A method, device and medium for early rockburst identification based on physical information injection

Abstract

This application discloses a method, device, and medium for early identification of rockbursts based on physical information injection, belonging to the technical field of underground engineering disaster monitoring and intelligent early warning. The method includes: collecting acoustic emission, microseismic, stress, and displacement data; combining these with static geological attributes such as lithology and joints to generate a high-dimensional feature cube through spatiotemporal alignment; injecting the cube's input physical information into a transformer for modeling, outputting intermediate features such as energy accumulation rate and a three-dimensional hazard confidence distribution; retrieving a rockburst case library based on the intermediate features, fusing similar cases to generate an explanation of the induction mechanism and prevention and control recommendations; extracting and visualizing the location of potential fracture surfaces; triggering edge alarms based on risk levels and recording execution feedback; monitoring the input distribution and physical residual distribution, and triggering incremental fine-tuning based on drift status and feedback. This application achieves the technical effects of reducing false alarm and false negative rates, improving early warning timeliness, and spatial positioning accuracy through the above method.

Description

Technical Field

[0001] This application relates to the technical field of underground engineering disaster monitoring and intelligent early warning, and in particular to a method, equipment and medium for early identification of rockbursts based on physical information injection. Background Technology

[0002] Rockburst (or rock burst) is a common and highly dangerous dynamic disaster in underground engineering tunnels of deep coal mines, metal mines, and other similar facilities. It manifests as the sudden instability of rock mass under stress, releasing enormous elastic energy instantaneously, which can easily lead to damage to tunnel support structures, mining equipment, and even casualties. Currently, monitoring and early warning of rockbursts mainly rely on various sensors installed around the tunnels and working faces. For example, triaxial stress gauges are used to monitor the stress state of the surrounding rock, acoustic emission (AE) sensors are used to capture elastic wave signals generated by rock fracturing, and microseismic monitoring systems are used to locate larger-scale rock fracturing events. Based on the data collected by these sensors, existing technologies generally use threshold-triggered alarms; that is, an alarm is issued when the monitored stress value, acoustic emission energy, or microseismic event intensity exceeds a preset safety threshold. In addition, some methods attempt to apply traditional machine learning algorithms (such as support vector machines and random forests) to statistically analyze monitoring data sequences in order to identify abnormal patterns. These technologies, to a certain extent, provide a guarantee for the safety of underground engineering and constitute the foundation for technological development in this field.

[0003] However, the aforementioned existing technologies have significant shortcomings. First, alarm methods relying on fixed thresholds are ineffective in practical applications. Due to the complex and variable geological conditions of deep rock strata, significant differences in lithology (such as hard rock, soft rock, and interbedded rock), varying degrees of tectonic joint development, and dynamic stress field transfer caused by mining activities, uniformly set thresholds are difficult to adapt to all working conditions. This results in high rates of false alarms (identifying non-dangerous signals as dangerous) and false alarms (failing to identify true dangers), making it difficult to meet high safety requirements in terms of accuracy and reliability of early warnings. Second, existing machine learning methods typically treat monitoring data as purely statistical sequences, failing to effectively incorporate the physical mechanisms of rock mass failure. These models neglect the constitutive relationship of the rock mass itself (stress-strain law), the tensor distribution characteristics of the in-situ stress field, and the key physical processes of energy accumulation and dissipation. Consequently, the models have poor generalization ability when facing complex geological conditions (such as high burial depth, strong tectonic stress, and lithological abrupt change zones), and the prediction results lack physical rationality, making it difficult to explain their underlying causes. Furthermore, with the advancement of mine digitalization, multi-source heterogeneous data such as acoustic emission, microseismic activity, stress, and displacement can be collected in real time. However, existing technologies lack the ability to efficiently integrate and deeply couple these multi-dimensional, spatiotemporally related data for modeling. This fails to fully utilize the overall value of the data for early and accurate risk identification and spatial positioning, and the timeliness and spatial accuracy of early warnings need to be improved.

[0004] Therefore, how to reduce the false alarm and missed alarm rates, improve the timeliness of early warning and the accuracy of spatial positioning has become an urgent technical problem to be solved. Summary of the Invention

[0005] This application provides a method, device, and medium for early rockburst identification based on physical information injection, in order to solve the following technical problems: how to reduce the false alarm and missed alarm rates, and improve the timeliness of early warning and spatial positioning accuracy.

[0006] In a first aspect, embodiments of this application provide a method for early identification of rockbursts based on physical information injection. The method includes: collecting acoustic emission waveform data, microseismic waveform data, triaxial stress data, and tunnel deformation and displacement data within an underground engineering area; combining the collected acoustic emission waveform data, microseismic waveform data, triaxial stress data, and tunnel deformation and displacement data with static geological attribute data, performing time synchronization and spatial coordinate alignment to generate high-dimensional feature cube data with a unified spatiotemporal reference; wherein the static geological attribute data includes lithology, structural joints, and strike / dip information; inputting the high-dimensional feature cube data into a preset physical information injection transformer for modeling to obtain intermediate feature vector data containing energy accumulation rate, stress concentration coefficient, and fracture evolution potential, as well as three-dimensional hazard confidence distribution data; and retrieving a preset rockburst case knowledge base based on the intermediate feature vector data to obtain similar historical case data. The similar historical case data and the intermediate feature vector data are concatenated and fused to generate natural language text data containing explanations of potential inducing mechanisms and suggestions for prevention and control measures. A three-dimensional risk heat map is generated based on the three-dimensional hazard confidence distribution data. Spatial location data of potential rupture surfaces is extracted from the three-dimensional risk heat map data, and the three-dimensional risk heat map data, spatial location data of potential rupture surfaces, and the natural language text data are visualized. Risk assessment is performed based on the three-dimensional hazard confidence distribution data. When the risk level reaches a preset threshold, an edge alarm signal is triggered, and on-site execution feedback data for the suggested prevention and control measures is recorded. The input distribution characteristics of the high-dimensional feature cube data and the physical residual distribution characteristics generated by modeling processing are monitored. The model drift state is determined based on the monitoring results, and an incremental fine-tuning process of the model is triggered according to the model drift state and the execution feedback data.

[0007] In one implementation of this application, the collected acoustic emission waveform data, microseismic waveform data, triaxial stress data, and tunnel deformation and displacement data are combined with static geological attribute data, and time synchronization and spatial coordinate alignment are performed to generate high-dimensional feature cube data with a unified spatiotemporal reference. Specifically, this includes: synchronizing the sensor nodes that collect the acoustic emission waveform data, microseismic waveform data, triaxial stress data, and tunnel deformation and displacement data based on a preset IEEE-1588PTP precise time protocol to obtain synchronization timestamp data; establishing a three-dimensional spatial coordinate framework data of the underground engineering area based on a preset RTK-SLAM algorithm and laser scanning point cloud data; and aligning the acoustic emission waveform data, microseismic waveform data, triaxial stress data, and tunnel deformation and displacement data with static geological attribute data. The deformation displacement data is spatiotemporally mapped based on the synchronous timestamp data and the three-dimensional spatial coordinate frame data to obtain spatially aligned monitoring point data; the acoustic emission waveform data and microseismic waveform data are processed by sliding window filtering, noise suppression, and energy envelope quantization at the edge computing nodes to obtain preprocessed waveform feature data; the triaxial stress data and tunnel deformation displacement data are processed by adaptive Kalman filtering to suppress noise and linear interpolation to compensate for missing values ​​to obtain preprocessed stress displacement feature data; the preprocessed waveform feature data, the preprocessed stress displacement feature data, and the static geological attribute data are organized into high-dimensional feature cube data indexed by time dimension, spatial dimension, and feature dimension according to a preset time window and spatial resolution.

[0008] In one implementation of this application, the high-dimensional feature cube data is input into a preset physical information injection transformer for modeling to obtain intermediate feature vector data containing energy accumulation rate, stress concentration coefficient, and fracture evolution potential, as well as three-dimensional hazard confidence distribution data. Specifically, this includes: performing frequency domain transformation on the high-dimensional feature cube data to obtain frequency domain feature data; inputting the frequency domain feature data into a multi-head self-attention layer of the physical information injection transformer; and explicitly introducing the rock mass elastoplastic constitutive residual data and the in-situ geostress tensor data as penalty terms when calculating the similarity between the query vector and the key vector in each attention head of the multi-head self-attention layer. Attention weight data conforming to mechanical constraints is calculated; wherein, the rock mass elastoplastic constitutive residual data is obtained by real-time calculation of the difference between predicted stress and actual stress based on constitutive equations obtained from field measurements or experiments; the value vector is weighted and aggregated based on the calculated attention weight data to obtain deep feature coupling data; the deep feature coupling data is input into the multi-expert gated subnetwork of the physical information injection converter; based on the lithological properties of the input data, the multi-expert gated subnetwork dynamically routes the deep feature coupling data to the corresponding expert subnetwork for processing to obtain the intermediate feature vector data and the three-dimensional hazard confidence distribution data.

[0009] In one implementation of this application, the deep feature coupling data is dynamically routed to the corresponding expert subnetwork for processing based on the lithological properties of the input data by the multi-expert gated subnetwork, thereby obtaining the intermediate feature vector data and the three-dimensional hazard confidence distribution data. Specifically, this includes: inputting the deep feature coupling data and the lithological label data from the static geological attribute data into the gated network of the multi-expert gated subnetwork; learning the coupling pattern between the deep feature coupling data and different lithological types based on the gated network, generating expert selection weight data; dynamically routing the deep feature coupling data to one or more of the best-matching expert subnetworks based on the expert selection weight data; and processing the routed deep feature coupling data based on the selected expert subnetwork, outputting the intermediate feature vector data and the three-dimensional hazard confidence distribution data.

[0010] In one implementation of this application, a preset rockburst case knowledge base is retrieved based on the intermediate feature vector data to obtain similar historical case data. The similar historical case data is then concatenated and fused with the intermediate feature vector data to generate natural language text data containing explanations of potential triggering mechanisms and suggestions for prevention and control measures. Specifically, this includes: performing a first-stage coarse-grained vector similarity retrieval in the rockburst case knowledge base based on the intermediate feature vector data to obtain a preliminary matching candidate case dataset; and performing a second-stage fine-grained retrieval on the preliminary matching candidate case dataset, wherein the fine-grained retrieval is based on static attribute vector similarity. Hierarchical matching calculations are performed on the similarity of the dynamic sequence vectors to filter out target case data with high similarity. The static attribute vectors include lithology, strike, and joint density, while the dynamic sequence vectors include energy decay curves and fracture evolution trajectories. The feature representations of the filtered target case data are concatenated with the intermediate feature vector data to form knowledge-enhanced feature input data. This knowledge-enhanced feature input data is then input into a language model that has been fine-tuned. Based on the language model, the knowledge-enhanced feature input data is processed to automatically generate natural language text data containing explanations of potential inducing mechanisms and suggestions for prevention and control measures.

[0011] In one implementation of this application, a three-dimensional risk heat map is generated based on the three-dimensional hazard confidence distribution data, and potential rupture surface spatial location data is extracted based on the three-dimensional risk heat map data. The three-dimensional risk heat map data, potential rupture surface spatial location data, and natural language text data are then visualized. Specifically, this includes: mapping the three-dimensional hazard confidence distribution data onto spatial voxels corresponding to a three-dimensional tunnel geographic information model to generate an initial three-dimensional risk heat map; performing a clustering analysis algorithm based on density and spatial connectivity on the initial three-dimensional risk heat map data to identify high-risk connected regions; calculating the minimum convex hull geometry data for each identified high-risk connected region, where the minimum convex hull geometry data represents the potential rupture surface spatial location data; and rendering and displaying a three-dimensional tunnel scene superimposed with the initial three-dimensional risk heat map data, the convex hull boundaries of the high-risk connected region data, and the natural language text data on a three-dimensional visualization platform.

[0012] In one implementation of this application, risk assessment is performed based on the three-dimensional hazard confidence distribution data. When the risk level reaches a preset threshold, an edge alarm signal is triggered, and on-site execution feedback data for the proposed prevention and control measures is recorded. Specifically, this includes: real-time assessment of the risk level of the three-dimensional hazard confidence distribution data within a continuous time window; determining that an alarm triggering condition is met when the risk level of a predetermined number of consecutive time windows reaches or exceeds a preset risk threshold; immediately sending the edge alarm signal to the edge alarm device deployed on-site in the roadway to trigger an audible and visual alarm when the alarm triggering condition is met; automatically displaying an alarm information window on the central dispatch platform, the alarm information window containing a high-risk area identifier, the proposed prevention and control measures, and key explanations in the natural language text data; and collecting and recording the operation type, operation time, and post-operation area stability observation results data performed by on-site personnel according to the proposed prevention and control measures to form the execution feedback data.

[0013] In one implementation of this application, the input distribution characteristics of the high-dimensional feature cube data and the physical residual distribution characteristics generated by the modeling process are monitored. Based on the monitoring results, the model drift state is determined, and an incremental fine-tuning process of the model is triggered according to the model drift state and the execution feedback data. Specifically, this includes: real-time calculation of a first distribution difference metric between the current input distribution characteristics of the high-dimensional feature cube data and the historical baseline input distribution characteristics; real-time calculation of a second distribution difference metric between the current physical residual distribution characteristics generated by the modeling process and the expected physical residual distribution characteristics; calculation of a comprehensive drift index based on the first and second distribution difference metric data; when the comprehensive drift index continuously exceeds a preset drift threshold for a predetermined duration, model drift is determined to have occurred, triggering a model degradation operation, automatically switching to a historically stable model version for inference, and generating model aging alarm information; collecting the execution feedback data, model prediction error data, and actual alarm event data, and labeling them to form an incremental training sample dataset; based on the incremental training sample dataset, when a model performance degradation is detected or a periodic update instruction is received, an incremental fine-tuning training process for the physical information injection converter is triggered to update the model parameters.

[0014] Secondly, embodiments of this application also provide an electronic device, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the computer program to implement one of the above-described methods for early rockburst identification based on physical information injection.

[0015] Thirdly, embodiments of this application also provide a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements a rockburst early identification method based on physical information injection as described in any of the above descriptions.

[0016] The rockburst early identification method, device, and medium based on physical information injection provided in this application embodiment have at least the following technical effects:

[0017] By embedding the rock mass elastoplastic constitutive residuals and in-situ stress tensors as explicit priors into the self-attention weight calculation, the model is forced to adhere to rock mass mechanical conservation and volume compatibility constraints, fundamentally avoiding misjudging non-hazardous signals such as environmental noise and local stress fluctuations as signs of rockburst. Simultaneously, a multi-expert gating mechanism dynamically adapts to the heterogeneous response characteristics of complex lithologies such as hard rock, soft rock, and interbedded rock, significantly reducing the risk of missed detections due to differences in geological conditions, and maintaining high generalization ability even under extreme conditions such as deep burial under high pressure and with well-developed fractures.

[0018] Based on a retrieval-enhanced generation (RAG) geological knowledge alignment mechanism, real-time monitoring features are semantically matched with a historical rockburst case database at multiple levels to automatically generate a natural language explanation containing "inducing mechanism - prevention and control measures". This design transforms abstract numerical early warnings into actionable engineering decision-making basis (such as "reinforcing anchor cable + directional pressure relief borehole parameters"), enabling dispatchers to simultaneously grasp the complete logical chain of "where is high risk, why is it high risk, and how to deal with it", significantly improving the efficiency of on-site adoption of early warning information and the accuracy of on-site response.

[0019] By leveraging sub-millisecond time synchronization and centimeter-level spatial registration of multi-source data such as acoustic emission and microseismic data in a unified three-dimensional coordinate system, combined with the high-resolution hazard confidence field output by the physical information injection model, millimeter-level spatial positioning of potential fracture surfaces is achieved. The system automatically extracts connected blocks of the risk heat volume and calculates the minimum convex hull, providing intuitive spatial coordinate guidance for advanced support layout and directional pressure relief drilling. This elevates the traditional threshold alarm's "regional level" response to "point-level" proactive prevention and control, significantly reducing the disaster response time window.

[0020] By continuously monitoring the drift of the input data distribution and physical residual distribution, and combining this with a closed-loop incremental learning mechanism based on on-site execution feedback, the model can dynamically adapt to changes in working conditions such as stress field evolution and rock structure variations caused by mining operations. While protecting data privacy, the federated learning framework enables the collaborative accumulation of heterogeneous geological experience across mining areas, ensuring that the system can continuously optimize its early warning performance even after large-scale deployment, thus fundamentally extending the technology's lifecycle.

[0021] High-reliability early warnings significantly reduce the frequency of manual inspections and production losses caused by accidental shutdowns; interpretable prevention and control recommendations shorten the solution development cycle and reduce the safety risks of experts frequently going deep into high-risk roadways; and the minute-level response capability at the edge avoids the need for emergency repairs due to the expansion of disasters. From the three dimensions of early warning accuracy, manpower allocation, and disaster response, we achieve an overall reduction in mine safety operation and maintenance costs. Attached Figure Description

[0022] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0023] Figure 1 A flowchart of an early rockburst identification method based on physical information injection provided in this application embodiment;

[0024] Figure 2 This is a schematic diagram of the internal structure of an early rockburst identification device based on physical information injection, provided as an embodiment of this application. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0026] This application provides a method, device, and medium for early rockburst identification based on physical information injection, in order to solve the following technical problems: how to reduce the false alarm and missed alarm rates, and improve the timeliness of early warning and spatial positioning accuracy.

[0027] The technical solutions proposed in the embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0028] Figure 1 This document provides a flowchart for early rockburst identification based on physical information injection, as an embodiment of this application. Figure 1 As shown in the figure, the rockburst early identification method based on physical information injection provided in this application embodiment specifically includes the following steps:

[0029] Step 1: Collect acoustic emission waveform data, microseismic waveform data, triaxial stress data, and tunnel deformation and displacement data within the underground engineering area.

[0030] Step 1.1: Acoustic emission waveform data acquisition.

[0031] Transmitted waveform data: refers to the elastic wave signals released by micro-fractures in rock, used to capture early signs of minute cracks within the rock mass. High-frequency sensors (such as the PAC-SAMOS system) capture signals at a sampling rate of ≥1MHz with a dynamic range of 100dB, ensuring high sensitivity and interference resistance.

[0032] In one specific instance, PAC-SAMOS acoustic emission sensors were deployed at 50m intervals on the roof of the E1203 working face roadway in Mine A. The sensors were installed to avoid known fracture zones and were aimed at areas of intact rock strata to capture micro-fracture events. After deployment, the system operated at a sampling rate of 1.2MHz and a dynamic range of 100dB to ensure the capture of elastic waves released from microcracks as small as 0.1mm. Real-time data was transmitted via explosion-proof armored cables to prevent underground gas or dust from affecting signal integrity. This deployment successfully captured multiple micro-fracture events within 24 hours, providing initial input for the system.

[0033] Step 1.2: Microseismic waveform data acquisition.

[0034] Microseismic waveform data: refers to the vibration signals generated by macroscopic fracturing events in the rock mass, monitored by broadband microseismic stations, with a frequency response range of 10Hz–1kHz and a sensitivity of 0.1m / s. 2 It is used to identify large-scale rock mass instability.

[0035] In a specific example, in Mine A, broadband microseismic stations were deployed at the intersection of roadways 300m from the working face, with a station spacing of 250m, covering the entire high-risk area. Sensitivity calibration (0.1m / s) was performed before installation. 2 The frequency response range is fixed at 10Hz–1kHz to accommodate deep, high-pressure rock formations. After deployment, the system successfully monitored an energy release event (such as rock mass displacement), and the real-time data was uploaded via the industrial ring network. This embodiment focuses on optimizing station spacing to ensure blind-spot-free monitoring and provide reliable input for subsequent analysis.

[0036] Step 1.3: Triaxial stress data acquisition.

[0037] Triaxial stress data refers to the measured values ​​of stress variation along the X / Y / Z axes, which are synchronously acquired by a fiber optic grating stress meter with a range of 0–60 MPa and an accuracy of ±0.5% FS. It is used to quantify the stress state of rock mass.

[0038] In a specific example, fiber optic stress gauges were installed at the coal-rock interface in Mine Area A. The sensor range was set to 0–60 MPa, and the accuracy was calibrated to ±0.5%FS. The installation location was chosen in a lithologically stable area, with a burial depth of 2m to avoid interference from mining activities. The system synchronizes X / Y / Z triaxial data (e.g., maximum principal stress 25 MPa) in real time and transmits it via cable. This embodiment emphasizes installation accuracy, successfully capturing stress concentration trends in the initial deployment and providing key parameters for the model.

[0039] Step 1.4: Acquisition of tunnel deformation and displacement data.

[0040] Tunnel deformation and displacement data: refers to the amount of tunnel convergence deformation, recorded by MEMS displacement gauges with a range of ±300mm and a resolution of 0.01mm, used to monitor the stability of the support structure.

[0041] In a specific example, MEMS displacement gauges were installed at key nodes of U-shaped steel supports in Mine A. The range was set to ±300mm, and the resolution to 0.01mm. During installation, the gauges were aligned with high-stress areas in the roadway, and the support interfaces were reinforced. The system records deformation in real time (e.g., a daily convergence of 0.5mm), and the data is uploaded via cable. This embodiment demonstrates optimized installation location, successfully identifying early creep and providing a displacement benchmark for risk warning.

[0042] Step 1.5: Data access and real-time upload.

[0043] Explosion-proof armored cable: A type of protective cable used in underground environments, featuring explosion-proof and mechanical damage-resistant properties to ensure the safe transmission of sensor data.

[0044] Downhole Industrial Ring Network: A high-speed network infrastructure that supports 100Mbps data transmission rates and is used to aggregate sensor data to edge computing nodes in real time.

[0045] In a specific example, in Mine A, acoustic emission, micro-vibration, stress, and displacement sensors were connected to the underground industrial ring network via explosion-proof armored cables. The system uploads data at a rate of 100 Mbps, with edge nodes verifying the packet loss rate in real time (target <0.1%). After deployment, the data transmission success rate reached 99.9% in the initial test, with no interruptions. This embodiment focuses on network configuration, ensuring minute-level data availability and supporting real-time system alerts.

[0046] Step 2: Combine the collected acoustic emission waveform data, microseismic waveform data, triaxial stress data, and tunnel deformation and displacement data with static geological attribute data, perform time synchronization and spatial coordinate alignment, and generate high-dimensional feature cube data with a unified spatiotemporal benchmark; wherein, the static geological attribute data includes lithology, structural joints, and strike and dip information.

[0047] Step 2.1: Based on the preset IEEE-1588PTP precise time protocol, synchronize the time of the sensor nodes that collect the acoustic emission waveform data, micro-vibration waveform data, triaxial stress data and roadway deformation displacement data to obtain synchronization timestamp data.

[0048] IEEE-1588PTP Protocol: Precision Time Protocol, which achieves sub-millisecond time synchronization through master-slave clock message exchange, solving the problem of clock drift in sensor nodes.

[0049] Synchronization timestamp data: All sensor data carries a unified time tag (format: YYYYMMDD-HHMMSS.FFF), with a time error ≤0.3ms.

[0050] A PTP master clock is deployed in the underground industrial ring network, periodically broadcasting Sync messages to the acoustic emission, micro-vibration, stress, and displacement sensor nodes. Nodes calculate path delays and calibrate their local clocks to ensure that all data stream timestamps are aligned. The master clock generates a reference time stamp (e.g., 20250222-022230.123) for subsequent spatiotemporal mapping.

[0051] In a specific example, a PTP master clock is deployed in Mine A, with sensor nodes accessing via Gigabit Ethernet. The master clock broadcasts a Sync message every 2 seconds, and the timestamp error is controlled to ±0.25ms after node calibration. Acoustic emission waveform data packets are tagged as 20250723-092015.456, and microseismic event data is synchronized as 20250723-092015.458, achieving sub-millisecond alignment across devices.

[0052] Step 2.2: Establish the three-dimensional spatial coordinate framework data of the underground engineering area based on the preset RTK-SLAM algorithm and laser scanning point cloud data.

[0053] RTK-SLAM algorithm: It integrates real-time dynamic positioning (RTK, accuracy ±2cm) and simultaneous localization and mapping (SLAM), and constructs an absolute three-dimensional coordinate system by matching LiDAR point clouds.

[0054] Three-dimensional spatial coordinate frame data: a tunnel spatial model divided into 1m×1m×1m grids, with the origin of the coordinates being the geodetic reference point.

[0055] A laser SLAM vehicle scans along the tunnel to generate a high-density point cloud (≥5000 points / ㎡), and an RTK base station provides geodetic coordinate correction. The point cloud is registered to a unified coordinate system, and a 3D mesh framework with absolute coordinates is output to provide spatial anchors for monitoring data.

[0056] In a specific example, a 32-line lidar SLAM vehicle (traveling speed 2 km / h) was used in Mine A to scan the E1203 working face roadway and generate a point cloud model. An RTK reference station was set up at the intersection of the roadways to register the point cloud to the origin of the coordinate system (X0, Y0, Z0) and output a mesh frame (e.g., the coordinates of mesh G1205 are (120.5, 45.2, -305.3)).

[0057] Step 2.3: Perform spatiotemporal mapping on the acoustic emission waveform data, micro-vibration waveform data, triaxial stress data, and roadway deformation and displacement data based on the synchronization timestamp data and the three-dimensional spatial coordinate frame data to obtain monitoring point data with spatially aligned positions.

[0058] Spatially aligned monitoring point data: The result of mapping the raw data to 3D mesh nodes, with missing points filled by interpolation.

[0059] Acoustic emission events: Based on the sensor array, the seismic source is located according to the time difference and mapped to the nearest grid center.

[0060] Microseismic events: The seismic source is obtained by spherical inversion of the data from 8 microseismometers and projected onto the 3D grid nodes.

[0061] Stress / displacement data: Associated with the corresponding mesh according to the installation coordinates, and missing points are interpolated using inverse distance weighted (IDW).

[0062] In a specific case, the microseismic event in mining area A was located at (X: 120.7, Y: 45.3, Z: -305.1) by spherical inversion and mapped to grid G1205; data for the stress gauge installation point (X: 120.5, Y: 45.0, Z: -305.0) was missing and was supplemented by IDW interpolation of the adjacent 3 points.

[0063] Step 2.4: Perform sliding window filtering, noise suppression, and energy envelope quantization on the acoustic emission waveform data and microseismic waveform data at the edge computing node to obtain preprocessed waveform feature data.

[0064] Sliding window filtering: Divide the waveform into fixed windows (e.g., 10ms) with an overlap rate of 50% to extract local features.

[0065] Energy envelope quantization: Calculate the root mean square (RMS) value of the signal within the window and convert the unit to microjoules (μJ).

[0066] Edge nodes (such as NVIDIA Jetson Orin) perform the following on acoustic emission / micro-vibration waveforms:

[0067] Noise reduction: 5-level decomposition of db4 wavelet basis, hard thresholding to filter out high-frequency noise <30dB;

[0068] Energy quantification: Calculate the RMS value for each window and output the energy time series curve.

[0069] In a specific example, the edge node divides the acoustic emission waveform into 10ms windows, filters out noise above 32kHz using db4 wavelets, and outputs an RMS energy sequence (e.g., energy value = 120μJ at t=092015.456).

[0070] Step 2.5: Perform adaptive Kalman filtering to suppress noise and linear interpolation to compensate for missing values ​​on the triaxial stress data and roadway deformation displacement data to obtain preprocessed stress displacement characteristic data.

[0071] Adaptive Kalman filtering: dynamically adjusts noise covariance to suppress temperature drift and transient interference.

[0072] Linear interpolation compensation: For missing points caused by network fluctuations, fill them linearly with adjacent data.

[0073] Stress and displacement sequences are filtered by Kalman filtering to suppress temperature drift, and missing segments (such as those caused by packet loss) are interpolated linearly over time to ensure data continuity.

[0074] In a specific case, stress data for mining area A was lost for 5 seconds due to a network interruption. The system supplemented the data by linear interpolation based on the previous value of 25.3 MPa and the subsequent value of 25.8 MPa. After filtering, the fluctuation range decreased from ±0.8 MPa to ±0.1 MPa.

[0075] Step 2.6: Organize the preprocessed waveform feature data, the preprocessed stress displacement feature data, and the static geological attribute data into high-dimensional feature cube data indexed by time dimension, spatial dimension, and feature dimension, according to a preset time window and spatial resolution.

[0076] High-dimensional feature cube data: a three-dimensional tensor with time (T), space (S), and feature (F) as axes (e.g., size T×S×F=100×500×20).

[0077] Time window: 10s window, 1s step sliding package;

[0078] Spatial resolution: 1m 3 The grid is the smallest unit;

[0079] Feature dimensions include waveform energy, stress components, displacement, and lithological coding.

[0080] Data is serialized into Apache Arrow format, supporting zero-copy loading on the GPU.

[0081] In a specific example, mining area A organizes the data within a 10-second window into a cube:

[0082] Timeline: 100 steps (10s ÷ 0.1s);

[0083] Spatial axis: 500 grids (500m tunnel);

[0084] Characteristic axes: acoustic emission energy, triaxial stress, displacement, sandstone lithology label (code=3).

[0085] Stored in Arrow format for use by the P-Former model.

[0086] Step 3: Input the high-dimensional feature cube data into a preset physical information injection converter for modeling to obtain intermediate feature vector data containing energy accumulation rate, stress concentration coefficient and fracture evolution potential, as well as three-dimensional hazard confidence distribution data.

[0087] Step 3.1: Perform frequency domain transformation on the high-dimensional feature cube data to obtain frequency domain feature data.

[0088] Frequency domain feature data: The time domain signal is converted into a frequency domain representation by Fast Fourier Transform (FFT) or Short Time Fourier Transform (STFT) to highlight frequency features (such as energy distribution) and make it easier for the model to capture periodic or transient events.

[0089] Frequency domain transformation is performed on waveform data (acoustic emission, microseismic activity) of high-dimensional feature cubes. FFT is used for global spectral analysis, and STFT is used for sliding-window local spectral analysis, outputting frequency domain tensors (such as amplitude and phase spectra). This helps identify high-frequency microfractures or low-frequency energy accumulation patterns, providing input for attention mechanisms.

[0090] In a specific example, STFT was applied to acoustic emission waveform data (cube characteristic axis) in mine A: the signal was segmented with a 256-point window and 50% overlap, the frequency domain amplitude spectrum was calculated, and a 20-dimensional characteristic vector with a frequency range of 0–64 kHz was output. This successfully captured the energy peak in the 15 kHz frequency band, indicating early fracture activity.

[0091] Step 3.2: Input the frequency domain feature data into the physical information injection into the multi-head self-attention layer of the converter.

[0092] Multi-head self-attention layer: The core component of Transformer, which calculates the association weights of the input sequence through multiple parallel attention heads (such as 8 heads) to enhance the feature interaction capability.

[0093] Frequency domain feature data is directly input into the P-Former's multi-head self-attention layer. Each attention head independently calculates the similarity of the query, key, and value vectors, initially modeling the dependencies between features (such as the temporal correlation between stress and energy). The input dimensions need to be aligned with the model (e.g., feature axis normalization) to ensure computational efficiency.

[0094] In a specific example, mine A inputs frequency domain features (size 100×500×128) into an 8-head self-attention layer. Each head processes a different frequency subspace (e.g., head 1 focuses on 0–10kHz), outputting preliminary coupling features to prepare for physical injection.

[0095] Step 3.3: When calculating the similarity between the query vector and the key vector for each attention head in the multi-head self-attention layer, the rock mass elastoplastic constitutive residual data and the in-situ geostress tensor data are explicitly introduced as penalty terms to calculate attention weight data that conforms to mechanical constraints; wherein, the rock mass elastoplastic constitutive residual data is obtained by calculating the difference between the predicted stress and the actual stress in real time based on the constitutive equation obtained from field measurements or experiments.

[0096] Rock mass elastoplastic constitutive residual data: the difference between predicted stress and actual stress calculated based on constitutive equations (such as the Hoek-Brown model), reflecting deviations from mechanical laws.

[0097] In-situ stress tensor data: Three-dimensional stress state measured on-site.

[0098] Penalty term: A constraint term added to the attention energy function to force the model to satisfy energy conservation and volume compatibility.

[0099] In a specific example, the constitutive residuals of the sandstone region in mining area A were calculated: predicted stress 25.3 MPa, measured stress 25.8 MPa, residual Ra. ij =0.5MPa; geostress tensor S ij σ z =20MPa. Let λ1=0.3, λ2=0.2, and add them to the attention weight calculation. Result: The model ignores noise interference and focuses on real energy accumulation events.

[0100] Step 3.4: Based on the calculated attention weight data, perform weighted aggregation on the value vector to obtain deep feature coupling data.

[0101] Deep feature coupling data: High-dimensional fusion of features across time and space is achieved by weighting and summing the value vectors using attention weights.

[0102] Based on the attention weights from step 3.3, the Value vector is weighted and aggregated. Features with high weights (such as high stress concentration areas) are enhanced, while those with low weights (such as noise) are suppressed. The output tensor integrates multi-source information (waveform, stress, static properties) to form a unified deep representation.

[0103] In a specific case, the aggregated features of mining area A enhanced the correlation between microseismic events and stress concentration (weight > 0.8), and the output coupled data was used for expert routing.

[0104] Step 3.5: Input the deep feature coupling data into the multi-expert gating subnetwork of the physical information injection converter.

[0105] Multi-expert gated subnetwork: This structure includes multiple expert subnetworks (such as hard rock experts and soft rock experts) and a gated network, dynamically adapting to different geological conditions.

[0106] A deep feature-coupled data input gating subnetwork is used. The gating network selects the most relevant expert based on input attributes (such as lithology) to resolve cross-scale differences (such as the brittleness of hard rock vs. the plasticity of soft rock). The expert subnetworks are trained independently to handle specific patterns.

[0107] In a specific example, three experts are assigned to Mine A: Expert 1 (hard rock), Expert 2 (soft rock), and Expert 3 (jointed rock). Coupled data is first input into a gating network to prepare for routing.

[0108] Step 3.6: Based on the lithological properties of the input data, the multi-expert gated subnetwork dynamically routes the deep feature coupling data to the corresponding expert subnetwork for processing to obtain the intermediate feature vector data and the three-dimensional hazard confidence distribution data.

[0109] Step 3.6.1: Input the depth feature coupling data and the lithological label data in the static geological attribute data into the gating network of the multi-expert gating sub-network.

[0110] Gated network: A neural network that learns the coupling patterns between features and lithology, and outputs expert-selected weights.

[0111] Deep feature-coupled data and lithological labels (from static attributes) are input into a gated network. The network learns mapping relationships (e.g., sandstone → high-energy accumulation mode) and generates weight vectors.

[0112] In a specific case, the input sandstone label (code=3) and coupling features for mining area A are given, and the gating network outputs weights [0.7, 0.2, 0.1], indicating that expert 1 is dominant.

[0113] Step 3.6.2: Based on the gated network, learn the coupling pattern between the deep feature coupling data and different lithological types, and generate expert selection weight data.

[0114] Expert-selected weighted data: The probability distribution output by the gating network indicates which expert the input data should be routed to.

[0115] The gated network generates weights using a softmax function, ensuring the sum is 1. The weights are based on feature-lithology correlations (e.g., high joint density routing to expert 3).

[0116] In a specific case, the gated network in Mine A learned the "high stress concentration + sandstone" pattern, outputting weights [0.85, 0.10, 0.05], and favoring expert 1.

[0117] Step 3.6.3: Based on the expert selection weight data, dynamically route the deep feature coupling data to one or more of the best-matching expert subnetworks.

[0118] Dynamic routing: Distributing data to one or more expert subnetworks based on weights.

[0119] Experts are selected based on their weights to determine the routing path (e.g., experts with a weight > 0.5 are selected). Data is distributed in parallel, and expert subnetworks process independently, outputting local features.

[0120] In a specific case, the weight of mining area A is [0.75, 0.20, 0.05], and the data is routed to expert 1 (hard rock expert) to handle fracture evolution potential.

[0121] Step 3.6.4: Based on the selected expert subnetwork, process the routed deep feature coupling data and output the intermediate feature vector data and the three-dimensional hazard confidence distribution data.

[0122] Expert subnetworks: specialized neural networks (such as CNNs or MLPs) that process features under specific geological conditions.

[0123] The selected expert subnetwork processes the input data and outputs intermediate feature vectors (such as energy accumulation rate) and a three-dimensional hazard confidence distribution (grid-level probability map). The confidence threshold is used for subsequent risk assessment.

[0124] In a specific case, Expert 1 of Mine A outputs: energy accumulation rate 0.92 (high risk), stress concentration factor 1.5, fracture evolution potential 0.8, and three-dimensional confidence distribution (e.g., confidence of grid G1205 = 0.95).

[0125] Step 4: Based on the intermediate feature vector data, retrieve the preset rockburst case knowledge base to obtain similar historical case data, and then concatenate and fuse the similar historical case data with the intermediate feature vector data to generate natural language text data containing explanations of potential triggering mechanisms and suggestions for prevention and control measures.

[0126] Step 4.1: Based on the intermediate feature vector data, perform a first-stage coarse-grained vector similarity retrieval in the rockburst case knowledge base to obtain a preliminary matching candidate case dataset.

[0127] Coarse-grained vector similarity retrieval: Use approximate nearest neighbor algorithms (such as FAISS) to quickly filter candidate cases and reduce the computational overhead of searching the entire database.

[0128] Candidate case dataset: A set of Top-N historical cases that have been initially matched (N≥10), with a similarity threshold >0.6.

[0129] Intermediate feature vectors (such as energy accumulation rate and stress concentration coefficient) are input into the FAISSGPU index library, and cosine similarity with all case vectors in the knowledge base is calculated. Top-N candidate cases are returned to ensure coverage of potentially similar scenarios.

[0130] In a specific case, the A mining area input the 128-dimensional intermediate feature vector into the FAISS index and retrieved the Top-15 candidate cases within 3 milliseconds (the highest similarity was 0.82, corresponding to the rockburst event at the E1201 working face in 2023).

[0131] Step 4.2: Perform a second-stage fine-grained search on the preliminary matched candidate case dataset. The fine-grained search is based on hierarchical matching calculation using static attribute vector similarity and dynamic sequence vector similarity to filter out target case data with high similarity. The static attribute vector includes lithology, strike, and joint density, and the dynamic sequence vector includes energy decay curve and fracture evolution trajectory.

[0132] Static attribute vector: Discrete geological feature encoding of the case (lithology: sandstone=3, strike: N30°E, joint density: 5 joints / m) 3 ).

[0133] Dynamic sequence vector: temporal characteristics of the case (energy decay curve: RMS value sequence, fracture evolution trajectory: displacement change sequence).

[0134] Hierarchical matching: First, calculate the cosine similarity of static attributes, and then evaluate the consistency of dynamic sequences through DTW (Dynamic Time Warping) distance.

[0135] Static matching: Filter cases with attribute similarity > 0.85;

[0136] Dynamic matching: Calculate the DTW distance for the remaining cases (the smaller the value, the more similar they are), and select the Top-K cases with the smallest distance (K≤5).

[0137] In a specific example, Mine A implemented the following for the Top-15 candidate cases:

[0138] Static filtering: Cases with joint density differences >10% were removed, leaving 8 cases;

[0139] Dynamic matching: The current fracture evolution trajectory has the smallest DTW distance (0.15) with case ID-202405, and is selected as the Top-1 target case.

[0140] Step 4.3: Concatenate the feature representations of the selected target case data with the intermediate feature vector data to form knowledge-enhanced feature input data.

[0141] Knowledge-enhanced feature input data: the result of concatenating the target case features and intermediate feature vectors, with the dimension expanded to more than 1.5 times that of the original features.

[0142] The static attribute vector (e.g., 128-dimensional) and dynamic sequence vector (e.g., 64-dimensional) of the Top-K target cases are concatenated to the end of the original intermediate feature vector to form a high-dimensional fusion input (e.g., 128+192=320-dimensional).

[0143] In a specific example, Mine A concatenates the Top-3 case vectors (64-dimensional × 3) with the original vector (128-dimensional) to generate a 320-dimensional feature input, which is then injected into the historical case context.

[0144] Step 4.4: Input the knowledge-enhanced feature input data into the language model that has been fine-tuned by instructions.

[0145] Instruction-based fine-tuned language large model: an LLM fine-tuned for natural language tasks (such as text generation), which maps feature vectors to the text space through an adaptation layer.

[0146] Knowledge-enhanced features are transformed into language model embedding dimensions (e.g., 768 dimensions) via a linear projection layer, and input prefix instructions (e.g., "Generate rockburst mechanism explanation:") guide the generation direction.

[0147] In a specific example, Mine A uses the LLaMA-7B model, adds an adaptation layer to map 320-dimensional features to 768-dimensional features, and inputs the command: "Analyze the causes of rockbursts based on the following features and provide measures:".

[0148] Step 4.5: Process the knowledge-enhanced feature input data based on the language big model to automatically generate natural language text data containing explanations of potential triggering mechanisms and suggestions for prevention and control measures.

[0149] Natural language text data: Structured text containing inducing mechanisms (such as "lateral structural stress dominance") and control measures (such as "φ75mm pressure relief hole").

[0150] The language model decodes and generates text, parsing key fields:

[0151] Inducing mechanism: Quantitative attribution (e.g., stress ratio);

[0152] Prevention and control measures: specific parameters (aperture diameter, spacing, support specifications);

[0153] Example: Associate with historical event IDs.

[0154] In a specific example, the output text for mine A is:

[0155] "Inducing mechanism: Lateral tectonic stress (accounting for 68%) superimposed on the energy accumulation of brittle sandstone; Prevention measures: Reinforce anchor cables (22mm diameter, 1.2m spacing) and construct pressure relief holes (φ75mm×6m, 5m×5m spacing). Reference case: 202405-E0502."

[0156] Step 5: Generate a three-dimensional risk heat map based on the three-dimensional hazard confidence distribution data, extract the spatial location data of potential rupture surfaces based on the three-dimensional risk heat map data, and visualize the three-dimensional risk heat map data, the spatial location data of potential rupture surfaces, and the natural language text data.

[0157] Step 5.1: Map the three-dimensional hazard confidence distribution data onto the spatial voxels corresponding to the three-dimensional tunnel geographic information model to generate initial three-dimensional risk heat map data.

[0158] 3D tunnel geographic information model: A 3D digital twin of the tunnel based on the integration of BIM (Building Information Modeling) and GIS (Geographic Information System), including spatial information such as support structure and rock strata boundaries.

[0159] Spatial voxel: A 1m×1m×1m cubic grid cell, serving as the basic unit for risk data mapping.

[0160] Initial 3D risk heatmap data: The risk confidence level (0.0-1.0) is mapped to the gradient color space distribution generated by the corresponding voxels, with red indicating high risk (>0.8) and blue indicating safe (<0.3).

[0161] Align the 3D hazard confidence distribution data (grid-level probability) output by P-Former with the spatial voxels of the tunnel geographic information model. Assign RGB color to each voxel based on its confidence value:

[0162] Low risk (0.0-0.3): Blue tones;

[0163] Medium risk (0.3-0.8): Yellow tones;

[0164] High risk (0.8-1.0): Red series;

[0165] Generate an initial heatmap for cluster analysis.

[0166] In a specific case, in mining area A, grid G1205 with a confidence level of 0.95 is mapped to RGB(255, 0, 0), and grid G1210 with a confidence level of 0.4 is mapped to RGB(255, 255, 0), forming an initial risk heat map.

[0167] Step 5.2: Perform a clustering analysis algorithm based on density and spatial connectivity on the initial three-dimensional risk heat map data to identify high-risk connected regions.

[0168] Density and Spatial Connectivity Clustering: An improvement to the DBSCAN algorithm that combines voxel density thresholds (e.g., 5 adjacent voxels) and spatial connectivity rules (shared faces / edges / vertices) to identify physically continuous high-risk blocks.

[0169] High-risk connected region data: a collection of spatial regions output by clustering, each region containing ≥10 high-risk voxels and satisfying geometric connectivity.

[0170] Perform the following operations on the initial heatmap:

[0171] Density filtering: Only voxels with a confidence level > 0.8 are retained;

[0172] Connectivity detection: Scan 26 neighborhoods (up, down, left, right, front, back + diagonal) to mark connected components;

[0173] Area merging: If the distance between two areas is less than 2m, they are merged into the same high-risk area.

[0174] Output the region ID and its list of voxels.

[0175] In one specific case, three high-risk connectivity zones were identified in mining area A:

[0176] Region ID-R1: Contains voxels {G1205, G1206, G1210} (volume 12m³) 3 );

[0177] Region ID-R2: Contains voxels {G1301, G1302} (volume 8m³) 3 ).

[0178] Step 5.3: Calculate the minimum convex hull geometry data for each of the identified high-risk connected regions, whereby the minimum convex hull geometry data represents the spatial location data of the potential rupture surface.

[0179] Minimal convex hull geometry: The minimum convex polyhedron that encloses all voxels of a high-risk connected region, and its surface represents the spatial location of potential rupture surfaces.

[0180] Potential fracture surface spatial location data: a set of three-dimensional coordinates of the convex hull vertices (e.g., [X1, Y1, Z1], [X2, Y2, Z2]...), which defines the fracture surface boundary.

[0181] Perform the following for each high-risk connected region:

[0182] Extract the coordinates of the center points of all voxels in the region;

[0183] Calculate the minimum convex hull using the Quickhull algorithm;

[0184] Output the convex hull vertex coordinate sequence and the topology of the triangular facet.

[0185] The convex hull surface is the potential energy release fracture surface.

[0186] In a specific example, mining area A calculates the convex hull for region ID-R1:

[0187] Vertex coordinates: [(120.5, 45.2, -305.3), (120.8, 45.5, -305.1), (121.0, 45.0, -305.5)];

[0188] Fracturing surface area: 15.2m² 2 .

[0189] Step 5.4: Render and display the three-dimensional tunnel scene superimposed with the initial three-dimensional risk heat map data, the convex hull boundary of the high-risk connected area data, and the natural language text data on the three-dimensional visualization platform.

[0190] Natural language text data: Explanation of the inducing mechanism and prevention and control measures generated in step 4 (e.g., "lateral structural stress is dominant, and φ75mm pressure relief hole is recommended").

[0191] Implement four-layer overlay rendering on Three.js or CesiumJS platforms:

[0192] Basic scenario: Lane BIM-GIS model (gray semi-transparent);

[0193] Risk popularity: The initial popularity map is rendered with a semi-transparent body (red→blue gradient);

[0194] Breach boundary: Line connecting the vertices of the convex hull (thick green solid line);

[0195] Text annotation: Floating knowledge cards associated with high-risk areas.

[0196] Users can explore details through cross-sectioning and rotation.

[0197] In a specific example, the dispatch screen in Mine A displays:

[0198] Region ID-R1: Red semi-transparent thermal volume + green convex hull boundary;

[0199] Click to bring up the text: "Inducing mechanism: transverse stress (68%) + brittle sandstone; Measures: φ75mm pressure relief holes spaced 5m apart";

[0200] Supports Z-axis sectioning for observation of deep stress distribution.

[0201] Step 6: Perform risk assessment based on the three-dimensional hazard confidence distribution data. When the risk level reaches a preset threshold, trigger an edge alarm signal and record the on-site implementation feedback data for the proposed prevention and control measures.

[0202] Step 6.1: Real-time assessment of the risk level of the three-dimensional hazard confidence distribution data within the continuous time window.

[0203] Risk level: Warning levels are divided according to confidence level (Level I: 0.6-0.7, Level II: 0.7-0.8, Level III: >0.8), and each level corresponds to a different response strategy.

[0204] Time window: The system defaults to 10 seconds as the evaluation unit, continuously monitoring the trend of confidence level changes.

[0205] Perform the following on the confidence data for each spatial grid:

[0206] Instantaneous rating: Match the risk level to the current value (e.g., 0.85 → Level III);

[0207] Trend analysis: Calculate the confidence slope for three consecutive windows (e.g., >0.02 / window decision acceleration risk).

[0208] In a specific case, the confidence level of grid G1205 in mining area A for three consecutive windows was 0.82→0.85→0.88, with a slope of 0.03 / window, which was determined to be a Level III acceleration risk.

[0209] Step 6.2: When it is detected that the risk level of a predetermined number of consecutive time windows has reached or exceeded the preset risk threshold level, it is determined that the alarm triggering condition is met.

[0210] Continuous pre-order quantity: The system's preset continuous alarm threshold (e.g., Level III requires 2 consecutive windows ≥ 0.8).

[0211] Dynamically detect two conditions:

[0212] Current level: Target grid confidence level ≥ threshold (e.g., Level III > 0.8);

[0213] Duration: The condition is met for N consecutive windows (N is configurable).

[0214] An alarm will be triggered if both conditions are met simultaneously.

[0215] In a specific case, setting a Level III alarm in Mine Area A requires two consecutive windows with a confidence level ≥ 0.8. When the confidence level of grid G1301 changes from 0.83 to 0.86 for two consecutive windows, the trigger condition is met.

[0216] Step 6.3: When the alarm triggering conditions are met, immediately send the edge alarm signal to the edge alarm device deployed at the roadway site to trigger the audible and visual alarm.

[0217] Edge alarm device: Hardware equipment (buzzer + tri-color LED) deployed on-site in the alleyway, supporting audible and visual alarms (red light flashing + 105dB buzzer).

[0218] Alarm signals are transmitted via the underground industrial ring network in milliseconds:

[0219] Equipment location: Signal-associated high-risk grid coordinates (e.g., G1205);

[0220] Graded response: Level III triggers full-power audio and visual effects, while Level II only flashes a yellow light.

[0221] In a specific case, G1205 in mining area A triggered a Level III alarm. The nearest edge device (coordinates 120.5, 45.2, -305.3) flashed a red light at a high frequency and the buzzer sounded every 3 seconds.

[0222] Step 6.4: An alarm information window will automatically pop up on the central dispatch platform. The alarm information window includes a high-risk area identifier, the prevention and control measures recommendations, and key explanations in the natural language text data.

[0223] Alarm Information Window: An interactive interface that automatically pops up on the dispatch platform, containing 3D positioning, prevention and control suggestions, and mechanism explanations.

[0224] The window content is presented in a structured manner, for example:

[0225] [High-risk area] E12-3# (G1205-G1210);

[0226] [Confidence Level] 0.88 (Level III Acceleration);

[0227]

Prevention and Control Measures

[0228] [Mechanism] Lateral tectonic stress dominates (70%);

[0229] In a specific example, a pop-up window in mining area A displays: "Area ID-R1 | Measures: Anchor cable reinforcement + pressure relief holes | Mechanism: Energy accumulation in brittle sandstone".

[0230] Step 6.5: Collect and record the types of operations performed by on-site personnel according to the prevention and control measures recommendations, the operation time, and the post-operation regional stability observation results to form the execution feedback data.

[0231] Execution feedback data: structured record fields (operation type, timestamp, stability metrics).

[0232] On-site personnel recorded the information using an explosion-proof mobile app:

[0233] Operation type: Predefined options (anchor cable reinforcement / pressure relief drilling / shotcrete support);

[0234] Effect observation: Select stability change (significant improvement / partial improvement / no change) from the drop-down menu.

[0235] Data is automatically synchronized to the central platform.

[0236] In a specific case, Mine A recorded: "20250723-0920 | 6 pressure relief holes @ 5m spacing | Energy decreased by 60% after 2 hours | Significant improvement".

[0237] Step 7: Monitor the input distribution characteristics of the high-dimensional feature cube data and the physical residual distribution characteristics generated by the modeling process, determine the model drift state based on the monitoring results, and trigger the incremental fine-tuning process of the model according to the model drift state and the execution feedback data.

[0238] Step 7.1: Calculate the first distribution difference measure data between the input distribution characteristics of the current high-dimensional feature cube data and the historical benchmark input distribution characteristics in real time.

[0239] Input distribution characteristics of high-dimensional feature cube data: refers to the statistical attributes (such as mean and variance) of the feature cube data generated in step 2.6, which are used to describe the distribution of real-time monitoring data.

[0240] Historical baseline input distribution characteristics: The baseline values ​​of data distribution recorded by the system during stable operation (such as after initial training) serve as a comparison reference.

[0241] First distribution difference measure data: an indicator that quantifies the difference between the current input distribution and the historical baseline. It is usually calculated using the Kullback-Leibler (KL) divergence. The larger the value, the more significant the difference.

[0242] Extract distribution features (such as the mean and standard deviation of each feature channel) from high-dimensional feature cube data (e.g., tensors of time × space × feature dimensions) in real time. Calculate the KL divergence by comparing it with historical benchmarks (distribution parameters stored in the database).

[0243] The KL divergence formula is simplified to the difference value (0 indicates no difference, >0.2 indicates significant drift).

[0244] The system updates the difference value every 10 minutes to ensure real-time performance.

[0245] This helps identify changes in the distribution of input data (such as new rock layers or sensor drift) and prevents the model from failing due to data offset.

[0246] In a specific example, mining area A calculates the KL divergence between the current input distribution (feature cube mean = 25.3, standard deviation = 5.1) and the historical baseline (mean = 24.8, standard deviation = 4.9). The difference is 0.18, which is below the threshold of 0.2, so no alarm is triggered. The example focuses on the KL divergence calculation process.

[0247] Step 7.2: Calculate in real time the second distribution difference measure data between the physical residual distribution characteristics generated by the current modeling process and the expected physical residual distribution characteristics.

[0248] The physical residual distribution characteristics generated by modeling: refers to the statistical properties (such as the distribution pattern of residual values) of the rock mass elastoplastic constitutive residuals in the output of the P-Former model (step 3), reflecting whether the model conforms to the laws of mechanics.

[0249] Expected physical residual distribution characteristics: The ideal residual distribution defined during the model training phase (such as a Gaussian distribution with a mean close to 0) indicates that the model strictly satisfies the physical constraints.

[0250] The second distribution difference measure is an index that quantifies the difference between the current residual distribution and the expected distribution. It is also calculated using KL divergence or Jensen-Shannon distance.

[0251] Monitor the physical residuals generated during P-Former inference (such as the difference between constitutive equation predictions and actual stresses). Extract the current residual distribution features (such as histogram statistics of the residual sequence) and compare them with the expected distribution (parameters saved during training).

[0252] Calculate the difference index to ensure that the model reasoning process always follows the laws of rock mechanics.

[0253] Excessively high discrepancy values ​​indicate that the physical constraints of the model have failed (e.g., inaccurate prediction of crack evolution).

[0254] Combined with step 7.1, this provides a dual drift detection mechanism.

[0255] In a specific example, the Jensen-Shannon distance is calculated between the current residual distribution (mean = 0.5 MPa, variance = 0.2) and the expected distribution (mean = 0.1 MPa, variance = 0.1) of mine area A. The difference is 0.25, exceeding the threshold of 0.2. This example focuses on a single operating point for comparing residual distributions.

[0256] Step 7.3: Calculate the comprehensive drift index data based on the first distribution difference measurement data and the second distribution difference measurement data.

[0257] Comprehensive drift index data: A weighted composite value (such as a weighted average) combining the differences between the first and second distributions, used to comprehensively assess the model's drift status. The formula is: Comprehensive index = w1 × first difference + w2 × second difference, with weights w1 and w2 set empirically (e.g., w1 = 0.6, w2 = 0.4).

[0258] The system automatically performs weighted summation of the first distribution difference measure (input data drift) and the second distribution difference measure (physical residual drift):

[0259] The weights are calibrated based on historical data (e.g., if the input data drift has a greater impact on the model, the weight is higher).

[0260] Output a comprehensive drift index (range 0-1), with a value >0.3 indicating high-risk drift.

[0261] This provides a global drift view, avoiding misjudgments based on a single metric.

[0262] In a specific example, mining area A has weights set as w1=0.6 and w2=0.4. The first difference is 0.18 and the second difference is 0.25. The comprehensive index is calculated as 0.6×0.18+0.4×0.25=0.208, which is lower than the threshold of 0.3.

[0263] Step 7.4: When the comprehensive drift index data continuously exceeds the preset drift threshold for a predetermined time, it is determined that model drift has occurred, triggering model degradation operation, automatically switching to a historical stable model version for inference, and generating model aging alarm information.

[0264] Preset drift threshold: The upper limit of drift tolerance defined by the system (e.g., comprehensive index > 0.3), determined based on historical tests.

[0265] Preset duration: The continuous time exceeding the threshold (e.g., 1 hour) to avoid accidental triggering due to instantaneous fluctuations.

[0266] Model downgrade operation: Automatically switch to a historical stable model version (such as the backup from last month) to ensure the system continues to run.

[0267] Model aging alarm information: Structured alarm messages sent to the operation and maintenance platform (such as alarm code "MODEL-DRIFT-ALERT").

[0268] Monitoring overall drift indicators:

[0269] If the threshold (e.g., 0.3) is exceeded for one consecutive hour, the model is considered to be drifting.

[0270] The system immediately switches to a historical stable model (such as an ONNX format backup) and sends an alarm message.

[0271] During the downgrade period, the backup model is used for inference to prevent risk warnings from failing.

[0272] This ensures high system availability and conforms to the "drift detector" design in the technical solution.

[0273] In a specific case, the comprehensive index of mining area A was >0.35 for 1 consecutive hour, triggering a downgrade: switching to the historical model version V2.1 and sending the alarm code "DRIFT-ALERT-001" to the operation and maintenance terminal.

[0274] Step 7.5: Collect the execution feedback data, model prediction error data, and actual alarm event data, and label them to form an incremental training sample dataset.

[0275] Execution feedback data: The operation type, timestamp, and stability results recorded in step 6.5 (e.g., "Energy decreased by 60% after pressure relief hole construction").

[0276] Model prediction error data: Quantified values ​​of the difference between the model output and the actual event (such as a rock eruption).

[0277] Actual alarm event data: Records details of the actual events that triggered the alarm.

[0278] Incremental training sample dataset: The labeled dataset, which includes input features and target labels (such as corrected risk confidence).

[0279] Collect and integrate three data sources:

[0280] Implementation feedback: The effectiveness of the measures input by on-site personnel through the APP.

[0281] Prediction error: Compare the model output with the actual monitoring data.

[0282] Alarm events: Store alarm trigger logs.

[0283] The labeled dataset is used for incremental training to ensure that the samples cover new scenarios.

[0284] In a specific example, mining area A collected data: feedback "stress reduction of 30% after anchor cable reinforcement", prediction error 0.1, alarm event ID-2025. After annotation, a sample set was formed containing 100 new entries. This example focuses on the data collection and annotation process.

[0285] Step 7.6: Based on the incremental training sample dataset, when a model performance degradation is detected or a periodic update instruction is received, the incremental fine-tuning training process of the physical information injection transformer is triggered to update the model parameters.

[0286] Incremental fine-tuning training process: Based on the existing model weights, use a new dataset for a limited number of training rounds (e.g., 10 rounds) to update the parameters without retraining the entire model.

[0287] Model performance degradation: assessed by validation set accuracy or residual metrics (e.g., accuracy degradation > 5%).

[0288] Regular update instructions: The system's preset timed trigger mechanism (e.g., once a week).

[0289] When a performance degradation is detected or an instruction is received:

[0290] Load the incremental dataset and fine-tune P-Former on the GPU server.

[0291] Optimize the multi-objective loss function (such as reconstruction error plus physical residual penalty).

[0292] The updated model is deployed back to edge nodes to support federated learning sharing.

[0293] In a specific example, the detection and verification accuracy of mining area A decreased by 7%, triggering fine-tuning: P-Former was trained using an incremental dataset (200 samples), the model was updated after 10 rounds, and federated learning was synchronized to other mining areas. The example focuses on performance detection and fine-tuning triggering.

[0294] The above are embodiments of the method proposed in this application. Based on the same inventive concept, embodiments of this application also provide an early rockburst identification device based on physical information injection, the structure of which is as follows: Figure 2 As shown.

[0295] Figure 2 This is a schematic diagram of the internal structure of a rockburst early identification device based on physical information injection, provided as an embodiment of this application. Figure 2 As shown, the device includes:

[0296] At least one processor 201;

[0297] And a memory 202 that is communicatively connected to at least one processor;

[0298] The memory 202 stores instructions executable by at least one processor, which are executed by at least one processor 201 to enable at least one processor 201 to:

[0299] Acoustic emission waveform data, microseismic waveform data, triaxial stress data, and tunnel deformation and displacement data are collected within the underground engineering area. These data are then combined with static geological attribute data, and time synchronization and spatial coordinate alignment are performed to generate high-dimensional feature cube data with a unified spatiotemporal reference. The static geological attribute data includes lithology, structural joints, and strike / dip information. The high-dimensional feature cube data is input into a preset physical information injection transformer for modeling, obtaining intermediate feature vector data containing energy accumulation rate, stress concentration coefficient, and fracture evolution potential, as well as three-dimensional hazard confidence distribution data. Based on the intermediate feature vector data, a preset rockburst case knowledge base is retrieved to obtain similar historical case data, and this similar historical case data is then compared with the intermediate feature vector data. The data is spliced ​​and fused to generate natural language text data containing explanations of potential inducing mechanisms and suggestions for prevention and control measures. A three-dimensional risk heat map is generated based on the three-dimensional hazard confidence distribution data. Spatial location data of potential rupture surfaces is extracted from the three-dimensional risk heat map data, and the three-dimensional risk heat map data, potential rupture surface spatial location data, and the natural language text data are then visualized. Risk assessment is performed based on the three-dimensional hazard confidence distribution data. When the risk level reaches a preset threshold, an edge alarm signal is triggered, and on-site feedback data on the implementation of the prevention and control measures suggestions is recorded. The input distribution characteristics of the high-dimensional feature cube data and the physical residual distribution characteristics generated by modeling processing are monitored. Based on the monitoring results, the model drift state is determined, and an incremental fine-tuning process for the model is triggered according to the model drift state and the implementation feedback data.

[0300] Some embodiments of this application provide corresponding to Figure 1 A non-volatile computer storage medium for early rockburst identification based on physical information injection, storing computer-executable instructions, wherein the computer-executable instructions are configured as follows:

[0301] Acoustic emission waveform data, microseismic waveform data, triaxial stress data, and tunnel deformation and displacement data are collected within the underground engineering area. These data are then combined with static geological attribute data, and time synchronization and spatial coordinate alignment are performed to generate high-dimensional feature cube data with a unified spatiotemporal reference. The static geological attribute data includes lithology, structural joints, and strike / dip information. The high-dimensional feature cube data is input into a preset physical information injection transformer for modeling, obtaining intermediate feature vector data containing energy accumulation rate, stress concentration coefficient, and fracture evolution potential, as well as three-dimensional hazard confidence distribution data. Based on the intermediate feature vector data, a preset rockburst case knowledge base is retrieved to obtain similar historical case data, and this similar historical case data is then compared with the intermediate feature vector data. The data is spliced ​​and fused to generate natural language text data containing explanations of potential inducing mechanisms and suggestions for prevention and control measures. A three-dimensional risk heat map is generated based on the three-dimensional hazard confidence distribution data. Spatial location data of potential rupture surfaces is extracted from the three-dimensional risk heat map data, and the three-dimensional risk heat map data, potential rupture surface spatial location data, and the natural language text data are then visualized. Risk assessment is performed based on the three-dimensional hazard confidence distribution data. When the risk level reaches a preset threshold, an edge alarm signal is triggered, and on-site feedback data on the implementation of the prevention and control measures suggestions is recorded. The input distribution characteristics of the high-dimensional feature cube data and the physical residual distribution characteristics generated by modeling processing are monitored. Based on the monitoring results, the model drift state is determined, and an incremental fine-tuning process for the model is triggered according to the model drift state and the implementation feedback data.

[0302] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments for IoT devices and media are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0303] The systems, media, and methods provided in this application are one-to-one correspondences. Therefore, the systems and media also have similar beneficial technical effects as their corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, the beneficial technical effects of the systems and media will not be repeated here.

[0304] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0305] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0306] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0307] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0308] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0309] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0310] Computer-readable media include both permanent and non-permanent, removable and non-removable media that can store information by any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0311] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0312] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for early identification of rockbursts based on physical information injection, characterized in that, The method includes: Acoustic emission waveform data, microseismic waveform data, triaxial stress data, and tunnel deformation and displacement data were collected within the underground engineering area. The collected acoustic emission waveform data, microseismic waveform data, triaxial stress data, and tunnel deformation and displacement data are combined with static geological attribute data, and time synchronization and spatial coordinate alignment are performed to generate high-dimensional feature cube data with a unified spatiotemporal benchmark; wherein, the static geological attribute data includes lithology, structural joints, and strike and dip information; The high-dimensional feature cube data is input into a preset physical information injection converter for modeling, to obtain intermediate feature vector data containing energy accumulation rate, stress concentration coefficient and fracture evolution potential, as well as three-dimensional hazard confidence distribution data. Based on the intermediate feature vector data, a preset rockburst case knowledge base is retrieved to obtain similar historical case data. The similar historical case data is then spliced ​​and fused with the intermediate feature vector data to generate natural language text data containing explanations of potential triggering mechanisms and suggestions for prevention and control measures. A three-dimensional risk heat map is generated based on the three-dimensional risk confidence distribution data. The spatial location data of potential rupture surfaces is extracted based on the three-dimensional risk heat map data. The three-dimensional risk heat map data, the spatial location data of potential rupture surfaces, and the natural language text data are then visualized. Risk assessment is performed based on the three-dimensional hazard confidence distribution data. When the risk level reaches a preset threshold, an edge alarm signal is triggered, and on-site feedback data on the implementation of the prevention and control measures is recorded. The system monitors the input distribution characteristics of the high-dimensional feature cube data and the distribution characteristics of the physical residuals generated by the modeling process. Based on the monitoring results, it determines the model drift state and triggers the incremental fine-tuning process of the model according to the model drift state and the execution feedback data.

2. The method for early rockburst identification based on physical information injection according to claim 1, characterized in that, The collected acoustic emission waveform data, microseismic waveform data, triaxial stress data, and tunnel deformation and displacement data are combined with static geological attribute data, and time synchronization and spatial coordinate alignment are performed to generate high-dimensional feature cube data with a unified spatiotemporal reference, specifically including: Based on the preset IEEE-1588PTP precise time protocol, the sensor nodes that collect the acoustic emission waveform data, micro-vibration waveform data, triaxial stress data and roadway deformation displacement data are synchronized in time to obtain synchronization timestamp data. The three-dimensional spatial coordinate framework data of the underground engineering area is established based on the preset RTK-SLAM algorithm and laser scanning point cloud data; The acoustic emission waveform data, microseismic waveform data, triaxial stress data, and roadway deformation and displacement data are spatiotemporally mapped based on the synchronous timestamp data and the three-dimensional spatial coordinate frame data to obtain monitoring point data with spatially aligned positions. The acoustic emission waveform data and microseismic waveform data are processed by sliding window filtering, noise suppression and energy envelope quantization at the edge computing node to obtain preprocessed waveform feature data. Adaptive Kalman filtering to suppress noise and linear interpolation to compensate for missing values ​​are performed on the triaxial stress data and roadway deformation and displacement data to obtain preprocessed stress and displacement characteristic data. The preprocessed waveform feature data, the preprocessed stress displacement feature data, and the static geological attribute data are organized into high-dimensional feature cube data indexed by time dimension, spatial dimension, and feature dimension, according to a preset time window and spatial resolution.

3. The method for early rockburst identification based on physical information injection according to claim 1, characterized in that, The high-dimensional feature cube data is input into a preset physical information injection transformer for modeling, obtaining intermediate feature vector data containing energy accumulation rate, stress concentration factor, and fracture evolution potential, as well as three-dimensional hazard confidence distribution data, specifically including: The high-dimensional feature cube data is subjected to frequency domain transformation to obtain frequency domain feature data; The frequency domain feature data is input into the physical information and injected into the multi-head self-attention layer of the converter; When calculating the similarity between the query vector and the key vector in each attention head of the multi-head self-attention layer, rock mass elastoplastic constitutive residual data and in-situ geostress tensor data are explicitly introduced as penalty terms to calculate attention weight data that conforms to mechanical constraints; wherein, the rock mass elastoplastic constitutive residual data is obtained by calculating the difference between the predicted stress and the actual stress in real time based on the constitutive equation obtained by field measurement or experiment. Based on the calculated attention weight data, the value vector is weighted and aggregated to obtain deep feature coupling data; The deep feature coupling data is input into the physical information injection converter's multi-expert gated subnetwork; Based on the lithological properties of the input data, the multi-expert gating subnetwork dynamically routes the deep feature coupling data to the corresponding expert subnetwork for processing, thereby obtaining the intermediate feature vector data and the three-dimensional hazard confidence distribution data.

4. The method for early rockburst identification based on physical information injection according to claim 3, characterized in that, Based on the lithological properties of the input data, the multi-expert gating subnetwork dynamically routes the deep feature coupling data to the corresponding expert subnetwork for processing, thereby obtaining the intermediate feature vector data and the three-dimensional hazard confidence distribution data, specifically including: The lithological label data from the deep feature coupling data and the static geological attribute data are input into the gating network of the multi-expert gating sub-network; Based on the gated network, the coupling pattern between the deep feature coupling data and different lithological types is learned, and expert selection weight data is generated. Based on the expert selection weight data, the deep feature coupling data is dynamically routed to one or more of the best-matching expert subnetworks; Based on the selected expert subnetwork, the routed deep feature coupling data is processed, and the intermediate feature vector data and the three-dimensional hazard confidence distribution data are output.

5. The method for early rockburst identification based on physical information injection according to claim 1, characterized in that, Based on the intermediate feature vector data, a preset rockburst case knowledge base is retrieved to obtain similar historical case data. This similar historical case data is then concatenated and fused with the intermediate feature vector data to generate natural language text data containing explanations of potential triggering mechanisms and suggestions for prevention and control measures. Specifically, this includes: Based on the intermediate feature vector data, a first-stage coarse-grained vector similarity retrieval is performed in the rockburst case knowledge base to obtain a preliminary matching candidate case dataset. The preliminary matched candidate case dataset is subjected to a second stage of fine-grained retrieval. The fine-grained retrieval is based on hierarchical matching calculation of static attribute vector similarity and dynamic sequence vector similarity to filter out target case data with high similarity. The static attribute vector includes lithology, strike, and joint density, and the dynamic sequence vector includes energy decay curve and fracture evolution trajectory. The feature representations of the selected target case data are concatenated with the intermediate feature vector data to form knowledge-enhanced feature input data; The knowledge-enhanced feature input data is then input into a large language model that has been fine-tuned by instructions. Based on the language big model, the knowledge-enhanced feature input data is processed to automatically generate natural language text data containing explanations of potential triggering mechanisms and suggestions for prevention and control measures.

6. The method for early rockburst identification based on physical information injection according to claim 1, characterized in that, A three-dimensional risk heat map is generated based on the three-dimensional hazard confidence distribution data. Spatial location data of potential rupture surfaces is extracted from the three-dimensional risk heat map data. The three-dimensional risk heat map data, spatial location data of potential rupture surfaces, and the natural language text data are then visualized. Specifically, this includes: The three-dimensional risk confidence distribution data is mapped onto the spatial voxels corresponding to the three-dimensional tunnel geographic information model to generate initial three-dimensional risk heat map data. Perform a clustering analysis algorithm based on density and spatial connectivity on the initial three-dimensional risk heat map data to identify high-risk connected regions. For each of the identified high-risk connected regions, its minimum convex hull geometry is calculated, and the minimum convex hull geometry represents the spatial location data of the potential rupture surface; The three-dimensional tunnel scene, which is overlaid with the initial three-dimensional risk heat map data, the convex hull boundary of the high-risk connected area data, and the natural language text data, is rendered and displayed on the three-dimensional visualization platform.

7. The method for early rockburst identification based on physical information injection according to claim 1, characterized in that, Risk assessment is performed based on the three-dimensional hazard confidence distribution data. When the risk level reaches a preset threshold, an edge alarm signal is triggered, and on-site implementation feedback data for the proposed prevention and control measures is recorded, specifically including: Real-time assessment of the risk level of the three-dimensional hazard confidence distribution data within a continuous time window; When the risk level of a predetermined number of consecutive time windows reaches or exceeds the preset risk threshold level, it is determined that the alarm triggering condition is met. When the alarm triggering conditions are met, the edge alarm signal is immediately sent to the edge alarm device deployed at the roadway site to trigger an audible and visual alarm. An alarm information window automatically pops up on the central dispatch platform. The alarm information window includes a high-risk area identifier, the prevention and control measures recommendations, and key explanations in the natural language text data. Collect and record the types of operations performed by on-site personnel based on the prevention and control measures recommendations, the operation time, and the post-operation regional stability observation results to form the execution feedback data.

8. The method for early rockburst identification based on physical information injection according to claim 1, characterized in that, The system monitors the input distribution characteristics of the high-dimensional feature cube data and the distribution characteristics of the physical residuals generated by the modeling process. Based on the monitoring results, it determines the model drift state and triggers an incremental fine-tuning process for the model according to the model drift state and the execution feedback data. Specifically, this includes: Real-time calculation of the first distribution difference measure between the current input distribution characteristics of the high-dimensional feature cube data and the historical baseline input distribution characteristics; Real-time calculation of the second distribution difference measure data between the physical residual distribution characteristics generated by the current modeling process and the expected physical residual distribution characteristics; Based on the first distribution difference measurement data and the second distribution difference measurement data, calculate the comprehensive drift index data; When the comprehensive drift index data continuously exceeds the preset drift threshold for a predetermined time, it is determined that model drift has occurred, triggering a model degradation operation, automatically switching to a historical stable model version for inference, and generating a model aging alarm message. Collect the execution feedback data, model prediction error data, and actual alarm event data, and label them to form an incremental training sample dataset; Based on the incremental training sample dataset, when a model performance degradation is detected or a periodic update instruction is received, an incremental fine-tuning training process for the physical information injection transformer is triggered to update the model parameters.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that, When the processor executes the computer program, it implements the early rockburst identification method based on physical information injection as described in any one of claims 1-8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements a rockburst early identification method based on physical information injection as described in any one of claims 1-8.

Images