Rock burst risk early warning system and method fusing microseismic signal and voiceprint recognition

Abstract

The application discloses a rock burst risk early warning system and method fusing microseismic signals and voiceprint recognition, signals are synchronously collected by a microseismic sensor network and a voiceprint sensor network, effective microseismic events and effective voiceprint events are obtained through noise reduction and short-time energy ratio method event detection; a microseismic feature vector is formed by extracting a focal coordinate, seismic moment, radiation energy and magnitude frequency coefficient of the microseismic event, a voiceprint feature vector is formed by extracting a root mean square value, spectral barycenter and mel frequency cepstrum coefficient of the voiceprint event; based on wave propagation time compensation, the two types of events are compactly matched within a preset space-time window, the matched voiceprint feature vector is aggregated by mean value, maximum value and standard deviation, and then is fused with the microseismic feature vector, and the fusion result is input into a machine learning model to output a rock burst risk probability, and compared with a three-level threshold to trigger a corresponding grade early warning. The application realizes the perception of rock mass rupture, and significantly improves the early warning accuracy and grading response ability.

Description

Technical Field

[0001] This invention relates to the field of rockburst disaster monitoring and early warning technology, specifically to a rockburst risk early warning system and method that integrates microseismic signals and acoustic signature recognition. Background Technology

[0002] Rockburst refers to a nonlinear dynamic failure phenomenon in which rock masses instantaneously release elastic strain energy due to excavation unloading or disturbance under deep, high-stress conditions, causing rock to burst, eject, or be thrown. It is characterized by its suddenness and destructiveness, seriously threatening the construction and operational safety of deep underground engineering projects. How to provide reliable, graded early warnings before rockbursts occur has always been a research hotspot and engineering challenge in the fields of rock mechanics and mining engineering.

[0003] Existing rockburst monitoring technologies can be broadly categorized into single-source monitoring and multi-source monitoring. Single-source monitoring, represented by microseismic monitoring, involves deploying three-component accelerometers or velocity meters around the tunnel or chamber to record low-frequency elastic waves generated by the rock mass during stress adjustment. Source parameters such as source coordinates, seismic moment, and radiated energy are then retrieved to determine the stress state and fracture scale of the rock mass. Acoustic emission monitoring, on the other hand, focuses on high-frequency transient elastic waves, reflecting the initiation and propagation of microcracks. However, single-sensor modes have inherent limitations: microseismic monitoring is sensitive to macroscopic fractures but has a delayed response to microcracks that are precursors to rockbursts; acoustic emission monitoring has high sensitivity but a short propagation distance and is susceptible to environmental noise interference. Neither can independently support highly reliable graded early warning systems.

[0004] To compensate for the shortcomings of single-source methods, multi-source fusion methods that jointly monitor various signals such as microseismic activity, acoustic emission, and electromagnetic radiation have emerged in recent years. While these methods improve the comprehensiveness of early warning, significant shortcomings remain in areas such as how to achieve refined spatiotemporal alignment of heterogeneous signals, how to extract discriminative features related to rock fracturing mechanisms from acoustic signals, and how to train stable early warning models under imbalanced sample conditions. Specifically, existing methods often use wide time windows for loose correlation, which can easily misclassify unrelated events as responses from the same rock mass; acoustic signal feature extraction often remains at the level of time-domain statistical parameters, failing to fully explore the fracturing mechanism information contained in the spectral envelope; machine learning models often process structured features and temporal features with a single structure, making it difficult to balance the discrimination efficiency of different feature spaces. Furthermore, the selection of thresholds often relies on experience and lacks a calibration process that matches the rock mass physical parameters, making it difficult to guarantee the early warning effect during actual deployment in engineering fields.

[0005] Therefore, how to achieve tight spatiotemporal alignment and feature fusion of microseismic signals and high-frequency acoustic features based on the idea of ​​voiceprint recognition, and to establish a rockburst risk early warning method with rigorous physical dimensions, traceable threshold calibration, and matching model structure and feature type, is a technical problem that urgently needs to be solved. Summary of the Invention

[0006] To address the shortcomings of the aforementioned background technologies, this invention proposes a rockburst risk early warning method that integrates microseismic signals and acoustic signature recognition. This method primarily solves the technical problems of insufficient accuracy and on-site operability of existing rockburst monitoring and early warning technologies, which are caused by factors such as a single sensing scale, loose spatiotemporal alignment of multi-source asynchronous signals, insufficient acoustic feature discrimination, fragmented modeling of structured features and temporal features, and the inability of single threshold early warning to match the requirements of graded engineering responses.

[0007] The method includes the following steps: Step S1: Simultaneously acquire micro-vibration signals and acoustic fingerprint signals through a micro-vibration sensor network and an acoustic fingerprint sensor network, and obtain valid micro-vibration events and valid acoustic fingerprint events through noise reduction and event detection; Step S2: Extract the microseismic feature vector of the effective microseismic event, including source coordinates, seismic moment, radiated energy and magnitude frequency coefficient, and the acoustic feature vector of the effective acoustic event, including root mean square value, spectral centroid and Mel frequency cepstral coefficient. Step S3: Match the effective voiceprint events with the effective micro-vibration events within a preset time window and a preset spatial window, and aggregate the matched voiceprint feature vectors to obtain an aggregated voiceprint feature vector; Step S4: The microseismic feature vector and the aggregated acoustic feature vector are fused into a fused feature vector. The rockburst risk probability is output by the machine learning model. The corresponding level of warning is triggered by comparing the rockburst risk probability with the preset warning threshold.

[0008] Preferably, the noise reduction in step S1 includes: using a notch filter with a cutoff frequency of 50Hz or 60Hz to eliminate power frequency interference on the micro-vibration signal, and using a Butterworth bandpass filter with a passband of 10Hz to 1kHz for bandpass filtering; and using spectral subtraction or wavelet thresholding to reduce noise on the acoustic signature signal.

[0009] Preferably, the event detection in step S1 uses the short-duration-long-duration energy ratio method, and the first event is calculated using the following formula. The short-time energy average of each sampling point Compared with long-term energy average : ; ; in, For the first Instantaneous energy at each sampling point It is a short-time smoothing factor. For long-term smoothing factor, Less than When the ratio An event is triggered when the value exceeds a preset threshold T, where the threshold T ranges from 2.5 to 4.0.

[0010] Preferably, the matching in step S3 includes: using the time of occurrence of the effective microseismic event. Using the earthquake source coordinates as a reference, calculate the distance d from the earthquake source coordinates to the acoustic sensor that triggered the valid acoustic event, and the wave propagation time compensation. : ; in, This represents the propagation speed of the acoustic signal within the rock mass. If the triggering time of the valid voiceprint event satisfy If d does not exceed the radius of the preset spatial window, then the match is considered successful. The length of the preset time window.

[0011] Preferably, the aggregation in step S3 includes simultaneously performing mean aggregation, maximum value aggregation, and standard deviation aggregation on the matched voiceprint feature vector, and the aggregated voiceprint feature vector is formed by concatenating the results of the mean aggregation, the maximum value aggregation, and the standard deviation aggregation.

[0012] Preferably, the fusion in step S4 is performed in either of the following two ways: Method 1: The microseismic feature vector and the aggregated acoustic signature feature vector are directly concatenated after Z-score normalization to obtain the fused feature vector; Method 2: After processing the microseismic feature vector and the aggregated acoustic signature feature vector separately through a dual-branch neural network, the outputs of the two branches are concatenated to obtain the fused feature vector. In the dual-branch neural network, the branch that processes the microseismic feature vector is a fully connected neural network, and the branch that processes the aggregated acoustic signature feature vector is a long short-term memory network or a one-dimensional convolutional neural network.

[0013] Preferably, the machine learning model described in step S4 undergoes a training phase before being put into use. The training phase includes: constructing a training sample set, wherein positive samples in the training sample set consist of the fused feature vector and corresponding rockburst labels within a preset time period before the occurrence of historical rockburst events, and negative samples consist of the fused feature vector and corresponding non-rockburst labels during non-rockburst periods; iteratively updating the parameters of the machine learning model until the loss function converges on the validation set.

[0014] Preferably, the preset warning threshold includes a first threshold, a second threshold, and a third threshold, wherein the first threshold is less than the second threshold, the second threshold is less than the third threshold, and triggering the corresponding level warning in step S4 includes: A yellow alert is triggered when the probability of rockburst risk is greater than the first threshold and not greater than the second threshold. The yellow alert includes pushing a prompt message to the monitoring terminal. An orange alert is triggered when the probability of rockburst risk is greater than the second threshold and not greater than the third threshold. The orange alert includes pushing a prompt message to the monitoring terminal and activating the on-site audible and visual warning device. A red alert is triggered when the probability of rockburst risk exceeds the third threshold. The red alert includes sending a prompt message to the monitoring terminal, activating the on-site audible and visual warning device, and triggering an automatic evacuation command.

[0015] The present invention also provides a rockburst risk early warning system that integrates microseismic signals and acoustic signature recognition, applicable to the above-mentioned method, including: a signal acquisition module, a feature extraction module, a spatiotemporal matching and aggregation module, and a risk prediction and early warning module; The signal acquisition module is used to synchronously acquire micro-vibration signals and acoustic fingerprint signals through a micro-vibration sensor network and an acoustic fingerprint sensor network, and obtain valid micro-vibration events and valid acoustic fingerprint events through noise reduction and event detection. The feature extraction module is used to extract the microseismic feature vector of the effective microseismic event, including source coordinates, seismic moment, radiation energy and magnitude frequency coefficient, and the acoustic feature vector of the effective acoustic event, including root mean square value, spectral centroid and Mel frequency cepstral coefficient. The spatiotemporal matching and aggregation module is used to match the effective voiceprint events with the effective micro-vibration events within a preset time window and a preset spatial window, and to aggregate the matched voiceprint feature vectors to obtain an aggregated voiceprint feature vector. The risk prediction and early warning module is used to fuse the microseismic feature vector and the aggregated acoustic feature vector into a fused feature vector, output the rockburst risk probability through a machine learning model, and trigger the corresponding level of early warning based on the comparison between the rockburst risk probability and a preset early warning threshold.

[0016] Preferably, the system further includes a model training module, which is used to construct a training sample set and train the machine learning model. The positive samples in the training sample set consist of the fused feature vector and the corresponding rockburst label within a preset time period before the occurrence of historical rockburst events, and the negative samples consist of the fused feature vector and the corresponding non-rockburst label during non-rockburst periods. The parameters of the machine learning model are iteratively updated until the loss function converges on the validation set.

[0017] The beneficial effects of the present invention include at least the following: First, by coordinating the monitoring of low-frequency micro-vibrations and high-frequency acoustic signatures, this invention can simultaneously capture macroscopic stress adjustment events and microscopic crack propagation events. Compared with single-source micro-vibration schemes, it can advance the rockburst precursor identification window by tens of minutes to several hours, significantly expanding the available time margin for handling.

[0018] Second, by using wave propagation time compensation and a tight spatiotemporal matching window, this invention can significantly reduce the correlation error of heterogeneous events, and the false correlation rate is greatly reduced compared with the traditional wide time window scheme, effectively suppressing the contamination of fusion features by unrelated acoustic events.

[0019] Third, by introducing acoustic features such as Mel frequency cepstral coefficients, this invention can characterize the spectral envelope differences of different fracture modes, enabling the early warning model to identify different fracture mechanisms such as tensile fracture and shear fracture, providing richer feature support for subsequent rockburst mechanism research and engineering disposal decisions.

[0020] Fourth, through the parallel structure of the dual-branch neural network, this invention enables the structured microseismic features and temporal acoustic features to be processed by their respective optimal sub-networks, maintaining high prediction stability even in engineering scenarios with imbalanced samples.

[0021] Fifth, by designing three-level thresholds (yellow, orange, and red) and differentiated response actions, this invention enables the output of the early warning system to be directly linked with different levels of on-site handling actions, such as push notifications from monitoring terminals, activation of on-site audio-visual warning devices, and triggering of automatic evacuation commands, thereby improving the operability of the early warning results at the engineering implementation level. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the overall process of the rockburst risk early warning method that integrates microseismic signals and acoustic fingerprint recognition provided in an embodiment of the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0024] This invention provides a rockburst risk early warning method that integrates microseismic signals and acoustic signature recognition. The overall process is as follows: Figure 1 As shown, it mainly includes four steps: signal acquisition and event detection, feature extraction, spatiotemporal matching and aggregation, and feature fusion and hierarchical early warning.

[0025] Step S1: Simultaneously acquire micro-vibration signals and acoustic fingerprint signals through a micro-vibration sensor network and an acoustic fingerprint sensor network, and obtain valid micro-vibration events and valid acoustic fingerprint events through noise reduction and event detection.

[0026] Specifically, in this embodiment of the invention, a microseismic sensor network and an acoustic fingerprint sensor network are deployed collaboratively within the monitoring area. The microseismic sensor network is used to capture elastic waves (P-waves and S-waves) generated by internal rock fractures. The microseismic sensors are selected from three-component accelerometers or velocities with a frequency response range of 5 Hz to 2 kHz. A three-dimensional sensor network is constructed in the monitoring area, such as in roadways, the roof and floor of mining areas, and behind tunnel faces. The sensor spacing is determined according to the monitoring accuracy requirements; the spacing in general areas is 15 to 20 meters, while in key monitoring areas such as stress concentration zones, geological structural zones, and behind tunnel faces, the spacing is increased to 5 to 10 meters. Installation is achieved through rock drilling, using epoxy resin or a special anchoring agent to tightly couple the sensors to the rock mass, ensuring signal transmission efficiency and avoiding near-field interference from roadway disturbance zones.

[0027] Acoustic sensor networks are used to capture sound and ultrasonic signals generated by processes such as micro-fracture, friction, and spalling on the surface and in shallow layers of rock masses. They are primarily composed of two types of sensors: the first type is a high-frequency acoustic emission sensor with a frequency range of 20 kHz to 1 MHz for high-frequency stress waves; the second type is an industrial microphone with a frequency response range of 20 Hz to 20 kHz for capturing audible sound waves. Through the complementary frequency bands of the two types of acoustic sensors, the acoustic signal covers both high-frequency micro-crack activity and low-frequency macro-fracture noise perceptible to the human ear.

[0028] In this embodiment of the invention, the synchronous acquisition of the micro-vibration sensor network and the acoustic fingerprint sensor network is achieved through the IEEE 1588 Precision Time Protocol or a satellite timing module, ensuring that the time deviation of all sensor nodes in the network is less than 1 ms. The sampling rate of the micro-vibration signal is set to 10 kHz to meet the requirements of elastic wave dominant frequency analysis; in the acoustic fingerprint signal, the sampling rate of the acoustic emission channel is set to 500 kHz to 1 MHz, and the sampling rate of the industrial microphone channel is not less than 44.1 kHz. All raw signals are packaged by the field pre-processing unit and then transmitted to the central processing server for subsequent analysis. In another embodiment of the invention, the triggering mechanism also includes two modes: continuous recording and event triggering. In continuous mode, all data is uploaded in real time; in event triggering mode, when any channel signal exceeds a threshold, a data buffer of 2 to 5 seconds before and after the event is automatically saved to reduce data redundancy.

[0029] At the working mechanism level, this invention combines low-frequency microseismic signals with high-frequency acoustic signature signals for joint monitoring, essentially constructing fracture information channels at both macroscopic and microscopic scales on the same rock mass. The dominant frequency band of the microseismic signal is concentrated between 5 Hz and 2 kHz, suitable for capturing macroscopic fracture events, but insensitive to the initiation of tiny cracks. The acoustic signature signal, especially the acoustic emission frequency band from 20 kHz to 1 MHz, can cover the energy frequency band radiated by the initiation and propagation of microcracks within the rock. The microseismic sensor corresponds to the macroscopic stress adjustment stage, while the acoustic signature sensor corresponds to the microscopic damage accumulation stage. Together, they form a continuous observation chain of "precursor-evolution-trigger" in time series, thus achieving scale complementarity for the entire process of rockburst formation at the physical mechanism level. This solves the problem that single-source monitoring cannot simultaneously detect macroscopic fractures and capture microscopic cracks.

[0030] The original microseismic signals and acoustic signature signals are inevitably affected by multiple sources of noise in the engineering field, such as power frequency interference, mechanical vibration, wind noise, and electromagnetic interference. If no effective noise reduction is performed, directly entering the event detection stage will lead to a large number of false alarms and missed alarms. Therefore, in this embodiment of the invention, a notch filter with a cutoff frequency of 50 Hz or 60 Hz and a bandwidth of ±2 Hz is used to eliminate power frequency interference such as power line interference for the microseismic signals, and a 4th to 8th order Butterworth bandpass filter with a passband of 10 Hz to 1 kHz is used for bandpass filtering to retain the effective microseismic frequency band and suppress out-of-band noise. For the acoustic signature signal, spectral subtraction or wavelet thresholding is used for noise reduction. Spectral subtraction is suitable for working conditions with relatively stable background noise, while wavelet thresholding is suitable for complex working conditions containing non-stationary impact noise. The two methods can be selected according to the noise characteristics of the site. In a strong interference environment, an LMS / NLMS adaptive filter can be used, with a reference noise channel as input, to cancel coherent noise in real time.

[0031] The denoised signal is fed into the short-time-long-time energy ratio method for event detection, automatically identifying valid microseismic events and acoustic emission events from the continuous data stream. In this embodiment of the invention, the short-time-long-time energy ratio method calculates the first event using an exponentially weighted moving average. The short-time energy average of each sampling point Compared with long-term energy average The calculation formula is as follows: ; ; In the above two formulas, For the first The instantaneous energy of a sampling point is defined as the signal amplitude value at that sampling point; It is a short-time smoothing factor, with a value ranging from 0.6 to 0.8, which makes the short-time energy average value respond quickly to the current energy change. This is a long-term smoothing factor, ranging from 0.95 to 0.998, which gives the long-term energy average a strong smoothing ability for short-term fluctuations, i.e., satisfies... Less than The relationship.

[0032] When the ratio The event is considered to have started when the ratio first exceeds a preset threshold T; the event is considered to have ended when the ratio falls below the threshold T and remains below it for a certain period of time. The threshold T ranges from 2.5 to 4.0, with a typical value of 3.0. In this embodiment, the threshold T is dynamically optimized in engineering practice, rather than being a fixed empirical constant. Initially, it is set empirically within the range of 2.5 to 4, with the specific value determined based on the on-site noise level. After the system is running, the threshold is quantified and calibrated using the "mean + k times standard deviation" method by statistically analyzing the distribution of the long-term and short-term average ratio STA / LTA of long-term noise data; ROC analysis is then performed in conjunction with historical labeled events for optimization. An adaptive mechanism is introduced during long-term operation to dynamically fine-tune the threshold based on the real-time noise level and the false alarm / missed alarm rate.

[0033] It should be noted that event detection is applied to both the microseismic signal channel and the acoustic signature signal channel, but the instantaneous energy calculation window length and threshold T used for each differ slightly to adapt to their respective signal bandwidths and noise levels. Events detected in the microseismic channel are called valid microseismic events, and events detected in the acoustic signature channel are called valid acoustic signature events. The output of each valid event includes the start time, end time, the sensor number, and the corresponding original waveform segment, which serve as input for subsequent feature extraction.

[0034] In another embodiment of the present invention, for suspected events after triggering, event descaling is further performed to eliminate non-fragmented interference. Microseismic events must simultaneously meet three conditions: first, the peak amplitude is not less than three times the background noise RMS to ensure the signal-to-noise ratio; second, the duration is between 10 milliseconds and 10 seconds, thereby eliminating transient interference such as electrical pulses and sparks less than 10 milliseconds, as well as continuous noise and multiple events superimposed for more than 10 seconds; third, spatial consistency is required, i.e., at least four sensors are triggered synchronously, and the deviation between the P-wave first arrival time difference recorded by each sensor and the theoretical propagation time difference calculated according to the regional wave velocity model is not greater than 10 milliseconds.

[0035] The amplitude selection for acoustic emission events also uses a criterion based on the 3σ criterion: ; In the formula, The peak amplitude of the event. The root mean square value of the background noise before the event is triggered, and the signal-to-noise ratio coefficient. The value is 3 to 4 for the high-frequency acoustic emission channel, 3 for the audible sound channel, and 2.5 for high-sensitivity scenarios. Duration of the acoustic emission event. Must meet The high-frequency acoustic emission channel is taken millisecond, Milliseconds, audible channel millisecond, The lower limit aims to eliminate transient interference with a duration of less than 0.1 milliseconds, such as electrical pulses and sparks; the upper limit aims to eliminate continuous noise and the superposition of multiple events.

[0036] Step S2: Extract the microseismic feature vector of effective microseismic events, including source coordinates, seismic moment, radiation energy, and magnitude frequency coefficient, and the acoustic feature vector of effective acoustic events, including root mean square value, spectral centroid, and Mel frequency cepstral coefficient.

[0037] Specifically, for the valid microseismic events selected in step S1, microseismic feature vectors are extracted; for the valid acoustic emission events selected in step S1, acoustic signature feature vectors are extracted. Both types of feature vectors are statistically analyzed within a preset sliding time window and spatial window. For example, in this embodiment, the time window length is 10 minutes and the sliding step size is 1 minute; the spatial window is a sphere with a radius of 80 to 100 meters centered on key areas such as mining areas or fault zones, or the monitoring section is divided into regular grid partitions of 50 meters × 50 meters × 50 meters for statistical analysis; for features requiring a sufficient sample size, such as the magnitude frequency coefficient b-value, the time window is adaptively extended until the cumulative number of events within the window is not less than 50.

[0038] The microseismic characteristic vector consists of four categories: location and energy parameters, statistical characteristics, waveform complexity characteristics, and spatiotemporal evolution characteristics. Among the location and energy parameters, the source coordinates... The solution is obtained by using the Geiger iterative method or the double-difference positioning method with the arrival time difference of P-wave and S-wave as input.

[0039] The region-specific P-wave and S-wave velocity models are calibrated through manual blasting. For example, 3-5 artificial blasting points are evenly distributed within the monitoring section, and the coordinates of these points are measured using RTK-GPS (error less than 0.5 meters). Before blasting, the synchronization deviation of all sensors across the network is confirmed to be less than 1 millisecond. After blasting, waveforms are collected, and the first arrival times of P-waves (accuracy no greater than 1 millisecond) and S-waves (accuracy no greater than 2 milliseconds) are manually recorded. Using the precise coordinates and first arrival times of the sensors as input, the P-wave and S-wave velocities are inverted using the least squares method to minimize the residual difference between theoretical and measured travel times. The model is then verified using independent blasting points that were not calibrated. Once the horizontal positioning error is no greater than 10 meters and the vertical error no greater than 15 meters, the model is activated. Recalibration is performed quarterly during the monitoring period. Regarding energy parameters, the seismic moment... With radiation energy Calculate according to the following formulas: ; ; In the formula, This refers to the density of the rock mass, expressed in kilograms per cubic meter. S-wave velocity, measured in meters per second; This is the distance from the epicenter, in meters; The far-field displacement spectrum is in the low-frequency range, meaning the frequency is much smaller than the source corner frequency. The asymptotic plateau value is determined in engineering by the geometric mean of the low-frequency band of the displacement spectrum; This is the radiation pattern factor, usually taken as 0.6; The far-field displacement spectrum in the low-frequency range (frequency much smaller than the source corner frequency) The asymptotic plateau value of the displacement spectrum is determined in engineering by the geometric mean of the low-frequency band of the displacement spectrum. The waveform represents velocity. Duration of the event; This is the derivative of the velocity waveform, i.e., acceleration.

[0040] In terms of statistical characteristics, the cumulative number of events The magnitude within the window is not less than the lower limit of magnitude. Total number of events; energy release rate The sum of energy released for each event within the window and the length of the time window. The ratio, that is The magnitude frequency coefficient b reflects the proportion of large and small events within a window, as determined by the Gutenberg-Richter formula. The fitting results show that a continuous decrease in the b-value is a precursor to high risk; apparent stress with apparent volume The fracture efficiency and damage extent are reflected by the following formula: ; ; In the formula, Shear modulus, in Pascals; The energy is radiated in Joules. Increased apparent stress indicates enhanced brittle fracture, while increased apparent volume indicates expansion of the unstable region.

[0041] Waveform complexity features include the frequency corresponding to the peak value of the amplitude spectrum, i.e., the dominant frequency. The frequency width, or bandwidth B, where the amplitude spectrum drops to -3dB from the peak value; and the corner frequencies obtained by fitting the displacement spectrum using the Brune model, which are positively correlated with the fracture size. And the energy ratio of P-waves to S-waves .

[0042] Spatiotemporal evolution characteristics based on event density With source mobility Reflecting the spatial concentration and migration trend of rupture activity, the concentration of seismic sources in a certain region usually indicates strain localization in that region, and the focal migration rate... The expression is: ; The voiceprint feature vector consists of three categories: time-domain features, frequency-domain features, and advanced acoustic features. In the time-domain features, the root mean square (RMS) reflects the signal energy; the event rate is the number of events per unit time; and the ring count is the number of oscillations where the waveform amplitude of a single event exceeds a preset threshold. In this embodiment, the threshold is taken as 3-5 times or 2-3 times the background noise RMS, corresponding to the high-frequency acoustic emission and audible sound channels respectively, and can be dynamically adjusted according to the on-site signal-to-noise ratio. The amplitude distribution is determined by skewness. and kurtosis Quantification: ; ; In the formula, n is the number of events within the time window. For the event amplitude, The mean amplitude is denoted as . A skewness greater than 0 indicates that the proportion of major events is relatively high, and a kurtosis greater than 0 indicates that the distribution has a sharp peak and thick tails, and the probability of a maximum value is higher than that of a normal distribution. The simultaneous increase of both usually corresponds to a continuous increase in rupture intensity and an increase in the risk of detonation. By combining the evolution of ring count and event rate, the stage characteristics of rupture from the initial stage through the middle stage to the detonation stage can be identified, which can serve as an important time-domain basis for early warning decision-making.

[0043] Regarding frequency domain characteristics, an FFT transform is performed on the entire waveform of each event, and the spectral centroid is... Spectral roll-off frequency and spectrum width Calculate or define them respectively according to the following formulas: ; ; In the formula, and The first Frequency variables and amplitudes at each frequency point This represents the total number of frequency points. The frequency at which the cumulative amplitude reaches 85% of the sum of the amplitudes at all frequencies is taken. The harmonic ratio (HR) is defined as the ratio of the integral of the power spectrum in the high-frequency band to the integral of the power spectrum in the low-frequency band. ; In the formula, The power spectral density is used; the upper and lower limits of the high-frequency integration are taken as follows: and The upper and lower limits of the low-frequency band integral are taken as follows: and ;in Typically, 20kHz or 0Hz is used. The threshold is determined by the sampling rate (e.g., 250kHz for a sampling rate of 500kHz). When HR is less than 1, shear fracture is the main type; when HR is greater than 3, tensile fracture is the main type; and when HR is between 1 and 3, tensile-shear combined fracture is the main type. The specific threshold needs to be calibrated on-site based on the lithology of the mining area and the sensor response characteristics.

[0044] In terms of advanced acoustic features, Mel frequency cepstral coefficients (MFCCs) are based on the critical band characteristics of human hearing and are sensitive to the "timbre" of broken sounds. They can effectively distinguish different sound sources such as tension, shearing, and friction. The extraction process includes pre-emphasis (coefficient of 0.97), framing (frame length 25 ms, frame shift 10 ms), Hamming windowing, 2048-point FFT, 20 to 40 Mel triangular filter banks, logarithmic energy extraction, and discrete cosine transform to obtain 12 to 16-dimensional coefficients. Linear predictive coefficients (LPCs) are obtained with the same pre-emphasis, framing, and windowing parameters. Through autocorrelation analysis and Levinson-Durbin recursive solution, 8th to 12th order coefficients are obtained, reflecting the spectral envelope of the broken sound source. Wavelet packet energy features are obtained by performing 3 to 5 layers of wavelet packet decomposition on the signal using the db4 wavelet basis to obtain sub-band energy and energy ratio, providing time-frequency localization information, which is suitable for fine analysis of non-stationary acoustic signals. The three types of features are used together to characterize the auditory properties, spectral structure, and time-frequency distribution of the voiceprint signal.

[0045] Step S3: Match valid voiceprint events with valid microseismic events within a preset time window and a preset spatial window, and aggregate the matched voiceprint feature vectors to obtain an aggregated voiceprint feature vector.

[0046] Specifically, the purpose of this step is to match microseismic events with acoustic emission events in order to proceed to the feature fusion stage. The prerequisite for matching is coordinate system one, meaning that the installation coordinates of all sensors are unified to the same three-dimensional rectangular coordinate system. This embodiment uses a mine-independent coordinate system, with the main shaft opening as the origin and the north, east, and vertical directions determined by the mine layout as the X, Y, and Z axes. Underground sensors obtain coordinates using a total station via traverse surveying (accuracy no greater than 0.5 meters). Surface sensors are first measured using RTK-GPS in the UTM coordinate system (horizontal accuracy ±1~2 cm, vertical accuracy ±2~3 cm), and then converted using a pre-calibrated UTM to mine-independent coordinate system conversion formula to ensure consistent coordinate accuracy across the entire network. Under the unified coordinate system and a synchronized clock across the entire network, each event is timestamped at the microsecond level.

[0047] Therefore, for each microseismic event, the corresponding acoustic emission event is searched using the following steps. Let the occurrence time of the microseismic event be... The coordinates of the earthquake source are ( The triggering time of the acoustic emission event is... The sensor coordinates are ( First, calculate the distance d from the microseismic source to the acoustic sensor and the wave propagation time compensation using the following formula. : ; ; In the formula, The propagation speed of acoustic waves in the rock mass is taken as 2000 to 4000 meters per second in this embodiment. The time difference is then calculated. ,when and When both events occur simultaneously, it is determined that the acoustic emission event and the microseismic event are successfully matched; in this embodiment, the time matching window... The spatial matching radius is between 0.005 and 0.01 seconds. The distance is 3 to 5 meters. After a successful match, the acoustic emission events matched for the microseismic event are simultaneously aggregated by mean, maximum value, and standard deviation according to the feature dimension. The three aggregated vectors are then concatenated to form the aggregated acoustic emission feature vector corresponding to the microseismic event. This ensures that the mean reflects the overall trend, the maximum value highlights extreme cases, and the standard deviation reflects the fluctuation of acoustic emission activity, thus avoiding information loss caused by a single aggregation method.

[0048] Step S4: Fuse the microseismic feature vector and the aggregated acoustic signature feature vector into a fused feature vector. The rockburst risk probability is output by the machine learning model. The corresponding level of warning is triggered based on the comparison between the rockburst risk probability and the preset warning threshold.

[0049] This embodiment provides two optional fusion methods, which should be selected based on the scale of the on-site data. One is feature-level stitching, which involves stitching together the microseismic feature vectors. With aggregated voiceprint feature vector Directly concatenated into a high-dimensional fused feature vector In this embodiment Includes but is not limited to , and , Includes but is not limited to RMS, , HR and MFCCs, assuming It is 10-dimensional. If it is 20-dimensional, then The dimensions are 30. In this embodiment, Z-score normalization or Min-Max normalization is applied to each dimension before fusion to avoid large numerical features dominating the model. This approach is applicable to structured feature models such as XGBoost and Random Forest.

[0050] Secondly, there is an automatic fusion method based on a dual-branch neural network. In this embodiment, the microseismic branch takes the microseismic feature vector as input, passes through 2 to 3 fully connected layers (64 and 32 neurons respectively), and uses ReLU as the activation function. The acoustic signature branch takes the acoustic signature feature sequence (20 dimensions × T time steps) as input, passes through 1 LSTM layer (64 units) or 1 one-dimensional convolutional layer (3 kernel length, 32 channels) to extract temporal dependencies. The outputs of the two branches are concatenated or weighted summed at the penultimate fully connected layer, followed by 1 to 2 fully connected layers and then a Softmax layer to output the rockburst risk probability. This method is suitable for application scenarios with large data scales and can adaptively learn the optimal fusion representation through the network.

[0051] Based on the fusion of feature vectors, this embodiment uses XGBoost as the prediction model, supplemented by LSTM as a temporal variant. The model training data consists of at least 1-2 years of field monitoring data, covering different stress states, lithological conditions, construction stages, and typical rockburst events. Rockburst event labels are constructed as positive samples based on feature vectors from 24 hours prior to the event's occurrence, and as negative samples based on feature vectors from any known rockburst event occurring more than 72 hours prior during normal times. Given the class imbalance caused by the rarity of rockbursts, a strategy combining Synthetic Minority Oversampling Technique (SMOTE) and weighted sampling is adopted. SMOTE, in the feature space, applies weighted sampling to each minority class sample... From its Randomly select one of the nearest neighbors ,according to New samples are synthesized, where λ is a random number in the interval [0-1]; weighted sampling assigns weights to each class during the training phase. Distribute inverse proportion to the sample size, i.e.: ; In the formula For the total number of samples, For the total number of categories, denoted as the number of samples in category j. Risk labels are categorized into three types: No risk (Category 0), no rockburst records, stable microseismic and acoustic signature activity; Low risk (Category 1), minor rockburst phenomena, no personnel or equipment damage; High risk (Category 2), moderate to severe rockbursts, causing equipment damage or work stoppage.

[0052] XGBoost is efficient at handling structured features, supports feature importance assessment, and is insensitive to missing values. It iteratively trains multiple decision trees using an additive model, minimizing the objective function in each training round to improve the overall model's prediction accuracy. Its objective function consists of two parts: a loss function and a regularization term. The former measures the error between the model's predictions and the true values, while the latter controls model complexity and prevents overfitting. The regularization term takes the form: ; where T is the number of leaf nodes of a single decision tree, is the output weight of the j-th leaf node, and γ and λ are hyperparameters; γT penalizes the scale of the tree and controls the complexity of the tree by penalizing the number of leaf nodes; L2 regularization is performed on the leaf node weights to prevent the output value of a single node from being too large. The hyperparameter search range is as follows: the number of trees is 100 to 500, the maximum depth is 6 to 10, the learning rate is 0.01 to 0.1, and the subsample ratio is 0.8 to 1.0. During training, the samples are divided into a training set, a validation set, and a test set in a ratio of 7:2:1. 5-fold cross-validation is used to select the optimal hyperparameters, and early stopping (early stopping rounds = 50) is used to prevent overfitting; the loss function uses cross-entropy or Focal Loss, and the optimizer uses Adam, with an initial learning rate of 1×10 - ³ and is adjusted using the cosine annealing strategy. Regularization is performed using Dropout (0.3~0.5) and L2 weight decay (1×10 -4 ). In addition to AUC-ROC, the recall rate and false positive rate of high-risk categories are emphasized as validation metrics. In this embodiment, it is required that the recall rate of high-risk categories on the test set is not less than 85% and the false positive rate is not higher than 15%.

[0053] When a large amount of historical data has been accumulated on site, LSTM can be used to model the time series fusion features. Its network structure is as follows: the input layer accepts sequences with a length T of 10 to 30 ; the hidden layer is composed of 2 layers of LSTM, with 64 units in each layer, and Dropout is set to 0.3; the output layer outputs the probabilities of three categories through Softmax; training generates time series samples in a sliding window manner, with a sliding step size of 1, and is supplemented by gradient clipping, and the clipping threshold is set to 5 for training.

[0054] During the online operation stage of the system, a set of the latest fusion features is taken every 5 to 10 minutes and input into the trained model to output the rock burst risk probability P ∈ [0, 1]. The warning threshold and response actions are set in three levels: when 0.70 < P ≤ 0.85, a yellow warning is triggered, and the corresponding section on the monitoring interface is highlighted and pushed to the duty personnel via text message; when 0.85 <p ≤ 0.95时触发橙色预警，系统自动生成预警报告，通知现场工程师进行人工核查并加强巡查；p>A red alert is triggered at 0.95, automatically activating the audible and visual alarms, sending evacuation instructions to relevant work teams, and initiating emergency plans, including ceasing operations and activating emergency ventilation. The threshold is not fixed; after system commissioning, it is adaptively fine-tuned based on recent false alarm and missed alarm rates: when the false alarm rate remains high, the threshold is increased in increments of 0.01 to 0.02; when the missed alarm rate is high, it is decreased in equal increments. Furthermore, adjustments are made simultaneously under conditions that significantly alter background noise levels, such as during the rainy season or after blasting.

[0055] The data processing and decision-making center simultaneously operates a 3D visualization platform and a feedback evolution module to support the long-term operation and continuous optimization of the system. The 3D visualization platform overlays a 3D mine model or BIM model, displaying sensor locations, microseismic sources, characteristic curves (b-value, energy release rate, acoustic event rate), and a risk probability heatmap in real time. The heatmap is generated based on spatial interpolation algorithms such as Kriging, displaying a gradient from green (low risk) to red (high risk) to aid in on-site decision-making. The feedback evolution module records the trigger time, location, and actual results of each warning, calculating the model's performance on the latest data weekly. If the AUC or high-risk recall rate decreases by more than 5%, model retraining is triggered. Monthly incremental learning adds new data to the training set, implementing incremental training or full retraining to adapt to changing operating conditions. Simultaneously, typical precursor patterns, such as a continuous decrease in b-value and a sudden increase in high-frequency acoustic signatures, are entered into the database to build a case knowledge base that can be accessed through similarity retrieval, providing historical comparison support for on-site handling.

[0056] The present invention also provides a rockburst risk early warning system that integrates microseismic signals and acoustic signature recognition, applicable to the above-mentioned method, including: a signal acquisition module, a feature extraction module, a spatiotemporal matching and aggregation module, and a risk prediction and early warning module; The signal acquisition module is used to synchronously acquire micro-vibration signals and acoustic fingerprint signals through a micro-vibration sensor network and an acoustic fingerprint sensor network, and obtain valid micro-vibration events and valid acoustic fingerprint events through noise reduction and event detection. The feature extraction module is used to extract the microseismic feature vectors of effective microseismic events, including source coordinates, seismic moment, radiation energy, and magnitude frequency coefficients, as well as the acoustic feature vectors of effective acoustic events, including root mean square value, spectral centroid, and Mel frequency cepstral coefficients. The spatiotemporal matching and aggregation module is used to match valid voiceprint events with valid microseismic events within a preset time window and a preset spatial window, and to aggregate the matched voiceprint feature vectors to obtain aggregated voiceprint feature vectors. The risk prediction and early warning module is used to fuse microseismic feature vectors and aggregated acoustic feature vectors into a fused feature vector, output the rockburst risk probability through a machine learning model, and trigger the corresponding level of early warning based on the comparison between the rockburst risk probability and the preset early warning threshold.

[0057] In this embodiment, the system also includes a model training module, which is used to construct a training sample set and train the machine learning model. The positive samples in the training sample set consist of fused feature vectors and corresponding rockburst labels within a preset time period before the occurrence of historical rockburst events, and the negative samples consist of fused feature vectors and corresponding non-rockburst labels during non-rockburst periods. The parameters of the machine learning model are iteratively updated until the loss function converges on the validation set.

[0058] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described; only preferred embodiments of the present invention are illustrated. The descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the present invention. As long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.

[0059] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this invention should be determined by the appended claims.

Claims

1. A rockburst risk early warning method integrating microseismic signals and acoustic signature recognition, characterized in that, Includes the following steps: Step S1: Simultaneously acquire micro-vibration signals and acoustic fingerprint signals through a micro-vibration sensor network and an acoustic fingerprint sensor network, and obtain valid micro-vibration events and valid acoustic fingerprint events through noise reduction and event detection; Step S2: Extract the microseismic feature vector of the effective microseismic event, including source coordinates, seismic moment, radiated energy and magnitude frequency coefficient, and the acoustic feature vector of the effective acoustic event, including root mean square value, spectral centroid and Mel frequency cepstral coefficient. Step S3: Match the effective voiceprint events with the effective micro-vibration events within a preset time window and a preset spatial window, and aggregate the matched voiceprint feature vectors to obtain an aggregated voiceprint feature vector; Step S4: The microseismic feature vector and the aggregated acoustic feature vector are fused into a fused feature vector. The rockburst risk probability is output by the machine learning model. The corresponding level of warning is triggered by comparing the rockburst risk probability with the preset warning threshold.

2. The rockburst risk early warning method integrating microseismic signals and acoustic signature recognition according to claim 1, characterized in that, The noise reduction in step S1 includes: using a notch filter with a cutoff frequency of 50Hz or 60Hz to eliminate power frequency interference on the micro-vibration signal, and using a Butterworth bandpass filter with a passband of 10Hz to 1kHz for bandpass filtering; and using spectral subtraction or wavelet thresholding to reduce noise on the acoustic signature signal.

3. The rockburst risk early warning method integrating microseismic signals and acoustic signature recognition according to claim 1, characterized in that, The event detection in step S1 uses the short-duration-long-duration energy ratio method, calculated according to the following formula: The short-time energy average of each sampling point Compared with long-term energy average : ； ； in, For the first Instantaneous energy at each sampling point It is a short-time smoothing factor. For long-term smoothing factor, Less than When the ratio An event is triggered when the value exceeds a preset threshold T, where the threshold T ranges from 2.5 to 4.

0.

4. The rockburst risk early warning method integrating microseismic signals and acoustic signature recognition according to claim 1, characterized in that, The matching in step S3 includes: using the time of occurrence of the effective microseismic event. Using the earthquake source coordinates as a reference, calculate the distance d from the earthquake source coordinates to the acoustic sensor that triggered the valid acoustic event, and the wave propagation time compensation. : ； in, This represents the propagation speed of the acoustic signal within the rock mass. If the triggering time of the valid voiceprint event satisfy If d does not exceed the radius of the preset spatial window, then the match is considered successful. The length of the preset time window.

5. The rockburst risk early warning method integrating microseismic signals and acoustic signature recognition according to claim 1, characterized in that, The aggregation in step S3 includes simultaneously performing mean aggregation, maximum aggregation, and standard deviation aggregation on the matched voiceprint feature vector. The aggregated voiceprint feature vector is formed by concatenating the results of the mean aggregation, the maximum aggregation, and the standard deviation aggregation.

6. The rockburst risk early warning method integrating microseismic signals and acoustic signature recognition according to claim 1, characterized in that, The fusion described in step S4 can be performed using either of the following two methods: Method 1: The microseismic feature vector and the aggregated acoustic signature feature vector are directly concatenated after Z-score normalization to obtain the fused feature vector; Method 2: After processing the microseismic feature vector and the aggregated acoustic signature feature vector separately through a dual-branch neural network, the outputs of the two branches are concatenated to obtain the fused feature vector. In the dual-branch neural network, the branch that processes the microseismic feature vector is a fully connected neural network, and the branch that processes the aggregated acoustic signature feature vector is a long short-term memory network or a one-dimensional convolutional neural network.

7. The rockburst risk early warning method integrating microseismic signals and acoustic signature recognition according to claim 1, characterized in that, The machine learning model described in step S4 undergoes a training phase before being put into use. The training phase includes: constructing a training sample set, wherein positive samples in the training sample set consist of the fused feature vector and corresponding rockburst label within a preset time period before the occurrence of historical rockburst events, and negative samples consist of the fused feature vector and corresponding non-rockburst label during non-rockburst periods; iteratively updating the parameters of the machine learning model until the loss function converges on the validation set.

8. The rockburst risk early warning method integrating microseismic signals and acoustic signature recognition according to claim 1, characterized in that, The preset warning threshold includes a first threshold, a second threshold, and a third threshold. The first threshold is less than the second threshold, and the second threshold is less than the third threshold. Triggering the corresponding level of warning in step S4 includes: A yellow alert is triggered when the probability of rockburst risk is greater than the first threshold and not greater than the second threshold. The yellow alert includes pushing a prompt message to the monitoring terminal. An orange alert is triggered when the probability of rockburst risk is greater than the second threshold and not greater than the third threshold. The orange alert includes pushing a prompt message to the monitoring terminal and activating the on-site audible and visual warning device. A red alert is triggered when the probability of rockburst risk exceeds the third threshold. The red alert includes sending a prompt message to the monitoring terminal, activating the on-site audible and visual warning device, and triggering an automatic evacuation command.

9. A rockburst risk early warning system integrating microseismic signals and acoustic signature recognition, applicable to the method described in any one of claims 1 to 8, characterized in that, include: The module includes a signal acquisition module, a feature extraction module, a spatiotemporal matching and aggregation module, and a risk prediction and early warning module. The signal acquisition module is used to synchronously acquire micro-vibration signals and acoustic fingerprint signals through a micro-vibration sensor network and an acoustic fingerprint sensor network, and obtain valid micro-vibration events and valid acoustic fingerprint events through noise reduction and event detection. The feature extraction module is used to extract the microseismic feature vector of the effective microseismic event, including source coordinates, seismic moment, radiation energy and magnitude frequency coefficient, and the acoustic feature vector of the effective acoustic event, including root mean square value, spectral centroid and Mel frequency cepstral coefficient. The spatiotemporal matching and aggregation module is used to match the effective voiceprint events with the effective micro-vibration events within a preset time window and a preset spatial window, and to aggregate the matched voiceprint feature vectors to obtain an aggregated voiceprint feature vector. The risk prediction and early warning module is used to fuse the microseismic feature vector and the aggregated acoustic feature vector into a fused feature vector, output the rockburst risk probability through a machine learning model, and trigger the corresponding level of early warning based on the comparison between the rockburst risk probability and a preset early warning threshold.

10. The rockburst risk early warning system integrating microseismic signals and acoustic signature recognition according to claim 9, characterized in that, The system also includes a model training module, which is used to construct a training sample set and train the machine learning model. The positive samples in the training sample set consist of the fused feature vector and the corresponding rockburst label within a preset time period before the occurrence of historical rockburst events, and the negative samples consist of the fused feature vector and the corresponding non-rockburst label during non-rockburst periods. The parameters of the machine learning model are iteratively updated until the loss function converges on the validation set.

Images