A Deep Well Rockburst Early Warning System and Method Based on Multi-Dimensional Monitoring

Abstract

The application discloses a deep well rock burst early warning system and method based on multidimensional monitoring, and belongs to the technical field of deep well rock burst early warning. The method collects multidimensional data such as stress, strain and microseismic through a monitoring network, and obtains an aligned multi-source data set through space-time registration. After dynamic noise suppression processing by matching physical characteristics, strong correlation features are screened through mutual information entropy. Then, frequency domain, time domain and time-frequency domain features are extracted through principal component extraction and phase space reconstruction, forming a multi-dimensional state space data set. The prediction model is inputted to obtain a danger level and trigger a corresponding early warning signal. Finally, the monitoring network layout and the prevention and control measures are dynamically adjusted based on the early warning signal. The method improves the early warning accuracy and real-time performance, and provides effective technical support for deep well rock burst prevention and control.

Description

Technical Field

[0001] This invention relates to the field of deep well rockburst early warning technology, and in particular to a deep well rockburst early warning system and method based on multi-dimensional monitoring. Background Technology

[0002] With the increasing depletion of shallow mineral resources, mining is gradually extending to deep wells. The geological and mechanical environment faced by deep well tunnels and mining areas is becoming increasingly complex. Rockburst, as a dynamic disaster with strong suddenness and great destructiveness, seriously threatens the safety of underground workers and the integrity of production equipment. Therefore, building an efficient rockburst early warning system has become a core requirement for safe mining in deep wells.

[0003] Existing rockburst early warning technologies are mostly based on single or a few monitoring parameters, judging dangerous conditions by setting fixed thresholds or empirical formulas. Some technologies attempt to integrate multi-source data, but mainly rely on simple data splicing or weighted analysis, lacking a systematic spatiotemporal correlation processing mechanism. At the same time, existing early warning models mostly use traditional statistical methods or single machine learning algorithms, which have limited adaptability to the complex dynamic environment of deep wells.

[0004] The current technology has three main problems: First, the monitoring dimension is too limited to fully reflect the multi-physics field evolution characteristics of rockbursts; second, the data processing methods lack specificity and do not take into account the differences in physical characteristics of different types of monitoring data, resulting in poor noise suppression and insufficient feature extraction; and third, the linkage between early warning and prevention and control measures is insufficient, making it impossible to dynamically adjust the monitoring strategy based on real-time early warning results, and making it difficult to achieve closed-loop management from early warning to prevention and control. Summary of the Invention

[0005] In view of this, embodiments of the present invention provide a deep well rockburst early warning system and method based on multi-dimensional monitoring to solve at least one of the above-mentioned technical problems.

[0006] To achieve the above objectives, firstly, a method for early warning of deep well rockburst based on multi-dimensional monitoring is provided, which includes the following steps:

[0007] By utilizing a monitoring network deployed in deep shaft tunnels and surrounding rock, multi-dimensional monitoring data, including stress values, strain values, microseismic event parameters, acoustic signals, temperature data, radon concentration values, electromagnetic radiation intensity, and three-dimensional deformation data of the tunnel surface, are collected in real time.

[0008] The multi-dimensional monitoring data is spatiotemporally registered to obtain a spatiotemporally aligned multi-source monitoring dataset.

[0009] Dynamic noise suppression based on physical characteristics is performed on the multi-source monitoring dataset to obtain a noise-suppressed dataset;

[0010] The mutual information entropy algorithm is used to filter features related to rockburst pressure from the noise-suppressed dataset to obtain a feature-optimized dataset.

[0011] By using principal component extraction and phase space reconstruction methods, frequency domain features, time domain features, and time-frequency domain features are extracted from the feature-optimized dataset to form a multidimensional state space dataset.

[0012] The multidimensional state space dataset is input into a pre-trained rockburst hazard level prediction model to obtain the rockburst hazard level.

[0013] Based on the rockburst hazard level, trigger the corresponding level of early warning signal;

[0014] Based on the corresponding level of early warning signals, the layout of the monitoring network and the prevention and control measures are dynamically adjusted.

[0015] Secondly, a deep well rockburst early warning system based on multi-dimensional monitoring is provided, which includes:

[0016] The data acquisition module includes a monitoring network deployed in the deep shaft tunnel and surrounding rock, used to collect multi-dimensional monitoring data in real time, including stress value, strain value, microseismic event parameters, acoustic signal, temperature data, radon concentration value, electromagnetic radiation intensity and three-dimensional deformation data of the tunnel surface.

[0017] The spatiotemporal registration module is used to perform spatiotemporal registration on the multi-dimensional monitoring data to obtain a spatiotemporally aligned multi-source monitoring dataset.

[0018] The noise suppression module is used to perform dynamic noise suppression based on physical characteristics on the multi-source monitoring dataset to obtain a noise-suppressed dataset.

[0019] The feature filtering module is used to filter features related to rockburst pressure from the noise-suppressed dataset using the mutual information entropy algorithm, so as to obtain a feature-optimized dataset.

[0020] The feature extraction module is used to extract frequency domain features, time domain features, and time-frequency domain features from the feature-optimized dataset using principal component extraction and phase space reconstruction methods to form a multidimensional state space dataset.

[0021] The level prediction module includes a pre-trained rockburst hazard level prediction model, which is used to input the multi-dimensional state space dataset into the model to obtain the rockburst hazard level.

[0022] The early warning triggering module is used to trigger an early warning signal of the corresponding level based on the rockburst hazard level.

[0023] The dynamic adjustment module is used to dynamically adjust the layout and control measures of the monitoring network based on the corresponding level of early warning signals.

[0024] The above technical solution has the following beneficial technical effects:

[0025] This technical solution forms a complete closed loop through multi-dimensional monitoring, spatiotemporal registration, noise suppression, feature screening and extraction, model prediction, and dynamic adjustment, achieving high precision, real-time performance, and intelligence in rockburst early warning. Multi-dimensional data acquisition comprehensively covers physical field information such as stress, strain, and microseismic activity. Combined with spatiotemporal registration technology, it ensures data spatiotemporal consistency, enhancing the early warning system's adaptability to complex geological environments. Dynamic noise suppression based on physical characteristics and mutual information entropy feature screening effectively remove interference and focus on key features, improving early warning accuracy. The combination of principal component extraction and phase space reconstruction mines the frequency domain, time domain, and nonlinear characteristics of the data, enhancing the system's ability to identify rockburst precursors. By outputting tiered early warning signals through predictive models and coordinating adjustments to monitoring layout and control measures, a dynamic response mechanism is formed, significantly shortening early warning response time, reducing false alarm rates, and providing reliable safety guarantees for deep well mining. Attached Figure Description

[0026] Figure 1 This is an overall flowchart of the deep well rockburst early warning method based on multi-dimensional monitoring according to an embodiment of the present invention;

[0027] Figure 2 This is a detailed flowchart of step S10 in an embodiment of the present invention;

[0028] Figure 3 This is a detailed flowchart of step S20 in an embodiment of the present invention;

[0029] Figure 4 This is a detailed flowchart of step S30 in an embodiment of the present invention;

[0030] Figure 5 This is a detailed flowchart of step S40 in an embodiment of the present invention;

[0031] Figure 6 This is a detailed flowchart of step S50 in an embodiment of the present invention;

[0032] Figure 7 This is a detailed flowchart of step S70 in an embodiment of the present invention;

[0033] Figure 8 This is a detailed flowchart of step S80 in an embodiment of the present invention;

[0034] Figure 9 This is a functional block diagram of a deep well rockburst early warning system based on multi-dimensional monitoring, according to an embodiment of the present invention.

[0035] Figure 10 This is a schematic diagram of the structure of a computer system according to an embodiment of the present invention. Detailed Implementation

[0036] The following description, in conjunction with the accompanying drawings, illustrates exemplary embodiments of the present invention, including various details to aid understanding. These details should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the invention. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0037] like Figure 1 As shown in the figure, this embodiment provides a method for early warning of deep well rockburst based on multi-dimensional monitoring, which includes the following steps:

[0038] S10: Utilizing a monitoring network deployed in deep shaft tunnels and surrounding rock, real-time collection of multi-dimensional monitoring data, including stress values, strain values, microseismic event parameters, acoustic signals, temperature data, radon concentration values, electromagnetic radiation intensity, and three-dimensional deformation data of the tunnel surface.

[0039] Specifically, a monitoring network deployed in deep shaft tunnels and surrounding rock is used to collect multi-dimensional monitoring data in real time. In a deep mining tunnel of a certain mine, a monitoring section is set up every 50 meters along the tunnel axis. Stress sensors and strain sensors are installed on the tunnel roof, sides, and floor of each section. A microseismic monitoring system and an acoustic monitoring device are set up every 100 meters. Temperature sensors and radon concentration sensors are arranged in a grid pattern within a 500-meter radius around the tunnel. A laser displacement sensor array is set up along the entire length of the tunnel to monitor the three-dimensional deformation of the surface. At the same time, electromagnetic radiation monitoring devices are added in key areas. All sensors transmit the collected stress values, strain values, microseismic event parameters, acoustic signals, temperature data, radon concentration values, electromagnetic radiation intensity, and three-dimensional deformation data of the tunnel surface to the ground monitoring center in real time through a combination of ZigBee wireless transmission and industrial Ethernet.

[0040] S20: Perform spatiotemporal registration on the multi-dimensional monitoring data to obtain a spatiotemporally aligned multi-source monitoring dataset.

[0041] Specifically, this step involves spatiotemporal registration of multi-dimensional monitoring data. A three-dimensional geological model spatiotemporal reference coordinate system is established, with the tunnel entrance as the origin (0,0,0), the tunnel axis as the X-axis, the perpendicular tunnel direction as the Y-axis, and the perpendicular tunnel floor as the Z-axis. Timestamp calibration is performed on the data from each sensor, and a GPS timing system is used to ensure that the time synchronization error of all sensors is less than 10ms. Cubic spline interpolation is used for time alignment of asynchronous data. A coordinate transformation matrix is ​​constructed to transform the original coordinates of each sensor to the reference coordinate system; for example, spherical coordinate transformation is used for microseismic sensors, and affine transformation is used for displacement sensors. Kriging interpolation is used to grid the spatially discrete point data, forming a 5m×5m×5m three-dimensional data field. The spatiotemporal registration error is calculated, and when the position error is less than 0.5m and the time error is less than 5ms, a spatiotemporally aligned multi-source monitoring dataset is output.

[0042] S30: Perform dynamic noise suppression based on physical characteristics on the multi-source monitoring dataset to obtain the noise-suppressed dataset.

[0043] Specifically, this step performs dynamic noise suppression based on physical characteristics on the multi-source monitoring dataset. For mechanical data, a stress-strain trend prediction model is established, and a drift compensation algorithm is activated when the deviation between the measured and predicted values ​​exceeds 3%. For microseismic and acoustic signals, wavelet packet decomposition combined with adaptive threshold denoising is used, and the number of decomposition layers is dynamically adjusted according to the signal energy distribution. For temperature and radon concentration data, baseline values ​​are calculated using a sliding window, and corrections are performed when fluctuations exceed ±0.5℃ or ±20Bq / m³ of the baseline. For electromagnetic radiation and displacement data, independent component analysis is used to separate interference signals. The signal-to-noise ratio and feature retention are calculated in real time before and after denoising. When the signal-to-noise ratio is ≥15dB and the feature retention is ≥90%, the noise-suppressed dataset is output.

[0044] S40: Use the mutual information entropy algorithm to filter features related to rockburst pressure from the noise-suppressed dataset to obtain a feature-optimized dataset.

[0045] Specifically, this step uses a mutual information entropy algorithm to screen features related to rockburst intensity. A mutual information entropy calculation model is constructed between these features and the rockburst hazard, for example:

[0046] ;

[0047] Where X represents each monitoring feature, and Y represents the rockburst hazard level. The mutual information entropy values ​​of features such as stress change rate, microseismic energy release rate, and acoustic wave velocity attenuation rate are calculated to be greater than 0.7 (feature contribution threshold), and these strongly correlated features are retained. Further calculation of the mutual information entropy redundancy between features is performed; when the similarity of the mutual information entropy values ​​of two features exceeds 0.85 (feature redundancy threshold), the feature with the clearer physical meaning is retained, ultimately forming an optimized dataset containing 12 key features.

[0048] S50: Using principal component extraction and phase space reconstruction methods, frequency domain features, time domain features, and time-frequency domain features are extracted from the feature-optimized dataset to form a multidimensional state space dataset.

[0049] Specifically, multidimensional features are extracted using principal component analysis (PCA) and phase space reconstruction methods. A covariance matrix is ​​constructed from 12 features, and eigenvalue decomposition is performed to extract the top 5 principal components with a cumulative variance contribution rate exceeding 85% (the PCA extraction threshold). The embedding dimension m=6 is calculated using the Cao method, and the time delay τ=5 is determined using mutual information to reconstruct the phase space. For each principal component, Fourier transform is performed to extract frequency domain features (e.g., dominant frequency, frequency band energy distribution), and time domain statistical features (mean, variance, skewness, etc.) are calculated. A 5-level decomposition using db4 wavelets is then used to extract time-frequency domain features, ultimately forming a multidimensional state space dataset containing 35 features.

[0050] S60: Input the multidimensional state space dataset into the pre-trained rockburst hazard level prediction model to obtain the rockburst hazard level.

[0051] Specifically, a multidimensional state-space dataset is input into a pre-trained rockburst hazard prediction model. The XGBoost algorithm (Extreme Gradient Boosting) is used to construct the prediction model, and the training set includes historical monitoring data and corresponding rockburst event records for the mine over the past three years. The model is trained using 5-fold cross-validation, with an adjusted learning rate of 0.1, a maximum tree depth of 6, and a minimum sample weight of 3. The model outputs four hazard levels: Level I (safe), Level II (low risk), Level III (medium risk), and Level IV (high risk). Model performance is evaluated using the ROC (Receiver Operating Characteristic) curve, with an AUC (Area Under the Curve) value of 0.92. Real-time multidimensional state-space data is input into the model to obtain the current rockburst hazard level.

[0052] S70: Trigger the corresponding level of early warning signal based on the rockburst hazard level.

[0053] Specifically, based on the level of rockburst hazard, corresponding warning signals are triggered. When the model output is Level I, the system displays a green safe status; when the output is Level II, a yellow warning is triggered, prompting enhanced monitoring; when the output is Level III, an orange warning is triggered, and a report recommending prevention and control measures is automatically generated; when the output is Level IV, a red warning is triggered, immediately cutting off the power supply to the hazardous area, activating the audible and visual alarm system, and notifying relevant personnel to evacuate urgently via SMS, APP push notifications, and other means.

[0054] S80: Based on the corresponding level of early warning signal, dynamically adjust the layout of the monitoring network and the prevention and control measures.

[0055] Specifically, the monitoring network layout and control measures are dynamically adjusted based on the early warning signal. When the early warning signal is Level I (safe), the normal layout and data acquisition frequency of the monitoring network are maintained (e.g., 5 minutes / time), and only routine equipment inspection and maintenance are performed; when it is Level II (low risk), the basic layout of the core monitoring area and secondary monitoring area is maintained, the data acquisition frequency of the peripheral monitoring area is increased by 20%, and the operating conditions of all sensors are checked simultaneously to ensure no faults; when it is Level III (medium risk), the sampling frequency of the sensors in the secondary monitoring area is increased by 50% (e.g., from 10 minutes / time to 4 minutes / time), the backup stress sensor and acoustic monitoring device in the core monitoring area are activated to enhance data redundancy, and support strength is strengthened (e.g., increasing the anchor bolt preload to the design value). At least one of the following prevention and control measures is required: (1.2 times the value) or stress-oriented stress relief (drilling stress relief at 3 high stress points); when it is Level IV (high risk), the spatial distribution density of sensors in the core monitoring area is increased to twice the original density (e.g., from 1 per 10 meters to 1 per 5 meters), the data acquisition frequency of the entire network is increased to real-time acquisition (1 second / time), and a combination of prevention and control measures of support strength enhancement, stress-oriented stress relief (drilling in the high stress area of ​​the entire region) and surrounding rock grouting reinforcement (using ultrafine cement grout) are simultaneously initiated. The implementation effect of all adjusted parameters and prevention and control measures (e.g. stress reduction, surrounding rock deformation rate) is fed back to the monitoring network in real time to form a dynamic optimization closed loop.

[0056] The advantages of this technical solution are as follows: after spatiotemporal registration of multi-dimensional monitoring data, the inherent correlation between different physical field information is deeply explored, realizing multi-dimensional cross-verification of early precursors of rockbursts. Compared with traditional single-dimensional monitoring, it can better capture the complex signal characteristics of the disaster incubation stage. The dynamic noise suppression based on physical characteristics not only filters out interference but also retains weak features related to rockbursts in various data. These features are often misjudged as noise by traditional noise reduction methods, thus effectively reducing false alarms and missed alarms. The dynamic closed-loop linkage between early warning signals, monitoring networks, and prevention and control measures enables the system to autonomously optimize parameters through continuous feedback, maintaining stable early warning reliability even in complex and changeable deep well geological environments. This full-process collaboration not only improves the accuracy and timeliness of early warnings but also forms an active response mechanism in disaster prevention and control, enhancing the comprehensive prevention and control capabilities for rockburst disasters.

[0057] like Figure 2 As shown, step S10 may specifically include the following steps:

[0058] S101: Zoning planning, which is based on the geomechanical parameters of deep well tunnels and the historical risk zoning of rockbursts, to determine the core monitoring area, secondary monitoring area and peripheral monitoring area of ​​the monitoring network.

[0059] In practice, geomechanical parameters such as stress distribution, surrounding rock strength, coal seam thickness, and fault distribution in the roadway are collected. Combined with historical data on the location, intensity, and impact range of rockburst events over the past decade, a risk matrix method is used for risk zoning. The core monitoring area is defined as regions with frequent historical rockbursts and stress values ​​exceeding 80% of the ultimate strength of the surrounding rock (e.g., fault junctions, abrupt changes in coal seam thickness). The secondary monitoring area is within 100-300 meters of the core area, with stress values ​​between 50% and 80% of the ultimate strength of the surrounding rock. The outer monitoring area is the roadway extension and surrounding areas outside the secondary area, with stress values ​​below 50% of the ultimate strength of the surrounding rock, forming a gradient monitoring coverage from high to low risk.

[0060] S102: Sensor deployment, which includes stress sensors, strain sensors, microseismic monitoring systems and acoustic monitoring devices in the core monitoring area, temperature sensors, radon concentration sensors and electromagnetic radiation monitoring devices in the secondary monitoring area, and a laser displacement sensor array in the peripheral monitoring area.

[0061] In practice, the core monitoring area adopts a dense layout, with one set of stress sensors and strain sensors installed every 5 meters on the roof, sides, and floor of the roadway, and one microseismic monitoring system (including a triangular array of three detectors) and acoustic monitoring device (with the transmitter and receiver positioned at corresponding locations on the sides of the roadway) deployed every 20 meters. In the secondary monitoring area, temperature sensors and radon concentration sensors are installed on the sidewalls of the roadway at 30-meter intervals, and one electromagnetic radiation monitoring device is installed every 50 meters. The monitoring points avoid areas obstructed by water or support structures. In the outer monitoring area, one set of laser displacement sensor arrays is installed every 100 meters along the roadway direction. Each set contains three laser transmitters (aimed at the roof, left side, and right side of the roadway, respectively) and one receiving terminal to ensure coverage of the three-dimensional deformation monitoring range of the roadway surface.

[0062] In some embodiments, stress sensors in the core monitoring area are arranged in a quincunx pattern along the sides and roof of the roadway, installed at depths suitable for rock stress monitoring requirements, and used to monitor changes in static stress in the rock mass in real time. They are connected to redundant ports of an industrial Ethernet switch via shielded cables. Strain sensors are installed on the surface of the roadway support structure and in the shallow part of the surrounding rock, alternating with the stress sensors according to monitoring accuracy requirements, and used to monitor the strain response of the surrounding rock and support structure. Data is transmitted in parallel through the same industrial Ethernet channel. Geophones of the microseismic monitoring system are arranged in groups along the roadway axis according to monitoring resolution requirements. Multiple geophones in each group are spatially distributed on the roadway floor and sides, installed at depths sufficient for microseismic signal capture, and used to capture microseismic events caused by rock mass fracturing. They are connected to a microseismic data acquisition host with time-stamping synchronization function via fiber optic links. The transmitter and receiver of the acoustic monitoring device are installed on opposite sides of the roadway, with spacing set according to acoustic wave propagation characteristics testing requirements. The transmitter is embedded in the rock mass at a certain depth, and the receiver is placed on the support surface to collect rock mass data. The sound wave propagation speed and attenuation characteristics are shared with the microseismic monitoring system via a redundant industrial Ethernet channel. Temperature sensors in the secondary monitoring area are installed at appropriate heights on the roadway sidewalls, spaced according to environmental temperature field monitoring requirements, and connected in series to a mining Ethernet node via a mine-grade flame-retardant cable for monitoring changes in the environmental temperature field. Radon concentration sensors are suspended at appropriate positions below the roadway roof, alternating with the temperature sensors, and transmit gas concentration data through the same mining Ethernet channel. Electromagnetic radiation monitoring devices are fixed at appropriate heights on the non-excavation sidewalls of the roadway, with the monitoring probe facing the roadway center, and connected to a regional data aggregator via shielded twisted-pair cables. The laser displacement sensor array in the peripheral monitoring area is set up along the roadway direction according to deformation monitoring range requirements. Each monitoring station contains multiple laser emitters, pointing to key deformation monitoring points on the roadway roof and both sides. The sensor body is fixed to the stable section of the roadway via an explosion-proof bracket, and laser reflectors are set at corresponding monitoring points. The sensors are connected to a self-organizing network gateway via a wireless radio frequency module for real-time capture of three-dimensional deformation on the roadway surface.

[0063] S103: Setting parameters, which sets the acquisition parameters for multi-dimensional monitoring data, including sampling frequency, trigger conditions and measurement range.

[0064] In specific implementation, the sampling frequency of the stress and strain sensors in the core monitoring area is set to 10Hz, and the trigger condition is that high-frequency acquisition (increased to 100Hz) is automatically started when the stress change rate exceeds 5% / min, with a range covering 0-100MPa; the sampling frequency of the microseismic monitoring system is set to 1kHz, and the trigger threshold is that the vibration amplitude exceeds 0.01mm / s; the sampling frequency of the acoustic wave monitoring device is set to 500Hz, with a range corresponding to acoustic wave velocities of 300-6000m / s. In the secondary monitoring area, the sampling frequency of the temperature sensor is set to 1Hz, with a range covering -5-60℃; the sampling frequency of the radon concentration sensor is set to 0.1Hz (once every 10 seconds), with a range of 0-1000Bq / m³; and the sampling frequency of the electromagnetic radiation monitoring device is set to 5Hz, with a range of 0-1000μV / m. In the peripheral monitoring area, the sampling frequency of the laser displacement sensor array is set to 0.5Hz, and the trigger condition is that automatic recording is performed when the single deformation exceeds 0.1mm, with a range covering 0-500mm.

[0065] S104: Data transmission, which constructs a hierarchical data transmission link based on the importance level of the monitoring area, wherein the data of the core monitoring area is transmitted through an industrial Ethernet channel with redundant backup, the data of the secondary monitoring area is transmitted through a mining Ethernet channel, and the data of the peripheral monitoring area is transmitted through a wireless self-organizing network, thereby realizing the real-time transmission of the multi-dimensional monitoring data.

[0066] In practice, the core monitoring area uses a dual-line redundant industrial Ethernet channel. The main line is an underground 10 Gigabit optical network, and the backup line is a fiber optic ring network. A priority scheduling mechanism ensures that stress, micro-seismic, and other data are transmitted first, with transmission delay controlled within 100ms. Data from the secondary monitoring area is connected to a star network via mining Ethernet switches. A data frame queuing mechanism is used to avoid transmission conflicts, with transmission delay controlled within 500ms. The laser displacement sensor array in the peripheral monitoring area is connected via a Zifeng wireless self-organizing network. Each sensor acts as a network node, and the data is aggregated to the regional gateway via multi-hop routing before being connected to the backbone network. The transmission cycle is dynamically adjusted according to the sampling frequency to ensure complete data upload.

[0067] S105: Data tagging, which timestamps and marks the geological coordinates of the multi-dimensional monitoring data.

[0068] In practice, the BeiDou time synchronization system is used to synchronize the time of all sensors, with the timestamp accuracy controlled within 1ms, to ensure that the time reference of data collected by different devices is consistent. Based on the coordinate system of the three-dimensional geological model of the tunnel, a unique geological coordinate is assigned to each sensor (for example, a stress sensor in the core area is marked as X=1250m, Y=25m, Z=5m, corresponding to 1250m in the tunnel axis, 25m on the left side, and 5m above the floor). The coordinate information is embedded in the metadata of the corresponding monitoring data to achieve precise correlation between data and spatial location.

[0069] The technical advantages of this solution lie in the following aspects: Based on geomechanical parameters and historical risk zoning, the monitoring area division enables dynamic focusing of monitoring resources. The combination of high-density key parameter monitoring in the core area and supplementary parameter monitoring in secondary and peripheral areas avoids resource redundancy while fully covering the multi-physical field characteristics of rockburst generation. The construction of hierarchical data transmission links not only ensures the real-time performance of key data through redundant transmission in the core area but also reduces data congestion through adaptive transmission methods in secondary and peripheral areas, improving the stability of data transmission across the entire system compared to a single transmission mode. The integration of timestamp synchronization and geological coordinate marking with hierarchical monitoring data strengthens the spatiotemporal correlation of different regions and data types, making cross-regional precursor signal coupling analysis possible in subsequent data processing.

[0070] like Figure 3 As shown, step S20 may specifically include the following steps:

[0071] S201: Establish a three-dimensional geological coordinate system (origin and three-axis definition), which establishes a spatiotemporal reference coordinate system based on the three-dimensional geological model of the deep well tunnel. The spatiotemporal reference coordinate system takes the tunnel starting point as the origin, the tunnel axis as the X-axis, the perpendicular tunnel direction as the Y-axis, and the perpendicular tunnel floor as the Z-axis, and associates it with absolute geological coordinates.

[0072] In some embodiments, a local coordinate system is formed with the tunnel excavation starting point (corresponding to wellhead coordinates X=523000m, Y=3762000m, Z=-850m) as the origin (0,0,0), the tunnel excavation direction (30° east of due north) as the X-axis, the direction perpendicular to the tunnel direction (30° west of northwest) as the Y-axis, and the direction perpendicular to the tunnel floor upwards as the Z-axis. Using three known high-level control points in the mining area (with coordinate accuracy better than 0.1m), a seven-parameter coordinate transformation method (including three translation parameters, three rotation parameters, and one scale parameter) is employed to link the local coordinate system with the National Geodetic Coordinate System 2000, enabling real-time conversion between local coordinates and absolute geological coordinates.

[0073] S202: Time calibration (unified time axis and asynchronous interpolation), which performs timestamp calibration on the multi-dimensional monitoring data, unifies the acquisition time of various monitoring devices used to collect the multi-dimensional monitoring data to the standard time axis of the spatiotemporal reference coordinate system, and uses linear interpolation to align the time dimension of asynchronous time data.

[0074] In some embodiments, all monitoring devices are connected to the mine's unified time synchronization system (based on BeiDou satellite time synchronization, with a synchronization accuracy of ±5ms), and the timestamp is automatically calibrated every hour. For asynchronous data caused by communication delays (such as peripheral monitoring data transmitted via ZigBee network), when the time interval between two adjacent valid data points exceeds twice the sampling period, a linear interpolation method is used to calculate the estimated value of the intermediate time: Let the data at time t1 be x1 and the data at time t2 be x2, then the interpolation result at time t∈(t1,t2) is x(t)=x1+(x2-x1)×(t-t1) / (t2-t1), ensuring that all data are aligned to the standard time axis of the reference coordinate system (based on Beijing time) in the time dimension.

[0075] S203: Spatial coordinate system transformation (matrix transformation), which is based on the spatiotemporal reference coordinate system, constructs a spatial coordinate transformation matrix for the location of each monitoring device, and transforms the original spatial data of different monitoring devices into three-dimensional coordinates under the spatiotemporal reference coordinate system.

[0076] In some embodiments, transformation models are established for the original coordinate systems of different types of monitoring devices: For microseismic monitoring systems, a transformation matrix from spherical coordinates (r, θ, φ) to rectangular coordinates (X, Y, Z) is used, where X = r × sinθ × cosφ + X0, Y = r × sinθ × sinφ + Y0, and Z = r × cosθ + Z0 ((X0, Y0, Z0) are the sensor installation position coordinates); for laser displacement sensors, an affine transformation matrix from local planar coordinates to three-dimensional reference coordinates is used, and translation, rotation, and scaling parameters are calculated using three known control points (e.g., the location of the roadway roof anchor bolts); for stress sensors, the three-dimensional coordinates of the installation position (obtained by a total station) are directly used as the original coordinates without additional transformation. All transformation matrices are optimized using the least squares method to ensure that the transformation residual is less than 0.1m.

[0077] S204: Spatial gridding processing (Kriging interpolation), which uses Kriging interpolation to perform spatial gridding processing on monitoring data with uneven spatial distribution, forming a spatially continuous monitoring data field. The monitoring data field is used to integrate the spatial correlation of multi-dimensional monitoring data.

[0078] In some embodiments, Kriging interpolation is used for spatial gridding. Based on the three-dimensional coordinates and monitoring values ​​of each monitoring point, a variogram model is constructed (e.g., a spherical model is selected, with a range of 50m and a sill value of 1.2 times the data variance). Interpolation calculations are performed on the core monitoring area using a 2m×2m×2m grid, the secondary monitoring area using a 5m×5m×5m grid, and the outer monitoring area using a 10m×10m×10m grid. During the interpolation process, a weighted approach with denser sampling points is used for key areas such as high-stress zones and areas with dense microseismic activity to ensure accurate representation of local details. Ultimately, a continuous three-dimensional monitoring data field is formed, which intuitively presents the spatial distribution and correlation of parameters such as stress, temperature, and displacement (e.g., the spatial overlap between stress concentration areas and temperature anomaly areas).

[0079] S205: Error Calculation and Verification (Time Deviation and Spatial Deviation), which calculates the registration error based on the time synchronization error and spatial coordinate deviation. The time synchronization error is the deviation between the timestamp of the calibrated monitoring data and the standard time axis, and the spatial coordinate deviation is the deviation between the transformed three-dimensional coordinates and the theoretical coordinates of the spatiotemporal reference coordinate system. When the registration error exceeds the registration accuracy threshold (i.e., the error exceeds the limit), the system returns to step S202 for recalibration, including time calibration or spatial transformation. When the registration error meets the requirements of the registration accuracy threshold (i.e., the error meets the standard), the system outputs a spatiotemporally aligned multi-source monitoring dataset (output spatiotemporally aligned dataset).

[0080] In some embodiments, the time synchronization error is calculated using the root mean square error (RMSE) between the calibrated timestamp and the standard time axis, and the spatial coordinate deviation is calculated using the mean Euclidean distance between the transformed coordinates and the theoretical coordinates (generated from the 3D model). The registration accuracy thresholds are set as follows: time synchronization error ≤ 10ms, spatial coordinate deviation ≤ 0.5m. When the calculation result exceeds the threshold (e.g., a batch of microseismic data has a spatial deviation of 0.8m due to detector drift), the system automatically returns to step S202 to recalibrate the time or update the transformation matrix; when the error meets the threshold requirements, a spatiotemporally aligned multi-source monitoring dataset containing timestamps (accurate to milliseconds) and 3D coordinates (accurate to 0.1m) is output, providing a unified spatiotemporal reference for subsequent data processing.

[0081] The beneficial effects of the above technical solution are as follows: the spatiotemporal reference coordinate system based on the three-dimensional geological model of the tunnel provides a unified reference framework for monitoring data of different types and locations, avoiding the correlation failure caused by the difference in coordinate systems of traditional multi-source data; the combination of timestamp calibration and spatial coordinate transformation effectively eliminates the problems of time asynchrony and spatial misalignment in data acquisition, enabling the originally scattered stress, microseismic, environmental and other multi-dimensional data to form an organic correlation in the same spatiotemporal dimension; spatial gridding further enhances the spatial continuity of the data, transforming the information of discrete monitoring points into a continuous data field that can be analyzed across the entire domain; and the dynamic verification mechanism of registration error ensures that the accuracy of the entire registration process is controllable, preventing invalid data from entering the subsequent processing stage.

[0082] like Figure 4 As shown, step S30 may specifically include the following steps:

[0083] S301: Classify and identify noise types. Based on the physical properties of multi-source monitoring data, classify and identify the noise types corresponding to each dimension of data. Among them, mechanical data includes stress and strain values, wave signals include micro-seismic event parameters and acoustic signals, environmental parameters include temperature data and radon concentration values, and electromagnetic and optical data include electromagnetic radiation intensity and three-dimensional deformation data of the tunnel surface.

[0084] Specifically, mechanical data (stress values, strain values) are mainly affected by sensor temperature drift and cable contact noise, manifesting as slow trend deviations or instantaneous pulse interference; wave-like signals (micro-vibration event parameters, acoustic signals) are easily affected by tunnel mechanical vibrations and blasting aftershocks, exhibiting broadband random noise superposition characteristics; environmental parameters (temperature data, radon concentration values) often generate baseline drift noise due to equipment heat dissipation and airflow disturbances; electromagnetic and optical data (electromagnetic radiation intensity, tunnel surface three-dimensional deformation data) are easily affected by industrial electromagnetic interference and measurement deviations caused by changes in illumination, manifesting as high-frequency electromagnetic noise or periodic optical interference signals. By analyzing the data spectral characteristics and fluctuation patterns, a mapping relationship between noise types and data categories is established, providing a basis for targeted noise reduction.

[0085] S302: Matching noise reduction algorithm, which uses a noise reduction algorithm that matches the physical characteristics of the data to suppress noise based on the identified noise type. It includes a drift compensation algorithm based on physical change trends for the mechanical data, an adaptive filtering algorithm for the wave signals, a baseline correction algorithm for the environmental parameters, and an interference subspace separation algorithm for the electromagnetic and optical data.

[0086] Specifically, for mechanical data, a stress-temperature drift model is established by collecting ambient temperature data from the same period. The drift curve is fitted using the least squares method and real-time compensation is performed. For example, when the temperature changes by 1°C, the stress value is linearly corrected by 0.2 MPa. For wave-like signals, an adaptive filter based on the LMS (Least Mean Square) algorithm is used to filter out 50Hz power frequency interference and 200-500Hz mechanical vibration noise by dynamically adjusting the filter coefficients, while retaining the effective frequency band of 2-50Hz for micro-vibration signals. For environmental parameters, a sliding window (window size set to 30 minutes) is used to calculate the baseline values ​​of temperature and radon concentration. When the instantaneous value deviates from the baseline by more than 3 times the standard deviation, baseline correction is performed. For electromagnetic and optical data, Principal Component Analysis (PCA) is used to separate the interference subspace in the electromagnetic radiation signal, eliminating the 150-300MHz frequency band interference generated by industrial equipment. Wavelet threshold denoising is used to separate the illumination interference component for laser displacement data.

[0087] Specifically, the interference subspace separation algorithm is a noise reduction method based on the signal subspace decomposition theory. Its principle is to decompose the mixed signals contained in electromagnetic and optical monitoring data into useful signal subspaces and interference signal subspaces through mathematical modeling, and then separate the two types of subspaces through spatial projection operations to filter out noise. Specifically, the algorithm first performs feature analysis on the collected data such as electromagnetic radiation intensity and three-dimensional deformation of the tunnel surface to identify useful signals related to rockburst (e.g., electromagnetic radiation anomalies caused by rock mass stress changes, characteristic signals of tunnel structural deformation) and interference signals (e.g., equipment electromagnetic interference, ambient light fluctuations, and noise from the measuring device itself). Then, through mathematical operations such as eigenvalue decomposition or singular value decomposition, it determines the spatial basis vectors of each type of signal and constructs the corresponding subspace model. Finally, through orthogonal projection, it removes the components belonging to the interference subspace from the mixed signal, retaining only the effective components of the useful signal subspace. This algorithm can selectively separate complex and dynamically changing interference in electromagnetic and optical data, and is especially suitable for scenarios with multiple sources of interference in deep well environments. It can suppress noise while preserving the feature information related to rockburst to the maximum extent, providing high-quality data support for subsequent feature extraction and early warning models.

[0088] S303: Monitoring variability and dynamic parameter adjustment. It monitors the data variability in real time during the noise suppression process and uses the data variability as feedback to dynamically adjust the parameters of the noise reduction algorithm so that the suppression intensity is adapted to the dynamic changes of noise.

[0089] Specifically, this step monitors data variability in real time and dynamically adjusts noise reduction parameters. The coefficient of variation (the ratio of standard deviation to mean) is used to quantify data variability. When the coefficient of variation for mechanical data exceeds 10%, the drift compensation time window is automatically reduced (from 1 hour to 30 minutes) to improve response speed. When the coefficient of variation for fluctuation signals suddenly increases, the convergence factor of the adaptive filter is increased (from 0.01 to 0.05) to enhance noise suppression. When the variability of environmental parameters decreases, the baseline correction threshold is relaxed (from 3 times the standard deviation to 4 times the standard deviation) to avoid overcorrection. When the variability of electromagnetic and optical data is abnormal, the number of principal components retained in the PCA is increased (from 3 to 5) to reduce effective signal loss, ensuring that the suppression strength adapts to the dynamic changes in noise in real time.

[0090] S304: Calculate the signal-to-noise ratio and feature retention, and determine the threshold. It calculates the signal-to-noise ratio and feature retention of the noise-suppressed data. When the signal-to-noise ratio reaches the signal-to-noise ratio threshold and the feature retention reaches the feature integrity threshold, the noise-suppressed dataset is output; otherwise, return to step S302 to readjust the parameters of the noise reduction algorithm.

[0091] Specifically, this step calculates the signal-to-noise ratio (SNR) and feature retention of the noise-suppressed data and performs threshold judgment. The SNR is calculated as the ratio of signal power to noise power. For mechanical data, the compensated stationary signal is used as the benchmark, and for wave-type signals, the characteristic frequency range of microseismic events is used as the effective signal. Feature retention is measured by the correlation coefficient of key features (e.g., peak stress, microseismic energy, temperature gradient) of the data before and after noise reduction. When the SNR reaches a preset SNR threshold (e.g., ≥20dB for mechanical data, ≥15dB for wave-type data) and the feature retention reaches a feature integrity threshold (e.g., correlation coefficient ≥0.9), the noise-suppressed dataset is output. If the thresholds are not met, the process returns to step S302 to readjust the parameters of the corresponding algorithm (e.g., filter order, compensation coefficient, etc.) until the threshold requirements are met.

[0092] The advantages of this technical solution are as follows: Classification-based noise identification based on the physical properties of the data avoids the neglect of differences in noise characteristics among different types of data by traditional uniform denoising methods, making noise suppression more targeted. Drift noise in mechanical data, interference noise in wave-like signals, baseline noise of environmental parameters, and interference signals in electromagnetic and optical data are accurately distinguished, avoiding the loss of effective features or noise residue caused by uniform denoising. The differentiated denoising algorithm matching the physical characteristics of the data ensures that key features related to rockburst (such as abrupt stress trends, dominant frequency characteristics of microseismic signals, and abnormal fluctuations in radon concentration) are fully preserved while suppressing noise. The dynamic parameter adjustment mechanism monitors data variability in real time, enabling the denoising intensity to adapt to dynamic changes in noise, effectively addressing the uncertainty of noise sources in deep well environments. The dual verification of signal-to-noise ratio and feature retention strictly controls the quality of the denoised data, preventing invalid data from entering subsequent processing stages.

[0093] like Figure 5 As shown, step S40 may specifically include:

[0094] S401: Construct a model that calculates the mutual information entropy between each feature in the noise-suppressed dataset and the rockburst hazard level. The mutual information entropy calculation model is based on the mutual information entropy formula in information theory and quantifies the degree of information correlation between feature variables and rockburst hazard level.

[0095] Specifically, when constructing the mutual information entropy calculation model between the features in the noise-suppressed dataset and the rockburst hazard, firstly, multi-dimensional monitoring data after noise suppression is collected as feature variables X. These feature variables include stress values, strain values, microseismic event parameters, acoustic signals, temperature data, radon concentration values, electromagnetic radiation intensity, and three-dimensional deformation data of the roadway surface. Simultaneously, the actual occurrence of rockbursts and the hazard level assessment results within the corresponding time period are compiled as rockburst hazard variables Y, which can be divided into three levels: low, medium, and high risk. Based on the mutual information entropy formula in information theory, I(X;Y)=H(X)+H(Y)-H(X,Y), where H(X) is the information entropy of feature variable X, H(Y) is the information entropy of rockburst hazard variable Y, and H(X,Y) is the joint information entropy of X and Y. In practice, for each feature variable, such as stress value, its historical monitoring data is discretized into several intervals, and the probability of occurrence in each interval is statistically analyzed to calculate H(X). Similarly, the probability of each hazard level of rockburst is statistically analyzed to calculate H(Y). Then, the joint probability of the characteristic variable being in a certain interval and the rockburst being in a certain hazard level is statistically analyzed to calculate H(X,Y). Through these calculations, a mutual information entropy calculation model that can quantify the degree of information correlation between the characteristic variable and the rockburst hazard level is constructed.

[0096] S402: Calculate the entropy value, which is calculated by the mutual information entropy calculation model for each feature. The mutual information entropy value represents the contribution of the feature to the prediction of rockburst hazard.

[0097] Specifically, using the mutual information entropy calculation model constructed above, the mutual information entropy value between each feature in the noise-suppressed dataset and the rockburst hazard is calculated. For example, for the stress value feature, its discretized interval data and the corresponding rockburst hazard level data are substituted into the model, and the mutual information entropy value of this feature is calculated according to the formula. The larger the mutual information entropy value, the higher the correlation between the feature and the rockburst hazard, and the greater its contribution to the prediction of rockburst hazard. Similarly, the mutual information entropy values ​​of all other features such as strain values, microseismic event parameters, and acoustic signals are calculated in turn, thereby obtaining a quantified result of the predictive contribution of each feature to the rockburst hazard.

[0098] S403: Filter contribution features. Based on the calculated mutual information entropy value, it filters features whose mutual information entropy value is greater than the feature contribution threshold to form a feature subset related to rockburst pressure.

[0099] Specifically, based on extensive historical monitoring data and analysis of rockburst cases, a feature contribution threshold is pre-set. This threshold can distinguish features that significantly contribute to the prediction of rockburst hazard. For example, statistical analysis determines the feature contribution threshold to be 0.6. Then, the calculated mutual information entropy values ​​of each feature are compared with this threshold, and features with mutual information entropy values ​​greater than 0.6 are selected. Assuming that the mutual information entropy values ​​of stress value, microseismic event parameters, and acoustic signal are 0.75, 0.82, and 0.68 respectively, all greater than the threshold of 0.6, these three features are retained, forming a subset of features strongly correlated with rockburst.

[0100] S404: Remove redundant features. This involves performing a redundancy analysis on the feature subset, removing redundant features with similar mutual information entropy values ​​and correlations higher than the feature redundancy threshold, to obtain the feature-optimized dataset.

[0101] Specifically, redundancy analysis is performed on the selected feature subset. The correlation coefficient between any two features in the feature subset is calculated, such as the correlation between stress values ​​and microseismic event parameters, and the correlation between stress values ​​and acoustic signals. Simultaneously, the mutual information entropy values ​​of these features are compared. When the mutual information entropy values ​​of two features are similar, and their correlation coefficient is higher than a pre-set feature redundancy threshold (e.g., 0.8), it indicates that these two features have high redundancy, meaning that the information they provide in predicting rockburst hazard overlaps significantly. For example, if the mutual information entropy values ​​of microseismic event parameters and acoustic signals are 0.82 and 0.80 respectively, which are relatively close, and their correlation coefficient is 0.85, which is higher than the feature redundancy threshold of 0.8, then one of the features is removed (any one can be removed depending on the actual situation). In this way, after removing redundant features, the feature-optimized dataset is obtained.

[0102] The aforementioned technical solution quantifies the correlation between feature variables and rockburst hazard levels by constructing a mutual information entropy calculation model. This enables accurate identification of features that are substantially significant for rockburst prediction, avoiding interference from irrelevant or low-contribution data. By calculating the mutual information entropy value, the predictive contribution of each feature is clarified, providing a quantitative basis for feature selection and ensuring that the selected feature subset focuses on highly correlated parameters. Through feature contribution threshold screening and redundancy analysis, redundant features are eliminated while retaining key information. This simplifies the dataset size to reduce the computational complexity of subsequent models and improves operational efficiency, while also reducing information overlap between features and avoiding the risk of model overfitting. Consequently, the accuracy and stability of rockburst hazard level prediction are improved, providing a more reliable data analysis foundation for subsequent early warning signal triggering and control measure adjustments. Ultimately, this achieves more accurate and efficient monitoring and early warning of deep well rockbursts.

[0103] like Figure 6 As shown, step S50 may specifically include the following steps:

[0104] S501: Construct a covariance matrix for the dataset after feature optimization, and extract principal components with variance contribution rates exceeding a preset threshold by performing eigenvalue decomposition on the covariance matrix to form a dimensionality-reduced feature subset.

[0105] Specifically, for the feature-optimized dataset (e.g., containing five feature variables: peak stress, microseismic energy, dominant acoustic frequency, radon concentration gradient, and electromagnetic radiation pulse frequency), the mean of each feature variable is first calculated. Then, a covariance matrix is ​​constructed using a formula. This matrix has a dimension of 5×5, and its elements represent the degree of co-variance between any two feature variables (e.g., the covariance between peak stress and microseismic energy is 12.8, reflecting a positive correlation between the two). After eigenvalue decomposition of the covariance matrix, five eigenvalues ​​are obtained. to The values ​​are 35.2, 18.7, 9.5, 4.3, and 2.1, respectively, and the corresponding eigenvectors represent the composition direction of the principal components. A preset threshold of 85% cumulative variance contribution rate is set. The calculated cumulative variance contribution rate of the first three eigenvalues ​​is (35.2+18.7+9.5) / (35.2+18.7+9.5+4.3+2.1)×100%=89.3%, which meets the threshold requirement. Therefore, the first three principal components are extracted to form a dimensionality-reduced feature subset (containing the three principal component variables).

[0106] S502: Based on the phase space reconstruction method, the embedding dimension and time delay parameters are determined according to the autocorrelation function and mutual information function of the reduced feature subset, and a high-dimensional phase space is constructed.

[0107] Specifically, based on time series data (sampling frequency 1Hz, duration 24 hours, total 86400 data points) composed of three principal component variables after dimensionality reduction, the time delay τ was calculated using the autocorrelation function: the autocorrelation coefficient was calculated for the principal component time series, and when the lag time was 5s, the autocorrelation coefficient first approached 0, initially determining τ=5s; then, the mutual information function was used for verification, calculating the mutual information value for different lag times, and it was found that the mutual information value was the smallest (0.32) when τ=5s, confirming the time delay parameter as 5s. The spurious neighbor method was used to determine the embedding dimension m: starting from m=1, the dimension was gradually increased, and the proportion of spurious neighbors in the high-dimensional space was calculated. When m=5, the proportion of spurious neighbors dropped to 3.2% (below the 5% threshold), therefore, the embedding dimension was determined to be 5. A high-dimensional phase space is constructed based on τ=5s and m=5, and each sample point is represented as (x(t), x(t+5), x(t+10), x(t+15), x(t+20)), where x(t) is the value of a certain principal component at time t.

[0108] S503: In the high-dimensional phase space, extract the frequency domain features, time domain features and time-frequency domain features of each feature respectively, wherein the frequency domain features are obtained by Fourier transform, the time domain features are calculated based on the statistical characteristics of time series, and the time-frequency domain features are extracted by wavelet transform.

[0109] Specifically, in the high-dimensional phase space, frequency domain features are extracted from the time series of each principal component: the time-domain signal is converted into a frequency domain spectrum using Fast Fourier Transform (FFT), and the principal frequency of principal component 1 is calculated to be 2.3 Hz and that of principal component 2 is 1.8 Hz, which are used as frequency domain features. Time domain features are extracted: based on the statistical properties of the time series, the mean (e.g., the mean of principal component 1 is 12.5 MPa), peak factor (the peak factor of principal component 2 is 3.8), kurtosis (the kurtosis of principal component 3 is 4.2), and root mean square value (the root mean square value of principal component 1 is 13.2 MPa) of each principal component are calculated, forming time domain features. Time-frequency domain features are extracted: the time series is decomposed into three levels using the db4 wavelet basis to obtain wavelet coefficients at different scales. The energy proportion at each scale (e.g., 25% at scale 1, 38% at scale 2, and 37% at scale 3) and wavelet entropy (the wavelet entropy of principal component 1 is 0.62) are calculated, which are used as time-frequency domain features.

[0110] S504: The extracted frequency domain features, time domain features, and time-frequency domain features are fused to form a multidimensional state space dataset.

[0111] Specifically, the extracted frequency domain features (3 principal frequency parameters), time domain features (3 principal components, each with 4 statistics, totaling 12), and time-frequency domain features (3 principal components, each with 4 wavelet parameters, totaling 12) are fused together. Through feature concatenation, a multidimensional state space dataset containing 3+12+12=27 feature parameters is formed. Each data sample corresponds to a state point in a high-dimensional phase space, which fully represents the multi-domain feature correlation of the monitored object in the spatiotemporal dimension.

[0112] By constructing a mutual information entropy calculation model, the degree of information correlation between each feature and the risk of rockburst is quantified, accurately selecting highly correlated features with high contribution, effectively eliminating the interference of irrelevant or low-correlation features on subsequent analysis, and improving the relevance and effectiveness of the feature set. At the same time, redundancy analysis is used to eliminate redundant features with similar mutual information entropy and high correlation, reducing data dimensionality and avoiding information duplication, reducing the complexity of subsequent feature extraction and model calculation, saving computing resources and improving processing efficiency. The final optimized feature dataset can retain key information and simplify the data structure, providing a high-quality data foundation for the construction of subsequent multi-dimensional state space datasets and the accurate output of prediction models, thereby improving the accuracy and timeliness of rockburst early warning.

[0113] like Figure 7 As shown, the pre-trained rockburst hazard level prediction model is constructed using a random forest algorithm optimized by combining a long short-term memory neural network with a genetic algorithm.

[0114] The rockburst hazard level prediction model may include:

[0115] The LSTM (Long Short-Term Memory) temporal feature extraction and classification layer is used to encode the temporal features in the multidimensional state space dataset bidirectionally using a bidirectional long short-term memory network, outputting a temporal feature vector, and then classifying the temporal feature vector through a fully connected layer to obtain the probability distribution of the rockburst hazard level in the temporal dimension.

[0116] The feature fusion and dimensionality reduction layer is used to concatenate the temporal feature vector with static features. The static features are frequency domain features and time-frequency domain features obtained from the feature-optimized dataset through principal component extraction and phase space reconstruction. The static features are weighted by an attention mechanism and dimensionality reduced by principal component analysis to output the dimensionality-reduced comprehensive feature vector.

[0117] A random forest classifier optimized by a genetic algorithm is used to optimize the hyperparameters of the random forest by taking the dimensionality-reduced comprehensive feature vector as input and using cross-validation accuracy as the fitness function, and outputting the probability distribution of the spatial dimension of the rockburst hazard level.

[0118] The model integration and output layer is used to fuse the probability distributions of rockburst hazard levels in the time-series dimension and the spatial dimension through a soft voting method, and output the rockburst hazard level.

[0119] In some embodiments, the LSTM temporal feature extraction and classification layer adopts a network structure including an input layer, a bidirectional LSTM layer, and a fully connected layer. The bidirectional LSTM layer contains forward LSTM units and backward LSTM units. The forward LSTM units encode the temporal features in the multidimensional state space dataset in chronological order, while the backward LSTM units encode them in reverse chronological order. By concatenating the output vectors from both directions to form a complete temporal feature vector, the temporal dependencies of the data are effectively captured. The fully connected layer contains two layers of linear transformations and activation functions. The first layer uses the ReLU (Rectified Linear Unit) activation function to perform a non-linear mapping on the temporal feature vector. The second layer uses the Softmax (Normalized Exponential Function) activation function to output the probability distribution of rockburst hazard levels (low, medium, and high) in the temporal dimension. The sum of the probability values ​​is 1, and each probability value corresponds to the likelihood of the corresponding level.

[0120] In some embodiments, the feature fusion and dimensionality reduction layer process is as follows: First, static features are selected from the frequency domain features and time-frequency domain features obtained by principal component extraction and phase space reconstruction of the feature-optimized dataset. These features do not change dynamically over time and mainly reflect the inherent attributes of the monitored object. The time-series feature vectors extracted by LSTM and output by the classification layer are concatenated with the above static features according to feature dimensions to form a fused feature matrix. Then, an attention mechanism is introduced, and the attention weights obtained through training are used to weight each feature dimension in the fused feature matrix, so that features with a greater impact on the early warning result receive higher weights, thereby enhancing the representation ability of key information. Finally, principal component analysis is used to reduce the dimensionality of the weighted fused feature matrix, retaining principal components with a cumulative variance contribution rate of over 90%, eliminating redundant information, and outputting a simplified comprehensive feature vector to provide high-quality input for the subsequent classifier.

[0121] In some embodiments, the construction process of the random forest classifier optimized by the genetic algorithm includes two parts: building the initial random forest model and optimizing the hyperparameters of the genetic algorithm. The initial random forest model contains 500 decision trees, and the training samples for each tree are selected from the training set through bootstrap sampling. Feature selection uses a random subset method. In the genetic algorithm optimization stage, the hyperparameters of the random forest (including the number of decision trees, maximum tree depth, and minimum number of leaf node samples) are used as optimization variables, and the classification accuracy of 5-fold cross-validation is used as the fitness function. The algorithm iteratively optimizes through selection operators (roulette wheel selection), crossover operators (single-point crossover), and mutation operators (random perturbation). Iteration stops when the fitness value does not improve significantly after 10 consecutive generations, and the optimal hyperparameter combination is output. The comprehensive feature vector output from feature fusion and dimensionality reduction layers is input into the optimized random forest classifier, and finally, the probability distribution of the three hazard levels in the spatial dimension is obtained. This distribution can effectively reflect the spatial differences in the degree of hazard in the monitored area.

[0122] In some embodiments, the model integration and output layer employ a soft voting method to achieve multi-dimensional result fusion. Specifically, the temporal dimension probability distributions output by the LSTM temporal feature extraction and classification layer are added to the spatial dimension probability distributions output by the random forest classifier optimized by the genetic algorithm, according to their respective levels, to obtain the cumulative probability value for each hazard level. For example, if the temporal dimension probabilities are 0.3 for low risk, 0.5 for medium risk, and 0.2 for high risk, and the spatial dimension probabilities are 0.2 for low risk, 0.4 for medium risk, and 0.4 for high risk, then the cumulative probabilities are 0.5 for low risk, 0.9 for medium risk, and 0.6 for high risk. The level with the highest cumulative probability value is selected as the final rockburst hazard level. If the cumulative probabilities are equal, the higher hazard level is prioritized as the output result to ensure the safety and rigor of the warning.

[0123] like Figure 8As shown, step S70 may specifically include the following steps:

[0124] S701: Preset correspondence between rockburst hazard levels and warning levels, where low hazard level corresponds to Level 1 warning, medium hazard level corresponds to Level 2 warning, and high hazard level corresponds to Level 3 warning.

[0125] Specifically, when S701 presets the correspondence between rockburst hazard levels and warning levels, the parameters are set through the system configuration module: Low-risk level is defined as a rockburst occurrence probability of less than 10% and an energy release rate of less than 50 J / h, corresponding to a Level 1 warning; Medium-risk level is defined as an occurrence probability of 10%-30% and an energy release rate of 50-150 J / h, corresponding to a Level 2 warning; High-risk level is defined as an occurrence probability of more than 30% and an energy release rate exceeding 150 J / h, corresponding to a Level 3 warning. This correspondence can be configured through the industrial control computer in the underground monitoring center. The configuration interface provides visual sliders and numerical input boxes. Operators can adjust the probability and energy thresholds according to the specific geological conditions of the mine. After setting, the parameters are stored in an intrinsically safe database server, supporting regular automatic backups. Modifications require administrator verification to ensure the security and traceability of the preset rules.

[0126] S702: Trigger the corresponding level of early warning signal according to the rockburst hazard level, including Level 1 early warning, which is a local area audible and visual alert signal; Level 2 early warning, which is a warning information pushed by the communication terminal of the monitoring area and accompanied by audible and visual alarm; and Level 3 early warning, which is an emergency alarm signal for the entire monitoring network coverage area, and simultaneously triggers personnel evacuation notification.

[0127] Specifically, when S702 triggers an early warning signal, the specific implementation methods for each level of early warning signal are as follows: For Level 1 early warning, explosion-proof audible and visual warning devices are deployed in local areas such as roadway intersections and working faces within the core monitoring area. These devices use explosion-proof housings and contain built-in yellow LED warning lights (flashing at a frequency of 2Hz) and a 110dB buzzer (emitting sound continuously at 3-second intervals). The warning message is "Level 1 Early Warning: There is a potential impact risk in a local area; please strengthen monitoring." For Level 2 early warning, text information (including the warning level, coordinates of the affected area, and risk warning) is pushed to underground workers within the secondary monitoring area and a 500-meter radius via an intrinsically safe mining communication terminal (such as a KT425 mining walkie-talkie). Simultaneously, secondary monitoring is activated. The red rotating warning light (60 revolutions per minute) and the 130dB intermittent alarm sound (1 second sound, 1 second pause) in the area are simultaneously displayed on the monitor of the ground monitoring center. The level 3 warning triggers the emergency alarm system covering the entire monitoring network area, including explosion-proof loudspeakers placed every 100 meters underground (playing "Level 3 Warning: Emergency Evacuation, immediately evacuate along the safety passage"), and the sound and light alarm device of the ground dispatch center (red flashing light and 150dB alarm sound). At the same time, a mandatory pop-up notification is sent to the terminals of all personnel in the high-risk area through the underground personnel positioning system. The notification includes the three nearest evacuation routes and the estimated evacuation time, and is linked with the mine emergency broadcast system to achieve full coverage.

[0128] S703: Records the trigger time, level, and coverage of the warning signal to form a warning log.

[0129] Specifically, when the S703 records early warning logs, the system automatically generates structured log files. The content includes: early warning trigger time (accurate to milliseconds), corresponding rockburst hazard level (e.g., high-risk level), early warning level (e.g., Level 3 early warning), coverage area (accurate to roadway number and coordinate range, e.g., "No. 3 coal seam transport roadway K0+200 to K0+500 section"), and triggering basis (e.g., "LSTM model outputs a high-risk probability of 85%, random forest model outputs a high-risk probability of 82%)." Logs are stored in a distributed database (e.g., HBase), with each log entry associated with a unique identifier (UUID), and are synchronously backed up to a remote ground server for a retention period of 3 years. The system also provides a log query interface, supporting filtering by time, level, region, and other criteria, and can export to Excel or PDF format for subsequent early warning effect evaluation and accident tracing analysis.

[0130] By pre-setting the correspondence between hazard levels and warning levels, the judgment criteria are clarified, and the standardization and reliability of warnings are improved. By adopting graded warning signals (low-risk local sound and light, medium-risk area push, high-risk full area alarm and evacuation linkage), the risk information can be accurately transmitted, balancing production interference and emergency efficiency. The warning log is recorded to form a closed loop, providing data support for evaluation and optimization, and improving the scientific nature, effectiveness and manageability of warnings as a whole, thus ensuring mine safety.

[0131] like Figure 9 As shown, step S80 may specifically include:

[0132] S801: If the warning signal is of high risk level, increase the spatial distribution density of monitoring points in the monitoring network and increase the data acquisition frequency. Select a combination of support strength enhancement, stress-oriented pressure relief and surrounding rock grouting reinforcement to implement prevention and control, and trigger the emergency evacuation order for personnel.

[0133] Specifically, when S801 issues a high-risk warning signal, the monitoring network is adjusted as follows: the density of stress sensors in the core monitoring area is increased from one every 5 meters to one every 2 meters; the spacing between temperature and radon sensors in the secondary monitoring area is reduced from 10 meters to 5 meters; and the number of sampling points in the peripheral laser displacement sensor array is increased by 50%. In terms of data acquisition frequency, stress and strain data are increased from 1Hz to 10Hz, and the sampling rate of micro-vibration and acoustic signals is increased from 2kHz to 10kHz to ensure that high-frequency dynamic changes are captured. During the implementation of prevention and control measures, the support strength was enhanced by using Φ22mm high-strength anchor bolts (original specification Φ20mm) arranged at 1.0m×1.0m intervals, and Φ17.8mm prestressed anchor cables were added (interval 2m×2m). Stress-directed pressure relief drilling with a diameter of 150mm and a depth of 15m was carried out, with a hole spacing of 3m, arranged in a fan shape on both sides of the roadway. Grouting reinforcement of the surrounding rock used cement-water glass double-liquid grout (volume ratio 1:1), with the grouting pressure controlled at 3-5MPa, the grouting hole depth 8m, and the spacing 5m. Emergency evacuation commands were sent to personnel terminals in high-risk areas via the underground personnel positioning system (KJ237 type), and the emergency broadcast system was activated to continuously broadcast evacuation routes. The dispatch center monitored the evacuation progress in real time until it was confirmed that all personnel had left the danger zone.

[0134] S802: If the warning signal is at the medium-risk level, maintain the basic layout of the monitoring network and appropriately increase the data acquisition frequency, and select at least one of the following for prevention and control: support strength enhancement or stress-oriented pressure relief.

[0135] Specifically, when responding to a medium-risk warning, the S802 monitoring network maintains a basic layout of 5 meters per monitoring point in the core area and 10 meters per monitoring point in the secondary area. The data acquisition frequency is appropriately increased, with mechanical data adjusted from 1Hz to 5Hz and fluctuation signals from 2kHz to 5kHz. If the control measures involve strengthening the support, Φ20mm anchor bolts are used to increase the spacing to 1.2m x 1.2m, and steel strips are added to the roadway roof. If stress-oriented pressure relief is chosen, pressure relief holes with a diameter of 120mm and a depth of 10m are constructed, spaced 5m apart, arranged in a single row on one side of the roadway. If both methods are combined, the above parameters are halved. During implementation, a designated person records parameters such as support installation torque and pressure relief hole discharge every 2 hours to ensure that the construction quality meets design requirements.

[0136] S803: If the warning signal is of low risk level, maintain the regular layout of the monitoring network and the data collection frequency, and select surrounding rock grouting reinforcement or conventional support and maintenance measures to implement prevention and control.

[0137] Specifically, for low-risk warnings, the S803 monitoring network maintains its standard layout (5 meters / sensor in the core area, 10 meters / sensor in the secondary area, and 20 meters / laser displacement point in the outer area), and the data acquisition frequency remains at its initial setting (1Hz for mechanical data, 2kHz for fluctuation data, and 5 minutes / time for environmental data). If rock grouting reinforcement is used as a control measure, low-concentration cement grout (water-cement ratio 1:1.5) should be selected, with a grouting pressure of 1-2 MPa, a grouting hole depth of 5m, and a spacing of 10m. If conventional support and maintenance are implemented, the anchor bolt nut torque should be checked every shift (ensuring ≥300N). m), anchor cable preload (≥150kN), timely re-tighten loose parts, and locally replace deformed support components.

[0138] S804: Feed back the monitoring adjustment parameters and prevention and control measures implementation results corresponding to each level to the monitoring network.

[0139] Specifically, the S804 feedback mechanism integrates monitoring adjustment parameters at each level (e.g., sensor density and sampling frequency for high-risk levels), control measure execution data (e.g., number of anchor bolts, total length of pressure relief holes, grouting volume), and implementation effects (e.g., stress value change rate and microseismic frequency attenuation rate after measures) into structured data, which is then transmitted to a ground database via industrial Ethernet. The system automatically generates a "Monitoring Adjustment and Control Effect Comparison Table," which includes multi-dimensional monitoring data comparison curves before and after the implementation of measures (e.g., stress time history curves and microseismic energy distribution maps), and calculates the risk reduction rate of the control measures (formula: (risk probability before measures - risk probability after measures) / risk probability before measures × 100%). This feedback data is used to dynamically optimize the monitoring network layout algorithm and prediction model parameters. For example, if the risk of a certain area decreases significantly after multiple stress relief implementations, the system will automatically increase the weight of that area in the model training, improving the accuracy of subsequent early warning and control.

[0140] For high-risk warnings, increasing the density of monitoring points and the frequency of data collection can enhance the real-time monitoring accuracy of dangerous areas. Combined with a combination of control measures such as support strength enhancement, stress-oriented pressure relief, and surrounding rock grouting reinforcement, the risk escalation can be quickly contained and personnel safety ensured. For medium-risk warnings, maintaining the basic monitoring layout and appropriately increasing the collection frequency can ensure monitoring effectiveness while avoiding resource waste. Selecting at least one of the measures of support strengthening or stress-oriented pressure relief can specifically control the spread of risk. For low-risk warnings, maintaining routine monitoring and control measures can achieve normalized risk management and reduce unnecessary cost inputs. The mechanism of feeding back the adjusted parameters and execution results to the monitoring network can form a closed-loop optimization, continuously improving the adaptability of the monitoring system and the effectiveness of control measures.

[0141] like Figure 9 As shown, this embodiment provides a deep well rockburst early warning system based on multi-dimensional monitoring, which includes:

[0142] The data acquisition module includes a monitoring network deployed in the deep shaft tunnel and surrounding rock, used to collect multi-dimensional monitoring data in real time, including stress value, strain value, microseismic event parameters, acoustic signal, temperature data, radon concentration value, electromagnetic radiation intensity and three-dimensional deformation data of the tunnel surface.

[0143] The spatiotemporal registration module is used to perform spatiotemporal registration on the multi-dimensional monitoring data to obtain a spatiotemporally aligned multi-source monitoring dataset.

[0144] The noise suppression module is used to perform dynamic noise suppression based on physical characteristics on the multi-source monitoring dataset to obtain a noise-suppressed dataset.

[0145] The feature filtering module is used to filter features related to rockburst pressure from the noise-suppressed dataset using the mutual information entropy algorithm, so as to obtain a feature-optimized dataset.

[0146] The feature extraction module is used to extract frequency domain features, time domain features, and time-frequency domain features from the feature-optimized dataset using principal component extraction and phase space reconstruction methods to form a multidimensional state space dataset.

[0147] The level prediction module includes a pre-trained rockburst hazard level prediction model, which is used to input the multi-dimensional state space dataset into the model to obtain the rockburst hazard level.

[0148] The early warning triggering module is used to trigger an early warning signal of the corresponding level based on the rockburst hazard level.

[0149] The dynamic adjustment module is used to dynamically adjust the layout and control measures of the monitoring network based on the corresponding level of early warning signals.

[0150] In a further embodiment, the data acquisition module specifically includes:

[0151] The monitoring zoning unit is used to determine the core monitoring area, secondary monitoring area and peripheral monitoring area of ​​the monitoring network based on the geomechanical parameters of deep well tunnels and the historical risk zoning of rockbursts;

[0152] The sensor configuration unit is used to install stress sensors, strain sensors, micro-vibration monitoring systems and acoustic monitoring devices in the core monitoring area, temperature sensors, radon concentration sensors and electromagnetic radiation monitoring devices in the secondary monitoring area, and a laser displacement sensor array in the peripheral monitoring area.

[0153] The parameter setting unit is used to set the acquisition parameters for multi-dimensional monitoring data, including sampling frequency, triggering conditions, and measurement range.

[0154] The transmission link unit is used to construct a hierarchical data transmission link based on the importance level of the monitoring area. The data of the core monitoring area is transmitted through an industrial Ethernet channel with redundant backup, the data of the secondary monitoring area is transmitted through a mining Ethernet channel, and the data of the peripheral monitoring area is transmitted through a wireless self-organizing network to realize the real-time transmission of the multi-dimensional monitoring data.

[0155] The spatiotemporal marking unit is used to synchronize the timestamps and mark the geological coordinates of the multi-dimensional monitoring data.

[0156] In a further embodiment, the spatiotemporal registration module specifically includes:

[0157] The coordinate system establishment unit is used to establish a spatiotemporal reference coordinate system based on the three-dimensional geological model of the deep well tunnel. The spatiotemporal reference coordinate system takes the tunnel starting point as the origin, the tunnel axis as the X-axis, the perpendicular tunnel direction as the Y-axis, and the perpendicular tunnel floor as the Z-axis, and is associated with absolute geological coordinates.

[0158] The time calibration unit is used to perform timestamp calibration on the multi-dimensional monitoring data, unify the acquisition time of various monitoring devices to the standard time axis of the spatiotemporal reference coordinate system, and use linear interpolation to align the time dimension of asynchronous time data.

[0159] The spatial transformation unit is used to construct a spatial coordinate transformation matrix for the location of each monitoring device based on the spatiotemporal reference coordinate system, and to transform the original spatial data of different monitoring devices to three-dimensional coordinates under the spatiotemporal reference coordinate system.

[0160] Spatial gridding unit is used to perform spatial gridding processing on monitoring data with uneven spatial distribution using the Kriging interpolation method to form a spatially continuous monitoring data field. The monitoring data field is used to integrate the spatial correlation of multi-dimensional monitoring data.

[0161] The error verification unit is used to calculate the registration error based on the time synchronization error and the spatial coordinate deviation. The time synchronization error is the deviation between the timestamp of the calibrated monitoring data and the standard time axis, and the spatial coordinate deviation is the deviation between the transformed three-dimensional coordinates and the theoretical coordinates of the spatiotemporal reference coordinate system. When the registration error exceeds the registration accuracy threshold, the time calibration or spatial transformation is re-executed. When the registration error meets the requirements of the registration accuracy threshold, the spatiotemporally aligned multi-source monitoring dataset is output.

[0162] In a further embodiment, the noise suppression module specifically includes:

[0163] The noise identification unit is used to classify and identify the noise type corresponding to each dimension of data based on the physical properties of multi-source monitoring data. Among them, mechanical data includes stress and strain values, wave signals include micro-vibration event parameters and acoustic signals, environmental parameters include temperature data and radon concentration values, and electromagnetic and optical data include electromagnetic radiation intensity and three-dimensional deformation data of the tunnel surface.

[0164] The algorithm matching unit is used to suppress noise by adopting a noise reduction algorithm that matches the physical characteristics of the data according to the identified noise type. It includes a drift compensation algorithm based on physical change trends for the mechanical data, an adaptive filtering algorithm for the wave signals, a baseline correction algorithm for the environmental parameters, and an interference subspace separation algorithm for the electromagnetic and optical data.

[0165] The parameter adjustment unit is used to monitor the data variability in the noise suppression process in real time, and uses the data variability as feedback to dynamically adjust the parameters of the noise reduction algorithm so that the suppression intensity is adapted to the dynamic changes of noise.

[0166] The threshold verification unit is used to calculate the signal-to-noise ratio and feature retention of the noise-suppressed data. When the signal-to-noise ratio reaches the signal-to-noise ratio threshold and the feature retention reaches the feature integrity threshold, the noise-suppressed dataset is output; otherwise, the noise reduction algorithm parameters are readjusted.

[0167] In a further embodiment, the feature filtering module specifically includes:

[0168] The model building unit is used to build a mutual information entropy calculation model between each feature in the noise-suppressed dataset and the rockburst hazard level. The mutual information entropy calculation model is based on the mutual information entropy formula in information theory and quantifies the degree of information correlation between feature variables and rockburst hazard level.

[0169] The entropy calculation unit is used to calculate the mutual information entropy value of each feature through the mutual information entropy calculation model. The mutual information entropy value represents the contribution of the feature to the prediction of rockburst hazard.

[0170] The feature filtering unit is used to filter features whose mutual information entropy values ​​are greater than the feature contribution threshold based on the calculated mutual information entropy values, forming a feature subset related to rockburst pressure.

[0171] The redundancy removal unit is used to perform redundancy analysis on the feature subset, remove redundant features with similar mutual information entropy values ​​and correlations higher than the feature redundancy threshold, and obtain the feature-optimized dataset.

[0172] In a further embodiment, the feature extraction module specifically includes:

[0173] The principal component extraction unit is used to construct a covariance matrix for the feature-optimized dataset, and extract principal components with variance contribution rates exceeding a preset threshold by performing eigenvalue decomposition on the covariance matrix to form a dimensionality-reduced feature subset.

[0174] The phase space construction unit is used to construct a high-dimensional phase space based on the phase space reconstruction method, by determining the embedding dimension and time delay parameters according to the autocorrelation function and mutual information function of the reduced feature subset.

[0175] The feature extraction unit is used to extract frequency domain features, time domain features and time-frequency domain features of each feature in the high-dimensional phase space, wherein the frequency domain features are obtained by Fourier transform, the time domain features are calculated based on the statistical characteristics of time series, and the time-frequency domain features are extracted by wavelet transform.

[0176] The feature fusion unit is used to fuse the extracted frequency domain features, time domain features, and time-frequency domain features to form a multidimensional state space dataset.

[0177] In a further embodiment, the pre-trained rockburst hazard level prediction model in the level prediction module is constructed using a random forest algorithm optimized by combining a long short-term memory neural network with a genetic algorithm.

[0178] In a further embodiment, the early warning triggering module specifically includes:

[0179] The level correspondence unit is used to preset the correspondence between the rockburst hazard level and the warning level, where the low hazard level corresponds to the first-level warning, the medium hazard level corresponds to the second-level warning, and the high hazard level corresponds to the third-level warning.

[0180] The signal triggering unit is used to trigger a corresponding level of early warning signal according to the rockburst hazard level. The level 1 early warning is an audible and visual alert signal for a local area; the level 2 early warning is an early warning information pushed by the communication terminal of the monitoring area and accompanied by an audible and visual alarm; and the level 3 early warning is an emergency alarm signal for the entire monitoring network coverage area, which also triggers a personnel evacuation notice.

[0181] The log recording unit is used to record the trigger time, level, and coverage of the warning signal, forming a warning log.

[0182] In a further embodiment, the dynamic adjustment module specifically includes:

[0183] The high-risk adjustment unit is used to increase the spatial distribution density of monitoring points in the monitoring network and increase the data acquisition frequency when the early warning signal is at a high-risk level. It selects a combination of support strength enhancement, stress-oriented pressure relief and surrounding rock grouting reinforcement to implement prevention and control, and triggers emergency evacuation instructions for personnel at the same time.

[0184] The medium-risk adjustment unit is used to maintain the basic layout of the monitoring network and appropriately increase the data acquisition frequency when the early warning signal is at the medium-risk level, and to select at least one of support strength enhancement or stress-oriented pressure relief to implement prevention and control.

[0185] The low-risk adjustment unit is used to maintain the normal layout and data acquisition frequency of the monitoring network when the early warning signal is at a low-risk level, and to select surrounding rock grouting reinforcement or conventional support and maintenance measures to implement prevention and control.

[0186] The feedback unit is used to feed back the monitoring adjustment parameters and the results of the implementation of prevention and control measures corresponding to each level to the monitoring network.

[0187] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements any of the methods described above. This invention also provides an electronic device. The electronic device of this invention includes: one or more processors; and a storage device for storing one or more programs, which, when executed by the one or more processors, cause the one or more processors to implement a deep well rockburst early warning method based on multi-dimensional monitoring provided by this invention. Figure 10 As shown, the computer system 800 of the electronic device includes a central processing unit (CPU) 801, which can perform various appropriate actions and processes according to programs stored in read-only memory (ROM) 802 or programs loaded from storage section 808 into random access memory (RAM) 803. The RAM 803 also stores various programs and data required for the operation of the computer system 800. The CPU 801, ROM 802, and RAM 803 are interconnected via bus 804. An input / output (I / O) interface 805 is also connected to bus 804. The following components are connected to the I / O interface 805: an input section 806 including a keyboard, mouse, etc.; an output section 807 including cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 808 including a hard disk, etc.; and a communication section 809 including a network interface card such as a LAN card, modem, etc. The communication section 809 performs communication processing via a network such as the Internet. A driver 810 is also connected to the I / O interface 805 as needed. Removable media 811, such as disks, optical discs, magneto-optical discs, semiconductor memories, etc., are installed on drive 810 as needed so that computer programs read from them can be installed into storage section 808 as needed.

[0188] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A deep well rock burst early warning method based on multi-dimensional monitoring, characterized in that, The method comprises the following steps: S10: Real-time collection of multi-dimensional monitoring data including stress value, strain value, microseismic event parameter, acoustic signal, temperature data, radon concentration value, electromagnetic radiation intensity and roadway surface three-dimensional deformation data by using a monitoring network arranged in a deep well tunnel and surrounding rock; S20: Time-space registration of the multi-dimensional monitoring data to obtain a time-space aligned multi-source monitoring data set; S30: Dynamic noise suppression based on physical characteristics of the multi-source monitoring data set to obtain a noise-suppressed data set; S40: Filtering of features related to rock burst intensity from the noise-suppressed data set by using a mutual information entropy algorithm to obtain a feature-optimized data set; S50: Extraction of frequency domain features, time domain features and time-frequency domain features from the feature-optimized data set by using a principal component extraction method and a phase space reconstruction method to form a multi-dimensional state space data set; S60: Input of the multi-dimensional state space data set into a pre-trained rock burst danger level prediction model to obtain a rock burst danger level; S70: Triggering of a warning signal of a corresponding level according to the rock burst danger level; S80: Dynamic adjustment of the layout of the monitoring network and prevention and control measures based on the warning signal of the corresponding level; The step S30 specifically comprises: S301: Classification and identification of noise types corresponding to each dimension of data based on the physical properties of the multi-source monitoring data, wherein the mechanical data includes stress value and strain value, the wave signal includes microseismic event parameter and acoustic signal, the environmental parameter includes temperature data and radon concentration value, and the electromagnetic and optical data includes electromagnetic radiation intensity and roadway surface three-dimensional deformation data; S302: Noise suppression by using a noise reduction algorithm matched with the physical characteristics of the data according to the identified noise types, which includes drift compensation algorithm based on physical change trend for the mechanical data, adaptive filtering algorithm for the wave signal, baseline correction algorithm for the environmental parameter and interference subspace separation algorithm for the electromagnetic and optical data; the adaptive filtering algorithm uses an adaptive filter based on least mean square algorithm; the baseline correction algorithm includes calculation of baseline values of temperature data and radon concentration value by using a sliding window, and baseline correction when instantaneous value deviates from the baseline value by more than 3 times of standard deviation; the interference subspace separation algorithm includes feature analysis of the collected electromagnetic radiation intensity and roadway surface three-dimensional deformation data to identify useful signals and interference signals related to rock burst; determination of spatial basis vectors of the useful signals and the interference signals by eigenvalue decomposition or singular value decomposition to construct corresponding subspace models; elimination of components belonging to the interference signal subspace in the mixed signals contained in the electromagnetic and optical data by orthogonal projection to only retain effective components of the useful signal subspace; S303: Real-time monitoring of data variability in the noise suppression process, and dynamic adjustment of parameters of the noise reduction algorithm by using the data variability as a feedback to adapt the suppression intensity to dynamic changes of noise. S304: Calculate the signal-to-noise ratio and feature retention of the noise-reduced data, and when the signal-to-noise ratio reaches the signal-to-noise ratio threshold and the feature retention reaches the feature integrity threshold, output the noise-reduced data set; otherwise, return to step S302 to adjust the parameters of the noise reduction algorithm again; wherein the feature retention is measured by calculating the correlation coefficient of the key features of the data before and after noise reduction, and the key features include stress peak, microseismic energy and temperature gradient; Wherein, step S50 specifically comprises: S501: Construct a covariance matrix for the feature-optimized data set, wherein the dimension of the covariance matrix is determined based on the number of feature variables, and the elements of the covariance matrix represent the degree of cooperative change between any two feature variables; perform eigenvalue decomposition on the covariance matrix to obtain a plurality of eigenvalues; set a preset threshold for the cumulative variance contribution rate, calculate the cumulative variance contribution rate of the first N eigenvalues, and when the cumulative variance contribution rate exceeds the preset threshold, extract the first N principal components to form a reduced dimension feature subset; wherein the preset threshold is 85% of the cumulative variance contribution rate; S502: Based on the phase space reconstruction method, determine the embedding dimension and time delay parameter according to the autocorrelation function and mutual information function of the reduced dimension feature subset, and construct a high-dimensional phase space; S503: In the high-dimensional phase space, extract the frequency domain feature, time domain feature and time-frequency domain feature of each feature respectively, wherein the frequency domain feature is extracted by: converting the time domain signal into a frequency domain spectrum through Fourier transform, and calculating the main frequency of each principal component; the time domain feature is extracted by: calculating the mean, peak factor, kurtosis and root mean square value of each principal component based on the time series statistical characteristics; the time-frequency domain feature is extracted by: using db4 wavelet basis to perform 3-layer wavelet decomposition on the time series, and calculating the energy proportion and wavelet entropy under each scale; S504: Fuse the extracted frequency domain feature, time domain feature and time-frequency domain feature to form a multi-dimensional state space data set.

2. The method of claim 1, wherein, Step S10 specifically comprises: S101: Determine the core monitoring area, secondary monitoring area and peripheral monitoring area of the monitoring network based on the geological mechanics parameters of deep well roadway and rock burst historical risk zoning; S102: Set stress sensors, strain sensors, microseismic monitoring systems and acoustic wave monitoring devices in the core monitoring area, set temperature sensors, radon concentration sensors and electromagnetic radiation monitoring devices in the secondary monitoring area, and set laser displacement sensor arrays in the peripheral monitoring area; S103: Set the acquisition parameters of multi-dimensional monitoring data, including sampling frequency, trigger condition and range; S104: Construct a hierarchical data transmission link based on the importance level of the monitoring area, wherein the data of the core monitoring area is transmitted through an industrial Ethernet channel with redundant backup, the data of the secondary monitoring area is transmitted through a mine Ethernet channel, and the data of the peripheral monitoring area is transmitted through a wireless ad hoc network, realizing real-time transmission of the multi-dimensional monitoring data; S105: Time stamp synchronization and geological coordinate labeling are performed on the multi-dimensional monitoring data.

3. The method of claim 1, wherein, Step S20 specifically comprises: S201: Establish a space-time reference coordinate system based on a deep well tunnel three-dimensional geological model, the space-time reference coordinate system takes the starting point of the tunnel as the origin, the tunnel axial direction as the X axis, the vertical tunnel direction as the Y axis, and the vertical tunnel floor as the Z axis, and is associated with the absolute geological coordinates; S202: Time stamp calibration is performed on the multidimensional monitoring data, the collection time of various monitoring devices used to collect the multidimensional monitoring data is unified to the standard time axis of the space-time reference coordinate system, and linear interpolation is used for time dimension alignment of time asynchronous data; S203: Based on the space-time reference coordinate system, a space coordinate conversion matrix of the setting position of each monitoring device is constructed, and the original space data of different monitoring devices is converted to three-dimensional coordinates under the space-time reference coordinate system; S204: The Kriging interpolation method is used for spatial gridding processing of the monitoring data with uneven spatial distribution, and a spatial continuous monitoring data field is formed, which is used for integrating the spatial correlation of the multidimensional monitoring data; S205: Calculate the registration error according to the time synchronization error and the spatial coordinate deviation, wherein the time synchronization error is the deviation value of the time stamp of the calibrated monitoring data from the standard time axis, and the spatial coordinate deviation is the deviation amount of the converted three-dimensional coordinates from the theoretical coordinates of the space-time reference coordinate system; when the registration error exceeds the registration accuracy threshold, return to step S202 to re-calibrate the time or space; when the registration error meets the requirements of the registration accuracy threshold, output the space-time aligned multi-source monitoring data set.

4. The method of claim 1, wherein, Step S40 specifically includes: S401: Construct a mutual information entropy calculation model between each feature in the noise suppressed data set and the rock burst danger, the mutual information entropy calculation model is based on the mutual information entropy formula in information theory, and quantifies the information correlation degree between the feature variable and the rock burst danger level; S402: Calculate the mutual information entropy value of each feature through the mutual information entropy calculation model, the mutual information entropy value represents the prediction contribution of the feature to the rock burst danger; S403: According to the calculated mutual information entropy value, filter the features with mutual information entropy values greater than the feature contribution threshold to form a feature subset related to the rock burst; S404: Redundancy analysis is performed on the feature subset, redundant features with similar mutual information entropy values and higher correlation than the feature redundancy threshold are removed, and a feature optimized data set is obtained.

5. The method of claim 1, wherein, The pre-trained rock burst danger level prediction model is constructed by using a long short-term memory neural network combined with a random forest algorithm optimized by a genetic algorithm.

6. The method of claim 1, wherein, Step S70 specifically includes: S701: Preset the corresponding relationship between the rock burst danger level and the warning level, wherein the low danger level corresponds to the first level warning, the medium danger level corresponds to the second level warning, and the high danger level corresponds to the third level warning; S702: Trigger the corresponding level warning signal according to the rock burst danger level, which includes a local area sound and light prompt signal for the first level warning, a communication terminal pushes the warning information and is accompanied by a sound and light alarm for the monitoring area for the second level warning, and an emergency alarm signal for the whole monitoring network coverage area for the third level warning, while triggering the personnel evacuation notification; S703: Record the triggering time, level and coverage of the early warning signal to form an early warning log.

7. The method of claim 6, wherein, Step S80 specifically includes: S801: If the early warning signal is of high risk level, increase the spatial distribution density of monitoring points in the monitoring network and increase the data acquisition frequency, select at least one of support strength reinforcement, stress directional pressure relief and surrounding rock grouting reinforcement for implementation of prevention and control, and trigger personnel emergency evacuation instruction at the same time; S802: If the early warning signal is of medium risk level, maintain the basic layout of the monitoring network and moderately increase the data acquisition frequency, and select at least one of support strength reinforcement or stress directional pressure relief for implementation of prevention and control; S803: If the early warning signal is of low risk level, keep the regular layout of the monitoring network and the data acquisition frequency, and select at least one of surrounding rock grouting reinforcement or regular support maintenance measures for implementation of prevention and control; S804: Feedback the monitoring adjustment parameters and the implementation results of the prevention and control measures corresponding to each level to the monitoring network.

8. A deep mine rock burst early warning system based on multi-dimensional monitoring, characterized in that, The system is used to execute the deep well rock burst early warning method based on multi-dimensional monitoring in any one of claims 1-7, and the system comprises: A data acquisition module, the data acquisition module comprising a monitoring network arranged in a deep well tunnel and surrounding rock, for real-time acquisition of multi-dimensional monitoring data including stress value, strain value, microseismic event parameter, acoustic signal, temperature data, radon concentration value, electromagnetic radiation intensity and tunnel surface three-dimensional deformation data; A time-space registration module, for time-space registration of the multi-dimensional monitoring data to obtain a time-space aligned multi-source monitoring data set; A noise suppression module, for dynamic noise suppression of the multi-source monitoring data set based on physical characteristics to obtain a noise suppressed data set; A feature screening module, for screening features related to rock burst intensity from the noise suppressed data set through mutual information entropy algorithm to obtain a feature optimized data set; A feature extraction module, for extracting frequency domain features, time domain features and time-frequency domain features from the feature optimized data set through principal component extraction method and phase space reconstruction method to form a multi-dimensional state space data set; A level prediction module, the level prediction module comprising a pre-trained rock burst danger level prediction model, for inputting the multi-dimensional state space data set into the model to obtain a rock burst danger level; An early warning triggering module, for triggering an early warning signal of a corresponding level according to the rock burst danger level; A dynamic adjustment module, for dynamically adjusting the layout of the monitoring network and the prevention and control measures based on the early warning signal of the corresponding level.

Images