interval of 60s. The acoustic emission sensor is a resonant sensor with a center response frequency of 35kHz and a sampling frequency of 1MHz; the acoustic emission signal processing frequency band is 20kHz to 80kHz. The sensor specifications are used to describe the feasible method; the sensor specifications can be adjusted according to the coal and rock medium, monitoring distance, and equipment calibration results.

[0035] Microseismic acquisition nodes, stress acquisition nodes, and acoustic emission acquisition nodes are connected to a precise time protocol network. The ground time server sends synchronization messages to the acquisition nodes, and the nodes calculate the clock deviation between their local clock and the time server. The acquisition nodes use the clock deviation to correct the acquisition timestamps. The precise time protocol is used to unify device clocks; continuous monitoring window segmentation and event merging are handled by the data processing program.

[0036] In this embodiment, the continuous monitoring window length is set to 5 minutes, and six continuous monitoring windows are combined to form a 30-minute multi-source time-series feature sequence. The 5 minutes following the end of the current continuous monitoring window are set as the target prediction time range. The continuous monitoring window adopts a left-closed, right-open time range. The time range starts at 10:00:00 and ends at 10:05:00, with data at timestamp 10:05:00 being included in the next continuous monitoring window. The left-closed, right-open rule avoids duplicate merging of boundary data.

[0037] Coal and rock mass stress data are categorized into corresponding continuous monitoring windows according to sampling timestamps. Within each continuous monitoring window, the mean stress, final stress value, rate of change of stress, and acceleration of stress change are calculated. If no new coal and rock mass stress data is available at a given time step, the data processing program uses the most recent valid value and writes a hold flag within the allowed hold time; if the allowed hold time is exceeded, a missing value flag is written, and zero values ​​are not used to replace missing measurements. Microseismic event data are categorized into continuous monitoring windows according to the event occurrence time, and the number of events, cumulative energy, maximum event energy, average dominant frequency, and event spatial density are calculated. Acoustic emission signal data are categorized into continuous monitoring windows according to the trigger time, and the number of triggers, cumulative energy, duration, ring count, and spectral entropy are calculated.

[0038] Microseismic signal denoising and acoustic emission signal denoising are performed within the effective response band of the sensor, respectively. The original signal undergoes five-level db4 wavelet decomposition, with detail coefficients shrunk using soft thresholding. For signal channel s, the wavelet threshold is calculated according to the following formula:

[0039] in, The wavelet threshold of signal channel s is represented. This represents the noise standard deviation estimated from the median of the highest frequency detail coefficients in signal channel s. This indicates the number of sampling points of signal channel s in the waveform segment. Represents the natural logarithm. The detail coefficients are processed using soft thresholding, and the reconstruction coefficients are calculated according to the following formula:

[0040] in, Indicates the coefficient number in signal channel s The corresponding wavelet detail coefficients, This represents the wavelet detail coefficients after soft-threshold shrinkage. Represents a symbolic function. This indicates that the larger of the two inputs is selected. After signal reconstruction, the ratio of the reconstructed signal energy to the original signal energy is calculated as the energy retention rate, which is calculated according to the following formula:

[0041] in, This represents the energy retention rate of signal channel s. Represents the energy of the reconstructed signal. This represents the original signal energy. When the energy retention rate is below 0.90, the data processing program lowers the wavelet threshold and reconstructs the signal until the energy retention rate reaches 0.90 or the preset number of iterations is reached.

[0042] The band energy of the microseismic signal was calculated in three frequency bands: 0.5Hz to 50Hz, 50Hz to 200Hz, and 200Hz to 500Hz. The band energy of the acoustic emission signal was calculated in three frequency bands: 20kHz to 30kHz, 30kHz to 45kHz, and 45kHz to 80kHz. The band boundaries were derived from sensor calibration results and historical event spectral distributions, and the microseismic signal and acoustic emission signal were not confined to the same frequency range. Time-domain characteristics included mean, variance, peak value, and root mean square value; frequency-domain characteristics included dominant frequency, energy proportion of each band, and spectral entropy.

[0043] All features are standardized using the mean and standard deviation calculated from the training data. For feature number j and time step t, the standardized feature is calculated according to the following formula:

[0044] in, This represents the standardized result of feature number j at time step t. This represents the original feature value of feature number j at time step t. Let represent the mean of feature number j in the training data. This represents the standard deviation of feature number j in the training data. The runtime phase uses the data saved during the training phase. and The standardized parameters are not recalculated using the current continuous monitoring window.

[0045] The principal component projection matrix is ​​obtained by decomposing the standardized feature covariance matrix of the training data. The principal component features are calculated according to the following formula:

[0046] in, This represents the principal component feature at time step t, where principal component number k is the principal component feature. Indicates the number of original features. Let represent the projection coefficients of the original feature number j corresponding to the principal component number k in the principal component projection matrix. Select the first K principal components such that their cumulative contribution rates satisfy:

[0047] in, This represents the cumulative contribution rate of the first K principal components. This represents the eigenvalue corresponding to principal component number k. The data processing program selects the smallest K value that satisfies the cumulative contribution rate condition. The K-dimensional principal component features from six consecutive time steps are stacked in chronological order to form a multi-source time-series feature sequence with a dimension of 6 x K.

[0048] The microseismic-acoustic emission waveform data tensor does not use the entire high-sampling-rate original waveform over a continuous 30 minutes. The data processing program extracts microseismic waveform segments 0.25 s before and 0.75 s after the microseismic event trigger time, and acoustic emission waveform segments 0.5 ms before and 2.5 ms after the acoustic emission trigger time. Each waveform segment undergoes bandpass filtering, baseline removal, and amplitude normalization. To form a fixed-dimensional input, the microseismic waveform segments and acoustic emission waveform segments are resampled to 1024 sampling points per channel. When the number of events within the continuous monitoring window exceeds a preset upper limit, a preset number of waveform segments are selected based on event energy and trigger time; for example, the preset upper limit is 20. When the number of events is less than the preset number, zero-filled waveforms are used, and a valid segment mask is simultaneously written. The valid segment mask is a binary vector with the same number of sampling points as the waveform segment, where positions containing original data are marked as 1, and positions with zero filling are marked as 0.

[0049] The microseismic three-component waveforms and acoustic emission waveforms are stacked in channels according to sensor number and signal type. The first dimension of the microseismic-acoustic emission waveform data tensor represents the channel, and the second dimension represents the sampling point. A valid segment mask serves as additional input to a one-dimensional convolutional neural network pattern recognition model to distinguish between zero-filled segments and true low-amplitude segments. The data processing program preserves the continuous monitoring window number, spatial monitoring unit number, and target prediction time range, ensuring a one-to-one correspondence between the multi-source temporal feature sequence, the microseismic-acoustic emission waveform data tensor, and the training labels.

[0050] The continuous digital calculation in this embodiment is given below. When the time server time is 10:00:00.000, the local time of the microseismic acquisition node is 10:00:00.0008, with a clock deviation of 0.0008s. When the microseismic acquisition node records an event at local time 10:02:15.4208, the corrected timestamp is 10:02:15.4200, and the event is included in the continuous monitoring window from 10:00:00 to 10:05:00. The stress acquisition nodes recorded 22.0 MPa, 22.4 MPa, 22.9 MPa, 23.5 MPa, and 24.2 MPa at 10:00:00, 10:01:00, 10:02:00, 10:03:00, and 10:04:00, respectively. The stress increment within 5 minutes was 2.2 MPa, and the average stress change rate calculated based on 5 minutes was 0.44 MPa / min.

[0051] For wavelet denoising, the noise standard deviation estimate is taken as 0.0269, the number of sampling points is taken as 1024, the general threshold calculation result is 0.1002, and the rounded threshold is 0.10. The wavelet detail coefficient sequence is 0.12, 0.08, 0.45, and -0.52, which are 0.02, 0, 0.35, and -0.42 after soft thresholding. The original sum of squares of detail coefficients is 0.4937, and the sum of squares of detail coefficients after shrinking is 0.2993. The reconstructed signal also includes unshrinked low-frequency approximation coefficients, and the sum of squares of low-frequency approximation coefficients is taken as 3.0000. The original signal energy is 3.4937, the reconstructed signal energy is 3.2993, and the energy retention rate is 0.9444, which meets the energy retention requirement of 0.90.

[0052] The four original features were, in order, the number of microseismic events (12), the rate of stress change (0.44 MPa / min), the number of acoustic emission triggers (85), and the proportion of high-frequency acoustic emission energy (0.36). The corresponding means for the training data were 8, 0.20 MPa / min, 50, and 0.20, respectively, and the standard deviations were 4, 0.12 MPa / min, 25, and 0.10, respectively. The standardized results were 1.00, 2.00, 1.40, and 1.60, respectively.

[0053] The projection coefficients of principal component number 1 are 0.50, 0.50, 0.50, and 0.50, respectively, and the projection result of principal component number 1 is 3.00. The projection coefficients of principal component number 2 are 0.50, -0.50, 0.50, and -0.50, respectively, and the projection result of principal component number 2 is -0.60. After performing the same processing on all six consecutive monitoring windows, a multi-source time-series feature sequence with a dimension of 6 x K is formed.

[0054] Example 2: This example uses the multi-source temporal feature sequence and microseismic-acoustic emission waveform data tensor formed in Example 1. The Long Short-Term Memory network temporal prediction model and the one-dimensional convolutional neural network pattern recognition model run in parallel. Both models correspond to the same spatial monitoring unit, the same continuous monitoring window, and the same target prediction time range.

[0055] The training data is divided into training, validation, and test sets according to monitoring time and mining stage. Samples with overlapping time ranges are retained in the same subset. Multiple consecutive windows formed before a rockburst event are also retained in the same subset. The aforementioned segmentation rules can avoid information duplication between the training and test sets due to adjacent overlapping windows.

[0056] The Long Short-Term Memory (LSTM) network time-series prediction model uses a two-layer LSM network. The first layer has a hidden state dimension of 128, and the second layer has a hidden state dimension of 64. The input sequence contains 6 time steps, each containing K principal component features. Training labels are determined based on confirmed risk outcomes within the target prediction time range. When no confirmed anomalies occur within the target prediction time range, the risk label is 0; when low-risk, medium-risk, or high-risk outcomes occur within the target prediction time range, the risk label is 0.35, 0.65, and 0.90, respectively. The risk label values ​​can be adjusted based on a risk level calibration set.

[0057] The second-layer Long Short-Term Memory (LSTM) network adopts a standard input gate, forget gate, and output gate structure. It updates the unit state and hidden state through a gating mechanism. The specific computational relationship is as follows:

[0058]

[0059]

[0060]

[0061]

[0062]

[0063] in, This represents the principal component eigenvector at time step t. This represents the hidden state vector of the previous time step. Represents the input gate vector. Represents the forget gate vector. Represents the candidate state vector. Represents the unit state vector. Indicates the output gate vector. This represents the hidden state vector at the current time step. , , and These represent the weight matrices for the corresponding gating calculations. , , and These represent the bias vectors used in the corresponding gating calculations. Represents a logical function. Represents the hyperbolic tangent function. This represents vector concatenation. This represents element-wise multiplication of vectors.

[0064] The hidden state at the final time step is processed by a fully connected layer and a logic function to obtain the predicted time series risk value:

[0065] in, This represents the predicted time-series risk value corresponding to the current continuous monitoring window. This represents the weight vector of the output layer of a Long Short-Term Memory (LSTM) network. This represents the transpose of a vector. This represents the bias of the output layer of the Long Short-Term Memory network. The model is trained using the mean squared error between the predicted and labeled values ​​as the loss function. Training stops and the model parameters with the lowest validation set loss are restored if the validation set loss does not decrease for ten consecutive training epochs.

[0066] The one-dimensional convolutional neural network pattern recognition model contains three convolutional layers with output channels of 32, 64, and 128 respectively, a kernel length of 3, and a stride of 1. Each convolutional layer is followed by batch normalization and a linear rectified activation function. The convolution calculation is as follows:

[0067] in, Indicates the convolutional layer number Output channel number Numbering at sampling locations activation value, This indicates the number of channels in the previous convolutional layer. Indicates the convolutional layer number The length of the convolution kernel, Indicates the convolutional layer number Middle connection input channel number With output channel number convolution kernel position number Corresponding weights Indicates the input channel number of the previous convolutional layer. Numbering at sampling locations activation value, Indicates the convolutional layer number Output channel number Corresponding bias, This indicates the corresponding batch normalization operation. This represents the linear rectification function.

[0068] The convolutional features are then input into a fully connected layer after global average pooling. The fully connected output for both classes is denoted as... and The probability value of the precursor pattern is calculated according to the following formula:

[0069] in, This represents the probability value of the precursor pattern corresponding to the current continuous monitoring window. This indicates the score for the category without warning signs. Indicates the score for the precursor category. This represents an exponential function. A valid segment mask in the microseismic-acoustic emission waveform data tensor is used to mask zero-padding segments. The model is trained using binary cross-entropy as the loss function. Positive class labels require the existence of a confirmed rockburst event within the target prediction time range; events that have already occurred within the input continuous monitoring window are not considered as positive class labels for future time ranges.

[0070] Both model outputs are probabilistically calibrated using a validation set. The calibration process uses a monotonic calibration function to map the original model outputs to a range of zero to one, with the mapping result corresponding to the risk meaning within the same target prediction time range. During the runtime phase, the number of correct and incorrect predictions is counted from historical warning periods with confirmed results. A prediction where the risk level is consistent with the confirmed result is recorded as a correct prediction; a prediction where the risk level is inconsistent with the confirmed result is recorded as an incorrect prediction. Warning periods without confirmed results are retained in the pending confirmation queue and do not participate in model reliability updates.

[0071] The Long Short-Term Memory (LSTM) network temporal prediction model is labeled L, and the one-dimensional convolutional neural network pattern recognition model is labeled C. Both the correct prediction prior count and the incorrect prediction prior count are set to 1. The model reliability is calculated using the following relationship:

[0072] in, This represents the model reliability corresponding to the continuous monitoring window t. This represents the correct prediction prior count corresponding to model identifier m. This represents the prior count of incorrect predictions corresponding to model identifier m. This indicates the number of correct predictions made by model identifier m within a preset historical range prior to the continuous monitoring window t. For example, the preset historical range is the most recent 50 continuous monitoring windows. This represents the number of incorrect predictions corresponding to model identifier m within a preset historical range prior to the continuous monitoring window t. The candidate fusion weights are obtained by normalizing the model reliability.

[0073] in, This represents the candidate fusion weights corresponding to the continuous monitoring window t. This indicates the model reliability of the Long Short-Term Memory (LSTM) network time-series prediction model. This represents the model reliability of a one-dimensional convolutional neural network pattern recognition model. The candidate fusion weights are exponentially smoothed.

[0074] in, This represents the current fusion weight corresponding to the continuous monitoring window t. This indicates the fusion weight of the previous continuous monitoring window. This represents a smoothing coefficient with a value greater than zero and less than or equal to one. The sum of the two candidate fusion weights is one, the sum of the fusion weights of the two previous consecutive monitoring windows is one, and the sum of the two current fusion weights remains one.

[0075] The comprehensive risk forecast value is calculated according to the following formula:

[0076] in, This represents the comprehensive risk prediction value for the current continuous monitoring window. This represents the current fusion weights of the Long Short-Term Memory network temporal prediction model. This represents the current fusion weights of the one-dimensional convolutional neural network pattern recognition model.

[0077] In a preferred embodiment, the model reliability is further constrained by the quality of the current input data. For the model identifier m, the data quality coefficient is calculated according to the following formula:

[0078] in, This represents the data quality coefficient corresponding to the continuous monitoring window t. Indicates the percentage of missing input data. Indicates time synchronization residual, Indicates the synchronous residual baseline value. Indicates the signal-to-noise ratio of the input signal. Indicates the signal-to-noise ratio reference value. Indicates the proportion of outliers. , , and Let represent the coefficients of the four data quality items, where all coefficients are non-negative and their sum is one. This indicates taking the smaller of the two inputs. The reliability of the modified model is calculated according to the following formula:

[0079] in, This represents the model reliability after data quality constraints. The corrected model reliability is used when calculating candidate fusion weights. Alternative .

[0080] The following is a continuous numerical calculation of model reliability and dynamic fusion. The Long Short-Term Memory (LSTM) network temporal prediction model correctly predicted 18 times and incorrectly predicted 2 times within the historical range, while the one-dimensional convolutional neural network (CNN) pattern recognition model correctly predicted 15 times and incorrectly predicted 5 times. The prior counts of correct and incorrect predictions for both models are set to 1. The model reliability of the LTM network temporal prediction model is 19 divided by 22, resulting in 0.8636; the model reliability of the one-dimensional CNN pattern recognition model is 16 divided by 22, resulting in 0.7273. The two candidate fusion weights are 0.5429 and 0.4571, respectively. The fusion weights of the previous continuous monitoring window are 0.60 and 0.40, respectively, with a smoothing coefficient of 0.40. The current fusion weights are 0.5771 and 0.4229, respectively.

[0081] In the data quality constraint calculation, the coefficients for all four data quality items are set to 0.25, the synchronization residual baseline is set to 1ms, and the signal-to-noise ratio baseline is set to 10. For the Long Short-Term Memory (LSTM) network, the missing data ratio is set to 0.05, the synchronization residual to 0.2ms, the signal-to-noise ratio to 12, and the outlier ratio to 0.02. The data quality coefficient is calculated as 0.25 x 0.95 + 0.25 x exponent - 0.2 + 0.25 x 1 + 0.25 x 0.98, resulting in a coefficient of 0.9372. For the One-Dimensional Convolutional Neural Network (BCNN), the missing data ratio is set to 0.20, the synchronization residual to 0.8ms, the signal-to-noise ratio to 6, and the outlier ratio to 0.10. The data quality coefficient is calculated as 0.25 x 0.80 + 0.25 x exponent - 0.8 + 0.25 x 0.60 + 0.25 x 0.90, resulting in a coefficient of 0.6873. The model reliability after data quality constraints is 0.8094 and 0.4999, respectively, and the normalized candidate fusion weights are 0.6182 and 0.3818, respectively.

[0082] The previous continuous monitoring window was smoothed using fusion weights of 0.60 and 0.40, and a smoothing coefficient of 0.40. The current fusion weights are 0.6073 and 0.3927, respectively. The current time series risk prediction value is set to 0.72, the current precursor pattern probability value is set to 0.84, and the comprehensive risk prediction value is calculated as 0.6073 multiplied by 0.72 plus 0.3927 multiplied by 0.84, resulting in a value of 0.7671.

[0083] Example 3: This example adds propagation delay compensation, precursor stage consistency verification, and partition hysteresis classification to Example 2. Precise time protocols can reduce clock differences between different acquisition nodes, but the propagation distance between different sensors and the event location varies, and the waveform arrival time still includes physical propagation time. Propagation delay compensation is used to convert the waveform arrival time into the event occurrence time.

[0084] The event localization module outputs the localization coordinate vector corresponding to event number e. Artificial seismic sources are deployed in the mine, and their coordinates and the arrival times of each sensor are recorded. The calibration propagation velocity of different spatial monitoring units is calculated based on the known propagation distance. Artificial seismic source calibration can be performed periodically according to the mining stage; recalibration is performed when changes in the coal and rock medium exceed the wave velocity update condition. The propagation delay and event occurrence time are calculated according to the following relationship:

[0085]

[0086] in, This indicates that the wave corresponding to event number e propagates to sensor number e. The propagation delay Indicates sensor number The three-dimensional coordinate vector, This represents the location coordinate vector of event number e. This indicates the calibration propagation speed in the coal and rock region where event number e is located. Indicates sensor number The recorded waveform arrival time, Indicated based on sensor number The event occurrence time is calculated in reverse. The event occurrence times obtained from multiple sensors are averaged using signal quality to form a unified event occurrence time.

[0087] After propagation delay compensation, the data processing program re-extracts microseismic waveform segments and acoustic emission waveform segments centered on the time of the unified event. Sensors with lower signal quality or whose propagation paths pass through known goaf areas can reduce their contribution to the time of the unified event. Propagation delay compensation does not change the original timestamp of the acquisition node; both the original timestamp and the corrected timestamp are stored in the historical database.

[0088] The consistency check of the precursor stage uses the continuous relationship between stress changes, acoustic emission activity, and microseismic activity in the coal and rock mass. The starting time of the sustained stress growth stage is denoted as... The start time of the acoustic emission enhancement phase is denoted as The start time of the microseismic activity enhancement phase is denoted as The stage of sustained stress growth can be determined by the stress change rate being higher than the stress change rate threshold for three consecutive time steps; the stage of enhanced acoustic emission activity can be determined by the number of acoustic emission triggers and the proportion of high-frequency energy being higher than the corresponding thresholds simultaneously; the stage of enhanced microseismic activity can be determined by the cumulative energy of microseismic events and the number of events being higher than the corresponding thresholds simultaneously.

[0089] The preset precursor sequence includes the stage of sustained stress growth, the stage of enhanced acoustic emission activity, and the stage of enhanced microseismic activity. The consistency coefficient of the precursor stages is calculated according to the following formula:

[0090] in, This represents the consistency coefficient of the precursor stage corresponding to the current continuous monitoring window. This indicates the precursor sequence indicator value, which satisfies... The value is 1 if the condition is met, and 0 if the condition is not met. This represents the reference time interval from the stage of sustained stress growth to the stage of enhanced acoustic emission activity. This represents the reference time interval from the stage of enhanced acoustic emission activity to the stage of enhanced microseismic activity. This indicates the permissible deviation scale for the first reference time interval. This indicates the permissible deviation scale for the second reference time interval. The reference time interval and permissible deviation scale are calculated from historically confirmed events.

[0091] The consistency coefficient for the precursor stage is used to make an incremental adjustment to the overall risk forecast value with a controlled upper limit. The adjusted overall risk forecast value is calculated according to the following formula:

[0092] in, This represents the overall risk forecast value after consistency correction during the precursor stage. This represents the comprehensive risk prediction value for the current continuous monitoring window. This represents the consistency correction coefficient, which is greater than zero and less than one. When the sequence of precursor stages is not satisfied, the consistency coefficient of the precursor stage is zero, and the overall risk prediction value does not increase.

[0093] Risk level thresholds are determined separately for each spatial monitoring unit. The threshold calibration dataset is not used for model parameter training. The threshold calibration process statistically analyzes the number of missed alarms, false alarms, and early warning times at different thresholds, and selects an escalation threshold based on mine risk management requirements. A de-escalation threshold is simultaneously set for each adjacent risk level, and the de-escalation threshold is lower than the escalation threshold.

[0094] In this embodiment, the upgrade threshold from low risk to medium risk is set to 0.50, and the downgrade threshold from medium risk to low risk is set to 0.46; the upgrade threshold from medium risk to high risk is set to 0.75, and the downgrade threshold from high risk to medium risk is set to 0.70. When the current risk level is medium risk and the comprehensive risk prediction value is 0.49, the comprehensive risk prediction value is lower than the upgrade threshold from low risk to medium risk, but higher than the downgrade threshold from medium risk to low risk, and the risk level remains medium risk. Subsequently, when the comprehensive risk prediction value decreases to 0.43, the comprehensive risk prediction value is lower than the downgrade threshold from medium risk to low risk, and the risk level is downgraded to low risk.

[0095] The continuous digital calculation for propagation delay compensation is as follows. The event location coordinates are 300m, 400m, and 0m, and the sensor coordinates are 0m, 0m, and 0m. The Euclidean distance between the two coordinates is 500m. The calibrated propagation speed is 2500m / s, and the propagation delay is 500 divided by 2500, resulting in a calculation of 0.200s. The arrival time of the waveform recorded by the sensor is 10:12:30.350, and the calculated event occurrence time is 10:12:30.150.

[0096] The arrival time recorded by another sensor, after propagation delay compensation, also yielded an event occurrence time close to 10:12:30.150. The data processing program used the corrected unified event occurrence time to extract waveform segments.

[0097] In the consistency calculation of the precursor stage, the starting time of the stress continuous growth stage is 10:00, the starting time of the acoustic emission enhancement stage is 10:06, and the starting time of the microseismic enhancement stage is 10:14. The first actual time interval is 6 minutes, and the second actual time interval is 8 minutes. The first reference time interval is 5 minutes, the second reference time interval is 8 minutes, the first allowable deviation scale is 4 minutes, and the second allowable deviation scale is 5 minutes. The precursor sequence meets the preset sequence, and the precursor stage consistency coefficient is an exponential function acting on -0.25, with a calculation result of 0.7788. The comprehensive risk prediction value is 0.68, the consistency correction coefficient is 0.12, and the corrected comprehensive risk prediction value is 0.68 plus 0.12 multiplied by 0.7788 multiplied by 0.32, with a calculation result of 0.7099.

[0098] Example 4: This example adds model divergence feedback, historical case local residual correction, and incremental learning protection to the existing example 3. The two models may output significantly different results within certain continuous monitoring windows. When the predicted time-series risk value is high while the probability value of the precursor pattern is low, the stress in the coal and rock mass may continue to rise, but local waveform precursors may not yet appear; when the probability value of the precursor pattern is high while the predicted time-series risk value is low, there may be local rupture, equipment vibration, or short-term construction disturbances. Directly using a weighted average will compress the model divergence information.

[0099] The model divergence is calculated using the following formula:

[0100] in, This indicates the model divergence within the current continuous monitoring window. This represents the predicted time-series risk value corresponding to the current continuous monitoring window. This represents the probability value of the precursor mode corresponding to the current continuous monitoring window. The divergence trigger threshold is set to 0.35, and the continuous window number threshold is set to 3. When the model divergence value of three consecutive continuous monitoring windows is greater than 0.35, the central computing node sends a sampling parameter adjustment command to the edge computing nodes. The sampling interval for coal and rock mass stress data is shortened from 60s to 10s, and the retention rate of original data for microseismic waveform segments and acoustic emission waveform segments is increased from 20% to 100%. When the model divergence value of two consecutive continuous monitoring windows is less than 0.20, the sampling parameters are restored to the normal configuration. The divergence resolution threshold is lower than the divergence trigger threshold, which can reduce repeated switching of sampling parameters.

[0101] The historical precursor feature database is indexed according to spatial monitoring units, fault distance intervals, mining stages, and risk levels. For example, geological conditions are divided into 0-20m, 20-50m, 50-100m, and greater than 100m based on fault distance; mining stages are divided into initial mining stage, normal mining stage, and final mining stage. Historical records include standardized feature vectors, model outputs, comprehensive risk prediction values, confirmed risk results, and data quality indicators. After the current monitoring window forms a standardized feature vector, the data processing program first filters candidate historical events with the same spatial monitoring unit or similar geological conditions, and then calculates the cosine similarity.

[0102] in, This represents the cosine similarity between the current continuous monitoring window and the historical event number e. This represents the standardized feature vector of the current continuous monitoring window. This represents the standardized feature vector corresponding to historical event number e. The data processing program selects a set of historical events with a cosine similarity of at least 0.75 and the highest ranking. Historical forecast residuals are defined as:

[0103] in, This represents the historical prediction residual for historical event number e. This represents the confirmed risk value corresponding to historical event number e. This represents the historical comprehensive risk forecast value corresponding to historical event number e. The current comprehensive risk forecast value has been corrected for local residuals, and the corrected comprehensive risk forecast value is calculated according to the following formula:

[0104] in, This represents the overall risk prediction value after local residual correction for historical cases. This represents the comprehensive risk prediction value for the current continuous monitoring window. This indicates that the input is restricted to the range of zero to one. Local residual correction is not performed when the historical event set is empty.

[0105] Newly added monitoring samples can only enter incremental training after obtaining confirmation tags. Confirmation tags can come from rockburst event records, hazard review results, and monitoring conclusions confirmed through prescribed procedures. Training samples are stored hierarchically according to spatial monitoring units, risk levels, and mining stages. Each incremental training batch contains new samples and historically representative samples. When the number of high-risk event samples is small, stratified sampling ensures that each training batch contains a predetermined proportion of high-risk event samples, for example, a predetermined proportion of 20%.

[0106] The online model is not directly updated on the running instance. The central compute node replicates the online model parameters to form a candidate model, which is then incrementally trained using 0.5 times the initial training learning rate. After the candidate model is trained, accuracy, precision, recall, and F1 score are calculated on the global benchmark validation set and the high-risk event validation subset, respectively. Preset deployment conditions include: the candidate model's global F1 score is not lower than the online model's global F1 score; the candidate model's high-risk event recall is not lower than the online model's high-risk event recall; and the increase in the false positive rate of the candidate model in stable operating condition samples does not exceed the allowable increase. When all three conditions are met, the candidate model replaces the online model. If any condition is not met, the online model continues to be used, and the candidate model's validation record is saved.

[0107] The continuous digital calculations for model divergence feedback are as follows. The predicted time-series risks for the three consecutive monitoring windows are 0.80, 0.82, and 0.84, respectively, and the probability values ​​for precursor modes are 0.32, 0.31, and 0.30, respectively. The divergence values ​​for the three models are 0.48, 0.51, and 0.54, respectively, all higher than 0.35, satisfying the divergence triggering condition. The edge computing node shortens the sampling interval for coal and rock mass stress data from 60s to 10s and increases the retention rate of original data for microseismic waveform segments and acoustic emission waveform segments to 100%. The model divergence values ​​for the subsequent two consecutive monitoring windows decrease to 0.18 and 0.16, both lower than 0.20, and the sampling parameters are restored to the normal configuration.

[0108] The continuous numerical calculation for local residual correction of historical cases is as follows. The current standardized feature vectors are 1 and 2, and the standardized feature vectors of a historical event are 2 and 3. The dot product of the two vectors is 8, with magnitudes of the square roots of 5 and 13 respectively. The cosine similarity is 8 divided by the square root of 65, resulting in 0.9923. After screening candidate historical events, the cosine similarities of the four historical events are 0.90, 0.80, 0.70, and 0.76, respectively. Historical events with a cosine similarity of 0.70 are below 0.75 and are not included in the historical event set. The cosine similarities of the three selected historical events are 0.90, 0.80, and 0.76, respectively, and the historical prediction residuals are 0.05, -0.02, and 0.04, respectively. The similarity-weighted residual is 0.90 multiplied by 0.05 plus 0.80 multiplied by -0.02 plus 0.76 multiplied by 0.04, resulting in 0.0594. The total cosine similarity is 2.46, and the weighted average residual is 0.0241. The current comprehensive risk prediction value is 0.71, and the comprehensive risk prediction value after local residual correction is 0.7341.

[0109] The continuous numerical calculations for determining whether a candidate model should go online are as follows. The online model has a global F1 score of 0.86, a high-risk event recall rate of 0.80, and a stable operating condition false alarm rate of 0.07. The candidate model has a global F1 score of 0.87, a high-risk event recall rate of 0.84, a stable operating condition false alarm rate of 0.075, a false alarm rate increase of 0.005, and an allowable increase of 0.01.

[0110] The candidate model meets all three criteria for going live, and it replaces the online model. Another candidate model has a global F1 score of 0.88 and a high-risk event recall of 0.77. Although the global F1 score is higher, the high-risk event recall is lower than the online model; therefore, this candidate model will not replace the online model.

[0111] The early warning response and system deployment for rockburst risk levels are categorized as no risk, low risk, medium risk, and high risk. Risk levels are output to the coal mine underground safety monitoring system. In a no-risk state, routine monitoring parameters are maintained; in a low-risk state, risk trends and model discrepancies are recorded; in a medium-risk state, the monitoring sampling interval is shortened and the proportion of original waveforms retained is increased; in a high-risk state, audible and visual alarm control signals and information push instructions are output. The mapping between risk levels and on-site handling procedures is configured by the mine safety management system, and the risk level determination algorithm does not depend on any specific information push method.

[0112] Edge computing nodes are deployed at downhole monitoring stations or locations close to the data source, performing timestamp correction, waveform slicing, wavelet denoising, feature extraction, data quality calculation, and raw data caching. Central computing nodes perform principal component projection, long short-term memory network inference, one-dimensional convolutional neural network inference, model reliability updates, dynamic fusion, risk level determination, and model version management. During communication interruptions, edge computing nodes continue to cache timestamped data; after communication is restored, edge computing nodes retransmit data in chronological order, and the retransmitted data is only used for historical reliability statistics and incremental training after obtaining the corresponding confirmation results.

[0113] Each early warning record should at least save the spatial monitoring unit number, the start and end times of the continuous monitoring window, the target prediction time range, the time-series risk prediction value, the precursor mode probability value, the fusion weights of the two models, the comprehensive risk prediction value, the rockburst risk level, the data quality coefficient, the model version, the threshold version, and the confirmation status. Saving the aforementioned fields enables the reproduction of the risk assessment process and supports subsequent threshold calibration, model reliability calculation, and candidate model validation.

[0114] The core of this invention is as follows: three types of monitoring data form corresponding continuous monitoring windows under a unified time reference; the same set of continuous monitoring windows respectively form multi-source time-series feature sequences and microseismic-acoustic emission waveform data tensors; two machine learning models predict the same target time range in parallel; model reliability is updated based on historical early warning cycles with confirmed results; the comprehensive risk prediction value is calculated based on the current fusion weights. Propagation delay compensation, data quality constraints, precursor stage consistency, model divergence feedback, partition hysteresis threshold, local residual correction of historical cases, and hierarchical memory playback are all additional technical features that do not change the core technical chain.

[0115] Training label confirmation uses time range binding records. The data processing program saves the start and end times of the input sequence, the start and end times of the target prediction time range, and the end time of the target prediction time range for each sample. After the field event records are entered into the historical database, the label confirmation program retrieves the corresponding target prediction time range according to the spatial monitoring unit number and the event occurrence time. When the event occurrence time falls within the target prediction time range and the event location is within the boundary of the spatial monitoring unit, the sample enters the pending verification state. The pending verification state is converted into a confirmation label after joint checks of event type, event energy, field records, and equipment status. Waveform records formed by equipment maintenance, blasting operations, and registered construction vibrations retain the interference type mark and are not directly used as positive rockburst labels. Samples without event records can only be used as negative samples after the monitoring equipment online rate meets the data integrity requirements and the target prediction time range has been verified.

[0116] Model parameters, standardized parameters, principal component projection matrices, and risk thresholds are all stored in a versioned manner. Each model version is associated with the start and end times of the training data, the set of spatial monitoring units, the list of feature fields, the version of the standardized parameters, the version of the principal component projection matrix, and the version of the labeling rules.

[0117] The online inference log saves all version numbers actually used. When sensor replacement, installation location change, mining stage transition, or monitoring frequency band adjustment meets the version update conditions, the data processing program creates a new parameter version, without directly overwriting the original parameters. Historical early warning cycles reproduce model output according to the version actually used at that time, and model reliability statistics do not directly merge prediction results generated by different labeling rules.

[0118] Anomaly data handling follows the principles of preserving measurement facts and limiting anomaly contributions. When data frame verification fails, the data processing program saves the error type and marks the corresponding data as invalid. When a sensor continuously outputs a fixed value, the data processing program determines the sensor's frozen state based on the duration of the fixed value and the device's range. When stress measurement exceeds the device's range, the data processing program records a range saturation flag and does not interpret the saturation value as the true stress peak. When microseismic waveforms or acoustic emission waveforms are clipped, the data processing program records the waveform clipping ratio and reduces the corresponding data quality coefficient. Anomaly data is not replaced with arbitrary constants. When the model input requires a fixed dimension, a validity mask is configured at zero-fill locations to enable the model to distinguish between missing locations and true zero-amplitude locations.

[0119] Risk threshold calibration is based on independent calibration data. The calibration data is compiled by spatial monitoring unit, statistically analyzing the number of missed alarms, false alarms, average early warning lead time, and risk level maintenance time under different candidate thresholds. Mine risk management requires setting different costs for missed alarms and false alarms. The threshold selection program chooses the threshold with the lower cost of false alarms from among the candidate thresholds that meet the lower limit of recall rate for high-risk events. After the upgrade threshold is determined, the downgrade threshold shifts towards lower values ​​based on the fluctuation range of the risk value under stable operating conditions. The threshold version and the model version are saved separately. After the model is updated, verification is first performed under the original threshold before determining whether threshold calibration needs to be re-executed.

[0120] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0121] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. It should be noted that any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A machine learning-based method for early warning of rockbursts in coal mines, characterized in that, include: Acquire microseismic event data, coal and rock mass stress data, and acoustic emission signal data within the space monitoring unit; configure a unified time reference for the microseismic event data, the coal and rock mass stress data, and the acoustic emission signal data; and collect multi-source monitoring data corresponding to the formation time of the continuous monitoring window. The multi-source monitoring data is preprocessed and feature extracted to form a multi-source time-series feature sequence and a microseismic-acoustic emission waveform data tensor. The multi-source time-series feature sequence includes microseismic statistical features, coal and rock mass stress features, and acoustic emission statistical features for multiple consecutive time steps. The microseismic-acoustic emission waveform data tensor includes microseismic waveform segments and acoustic emission waveform segments corresponding to the continuous monitoring window. The multi-source time-series feature sequence is input into a long short-term memory network (LSTM) time-series prediction model to obtain the time-series risk prediction value within the target prediction time range after the current continuous monitoring window ends. The microseismic-acoustic emission waveform data tensor is input into a one-dimensional convolutional neural network (CNN) pattern recognition model to obtain the precursor mode probability value within the target prediction time range. From historical early warning periods with confirmed results, the model reliability of the LSTM time-series prediction model and the one-dimensional CNN pattern recognition model are determined respectively, and the fusion weights of the two models are updated based on the model reliability. A comprehensive risk prediction value is calculated based on the time-series risk prediction value, the precursor mode probability value, and the fusion weights, and the rockburst risk level is determined based on the comprehensive risk prediction value.

2. The method for early warning of coal mine rockburst based on machine learning according to claim 1, characterized in that, Configuring a unified time reference for the microseismic event data, the coal and rock mass stress data, and the acoustic emission signal data includes: performing clock correction on the microseismic acquisition nodes, stress acquisition nodes, and acoustic emission acquisition nodes using a precise time protocol; writing the corrected acquisition time into a data frame; dividing the continuous monitoring window according to a preset monitoring cycle; maintaining the coal and rock mass stress data to the corresponding time step according to the timestamp; merging the microseismic event data and acoustic emission signal data to the corresponding time step according to the event occurrence time; and writing a missing flag for missing time steps.

3. The method for early warning of coal mine rockburst based on machine learning according to claim 1, characterized in that, The multi-source monitoring data is preprocessed and its features are extracted, including: performing wavelet decomposition on the signal within the effective response frequency band obtained from sensor calibration, performing soft thresholding on the wavelet detail coefficients and reconstructing the signal; extracting the mean, variance, peak value, root mean square value, dominant frequency, frequency band energy and spectral entropy from the reconstructed signal; standardizing the extraction results; projecting the principal component projection matrix obtained from the training data, selecting the principal components whose cumulative contribution rate reaches a preset contribution rate threshold, and forming the multi-source time-series feature sequence.

4. The method for early warning of coal mine rockburst based on machine learning according to claim 1, characterized in that, The training samples of the Long Short-Term Memory Network time-series prediction model consist of multi-source time-series feature sequences corresponding to multiple consecutive monitoring windows. The training labels are determined by the confirmed risk results within the target prediction time range after the input sequence ends. The training set, validation set, and test set are divided according to the monitoring time or mining stage so that samples with overlapping time ranges are located in the same data subset. The long short-term memory network time-series prediction model outputs a time-series risk prediction value with a value between zero and one.

5. The method for early warning of coal mine rockburst based on machine learning according to claim 1, characterized in that, The formation of the microseismic-acoustic emission waveform data tensor includes: extracting microseismic waveform segments and acoustic emission waveform segments according to the event trigger time; performing bandpass filtering, amplitude normalization, resampling, and fixed-length processing on the waveform segments; stacking the processed waveform segments into channels according to the sensor number and signal type; the positive class label determination rule of the one-dimensional convolutional neural network pattern recognition model is: if a confirmed rockburst event exists within the target prediction time range after the end of the continuous monitoring window to which the waveform segment belongs, it is marked as positive.

6. The method for early warning of coal mine rockburst based on machine learning according to claim 1, characterized in that, For the model identifier m, which is either the Long Short-Term Memory Network Temporal Prediction Model Identifier L or the One-Dimensional Convolutional Neural Network Pattern Recognition Model Identifier C, the model reliability is calculated based on the historical early warning period for which confirmed results have been obtained: ； in, This represents the model reliability corresponding to the continuous monitoring window t. This represents the correct prediction prior count corresponding to model identifier m. This represents the prior count of incorrect predictions corresponding to model identifier m. This indicates the number of correct predictions corresponding to model identifier m within a preset historical range prior to the continuous monitoring window t. This represents the number of incorrect predictions corresponding to model identifier m within a preset historical range prior to the continuous monitoring window t; candidate weights are calculated based on model reliability, and exponential smoothing and normalization are performed on the candidate weights to make the sum of the fusion weights of the long short-term memory network time-series prediction model and the fusion weights of the one-dimensional convolutional neural network pattern recognition model equal to one.

7. The method for early warning of coal mine rockburst based on machine learning according to claim 6, characterized in that, The data quality coefficient is calculated based on the input data corresponding to the model identifier m within the current continuous monitoring window. The data quality coefficient is determined by data integrity, time synchronization quality, signal-to-noise ratio, and outlier ratio. The model reliability is corrected using the data quality coefficient, and the candidate weights are calculated based on the corrected model reliability.

8. A method for early warning of coal mine rockburst based on machine learning according to claim 2, characterized in that, After clock calibration using a precise time protocol, the propagation delay is calculated based on the microseismic event location coordinates, sensor three-dimensional coordinates, and the propagation velocity in the coal and rock medium: ； ； in, This indicates that the wave corresponding to event number e propagates to sensor number e. The propagation delay Indicates sensor number The three-dimensional coordinate vector, This represents the location coordinate vector of event number e. This indicates the calibration propagation speed in the coal and rock region where event number e is located. Indicates sensor number The recorded waveform arrival time, Indicated based on sensor number The event occurrence time is calculated inversely; the microseismic waveform segment and acoustic emission waveform segment are aligned according to the event occurrence time.

9. A method for early warning of coal mine rockburst based on machine learning according to claim 1, characterized in that, The starting times of the stress growth phase, the acoustic emission enhancement phase, and the microseismic enhancement phase are determined respectively. The consistency coefficient of the precursor phase is calculated based on the degree of matching between the three starting times and the preset precursor sequence. The comprehensive risk prediction value is incrementally corrected based on the consistency coefficient of the precursor phase.

10. A method for early warning of coal mine rockburst based on machine learning according to claim 1, characterized in that, Calculate the model divergence between the predicted time-series risk value and the probability value of the precursor pattern; when the model divergence exceeds the divergence trigger threshold within multiple consecutive monitoring windows, send a sampling parameter adjustment instruction to the monitoring and acquisition node. The sampling parameter adjustment instruction is used to shorten the sampling interval of coal and rock mass stress data and increase the retention ratio of original data of microseismic waveform segments and acoustic emission waveform segments.

Images