Deep geologic body occurrence characteristic quantitative representation modeling method and system

Abstract

The invention provides a deep geologic body occurrence characteristic quantitative representation modeling method and system, and the model focuses on the multi-physics field occurrence environment of rock mass material composition, structural characteristics, stress distribution and the like under the continuous mining condition of a deep underground metal mine. An information expression framework capable of crossing a'microcosmic-experimental-engineering 'multi-level spatial scale is constructed, microcosmic rock core mineral components and crack information, laboratory mechanical response parameters and structural surface shear characteristics and macroscopic fault system and stress disturbance data reflected by microseismic events are introduced, a heterogeneous data fusion method is comprehensively applied, and a microcosmic-experimental-engineering multi-level spatial scale is constructed. A three-dimensional tensor model with dynamic evolution capability and physical interpretation capability is established, high-precision, dynamic and multi-scale comprehensive modeling of a complex deep geologic body occurrence environment is realized, and the model has the characteristics of clear information hierarchy, high fusion precision, strong physical interpretability and the like. And dynamic perception and scientific research and judgment of geologic mass mechanics and structure states can be realized in a complex deep environment.

Description

Technical Field

[0001] This application relates to the technical field of rock mass physical modeling, and particularly to a method and system for quantitatively characterizing the occurrence characteristics of deep geological bodies. Background Art

[0002] With the continuous increase in the depth of mineral resource development, the continuous mining of deep metal mines has become the mainstream direction of current resource utilization. However, the occurrence environment of deep geological bodies is complex and variable, and the rock mass shows significant spatial heterogeneity and scale coupling characteristics in terms of material composition, structural features, stress distribution, and its dynamic response, severely restricting the safety and efficiency of mine mining. How to accurately identify and quantitatively describe the physical and mechanical occurrence environment of deep geological bodies and construct a three-dimensional geological model with engineering guiding significance has become one of the core issues in current mine intelligent construction and disaster prevention and control.

[0003] Most traditional geological modeling methods are based on drilling, geological logging, and artificial experience for geometric modeling. Although they have certain effects in describing the overall stratigraphic structure, they show great limitations when facing practical problems such as the development of fracture structures, sudden changes in rock mass mechanical properties, and the evolution of deep stress fields. On the one hand, such models lack the ability to express microstructural features and cannot cover microscale elements such as mineral composition and pore-fracture distribution. On the other hand, their mechanical properties are often assigned by experience or processed by regional averaging, making it difficult to reflect the heterogeneity and anisotropy of rock mass mechanical parameters at different spatial positions, resulting in insufficient model interpretability and predictive ability. In addition, current mainstream three-dimensional geological modeling software mainly focuses on geometric modeling and static expression, unable to achieve dynamic updates of mechanical states and response simulations under mining disturbances, and difficult to meet the real-time perception and dynamic adjustment requirements under deep continuous mining conditions. Summary of the Invention

[0004] This application proposes a method and system for quantitatively characterizing the occurrence characteristics of deep geological bodies, which can solve one of the problems existing in the background art.

[0005] To achieve the above object, this application adopts the following technical solutions:

[0006] In the first aspect, a method for quantitatively characterizing the occurrence characteristics of deep geological bodies is provided. The modeling method includes:

[0007] Collecting multi-scale occurrence information of deep geological bodies, where the multi-scale occurrence information includes: mineral structure information at the microscale, physical and mechanical parameter information at the experimental scale, and geological response information at the engineering scale;

[0008] The mapping relationship of the occurrence information between scales established in the spatial dimension; the multi-scale occurrence information is correspondingly mapped to the point layer, line layer, surface layer and volume layer of the nested model constructed based on the three-dimensional coordinate grid, wherein the point layer corresponds to the mineral structure information and the physics parameter information, the surface layer corresponds to the surface structure obtained by processing the geological response information, the line layer is the intermediate layer between the point layer and the surface layer and corresponds to the linear information used to describe the structural continuity and directionality, the volume layer corresponds to the volume information used to reflect the regional state, and the volume information is the aggregation information of the point layer, the line layer and the surface layer in the spatial dimension;

[0009] Based on the geological response information, a three-dimensional coordinate set of earthquake sources is obtained, and a spatial clustering algorithm is used to identify high-density earthquake source clusters from the three-dimensional coordinate set of earthquake sources; a three-axis tensor inversion method based on the first motion polarity of seismic waves is used to perform earthquake source mechanism inversion on the events corresponding to the high-density earthquake source clusters, identify active tectonic units from the high-density earthquake source clusters, obtain the attributes of the active tectonic units, and analyze and obtain the time dimension of the identified active tectonic units. The active intensification time-space segment, the active tectonic unit is the identified surface structure; the identified active tectonic unit, the attributes of the active tectonic unit and the active intensification time-space segment are recorded in the structure atlas;

[0010] Extract modal feature vectors from the geological response information, combine the modal feature vectors with relative time encoding to form a modal vector sequence, and form an input tensor with all modal vector sequences and input it into the Transformer model for encoding and decoding to obtain a context-aware feature tensor, and input the context-aware feature tensor into a multi-layer perceptron MLP for state mapping to obtain a state prediction tensor; and

[0011] Based on the nested model, the structure atlas and the state prediction tensor, a three-dimensional occurrence environment model updated with time is constructed.

[0012] Based on the above technical solutions, the model focuses on the multi-physical field occurrence environment of the deep underground metal mine under continuous mining conditions, including the material composition, structural characteristics, and stress distribution of the rock mass. It constructs an information expression framework that can span multiple hierarchical spatial scales from "microscopic - experimental - engineering". By introducing microscopic core mineral composition and fracture information, laboratory mechanical response parameters and shear characteristics of structural planes, as well as macroscopic fault systems and stress perturbation data reflected by microseismic events, and comprehensively using heterogeneous data fusion methods, a three-dimensional tensor model with dynamic evolution ability and physical interpretability is established to achieve high-precision, dynamic, and multi-scale comprehensive modeling of the occurrence environment of complex deep geological bodies. This model has the characteristics of clear information hierarchy, high fusion accuracy, and strong physical interpretability, and can realize the dynamic perception and scientific judgment of the mechanical and structural states of geological bodies in complex deep environments, providing core support for rock mass disaster prediction and intelligent decision-making of underground engineering.

[0013] In a possible design approach of the first aspect, an occurrence information mapping relationship between scales in the spatial dimension is established, specifically including:

[0014] Convert the mineral structure information into equivalent physical and mechanical parameters; and

[0015] Based on the equivalent medium hypothesis, construct a mapping matrix between the physical and mechanical parameters and the geological response information.

[0016] In a possible design approach of the first aspect, a three-axis tensor inversion method based on the first motion polarity of seismic waves is used to perform focal mechanism inversion on the events corresponding to the high-density seismic source clusters, and active tectonic units are identified from the high-density seismic source clusters to obtain active tectonic unit attributes, specifically including:

[0017] According to the three-dimensional coordinates of the high-density seismic source cluster, determine the focal mechanism tensor;

[0018] Perform eigenvalue decomposition on the focal mechanism tensor to obtain the maximum compression direction eigenvector and the maximum tension direction eigenvector, and determine two plane structures as candidate plane solutions. The maximum compression direction eigenvector and the maximum tension direction eigenvector correspond to the stress principal axes;

[0019] Based on the model tensor expression approximated by the focal mechanism vector and defined by the normal vector and the slip direction vector, obtain the normal vector and the slip direction vector expressed by the maximum compression direction eigenvector and the maximum tension direction eigenvector;

[0020] From the normal vector and the slip direction vector, calculate the geometric parameters of the candidate plane structure. The geometric parameters include: strike, dip angle, and slip angle;

[0021] Compare the candidate plane structure with the observed first motion polarity on the epicenter projection map, and screen out the unique plane solution from the candidate plane solutions; and

[0022] Use the slip angle to determine the plane structure type of the unique plane solution,

[0023] The active tectonic unit attributes include: three-dimensional coordinates, stress principal axes, the geometric parameters, and the plane structure type.

[0024] In a possible design of the first aspect, analyzing to obtain the active intensification spatio-temporal segment of the identified active tectonic unit in the time dimension specifically includes:

[0025] Determine the energy release value and occurrence time of the event according to the geological response information;

[0026] Through the sliding window method, analyze whether the activity degree of the plane structure intensifies in the time dimension. If so, extract the energy release video information,

[0027] The active tectonic unit attributes further include: energy release time-frequency information.

[0028] In a possible design of the first aspect, the Transformer model introduces the Mohr-Coulomb shear failure criterion as a physical constraint, and the shear failure criterion is defined as follows:

[0029]

[0030] Where: τ is the shear stress, σ_n is the normal stress, c is the cohesion, is the internal friction angle.

[0031] In a possible design of the first aspect, the shear stress is equivalent to the equivalent shear stress, which is used to compare with the shear stress reference value of the Mohr-Coulomb model. When the comparison result meets the preset conditions, at least one of the following feedback mechanisms is executed: reducing the deviation mode channel weight, starting the redundant data path compensation,

[0032] The reducing the deviation mode channel weight includes: statistically analyzing the sensitivity of the comparison result in each modal feature vector, and screening out the most sensitive modal feature vector; and, down-regulating the weight coefficient corresponding to the most sensitive modal feature vector,

[0033] The starting the redundant data path compensation includes: fusing and enhancing the redundant feature vector with the context-aware feature tensor.

[0034] In a possible design of the first aspect, the Transformer model adopts the multi-head attention mechanism.

[0035] In a possible design of the first aspect, the modeling method further includes: outputting the structural state parameters, risk indicators, and historical evolution sequence results to the monitoring platform through a standardized interface to support remote online analysis, early warning, and dispatching linkage control.

[0036] In a possible design of the first aspect, the modeling method further includes: performing standardization and spatial alignment processing on the multi-scale occurrence information.

[0037] In a possible design of the first aspect, the mineral structure information includes: mineral particle fabric, grain morphology index, microfracture density, and interfacial bonding strength; the physical and mechanical parameter information includes: elastic modulus, Poisson's ratio, longitudinal and transverse wave velocities, shear strength, peak strength, and residual strength; the geological response information includes: microseismic data, energy release data, strain, temperature, humidity, and resistance, and the state prediction tensor involves the following state parameters: structural principal stress value, shear energy density, displacement rate, damage factor, and local failure probability.

[0038] In a second aspect, a system for quantitatively characterizing the occurrence characteristics of deep geological bodies is provided, and the system includes:

[0039] A multi-scale information acquisition module, configured to acquire multi-scale occurrence information of a deep geological body, where the multi-scale occurrence information includes: mineral structure information at the microscale, physical and mechanical parameter information at the experimental scale, and geological response information at the engineering scale;

[0040] A parameter mapping module, configured to establish a mapping relationship of occurrence information between scales in the spatial dimension; mapping the multi-scale occurrence information to the point layer, line layer, surface layer, and volume layer of a nested model constructed based on a three-dimensional coordinate grid, where the point layer corresponds to the mineral structure information and the physics parameter information, the surface layer corresponds to the surface structure obtained by processing the geological response information, the line layer is the intermediate layer between the point layer and the surface layer and corresponds to the linear information used to describe the structural continuity and directionality, and the volume layer corresponds to the volume information used to reflect the regional state, and the volume information is the aggregated information of the point layer, the line layer, and the surface layer in the spatial dimension;

[0041] A structure recognition module, which is used to obtain a three-dimensional coordinate set of seismic sources based on the geological response information, and adopt a spatial clustering algorithm to identify high-density seismic source clusters from the three-dimensional coordinate set of seismic sources; adopt a three-axis tensor inversion method based on the first motion polarity of seismic waves to perform seismic source mechanism inversion on the events corresponding to the high-density seismic source clusters, identify active tectonic units from the high-density seismic source clusters, obtain active tectonic unit attributes, and analyze and obtain the active intensification time-space segments of the identified active tectonic units in the time dimension. The active tectonic unit is an identified surface structure; record the identified active tectonic units, the active tectonic unit attributes, and the active intensification time-space segments into a structure atlas;

[0042] A heterogeneous data fusion module, which is used to extract modal feature vectors from the geological response information, combine the modal feature vectors with relative time encoding to form a modal vector sequence, form an input tensor with all modal vector sequences and input it into a Transformer model for encoding and decoding to obtain a context-aware feature tensor, and input the context-aware feature tensor into a multi-layer perceptron (MLP) for state mapping to obtain a state prediction tensor; and

[0043] A three-dimensional modeling module, which is used to construct a three-dimensional occurrence environment model updated with time based on the nested model, the structure atlas, and the state prediction tensor.

[0044] In a third aspect, an electronic device is provided. The electronic device includes: a processor and a memory coupled to the processor. The memory is used to store a computer program; the processor is used to execute the computer program stored in the memory so that the electronic device executes the modeling method according to any possible implementation manner in the first aspect.

[0045] In a fourth aspect, a computer-readable storage medium is provided, including a computer program or instruction. When the computer program or instruction runs on a computer, the computer is made to execute the modeling method according to any possible implementation manner in the first aspect.

[0046] In a fifth aspect, a computer program product is provided, including: a computer program or instruction. When the computer program or instruction runs on a computer, the computer is made to execute the modeling method according to any possible implementation manner in the first aspect. Description of the Drawings

[0047] To more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings required for use in the embodiments or related technical descriptions. Obviously, the drawings described below are only some embodiments of the embodiments of the present application. For those of ordinary skill in the art, other drawings can also be obtained based on these drawings without creative efforts.

[0048] Figure 1 It is a structural diagram of the deep geological modeling process provided by the embodiments of the present application, showing the overall modeling process of the present invention;

[0049] Figure 2 It is a structural block diagram of the in-situ rock mass information acquisition system provided by the embodiments of the present application, showing the collaborative relationship between multiple types of sensors (such as microseismic meters, acoustic emission sensors, strain gauges, etc.) deployed in key areas such as goafs and high-stress zones. The signal flows from the sensing end through the synchronous trigger module and the signal processing module to the edge computing platform and the data fusion module;

[0050] Figure 3 It is a schematic diagram of the system working process and module interfaces provided by the embodiments of the present application, showing the whole process from "field multi-modal perception → heterogeneous data preprocessing → parameter mapping → model fusion → 3D modeling → output service" and the interface logic between modules, constructing a complete integrated closed-loop path of "perception - modeling - visualization - service";

[0051] Figure 4 It is a flow chart of heterogeneous data fusion processing in engineering applications provided by the embodiments of the present application, showing the data fusion strategy constructed by the system for multi-source heterogeneous information such as microseismic, acoustic emission, resistivity, temperature and humidity based on in-situ rock mass perception. The left side of the figure is a schematic diagram of a typical monitoring scenario, where multiple types of sensor nodes are deployed in high-stress concentration areas to form a dense in-situ perception network. After the sensors collect signals, the synchronous trigger module ensures time consistency, and then the signals are input into the fusion process module on the right side. Detailed implementation manners

[0052] In order to make the purpose, technical solutions and advantages of the present application more clear and understandable, the following further details the present application in combination with the drawings and embodiments. It should be understood that the specific embodiments described here are only used to explain the present application and are not used to limit the present application.

[0053] It should be noted that although the functional modules are divided in the schematic diagram of the device and the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in a different module division in the device or a different order in the flowchart. Terms such as "first" and "second" in the specification, claims and the above-mentioned drawings are used to distinguish similar objects and do not necessarily describe a specific order or sequence.

[0054] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which this application belongs. The terms used herein are only for the purpose of describing the embodiments of this application and are not intended to limit this application.

[0055] The following uses a specific embodiment to exemplarily illustrate the method for quantitatively characterizing and modeling the occurrence characteristics of deep geological bodies in this application.

[0056] The purpose of this embodiment is to provide a cross-scale quantitative characterization and modeling method for the occurrence characteristics of deep geological bodies, aiming to systematically solve the technical bottlenecks of existing deep geological modeling methods in terms of scale connection, physical consistency, data fusion and dynamic response capabilities. Specifically, this model focuses on the multi-physical field occurrence environment of the material composition, structural characteristics and stress distribution of rock masses under continuous mining conditions of deep underground metal mines, and constructs an information expression framework that can span multiple hierarchical spatial scales of "microscopic - experimental - engineering". By introducing microscopic core mineral composition and fracture information, laboratory mechanical response parameters and shear characteristics of structural planes, as well as macroscopic fault systems and stress perturbation data reflected by microseismic events, and comprehensively using heterogeneous data fusion methods, a three-dimensional tensor model with dynamic evolution ability and physical interpretability is established to achieve high-precision, dynamic and multi-scale comprehensive modeling of the complex deep geological body occurrence environment.

[0057] As Figure 1 shown, the modeling method includes the following steps:

[0058] S1. Obtain multi-scale geological body occurrence information and construct an initial data atlas.

[0059] In this step, based on the typical occurrence characteristics of deep geological bodies, multi-scale information from microscopic to macroscopic and from point locations to regions is obtained:

[0060] At the microscopic level, information such as mineral grain fabric, grain size, and microfracture distribution is obtained, mainly relying on means such as laboratory scanning electron microscopy (SEM) and X-ray diffraction XRD;

[0061] At the mesoscale level, physical and mechanical parameters such as elastic modulus, shear strength, and wave velocity are extracted by combining on-site core test data;

[0062] At the engineering scale level, relying on the microseismic monitoring system, acoustic emission network, strain gauges, temperature and humidity sensors and gas probes deployed underground, information reflecting the activities of large-scale structures (faults, joints, etc.) is continuously collected.

[0063] After preprocessing, the above heterogeneous data form an initial data map covering dimensions such as material, structure, stress, and environment, providing basic support for subsequent modeling.

[0064] S2. Build a spatial scale mapping mechanism to achieve data level alignment.

[0065] This step relies on the data map established in S1 to further establish the mapping relationship between the scale parameters:

[0066] The experimentally obtained rock physical parameters (e.g., the relationship between micro-crack density and experimental elastic modulus) are deduced upward;

[0067] Using the structural model and fault plane extraction algorithm, the source parameters of the microseismic signal inversion are mapped to the spatial grid;

[0068] The parameters between different scales (micro-experimental-engineering) are spatially layered and matched to form a cross-scale spatial data structure of "point-line-surface-volume".

[0069] Solve the problems of inconsistent dimensions and uncoordinated spatial distribution among heterogeneous source data, and provide a unified input template for subsequent multi-source fusion.

[0070] S3. Identify key tectonic units and structural activity areas, and extract evolutionary characteristics.

[0071] Based on the spatial data structure constructed in S2, graph theory and spatiotemporal clustering algorithms are used to identify key structural surfaces and fracture systems:

[0072] Density clustering and slip direction fitting are used on microseismic data to identify active structural surfaces;

[0073] Extract the concentrated area of ​​earthquake source based on the acoustic emission energy and frequency characteristics, and determine the rupture evolution path;

[0074] The three-dimensional structural inversion method is introduced to quantitatively model the structural evolution trend in different time periods and form a dynamic structural response network map.

[0075] The dynamic characteristics of the geological body structural response were further explored to provide behavioral mechanism support for subsequent physical constraint modeling.

[0076] S4. Construct a multimodal physical model fusion mechanism to realize data-driven and physical constraint coupled modeling.

[0077] The standardized data structures constructed by S1–S3 are fused.

[0078] After identifying the main tectonic activities and mechanical evolution paths, the system introduces a physical model for modeling constraints and integrates multi-modal data:

[0079] At the basic level, strength criteria such as Mohr-Coulomb and Hoek-Brown are used to judge the stability of the fault zone;

[0080] In the high-level fusion, a multi-source Transformer fusion mechanism is introduced. First, feature encoding and time embedding are performed on each modal data to form a modal input sequence with time tags; then, the multi-head attention mechanism is used to learn the non-linear interaction relationships between modalities;

[0081] The system compares the fusion result with the predicted value of the physical model, and uses a residual feedback mechanism to dynamically adjust the fusion weight. It realizes the transformation of the modeling mode from "pure data-driven" to "physical-data collaborative-driven", ensuring the physical credibility of the model.

[0082] S5. Construct a cross-scale three-dimensional dynamic occurrence environment model to realize the visual reconstruction of the evolution of the structural body.

[0083] After the fusion modeling is completed, the system constructs a multi-level three-dimensional model based on a unified spatial reference coordinate:

[0084] Each key fracture structure and physical parameter are expressed in the three-dimensional model in the form of grid cell nesting;

[0085] The model has time tags and a dynamic update mechanism, and can update the evolution state with data collection and structural activities;

[0086] At the same time, the system integrates a real-time visualization interface to realize the visualization of the rock mass state, the visual analysis of the risk trend, and the backtracking of the evolution path. It realizes the reconstruction of the "visible, measurable, and traceable" characteristics of the complex deep occurrence environment, providing dynamic support for downstream prediction, warning, and scheduling.

[0087] S6. Output the core parameters of the model and support the real-time call of the engineering monitoring system.

[0088] The final output of the model includes but is not limited to: the stress state parameters of the partitioned rock mass; the state indicators of the three-dimensional structural units (deformation amount, shear stress, failure probability); the classification of the danger levels of each region, the trend of the instability probability of the structural plane; the synchronous timestamp, the signal quality weight, the physical model residual index, etc.

[0089] The parameters support uploading to the monitoring cloud platform in multiple ways such as RS485, CAN, and 4G, and can be docked with the engineering intelligent scheduling system, warning system, etc., to realize the real-time call and engineering response of "model as a service".

[0090] The following elaborates on each step in detail.

[0091] Step S1

[0092] It aims to systematically obtain multi-dimensional occurrence information of deep geological bodies from the micro-structure to the engineering scale, and construct an original data foundation system to support subsequent fusion modeling and dynamic reconstruction.

[0093] The system designs multi-level sensing units around the three major occurrence elements of "micro-structure - structure - stress", realizes the unification of spatial coverage, physical field coordination and time-series acquisition, and forms a structured data atlas with engineering adaptability, where:

[0094] Microscopic level: Relying on core samples, through means such as X-ray diffraction (XRD), scanning electron microscopy (SEM), and thin section identification, information on mineral composition, particle size distribution, pore and fracture morphology, and micro-structure texture is obtained. The system parameterizes them into grain morphology indexes (aspect ratio, particle distribution function), micro-fracture density, interface bonding strength, etc., as the input source for subsequent mechanical property mapping.

[0095] Laboratory scale: Using methods such as triaxial compression, wave velocity test, ultrasonic CT, etc., parameters such as elastic modulus, Poisson's ratio, longitudinal and transverse wave velocities, peak strength and residual strength of representative rock samples are obtained. And the equivalent medium hypothesis is introduced to normalize the test data to the macroscopic volume unit, and the test data are all converted into the average or equivalent properties under a unified volume scale. This normalization provides a common physical benchmark for subsequent cross-scale data docking, making the experimental data comparable and consistent. It provides a reference baseline for subsequent spatial scale conversion.

[0096] Engineering scale: A multi-source sensor array is arranged around the roadway perimeter, goaf boundary and high stress concentration zone, including: microseismic sensors (frequency band 0.1 - 50 Hz, sensitivity ≤ 10 -6 m / s 2 ) for recording structural slip and focal mechanism inversion; acoustic emission sensors (frequency response 20 - 400 kHz) for monitoring local fracture activation and energy release; fiber Bragg grating strain gauges (accuracy ±1 με) for recording the continuous deformation trend of surrounding rock; resistivity probes (range 10 - 10 5 Ω·m) for identifying the water content and connectivity of structures; multi-component gas sensors (response time < 10 s, CH4 / CO2 / H2 resolution better than 0.01%) for identifying fracture penetration and escape anomalies. A multi-field information sensing network with penetrability in space, complementarity in physical properties and synchrony in time-series response is formed.

[0097] After the original data is collected, the system performs time marking, physical dimension classification, and synchronous recording of environmental parameters (such as temperature, humidity, power fluctuations, etc.), and constructs a structured raw data matrix X_raw: X_raw ∈ R^{n×t×d}, where n is the number of sensors, t is the sampling time series, d is the multi-modal physical dimension, and R is the set of real numbers, indicating that each element in this tensor (or matrix, array) is a real value.

[0098] Step S2

[0099] Based on the multi-scale raw perception data obtained in S1, establish the mapping relationship of each physical quantity in the spatial and scale dimensions, and realize the unified structural expression between data from different sources and different levels. The system combines the laws of rock mechanics experiments, engineering experience relationships, and spatial nesting models to construct a three-level parameter connection path of "microscopic - experimental - engineering", and completes data normalization, grid projection, and structural coupling to generate an inputable cross-scale structural data template.

[0100] In the microscopic-experimental scale parameter mapping, the system constructs a transformation function group based on empirical formulas such as fracture rate - elastic modulus, mineral composition - shear strength, and texture direction - Poisson's ratio.

[0101] Convert microscopic characterization indicators into equivalent physical parameters. Taking the elastic modulus E as an example, it can be expressed as

[0102] E eq =α1·ρ f +α2·AR+α3·CI+ε

[0103] where, E eq is the equivalent elastic modulus, ρ f is the fracture density, AR is the aspect ratio of particle size, CI is the interface bond strength index, ε is the regression residual, α1 is the influence weight of fracture density on the equivalent elastic modulus, α2 is the influence weight of aspect ratio of particle size on the equivalent elastic modulus, and α3 is the influence weight of interface bond strength index on the equivalent elastic modulus. These influence weights are the regression coefficients or fitting parameters of the empirical regression model, used to quantify the contribution weights of different microscopic characteristics to the equivalent elastic modulus. These coefficients are determined through actual experimental data or statistical regression analysis.

[0104] (Equivalent medium) In the experimental-engineering scale conversion, considering the influence of sampling representativeness and scale effect, the system introduces the hypothesis of "equivalent structural unit" and constructs a mapping matrix:

[0105] M _eng =W1·M -lab +W2·Δσ+W3·H -geo

[0106] where: M _engis the engineering scale parameter set, M _lab is the experimental scale parameter, Δσ represents the regional stress gradient, H _geo represents the geological bedding heterogeneity function, W_i is the empirical weighting coefficient, and i takes 1, 2, 3.

[0107] In terms of the spatial expression structure, the system constructs a multi-level nested model based on a three-dimensional coordinate grid, where:

[0108] Node attribute vector: A self-defined variable used to represent the characteristics of micro data (such as mineral composition, grain size, microfracture density, etc.) and laboratory data (such as elastic modulus, Poisson's ratio, wave velocity, etc.) after normalization processing. Each node attribute vector is a multi-dimensional vector, and its components are obtained through data preprocessing and mapping function calculations.

[0109] Line element: Used to describe the linear part in the model that reflects the continuous structural characteristics in the rock mass, such as the extension trend of faults, joints or fractures. The line element data in this embodiment, including continuous structural characteristics such as faults and joints, can either directly come from existing observational data at the micro, experimental and engineering scales, such as core thin section analysis, rock mechanics test records, underground geological logging and fault mapping, or can be identified through calculation methods such as spatial clustering, trend fitting or density analysis of microseismic, acoustic emission and other monitoring data. Among them, the known or visible joint trends, fracture trajectories, etc. can usually be directly extracted, while the hidden structures or linear features that are not significantly exposed need to be determined by the calculation and fitting of monitoring data. Therefore, the formation process of line element data includes both direct acquisition and calculation identification and fitting reconstruction based on dynamic monitoring data, and the two complement each other to jointly support the complete construction of the three-dimensional structure model. As an intermediate level connecting nodes (micro / experimental data) and surface elements (macroscopic structural surfaces), it helps to establish continuity and directionality descriptions in three-dimensional space and provides support for the data connection of subsequent structural surfaces and volume elements.

[0110] Surface element: It mainly maps the continuous two-dimensional structural characteristics obtained by spatial interpolation of sensor data, such as fault surfaces, joint surfaces and other structural surface information, and its attributes include geometric parameters and corresponding dynamic monitoring data. The surface elements described in this embodiment mainly refer to the known or fitted geometric structural surfaces inside the rock mass, including fault surfaces, joint surfaces and other planar structures, as the basic geometric units of the three-dimensional geological model. These surface elements can come from geological logging, experimental observations or spatial fitting reconstruction. On this basis, through the clustering analysis, focal mechanism inversion and energy release trend identification of microseismic, acoustic emission and other monitoring data, the key areas with current slip activities or failure trends are further determined and marked as "possible active tectonic units". Therefore, the active tectonic unit is the recognition result based on the dynamic behavior characteristics of the surface element and is used to support risk monitoring and early warning decisions.

[0111] Volume element: It represents the volume data at the engineering scale. By aggregating and weighting point, line, and surface data in space, regional average or overall state information is obtained.

[0112] Microscopic and experimental data are mapped into the "node attribute vector".

[0113] Engineering perception data serves as the state parameter of "surface and volume elements".

[0114] Structural surfaces, faults, and joint groups are implanted into the main grid structure in the form of mosaic layers. As specific instances of surface elements, they are implanted into the main grid structure in the form of mosaic layers to form the surface element data in the model, which is used to express the main planar structural features in the rock mass. Form the surface element data in the model, which is used to express the main planar structural features in the rock mass.

[0115] Overall, a "point-line-surface-volume" structured data expression system is formed, supporting the independent invocation and cross-layer reference of data at each level.

[0116] In the data unification stage, the system performs dimensionless normalization on each physical quantity, using a combination strategy of Z-score standardization and Min-Max scaling to ensure that different modal data have a unified input scale.

[0117] Step S3

[0118] Furthermore, to ensure that the constructed model has sufficient spatial resolution and physical interpretability, the system introduces a multi-stage and multi-algorithm combination structure recognition strategy in step S3:

[0119] 1. Microseismic event density clustering to identify potential structural active zones:

[0120] Preprocess the microseismic data to obtain the three-dimensional coordinate set of the earthquake hypocenters {P i (x i , y i , z i )}, and introduce the density-based spatial clustering algorithm (DBSCAN) to identify the dense areas of the spatial distribution of earthquake hypocenters. This algorithm uses event density as the clustering basis, overcoming the disadvantages of traditional clustering such as the need to specify the number of clusters and being sensitive to shapes.

[0121] The clustering is based on the following two core parameters: neighborhood radius ε: defining the "proximity" relationship; minimum number of neighbors MinPts: defining the "dense area" determination criterion.

[0122] The clustering rule is as follows: For any point P i , if the number of neighbors in its ε-neighborhood satisfies ≥ MinPts, it is defined as a "core point"; all points connected to the core point and reachable recursively belong to the same cluster.

[0123] In the DBSCAN algorithm, "recursively reachable" describes the process of associating a point with a starting core point through a series of consecutive direct neighborhood connections. In short, if there is a connection chain P→Q→…→R starting from point P, where each pair of adjacent points satisfies: the latter point is within the ε-neighborhood of the former point (and the previous points are all core points to ensure the quality of their neighborhoods), then we say that point R is "recursively reachable" from P. The specific processing steps are as follows: Core point determination: For each data point, calculate the number of neighbors within its ε-neighborhood. If the number is not less than the set MinPts, then this point is a core point. Direct neighborhood expansion: Starting from a core point P, mark all points within the ε-neighborhood of P as directly reachable points. Recursive expansion process: For each point within the neighborhood of P (assuming one of the points is Q), if Q is also a core point, then check the ε-neighborhood of Q and add all points within Q's neighborhood that have not been assigned to the current cluster to the cluster. Continue to repeat this process for the newly added core points until no further expansion is possible. Cluster construction: Throughout the process, starting from P, all points that can be reached through continuous direct neighborhood expansion and are connected by such a chain are grouped into the same cluster, that is, these points are all "recursively reachable". This "recursion" means that it does not directly require a point to fall directly within the ε-neighborhood of the starting point P, but allows reaching the final point through one or more intermediate core points, thereby constructing a density-connected region. For example: P is a core point, and Q is within its neighborhood; Q is also a core point, and R is within its neighborhood; although R is not directly within the neighborhood of P, because there is a continuous connection from P to Q and then from Q to R, we consider R to be recursively reachable from P. This mechanism enables DBSCAN to capture local density-connected regions and not just rely on the direct distance relationship between points and the starting point. Points that do not belong to any cluster are regarded as "noise" or "isolated seismic sources".

[0124] Through this method, several high-density seismic source clusters Cj are obtained, and each cluster corresponds to a possible active tectonic unit, such as a buried fault, joint group, fracture intersection zone, etc.

[0125] 2. Determining the slip direction and rupture property through focal mechanism inversion:

[0126] For each identified seismic source cluster, the system further selects typical events with large magnitudes and complete signals for focal mechanism inversion. Using the three-axis tensor inversion method based on the P-wave first motion polarity, the stress principal axis and slip plane solutions are constructed:

[0127] The focal mechanism tensor is expressed as:

[0128]

[0129] Among them: Each component such as Mxx, Mxy represents the moment contribution of the seismic source in different coordinate directions; Mxx, Myy, Mzz: are the diagonal components along the X, Y, and Z directions respectively; Mxy, Mxz, Myz: are the non-diagonal components, representing the coupling effects in different directions.

[0130] Since what is obtained by inversion is the deviation (or pure shear) effect of the seismic source, the zero-trace condition, i.e., Mxx + Myy + Mzz = 0, is usually required to ensure that this tensor only reflects shear failure rather than volume change.

[0131] This tensor can calculate the strike, dip, and slip angle of the fault plane, and combine with the clustering results to achieve the coupled annotation of the structural plane and the motion attributes. Specifically as follows:

[0132] Calculate the stress principal axis direction from the tensor: Perform eigenvalue decomposition on the tensor M to obtain three eigenvalues λ1, λ2, λ3 and the corresponding eigenvectors v1, v2, v3. Usually, we stipulate that: v1 corresponds to the maximum compression direction (P axis); v2 is the neutral axis (B axis); v3 corresponds to the maximum tension direction (T axis).

[0133] Derive the fault plane normal vector n and the slip direction s:

[0134] The system constructs two candidate fault plane solutions (the main fault plane and the auxiliary plane) based on the combination of the principal stress directions. Each fault plane is represented by a set of normal vector n and slip direction vector s. The system approximates M as a double couple tensor, and its model tensor M model is expressed as:

[0135]

[0136] n: The unit normal vector, representing the normal direction of the fault plane;

[0137] s: The unit slip direction vector, representing the sliding direction of the rock mass on the fault;

[0138] The tensor product (outer product) of vectors.

[0139] Construct the formula det(M - λI) = 0, where λ is the eigenvalue, I is the identity matrix, and det represents the determinant. Solve this third-order characteristic equation to obtain three eigenvalues. Usually, the sorting is as follows: λ1: The maximum eigenvalue, corresponding to the maximum compressive stress P axis, λ2: The intermediate eigenvalue, corresponding to the B axis, λ3: The minimum eigenvalue, corresponding to the maximum tensile stress T axis. For each eigenvalue λ i , solve the following homogeneous linear equations: (M - λ i I)v i = 0, and the corresponding eigenvector v i can be obtained.

[0140] Combining the two equations, we can calculate:

[0141] Main fault plane normal vector:

[0142]

[0143] Slide direction vector:

[0144]

[0145] Calculate the geometric parameters (φ, δ, λ) of the fault plane from the vector:

[0146] The system constructs the normal vector n = (n x ,n y ,n z ) and sliding direction s=(s x ,s y ,s z ) is introduced into the following geometric formula to calculate the spatial geometric parameters of the fault plane item by item:

[0147] Inclination angle δ: δ = arccos(|n z ∣)

[0148] Towards φ: φ = arctan2(n y ,n x )

[0149] Slip angle λ: The system first constructs the unit vector of the direction of movement: t = (-sinφ, cosφ, 0)

[0150] Then calculate the slip angle: λ = arctan2(s z ,s·t)

[0151] Determine a unique fault plane solution:

[0152] Since the tensor structure allows for two plane solutions, the system compares the calculated plane solutions with the observed first motion polarities on the epicenter projection map. As the solution of the focal mechanism tensor usually yields two sets of symmetric fault plane parameters (the main fault plane and the auxiliary plane), to determine the unique reasonable solution, the system projects the spatial parameters (strike, dip, slip angle) of these two sets of plane solutions onto the epicenter projection map (equivalent to the spherical projection map) and marks the different polarity regions (compression zone and dilation zone) divided by the two sets of plane solutions on the map. At the same time, the first motion polarities of the P-waves recorded by the monitoring array (i.e., whether the first-arriving wave is compressive or dilative) are marked with the corresponding azimuth and polarity symbols on the same projection map. The system compares the consistency between the polarity distributions corresponding to the two sets of plane solutions and the actually observed first motion polarities. If the polarity distribution of a certain plane solution is consistent with or has the smallest error compared to most of the actually observed polarities, then that plane is determined to be the unique main fault plane solution for subsequent fault activity analysis and modeling. The unique combination of main fault plane parameters (φ, δ, λ) that conforms to the observed data is selected.

[0153] Combined with the clustering results of the focal clusters, perform structural plane coupling annotation:

[0154] The system has completed the spatial clustering of focal events in the preprocessing stage, and each clustering unit represents a focal cluster. To establish the association between the structural planes and the focal distribution, the system performs the following operations:

[0155] Analysis of parameter consistency within the cluster: The system statistically analyzes the strike, dip, and slip angle of multiple typical events within the focal cluster. If the standard deviation is lower than the set threshold, for example, ±15°, then it is considered that this cluster corresponds to a set of structural planes.

[0156] Fault type identification: The system determines the tectonic attribute of this structural plane based on the average slip angle λ: λ ≈ +90°: reverse fault; λ ≈ -90°: normal fault; λ ≈ 0° / ±180°: strike-slip fault.

[0157] Annotation output: The system writes the geometric parameters (φ, δ, λ) and the motion attribute (fault type) of this structural plane as the structural plane annotation result into the structural recognition atlas, forming a coupled representation of the focal and fault structures.

[0158] 3. Analysis of energy release and temporal variation:

[0159] Combined with the energy release value E of each microseismic or acoustic emission event i and the event occurrence time t i , the system constructs the energy accumulation curve within the unit structural plane:

[0160]

[0161] where N t represents the cumulative number of events before the cut-off time t.

[0162] By using the sliding window method (such as 1h, 6h, 24h), analyze whether the activity of each tectonic unit intensifies in the time dimension, and the following can be identified: local stress accumulation; increased slip frequency; main-aftershock sequence characteristics; and activity linkage with the surrounding area. If continuous high-frequency events are accompanied by energy mutations or "repeating hypocenters", they are classified as "high-risk structural activity areas" for nested calls by the S5 dynamic model.

[0163] Among them, the steps of the method for identifying the intensification of tectonic unit activity are as follows:

[0164] Construct time series data: The system sorts each seismic source event in chronological order and divides it into the corresponding tectonic unit (seismic source cluster) according to the clustering results. For each tectonic unit, the system extracts the triggering times of the seismic source events belonging to it to form the event time series of the tectonic unit:

[0165] T = {t1, t2,..., t N}, t i ∈[T start , T end

[0166] Among them, T start represents the start time of the analysis or monitoring, that is, the starting point of the set or recorded time interval, and T end represents the end time of the analysis or monitoring, that is, the termination point of the set or recorded time interval.

[0167] Set the sliding window parameters: The system presets a time window length Δt, which can be selected from common scales such as 1 hour (1h), 6 hours (6h), or 24 hours (1d). At the same time, set the sliding step δt (such as sliding 10 minutes or 1 hour each time).

[0168] The sliding window scanning interval is: [t k , t k +Δt), t k = T start + k·δt

[0169] Conduct sliding count analysis: For each tectonic unit, within each sliding window, count the number of seismic source events in this window: N k = |{ti ∈ T ∣ t k ≤ t i < t k +Δt}|, and obtain the time-activity frequency sequence of each tectonic unit:

[0170] {(t k , N k )}, k = 1, 2,..., K

[0171] ​To identify the trend of the change in activity level, the system can evaluate whether the activity level is "intensifying" in the following ways:

[0172] Growth rate analysis: Compare the growth rate of the number of events between consecutive time windows:

[0173]

[0174] If r k is continuously positive and increasing, it indicates that the activity frequency is rising.

[0175] Moving average frequency Perform a moving average on N k and observe whether there is an overall upward trend:

[0176]

[0177] Global statistic (calculated among all N k ): μ N : Historical average frequency (global statistical value); σ N : Standard deviation of historical frequency.

[0178] Judgment criterion: If a certain period of time satisfies: r k > 0, and r k > 0 continuously exceeds 3 windows, it is considered that there is an obvious frequency increase in the structural unit, and it is marked as a high-frequency candidate section.

[0179] The system converts the magnitude M i to the seismic source energy E i , using the empirical formula: log 10 E i = 1.5Mi + 4.8

[0180] where E i is in joules (J), applicable to the magnitude range M i ∈[-1, 5]. The system sums up the energy within each sliding window:

[0181]

[0182] And calculates the energy change amount between adjacent windows:

[0183] ΔE k = E k - E k-1

[0184] The system calculates the standard deviation σ k of all E E , if: ΔE k > 2σE, it is regarded as an energy mutation event, indicating that the structural unit is rapidly releasing stress.

[0185] The system compares the event waveforms within the candidate section. For any event pair (i, j), the waveform cross-correlation coefficient is calculated at the same station. If the following conditions are met:

[0186]

[0187] where: x i (t) refers to the amplitude function of the original vibration signal collected at a certain seismic monitoring station for a certain source event i over time, and x j (t) refers to the amplitude function of the original vibration signal collected at a certain seismic monitoring station for a certain source event j over time. If R i,j > 0.9 (high waveform similarity) and the spatial distance dist(i, j) < 500 m (close spatial distance), it is marked as a repeated source pair, indicating local repeated slip on the fault plane.

[0188] Joint judgment of increased activity: Within any sliding window segment, if the following three conditions are simultaneously met: "high-frequency activity", "energy mutation", and "repeated sources", the system determines that the tectonic unit enters a state of increased activity and records this spatio-temporal segment into the structure map for subsequent risk assessment:

[0189]

[0190] Output the structure map and parameterize the encoding: The system finally outputs the structure recognition map S(x, y, z, t). Each structural unit in the map contains the following attribute fields: three-dimensional spatial position and extension range; geometric parameters of the tectonic surface (dip angle, strike, thickness); activity level (low / medium / high); average slip direction and stress principal axis; peak energy and frequency characteristics; spatial coupling degree with other structures; time identifier and status update period, where:

[0191] Extension range:

[0192] Data source: The three-dimensional positioning coordinates of all source events within this structural unit, denoted as Output by the source location system. Processing method: Calculate the minimum-maximum values of this point set in the three coordinate axis directions to form a three-dimensional bounding box:

[0193] x min ~x max , y min ~y max , z min ~z max

[0194] Use principal component analysis (PCA) to identify the direction axis, fit the point cloud, and obtain the direction and length of the tectonic principal axis.

[0195] Frequency characteristics:

[0196] Data source: The waveform signal x i (t) of each seismic source event within the structural unit, which is collected by seismic monitoring stations. Processing method: Perform a fast Fourier transform (FFT) on the waveform to obtain the amplitude-frequency characteristics; extract the main frequency f peak of each spectrum, the frequency centroid f centroid and the bandwidth; in terms of the dimensions of the structural unit, take the mean and standard deviation of each index to form an overall spectrum description.

[0197] Spatial coupling degree with other structures:

[0198] When the system identifies a structural unit, it has extracted the three-dimensional boundary range (i.e., the spatial extension range) of each structural unit and the normal vector direction of the tectonic surface. The system compares the three-dimensional boundaries of any two structural units and calculates the shortest boundary distance d ij between them, and at the same time calculates the cosine of the angle (direction similarity) c ij between their tectonic surface normal vectors = |n i · n j |. On this basis, the system uses the following formula to calculate the spatial coupling degree of two structural units:

[0199]

[0200] where d ij is the closest boundary distance between structural units i and j; n i , n j are the unit normal vectors of their tectonic surfaces; c ij is the direction similarity, and the larger the value, the more consistent the tectonic surface orientations; w1 and w2 are the weighting coefficients for fusing distance and direction, satisfying w1 + w2 = 1.

[0201] Finally, the system aggregates the C ij between each structural unit and all other units, and writes the average coupling degree or the maximum coupling degree as the "spatial coupling degree" field of this structural unit into the atlas, which is used to represent the association strength of this structure in the spatial network.

[0202] The structural recognition results are transmitted to the fusion modeling module (S4) and the 3D modeling module (S5) through a standard interface to form a "geometry-physics-time" integrated structural control input, significantly enhancing the spatial accuracy and dynamic response ability of the model.

[0203] The structural recognition chain of "source location → spatial clustering → physical inversion → time analysis → coded output result" enables this embodiment to not only extract structural units from observation data, but also enhance the model interpretability with physical constraints, strengthen the early warning ability with dynamic evolution, and realize a truly computable, predictable, and integrable deep structure recognition mechanism.

[0204] Step S4

[0205] Furthermore, S4 realizes the unified characterization, coupled modeling, and physical consistency constraint of heterogeneous modal data (vibration, acoustic emission, microseismic, strain, resistivity, gas concentration, etc.).

[0206] Modal input embedding and time alignment:

[0207] Construct modal embedding vectors for sensor signals from different sources. Let the original data be X_raw ∈ R^{n×t×d}, where n is the number of sensors, t is the time length, and d is the physical attribute dimension. The system sends each modal signal into an independent feature encoder (such as 1D-CNN, GRU, or MLP) to extract the modal feature vector f_i(t), and then adds it to the relative time encoding τ(t) to form a unified input sequence: Z_i(t) = f_i(t) + τ(t).

[0208] All modal vector sequences Z_i(t) form an input tensor Z ∈ R^{N×T×D} as the input of the Transformer model.

[0209] Among them, each modal is input:

[0210] Waveform modality: For the source waveform signal x i (t) from the station, the system uses a one-dimensional convolutional neural network (1D-CNN) to extract its local time structure features. After filtering and normalization processing, the waveform signal is input into the network as a one-dimensional sequence of length L, and finally outputs a waveform feature vector with a dimension of d1

[0211] Static modality: The system forms a feature vector si ∈ R from static attributes such as the source location, structural extension range, principal stress axis direction, maximum energy value, and average dominant frequency m , and inputs it into a multi-layer perceptron (MLP) structure for feature transformation and compression, and outputs a static coding vector with a dimension of d2

[0212] Time series modality: For the trigger time series of multiple source events within the structural unit, the system constructs a structural sliding window representation {t k} Combine it with indicators such as energy and activity level to form an input matrix, and send it to a gated recurrent neural network (GRU) for time series modeling. Finally, obtain a time-dependent feature vector with dimension d3 that reflects the activity evolution trend

[0213] The multi-head attention mechanism realizes modality interaction:

[0214] The core of the Transformer model is the multi-head self-attention mechanism (Multi-Head Self-Attention), which is used to extract key dependencies between modalities and within time series. The calculation method of the attention mechanism is as follows:

[0215]

[0216] Where: Q = ZW_Q, K = ZW_K, V = ZW_V are the query, key, and value matrices, W_Q, W_K, W_V are trainable weights, and d_k is the key dimension.

[0217] Through parallel calculation of multiple attention heads, the model can capture modality interaction relationships and time linkage mechanisms from multiple subspaces.

[0218] Fusion feature decoding and state vector generation:

[0219] After completing the feature fusion process, the system inputs the fused feature tensor into several Transformer encoding Encoder layers to further model the context dependencies of the structural units in the time dimension and modality dimension. The output result of this stage is the fused context-aware feature tensor, denoted as: Z fused ∈R^{m×t×d}, where: m represents the number of structural units; t represents the number of steps in the time series, usually corresponding to the number of sliding time windows within the monitoring period; d represents the feature dimension, depending on the dimension setting of the Transformer output.

[0220] The above tensor Z fused represents the expression result of the system's fused behavior features of each structural unit at each time step.

[0221] To achieve interpretable modeling of the structural state, the system sends this feature tensor into a multi-layer perceptron (MLP) module for state mapping, and outputs a multi-dimensional state index tensor Y∈R^{m×t×s} of the structure. Where: s represents the state index dimension, that is, the number of state variables corresponding to each structural unit at a certain moment. The state variables may include but are not limited to: structural principal stress value, shear energy density, displacement rate, damage factor, local failure probability, etc.

[0222] The processing procedure of the MLP module is as follows:

[0223] For each structural unit i and the feature vector Z_fused[i,t,:] ∈ R^d at time step t, the system first inputs a layer of linear transformation and activation function: h_{i,t} = ReLU(W1·Z_fused[i,t,:] + b1), where W1 ∈ R^{d mid ×d}, is the intermediate hidden layer mapping matrix, d mid is the intermediate dimension, and the activation function uses ReLU. Subsequently, the system performs a second layer of linear mapping to obtain the final state output: y_{i,t} = W2·h_{i,t} + b2, where W2 ∈ R^{s×d mid}, and y_{i,t} ∈ R^s represents the state index vector of this structural unit at this time step.

[0224] The above process is independently executed for each structural unit and each time step, and finally an output tensor is generated: Y ∈ R^{m×t×s}.

[0225] Introduce physical model constraints to improve the credibility of fusion:

[0226] To avoid the phenomenon of "statistically correct but physically wrong" in the deep network due to data noise, the system introduces the Mohr-Coulomb shear failure criterion as a physical constraint. This criterion is defined as follows:

[0227]

[0228] where: τ is the shear stress, σ_n is the normal stress, c is the cohesion, is the internal friction angle.

[0229] The stress state in the fusion output is converted into an equivalent shear stress τ_fused, and the residual with the Mohr-Coulomb model τ_model is compared:

[0230] r(t) = |τ_fused(t) - τ_model(t)|

[0231] If r(t) > δ continuously and lasts for the T_crit time window, the system executes the following feedback mechanism: reduce the weight of the deviation mode channel and start the redundant data path compensation.

[0232] Among them, the threshold parameter δ is the equivalent shear stress residual tolerance set by the system, which is used to determine whether the deviation between the fusion shear stress and the theoretical model shear stress reaches the warning level. Its definition is as follows: δ ∈ R + , with the unit of MPa, representing the maximum shear stress residual limit allowed by the system. The system is default set to an empirical value, such as: 0.5 MPa, and can also be dynamically set according to the historical data residual distribution:

[0233] δ = μ r + k·σ r

[0234] Where: μ r is the residual mean value, σ r is the standard deviation, and k is the adjustment coefficient (it is recommended to take 2 - 3).

[0235] When the system detects that the residual between the fusion shear stress and the theoretical model shear stress continuously exceeds the set threshold δ and the duration exceeds the set threshold T_crit, the system determines that there is a significant abnormality in the current structural state. To improve the robustness and adaptive ability of the model, the system triggers the following two feedback mechanisms:

[0236] Reduce the weight of the deviation mode channel:

[0237] When the system detects an obvious difference between the shear stress prediction result and the theoretical model (i.e., the residual r(t) > δ and the duration exceeds T_crit), the system will first determine which modal channel the residual mainly comes from. At this time, the system executes the mechanism of "reducing the weight of the deviation mode channel", and its specific implementation process is as follows: Before each fusion of multi-modal features (such as waveform features, spatial position features, time series features), the system records the weight vectors α = [α wave , α pos , α time of each mode, and these weights can come from the output of the attention mechanism or the feature channel weighting coefficient. When the residual exceeds the threshold, the system, according to the internal tracking module, statistically analyzes the sensitivity of the residual on each modal feature, such as through reverse gradient, attention shift or error attribution analysis, to identify the mode that causes the largest prediction error, that is, the "deviation mode". The system dynamically down-regulates the weight coefficient of this "deviation mode", for example: multiplying the original weighted value by a coefficient γ ∈ [0.3, 0.7] to achieve an inhibitory effect: α k ← γ·α k . After all modal weights are renormalized, they participate in the next round of feature fusion to reduce the dominant role of this mode in the final prediction result.

[0238] In the above way, the system can effectively reduce the influence of abnormal modes on the output result, prevent the continuous expansion of single-modal errors, and improve the overall prediction stability.

[0239] Start redundant data path compensation:

[0240] After suppressing the deviation mode, the system will also activate the "redundant data path compensation" mechanism to introduce additional supplementary feature information and enhance the model's ability to express the structural state. The specific implementation process is as follows: During system initialization, in addition to constructing the main feature extraction channels (such as the main waveform window and current time segment position estimation), several redundant path input modules will be reserved, such as: historical waveform segments of the previous cycle; secondary station positioning coordinates; auxiliary spectral features; auxiliary geological structure labels, etc. These redundant data are in an "inactive" state under normal circumstances and do not enter the main channel to participate in the fusion calculation. Once the system detects a continuous residual exceeding the threshold event, the activation mechanism will be triggered to send these reserved features into a specific "redundant channel encoder" (such as a spare 1D-CNN, MLP, etc.) to extract new redundant features f_redundant. The system will splice or weighted fuse the redundant features with the original fusion features f_fused to generate an enhanced fusion feature vector f_new:

[0241]

[0242] This enhanced feature f_new will be sent to the subsequent state prediction module to participate in the re-judgment of indicators such as shear stress or failure probability.

[0243] By introducing additional data channels, the system can still maintain a certain judgment ability when facing information distortion or data quality degradation in the main channel, thereby improving the overall robustness of the system.

[0244] The system realizes the dynamic adaptive adjustment of the model after the abnormal criterion is triggered through the linkage of two mechanisms: "reducing the weight of the deviation mode channel" and "activating the redundant data path compensation", effectively avoiding the expansion of prediction deviation and enhancing the stability and engineering adaptability of structural state recognition.

[0245] Adjust Transformer parameters through residual backpropagation optimization

[0246] Residual control and dynamic parameter adjustment:

[0247] The system introduces the physical residual r(t) into the loss function as a regularization term:

[0248] L_total = L_pred + λ·L_phys

[0249]

[0250] Among them, λ is the regularization coefficient to control the physical consistency weight, and L_total is the total loss function during the training process; L_pred is the basic prediction loss term, which can be the cross-entropy loss in the structural state classification task or the mean square error (MSE) in the regression task; L_phys is the physical residual regularization term, which represents the cumulative square difference between the shear stress output by the model and the theoretical shear stress. The formula of L_phys defines the sum of the squares of the physical residuals over the entire time series and is usually used as part of the loss function to evaluate the overall deviation of the model in physical consistency. r(t) represents the physical residual at a certain moment t, which is the difference between the model prediction value and the physical theoretical value (such as the Mohr-Coulomb failure criterion), r(t) 2 It means squaring the residual to avoid positive and negative cancellation and at the same time emphasizing the impact of larger errors. By jointly optimizing the prediction error and the physical residual, the system fusion result can not only accurately reflect the modal response trend but also have mechanical interpretability under the shear failure criterion.

[0251] This step forms a dual modeling framework with Transformer deep feature learning as the main line and physical mechanics criteria as the control boundary, effectively improving the robustness and generalization ability of multi-modal data in the non-stable state, and providing dynamic and reliable state input for the S5 three-dimensional structure reconstruction.

[0252] Step S5

[0253] Implementation method of the three-dimensional structure reconstruction module:

[0254] After completing the structural unit identification and state prediction, the system enters the structural model reconstruction stage. In this stage, combined with the generated structural identification atlas, time series state indicators and physical boundary constraints, a three-dimensional structure evolution model that can be updated over time is reconstructed. The specific implementation process is as follows:

[0255] The system takes each structural unit in the S3 structural identification atlas S(x, y, z, t) as the basis of the modeling unit, and extracts its spatial geometric parameters (such as three-dimensional boundaries, trends, dips, thicknesses) as the basic components of geometric modeling. Map the index values (such as shear stress, failure probability, etc.) in the S4 state prediction tensor Y∈R^{m×t×s} to the corresponding structural units, and form a "structural state visualization slice" or a three-dimensional dynamic assignment model through color mapping, voxel superposition, etc.

[0256] The system synchronously introduces the theoretical boundaries provided by the Mohr-Coulomb model to constrain the physical rationality of the three-dimensional structure in terms of spatial distribution and cross-section strength, ensuring that the reconstructed model does not violate the basic laws of geomechanics.

[0257] If there is a high degree of coupling between structural units (i.e., C i,jWhen θ) is satisfied, the system will automatically connect adjacent structural planes, construct a structural network, and generate a continuous structural plane model through interpolation methods.

[0258] The final output is a set of 3D structural models that can be updated over time.

[0259] Step S6

[0260] Output structural state parameters to support online monitoring and remote call of the project. Finally, the system will output the results after fusion and modeling as structured data, mainly including: the current physical state indicators (stress, strain, risk level) of each grid cell; the model confidence weight and data quality identification; the geological structure relationship where the structural unit is located; multi-modal signal sources and their attention contribution degrees.

[0261] During the multi-modal input modeling process, the system not only utilizes data features from different sources (such as seismic source waveforms, spatial positions, time evolution information), but also needs to judge and dynamically weight the reliability and contribution degrees of each modality. For this purpose, the system introduces two mechanisms, "data quality identification" and "model confidence weight", to guide the weighted allocation in the modality fusion process. First, the system evaluates the quality of each modality input data through a preprocessing module. According to factors such as the integrity of the waveform signal, noise level, data missing situation, and time synchronization, the system generates a quality score value q k ∈[0,1], called the data quality identification, indicating the reliability degree of this modality within the current spatio-temporal range. The closer the value is to 1, the more credible the data is; secondly, during the model training process, the system automatically learns the importance of each modality feature in the prediction task through an attention mechanism (such as channel attention or multi-head attention structure), calculates its attention contribution degree, denoted as α k ∈[0,1]. This weight represents the influence degree of this modality feature in the current structural unit state prediction. Finally, in the multi-modal feature fusion stage, the system comprehensively considers the quality score and attention contribution degree of each modality, calculates the fusion weighted coefficient ω k for controlling its proportion in the final feature vector, which is specifically expressed as follows:

[0262] ω_k = q_k·α_k

[0263] This mechanism enables the model to adaptively adjust the information fusion structure in the face of different data qualities or modality uncertainties, effectively improving the overall stability of the model and the credibility of the prediction results.

[0264] All data is uploaded to the mine monitoring platform through standard communication interfaces (such as RS485, CAN, 4G, etc.), and can be docked with existing sensing systems, dispatching systems, and AI early warning modules in real time to achieve model-driven remote monitoring, risk perception, and intelligent intervention.

[0265] In summary, in this embodiment, by constructing a geological body perception system covering multiple levels of micro-experiment-engineering, combined with cross-scale parameter mapping, key structure identification, and heterogeneous data fusion mechanisms, the multi-modal perception, dynamic modeling, and three-dimensional reconstruction of the geological body state under deep complex occurrence environments are systematically realized. The proposed fusion framework based on multi-source Transformers not only effectively explores the potential coupling relationships between different physical modalities but also realizes the collaborative modeling of data-driven and physical constraints by introducing classical mechanical models such as shear failure, thereby improving the credibility and engineering applicability of the model output results in unstable environments.

[0266] This embodiment also provides a system for implementing the method of the cross-scale quantitative characterization model of the physical and mechanical occurrence environment of the deep geological body, including a multi-scale information acquisition module, a parameter mapping module, a structure identification module, a heterogeneous data fusion module, a three-dimensional modeling module, and a dynamic output module; the above-mentioned modules are connected in sequence according to the occurrence environment modeling process and operate collaboratively. The multi-scale information acquisition module is used to obtain original occurrence data such as the material, structure, and stress state of the geological body at the micro, experimental, and engineering scales. The information collected includes mineral texture, fracture morphology, stress changes, and environmental parameters, etc., and the classification and arrangement of time tags and physical dimensions are completed; the parameter mapping module is used to establish the mapping relationship between parameters at different scales, unify the consistency of various data in spatial distribution, scale expression, and physical meaning, and construct a "point-line-plane-body" nested structure template; the structure identification module is used to identify key structural units such as tectonic planes and joint groups based on data such as microseismic and acoustic emission, extract their spatio-temporal evolution characteristics, and form a structural activity map; the heterogeneous data fusion module is used to integrate multi-modal physical field data, adopt a multi-source fusion mechanism based on Transformers and a physical model constraint method to generate a stable and reliable occurrence state estimation value; the three-dimensional modeling module is used to map the fusion result to a unified coordinate system, construct a three-dimensional rock mass structure model that supports dynamic updates, and realize the visual expression of the attributes and states of structural units; the dynamic output module is used to output results such as structural state parameters, risk indicators, and historical evolution sequences to the monitoring platform through a standardized interface, supporting remote online analysis, early warning, and dispatching linkage control.

[0267] The overall system has the advantages of strong modeling continuity, high data fusion ability, accurate structure identification, and clear three-dimensional visualization, etc. It can significantly enhance the ability to identify the geological structure evolution process and judge the risk trend, providing key technical support for occurrence environment modeling, intelligent decision-making, and disaster warning in scenarios such as deep mines and underground projects.

[0268] An embodiment of the present application further provides an electronic device, including: a processor, and a memory coupled to the processor, where the memory is used to store a computer program; the processor is used to execute the computer program stored in the memory so that the electronic device executes the method described in any one of the above embodiments.

[0269] The electronic device may be a computing device such as a desktop computer, a notebook, a palm computer, and a cloud server. The electronic device may include, but is not limited to, a processor and a memory.

[0270] The so-called processor may be a central processing unit (CPU), or may also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc. The processor is the control center of the electronic device and connects various parts of the entire device through various interfaces and lines.

[0271] The memory may be used to store the computer program. The processor realizes various functions of the electronic device by running or executing the computer program stored in the memory and calling the data stored in the memory.

[0272] The memory may mainly include a program storage area and a data storage area. Among them, the program storage area may store an operating system, application programs required for at least one function, etc.; the data storage area may store data created according to the use of the mobile phone, etc. In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, at least one magnetic disk storage device, a flash memory device, or other volatile solid-state storage devices.

[0273] An embodiment of the present application further provides a storage medium, which is a computer-readable storage medium. The computer program is stored in the computer-readable storage medium. When the computer program is executed by a processor, the steps of the above various method embodiments can be implemented. Among them, the computer program includes computer program code, and the computer program code can be in the form of source code, object code, executable file or some intermediate form, etc. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disc, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electrical carrier signal, telecommunication signal, and software distribution medium, etc.

[0274] An embodiment of the present application further provides a computer program product, including: a computer program or instruction. When the computer program or instruction runs on a computer, the computer is enabled to execute the method of any one of the above possible implementation manners.

[0275] The above is the preferred implementation manner of the present application. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present application, several improvements and refinements can be made, and these improvements and refinements are also regarded as the protection scope of the present application.

Claims

1. A method for quantitative characterization and modeling of the occurrence characteristics of deep geological bodies, characterized in that: The modeling method comprises: Acquire multi-scale occurrence information of deep geological bodies, wherein the multi-scale occurrence information includes: mineral structure information at a microscopic scale, physical and mechanical parameter information at an experimental scale, and geological response information at an engineering scale; Establishing a mapping relationship of the occurrence information between scales in the spatial dimension; mapping the multi-scale occurrence information to the point layer, line layer, surface layer and volume layer of the nested model constructed based on the three-dimensional coordinate grid, wherein the point layer corresponds to the mineral structure information and the physical parameter information, the surface layer corresponds to the surface structure obtained by processing the geological response information, the line layer is an intermediate layer between the point layer and the surface layer, corresponding to the linear information used to describe the continuity and directionality of the structure, the volume layer corresponds to the volume information used to reflect the regional state, and the volume information is the aggregated information of the point layer, the line layer and the surface layer in the spatial dimension; Based on the geological response information, a three-dimensional coordinate set of earthquake sources is obtained, and a spatial clustering algorithm is used to identify a high-density earthquake source cluster from the three-dimensional coordinate set of earthquake sources; a three-axis tensor inversion method based on the first motion polarity of seismic waves is used to perform focal mechanism inversion on events corresponding to the high-density earthquake source cluster, and active tectonic units are identified from the high-density earthquake source cluster to obtain the attributes of the active tectonic units, and the active and intensified spatiotemporal segments of the identified active tectonic units in the time dimension are analyzed, and the active tectonic units are identified as surface structures; the identified active tectonic units, the attributes of the active tectonic units, and the active and intensified spatiotemporal segments are recorded in a structural map; Extracting modal feature vectors from the geological response information, combining the modal feature vectors with relative time codes into a modal vector sequence, inputting all modal vector sequences into input tensors into a Transformer model for encoding and decoding to obtain a context-aware feature tensor, and inputting the context-aware feature tensor into a multi-layer perceptron MLP for state mapping to obtain a state prediction tensor; and Based on the nested model, the structural map and the state prediction tensor, a three-dimensional occurrence environment model updated over time is constructed.

2. The modeling method according to claim 1, characterized in that: The mapping relationship of endowment information between scales in the spatial dimension is established, including: Converting the mineral structure information into equivalent physical and mechanical parameters; and Based on the equivalent medium assumption, a mapping matrix between the physical and mechanical parameters and the geological response information is constructed.

3. The modeling method according to claim 1, characterized in that: The three-axis tensor inversion method based on the first motion polarity of seismic waves is used to perform focal mechanism inversion on the events corresponding to the high-density source clusters, and active tectonic units are identified from the high-density source clusters to obtain the properties of the active tectonic units, including: Determining a focal mechanism tensor according to the three-dimensional coordinates of the high-density source cluster; Performing eigenvalue decomposition on the focal mechanism tensor to obtain an eigenvector in the maximum compression direction and an eigenvector in the maximum tension direction, and determining two surface structures as candidate surface solutions, wherein the eigenvector in the maximum compression direction and the eigenvector in the maximum tension direction correspond to the stress principal axis; Based on the model tensor expression defined by the normal vector and the sliding direction vector approximated by the focal mechanism vector, obtaining the normal vector and the sliding direction vector expressed by the maximum compression direction eigenvector and the maximum tension direction eigenvector; Calculating the geometric parameters of the candidate surface structure from the normal vector and the sliding direction vector, the geometric parameters including: strike, dip and sliding angle; Comparing the candidate surface structure with the observed first-motion polarity on the epicenter projection map, and selecting a unique surface solution from the candidate surface solutions; and Using the slip angle, determining the surface structure type of the unique surface solution, The attributes of the active structural unit include: three-dimensional coordinates, stress principal axis, geometric parameters and surface structure type.

4. The modeling method according to claim 3, characterized in that: The analysis obtains the active and intensified time and space segments of the identified active structural unit in the time dimension, specifically including: Determine the energy release value and occurrence time of the event according to the geological response information; Through the sliding window method, we analyze whether the activity of the surface structure in the time dimension is increasing. If so, we extract the time-frequency information of energy release. The active structural unit attributes also include: energy release time-frequency information.

5. The modeling method according to claim 1, characterized in that: The Transformer model introduces the Mohr-Coulomb shear failure criterion as a physical constraint, and the shear failure criterion is defined as follows: τ=c+σ_n·tan(φ) Among them: τ is shear stress, σ_n is normal stress, c is cohesion, and φ is the internal friction angle.

6. The modeling method according to claim 5, characterized in that: The shear stress is equivalent to an equivalent shear stress for comparison with the shear stress reference value of the Mohr-Coulomb model. When the comparison result meets the preset conditions, at least one of the following feedback mechanisms is executed: reducing the weight of the deviation modal channel, starting redundant data path compensation, The reducing the weight of the deviation modal channel includes: counting the sensitivity of the comparison result to each modal feature vector, selecting the most sensitive modal feature vector; and lowering the weight coefficient corresponding to the most sensitive modal feature vector. The starting redundant data path compensation includes: fusing and enhancing the redundant feature vector with the context-aware feature tensor.

7. The modeling method according to claim 1, characterized in that: The Transformer model adopts a multi-head attention mechanism, and the modeling method also includes: standardizing and spatially aligning the multi-scale endowment information.

8. The modeling method according to claim 1, characterized in that: The modeling method also includes: outputting structural state parameters, risk indicators and historical evolution sequence results to a monitoring platform through a standardized interface, supporting remote online analysis, early warning and dispatch linkage control.

9. The modeling method according to any one of claims 1 to 8, characterized in that: The mineral structure information includes: mineral particle structure, grain morphology index, microcrack density and interface bonding strength; the physical and mechanical parameter information includes: elastic modulus, Poisson's ratio, longitudinal and transverse wave velocity, shear strength, peak strength and residual strength; the geological response information includes: microseismic data, energy release data, strain, temperature, humidity and resistance, and the state prediction tensor involves the following state parameters: structural principal stress value, shear energy density, displacement rate, damage factor and local damage probability.

10. A system for quantitative characterization of the occurrence characteristics of deep geological bodies, characterized in that: The system comprises: A multi-scale information acquisition module is used to acquire multi-scale occurrence information of deep geological bodies, wherein the multi-scale occurrence information includes: mineral structure information at a microscopic scale, physical and mechanical parameter information at an experimental scale, and geological response information at an engineering scale; A parameter mapping module is used to establish a mapping relationship of the occurrence information between scales in the spatial dimension; the multi-scale occurrence information is mapped to the point layer, line layer, surface layer and volume layer of the nested model constructed based on the three-dimensional coordinate grid, wherein the point layer corresponds to the mineral structure information and the physical parameter information, the surface layer corresponds to the surface structure obtained by processing the geological response information, the line layer is an intermediate layer between the point layer and the surface layer, corresponding to the linear information used to describe the continuity and directionality of the structure, the volume layer corresponds to the volume information used to reflect the regional state, and the volume information is the aggregated information of the point layer, the line layer and the surface layer in the spatial dimension; A structural identification module is used to obtain a three-dimensional coordinate set of earthquake sources based on the geological response information, and to identify a high-density earthquake source cluster from the three-dimensional coordinate set of earthquake sources by using a spatial clustering algorithm; to perform focal mechanism inversion on events corresponding to the high-density earthquake source cluster by using a three-axis tensor inversion method based on the first motion polarity of seismic waves, to identify active tectonic units from the high-density earthquake source cluster, to obtain the attributes of the active tectonic units, and to analyze and obtain the active and intensified spatiotemporal segment of the identified active tectonic units in the time dimension, wherein the active tectonic units are identified surface structures; and to record the identified active tectonic units, the attributes of the active tectonic units, and the active and intensified spatiotemporal segment into a structural map; A heterogeneous data fusion module is used to extract modal feature vectors from the geological response information, combine the modal feature vectors with relative time codes into a modal vector sequence, combine all modal vector sequences into input tensors and input them into a Transformer model for encoding and decoding to obtain a context-aware feature tensor, input the context-aware feature tensor into a multi-layer perceptron MLP for state mapping to obtain a state prediction tensor; and A three-dimensional modeling module is used to construct a three-dimensional environment model updated over time based on the nested model, the structural map and the state prediction tensor.

Images