A deep learning-based three-dimensional transparent modeling system for fractured rock mass

Abstract

This invention provides a deep learning-based 3D transparent modeling system for fractured rock masses, belonging to the field of intelligent control technology for geological exploration and engineering machinery. It includes: a data synchronization module for receiving and aligning heterogeneous waveform data; a stress mapping module for calculating the unloading stress gradient based on machinery operating status data and reducing the initial elastic modulus to generate a priori dynamic elastic modulus field; a decoupling solution module for calling a physical information network base to solve the mechanical and electromagnetic branches respectively, outputting a multi-field coupled tensor; a boundary feedback module for correcting the above multi-field coupled tensor using drilling measurement data, calculating a comprehensive geological anomaly index, and generating advance control commands; and a voxel rendering module for generating a 3D transparent geological model based on the corrected tensor and the comprehensive geological anomaly index. This invention introduces a priori physical field and a multi-field decoupling mechanism, improving the model solution convergence speed and realizing 3D visualization of complex geological anomaly areas and adaptive deceleration control of equipment.

Description

Technical Field

[0001] This invention relates to the field of geological exploration and intelligent control technology of engineering machinery, and in particular to a three-dimensional transparent modeling system for fractured rock masses based on deep learning. Background Technology

[0002] In underground engineering tunneling operations, obtaining the internal structural properties of the rock mass ahead of the working face is fundamental to ensuring construction safety. Traditional three-dimensional geological modeling methods mainly rely on static geological borehole data from the exploration stage, combined with single physical field exploration methods for spatial interpolation and numerical inversion. When facing fractured rock masses and abruptly changed strata, existing technologies have obvious shortcomings. Conventional modeling methods treat the rock mass to be tunneled as a static medium, ignoring the rock mass unloading effect caused by mechanical cutting operations. Mechanical unloading will cause stress redistribution and nonlinear degradation of elastic modulus in the rock mass ahead. Since the unloading effect is not introduced as a priori physical condition into the network inversion calculation, the physical parameters output by the existing geological model deviate significantly from the actual rock mass mechanical state.

[0003] Furthermore, single-physics field detection in complex and fractured strata can be subject to environmental noise interference, resulting in multiple solutions. Existing multiphysics field fusion deep learning models often fail to separate the physical signals by frequency. The low-frequency characteristics of mechanical response and the high-frequency characteristics of electromagnetic wave reflection overlap in the same network structure, causing frequent oscillations in the network parameter optimization process. This reduces the convergence speed and solution accuracy of multiphysics field coupled calculations. At the same time, most existing geological modeling systems are limited to generating visual images and have not established a data feedback link between geological exploration numerical extrapolation and the underlying execution unit of tunneling equipment. When approaching high-risk geological anomaly areas, the system cannot automatically calculate the safe buffer distance and generate equipment control laws based on the three-dimensional spatial boundary of the anomaly, resulting in tunneling machinery lacking adaptive deceleration and risk avoidance capabilities in abruptly changing strata.

[0004] Therefore, this application provides a deep learning-based 3D transparent modeling system for fractured rock masses to meet the requirements. Summary of the Invention

[0005] The purpose of this invention is to provide a three-dimensional transparent modeling system for fractured rock masses based on deep learning in order to solve the above-mentioned problems. By introducing a priori unloading stress field and multi-field decoupling solution, and combining abnormal boundary feedback to generate advance control commands, the problems mentioned in the background art are solved.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] A deep learning-based 3D transparent modeling system for fractured rock masses, the system comprising: The data synchronization module is used to receive mechanical operating status data and heterogeneous physical field waveform data, perform time alignment operation on the heterogeneous physical field waveform data and convert it into a one-dimensional time series matrix; The stress mapping module is used to acquire the mechanical operating state data and calculate the unloading stress gradient, and use the unloading stress gradient to perform order reduction calculation on the initial elastic modulus to generate a priori dynamic elastic modulus field. The decoupled solution module is used to call the physical information network base containing mechanical and electromagnetic branches, receive the prior dynamic elastic modulus field and the one-dimensional time series matrix, solve the mechanical and electromagnetic branches respectively, and splice the solution convergence results to output a multi-field coupling tensor. The boundary feedback module is used to convert the drilling measurement data into boundary constraint terms and feed them back to the decoupled solution module to obtain the corrected multi-field coupling tensor. Based on the corrected multi-field coupling tensor, the module calculates the comprehensive geological anomaly index and generates control commands based on the comprehensive geological anomaly index. The voxel rendering module is used to receive the corrected multi-field coupling tensor and the comprehensive geological anomaly index, convert the corrected multi-field coupling tensor into rock physics parameters and write them into voxel nodes, and fuse the comprehensive geological anomaly index to generate a three-dimensional transparent geological model.

[0008] In the above technical solution, the rock mechanical degradation phenomenon caused by mechanical excavation disturbance is transformed into prior physical field input data, providing boundary conditions with a physical basis for deep learning networks. By setting up a physical information network to process mechanical and electromagnetic data separately, and introducing drilling measurement data to perform dynamic correction, the calculation error of relying solely on static survey data is avoided. The data flow mechanism from multi-field decoupling calculation to three-dimensional voxel modeling improves the convergence speed of multi-field coupled tensor calculation in complex strata environments.

[0009] Furthermore, the heterogeneous physical field waveform data includes continuous vibration signals and electromagnetic reflection wave signals; the data synchronization module is specifically used to: obtain the step time point as the time anchor point, establish an index association based on the absolute timestamp, the step time point and the preset time sliding window, and arrange the continuous vibration signal and the electromagnetic reflection wave signal vertically according to the amplitude data of discrete sampling points to construct the one-dimensional time series matrix.

[0010] By adopting the above-mentioned further solution, in order to address the problem of inconsistent sampling frequencies of different physical field devices, data is extracted by using the time point of the work step as a reference and in conjunction with a preset time sliding window, and uniformly converted into a one-dimensional time series structure. The above-mentioned time alignment operation eliminates the misalignment of the input feature time axis caused by asynchronous sampling, and ensures the time consistency of the input matrix of the downstream neural network.

[0011] Furthermore, the stress mapping module calculates the unloading stress gradient, specifically including: extracting the propulsion force and advance speed from the mechanical operating status data, and using the product of the propulsion force and advance speed as the work parameter in the propulsion direction by combining the propulsion force work conversion coefficient; extracting the cutting head torque and angular velocity from the mechanical operating status data, and using the product of the cutting head torque and angular velocity as the work parameter in the cutting rotation direction by combining the torque work conversion coefficient; adding the work parameter in the propulsion direction and the work parameter in the cutting rotation direction, and dividing by the volume of rock broken by cutting per unit time to calculate the unloading stress gradient.

[0012] By adopting the above-mentioned further scheme, based on the correlation between rock fracture work and internal stress release, the multi-degree-of-freedom propulsion work and cutting rotation work are superimposed on a scalar, and normalized according to the actual cutting and crushing volume. The superposition and normalization operations convert multi-dimensional mechanical operation parameters into single-dimensional unloading stress gradient data, reducing the error accumulation caused by multi-level formula conversion links.

[0013] Furthermore, the stress mapping module utilizes the unloading stress gradient to perform a reduced-order calculation on the initial elastic modulus to generate a priori dynamic elastic modulus field. Specifically, this includes: using the ratio of the unloading stress gradient to the benchmark value of the uniaxial compressive strength of the rock mass as the internal damage driving force, substituting it into the exponential damage evolution equation constructed based on the brittle damage evolution parameters of the rock mass to calculate the dynamic elastic modulus at the spatial coordinate nodes; and calling a three-dimensional inverse distance weighted interpolation algorithm to smoothly map and transition the dynamic elastic modulus values ​​at the spatial coordinate nodes to the unsampled surrounding grid nodes to generate the priori dynamic elastic modulus field.

[0014] Using the above-mentioned further scheme, the nonlinear attenuation value of the elastic modulus of brittle rock mass caused by unloading is calculated using the exponential damage evolution equation. Combined with the three-dimensional inverse distance weighted interpolation algorithm, the dynamic elastic modulus of local spatial coordinate nodes is smoothly mapped to the surrounding grid area, thus constructing a continuous non-uniform reference physical field for subsequent physical partial differential equation solving.

[0015] Furthermore, the physical information network base includes a shallow feature extraction layer, as well as parallel low-frequency mechanical deep fully connected layers and high-frequency electromagnetic deep fully connected layers; the mechanical branch outputs a displacement response field and an equivalent elastic modulus feature matrix through the low-frequency mechanical deep fully connected layer; the electromagnetic branch outputs an electric field response field and a relative permittivity feature matrix through the high-frequency electromagnetic deep fully connected layer; the multi-field coupling tensor is generated by merging the equivalent elastic modulus feature matrix and the relative permittivity feature matrix along the feature dimension.

[0016] By adopting the above-mentioned further scheme, in order to address the low-frequency macroscopic response characteristics exhibited by the mechanical field and the high-frequency scattering characteristics exhibited by the electromagnetic wave, a parallel deep fully connected layer is established to perform frequency separation calculations. The separation processing path prevents redundant oscillations of low-frequency signals and smoothing filtering of high-frequency signal characteristics in a single network structure, thus ensuring the data purity of the equivalent elastic modulus characteristic matrix and the relative permittivity characteristic matrix.

[0017] Furthermore, when the decoupled solution module solves the mechanical branch and the electromagnetic branch, it constructs branch loss functions containing observation data residuals and physical equation residuals for iterative optimization. The physical equation residuals of the mechanical branch are constructed based on the partial differential equations of elastic dynamics, and the prior dynamic elastic modulus field is used as the reference physical coefficient of the partial differential equations of elastic dynamics. The physical equation residuals of the electromagnetic branch are constructed based on the partial differential equations of electromagnetic wave propagation.

[0018] By adopting the above-mentioned further scheme, the residual term of the partial differential equation is embedded in the network loss function, so that the iterative direction of the network parameter gradient optimization follows the laws of elastic dynamics and electromagnetic wave propagation. The prior dynamic elastic modulus field is introduced as the benchmark physical coefficient, which narrows the search range of the solution space and ensures that the output results conform to the laws of geological engineering under the condition of limited measured samples in deep rock mass.

[0019] Furthermore, the boundary feedback module calculates the comprehensive geological anomaly index, specifically including: calculating the degree of weakening of the equivalent elastic modulus feature matrix relative to the background elastic modulus benchmark value of the geological rock mass without anomalies, and the degree of increase of the relative permittivity feature matrix relative to the background relative permittivity benchmark value; after normalizing the degree of weakening and the degree of increase, performing weighted fusion according to the preset mechanical anomaly weight coefficient and electromagnetic anomaly weight coefficient to obtain the comprehensive geological anomaly index.

[0020] Using the above-mentioned further scheme, the geological fracture zone and water-bearing area are characterized by weakened mechanical strength and increased relative permittivity. By extracting the deviation of the equivalent elastic modulus and relative permittivity from the benchmark value and performing weighted fusion, the environmental noise present in the single physical field observation data can be filtered out, and a comprehensive evaluation index for quantifying the severity of geological anomalies can be obtained.

[0021] Furthermore, the boundary feedback module generates control commands, specifically including: extracting the comprehensive geological anomaly index that is greater than a preset risk threshold for spatial clustering, and extracting the three-dimensional geometric boundary of the continuous geological anomaly body; based on the shortest spatial Euclidean distance between the three-dimensional geometric boundary and the spatial coordinates of the tunnel boring machine cutter head, combined with the set minimum safety buffer distance, calculating the dynamic advance speed feedback control law based on the monotonic control relationship, and then generating the target advance speed command.

[0022] By adopting the above-mentioned further scheme, the three-dimensional geometric boundary of the abnormal area is determined by the spatial clustering algorithm, and the linkage feedback control logic between the safety buffer distance and the spatial coordinates of the tunneling machine cutting head is established. The dynamic advance speed feedback control law is directly generated by spatial Euclidean distance calculation, realizing the direct conversion of geological prediction values ​​into equipment control parameters, and ensuring that the tunneling equipment can autonomously output deceleration commands when approaching the risk area.

[0023] Furthermore, the voxel rendering module generates a 3D transparent geological model, specifically including: using a nonlinear transfer function to convert the comprehensive geological anomaly index into the opacity of voxel nodes and filtering out background geological data; combining a pseudo-color mapping table to convert the comprehensive geological anomaly index into red, green, and blue three-channel color values, and combining them to form the four-channel optical properties of the light sampling points; emitting virtual light rays from a virtual camera into the 3D voxel field, iteratively synthesizing and accumulating the four-channel optical properties of the light sampling points along the light propagation direction; stopping the accumulation calculation when the accumulated opacity reaches the saturation limit, and outputting a 3D visualization image.

[0024] The above-mentioned further scheme utilizes transfer functions and color mapping tables to map the numerical comprehensive geological anomaly index into the optical properties of voxel nodes, and relies on the iterative color synthesis accumulation mechanism of virtual rays to process three-dimensional data. The accumulation process can present the density gradient relationship and geometric distribution of geological anomalies in three-dimensional space, and generate a three-dimensional visualization image that removes non-anomaly background interference.

[0025] Furthermore, the boundary feedback module is also used to monitor the spatial variation gradient of the comprehensive geological anomaly index along the advancement direction. When the spatial variation gradient exceeds the preset safety baseline, the feedback control module reduces the preset time sliding window size to improve the spatial sampling resolution of the heterogeneous physical field waveform data.

[0026] Using the above-mentioned further scheme, a negative feedback control mechanism is established between the time sliding window size and the geological spatial change gradient. When the spatial change gradient of the comprehensive geological anomaly index exceeds the safe baseline, the sliding window size is actively reduced to increase the slice extraction density of the time series matrix, thereby improving the spatial resolution of the physical field waveform data for the stratigraphic abrupt change area and avoiding the smoothing elimination of small fault features by the wide time window.

[0027] Compared with the prior art, the present invention has at least the following beneficial effects: 1. This invention calculates the unloading stress gradient by extracting mechanical operating state data, and generates a priori dynamic elastic modulus field by reducing the order of the initial elastic modulus based on the damage evolution equation. This transforms the rock mechanics attenuation caused by mechanical tunneling disturbance into priori physical parameters, providing benchmark conditions that conform to engineering reality for the subsequent physical information network. This solves the problem that traditional static modeling is detached from the dynamic changes in construction, thereby narrowing the solution space search range in the network solution process and improving the convergence speed and physical accuracy of multi-field coupled tensor calculation.

[0028] 2. This invention employs a physical information network base containing mechanical and electromagnetic branches for separate solution, and embeds the corresponding elastic dynamics and electromagnetic wave propagation partial differential equation residuals into the branch loss function. Simultaneously, it introduces drilling measurement data as boundary constraint terms for feedback correction. Parallel frequency decoupling calculations prevent the characteristic aliasing of low-frequency mechanical response and high-frequency electromagnetic scattering in the same network structure. The dual constraint mechanism of physical equation residuals and measured boundary data ensures that the iteration of network model parameters strictly follows the fundamental physical laws, thereby improving the numerical accuracy of geological parameter inversion under complex strata conditions.

[0029] 3. This invention calculates a comprehensive geological anomaly index based on multi-field coupling tensors, uses a voxel rendering module to map the comprehensive geological anomaly index to optical properties to generate a three-dimensional transparent geological model, and simultaneously generates a target advance speed command based on the three-dimensional geometric boundary of the geological anomaly and the Euclidean distance between the cutting head and the cutting head. The above processing mechanism connects the data links of geological state deduction, spatial three-dimensional visualization and equipment bottom-level control. While filtering out background data and intuitively presenting the internal geological mutation morphology, it ensures that the tunneling equipment can perform autonomous deceleration response when approaching continuous geological anomalies, thus ensuring the safety of operation in abrupt geological environments. Attached Figure Description

[0030] The accompanying drawings, which are incorporated herein and form part of the specification, illustrate embodiments of the invention and, together with the specification, further serve to explain the principles of the invention and enable those skilled in the art to practice and use the invention.

[0031] Figure 1 This is a schematic diagram of the system architecture of the present invention; Figure 2 This is a schematic diagram of the method flow of the present invention; Figure 3 This is a schematic diagram of the internal processing logic of the data synchronization module of the present invention; Figure 4 This is a schematic diagram of the underlying physical calculation logic of the stress mapping module of the present invention; Figure 5 This is a schematic diagram of the network architecture and computing power scheduling logic of the decoupled solution module of the present invention; Figure 6 This is a schematic diagram of the closed-loop control logic of the boundary feedback module of the present invention; Figure 7 This is a schematic diagram of the internal rendering calculation process of the voxel rendering module of the present invention; Figure 8 This is a schematic diagram comparing the spatial inversion distribution of the equivalent elastic modulus in front of the working face of different detection algorithms of the present invention.

[0032] Figure label: 10. Data synchronization module; 20. Stress mapping module; 30. Decoupling solution module; 40. Boundary feedback module; 50. Voxel rendering module.

[0033] As shown in the figure, specific structures and devices are marked in the figure to clearly illustrate the structure of the embodiments of the present invention. However, this is only for illustrative purposes and is not intended to limit the present invention to this specific structure, device and environment. Those skilled in the art can adjust or modify these devices and environments according to specific needs. Detailed Implementation

[0034] The following is a detailed description of a deep learning-based 3D transparent modeling system for fractured rock masses provided by the present invention, with reference to the accompanying drawings and specific embodiments. It should be noted that, to make the embodiments more detailed, the following embodiments are the best and preferred embodiments; those skilled in the art can also use other alternative methods to implement some well-known technologies; and the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0035] It should be noted that the use of terms such as "an embodiment," "an embodiment," "an exemplary embodiment," and "some embodiments" in the specification indicates that the described embodiment may include a specific feature, structure, or characteristic, but not every embodiment necessarily includes that specific feature, structure, or characteristic. Furthermore, when a specific feature, structure, or characteristic is described in connection with an embodiment, implementing such a feature, structure, or characteristic in conjunction with other embodiments (whether explicitly described or not) should be within the knowledge of those skilled in the art.

[0036] Generally, terms can be understood at least partly from their use in context. For example, depending at least partly on the context, the term "one or more" as used herein can be used to describe any feature, structure, or characteristic in a singular sense, or a combination of features, structures, or characteristics in a plural sense. Additionally, the term "based on" can be understood not necessarily to convey an exclusive set of factors, but rather, alternatively, depending at least partly on the context, to allow for the presence of other factors that are not necessarily explicitly described.

[0037] See attached document Figure 1This invention provides a three-dimensional transparent modeling system for fractured rock masses based on deep learning. The system includes: a data synchronization module 10, a stress mapping module 20, a decoupling solution module 30, a boundary feedback module 40, and a voxel rendering module 50.

[0038] The data synchronization module 10 receives mechanical operation status data and heterogeneous physical field waveform data collected by the front-end heterogeneous sensing device cluster. It performs timing alignment operation on the waveform data to be processed according to a unified hardware clock protocol and converts it into a time series matrix and stores it in memory.

[0039] The stress mapping module 20 establishes a data connection with the data synchronization module 10, extracts the tunneling machine operating parameters that have been synchronized in time and space, calculates the unloading stress gradient in the spatial domain in front of the working face using the rock breaking machinery work relationship, and performs a reduction calculation on the initial elastic modulus of the specified spatial region accordingly, outputting the a priori dynamic elastic modulus field to the system network space.

[0040] The decoupling solution module 30 calls the offline trained physical information network base and synchronously receives the prior dynamic elastic modulus field and time series matrix. The internal algorithm logic of the decoupling solution module 30 independently solves the low-frequency mechanical branch and the high-frequency electromagnetic branch in the time domain. After obtaining the static convergence result, it performs tensor splicing operation in the spatial domain and outputs a multi-field coupled tensor containing multi-dimensional physical property parameters.

[0041] The boundary feedback module 40 is used to process the data feedback information generated by the support process, and to form geological anomaly identification, tunneling equipment control feedback and support design optimization parameters based on the decoupling solution results. The boundary feedback module 40 includes a drilling boundary constraint unit and an anomaly control feedback unit. The drilling boundary constraint unit acquires the drilling measurement data and converts it into physical strength reference values, constructs a Dirichlet boundary loss term and feeds it back to the decoupling solution module 30. The anomaly control feedback unit receives multi-field coupled tensors to calculate the comprehensive geological anomaly index, and then generates spatial constraint conditions and control commands such as avoidance and deceleration. It also combines the rock mass integrity index to output recommended support design parameters such as the spacing between anchor cables or anchor bolts and the support strength, so as to realize the dynamic matching of the roadway surrounding rock support scheme.

[0042] The voxel rendering module 50 receives multi-field coupling tensors, comprehensive geological anomaly indices, and boundary information of continuous geological anomalies. The voxel rendering module 50 uses rock physics conversion rules, such as using the Topp empirical formula to convert dielectric constant and volumetric water content, and converts integrity index based on the empirical relationship model of rock mass elastic modulus. It converts the independent parameters in the multi-field coupling tensors into rock mass integrity index and volumetric water content respectively. The voxel rendering module 50 generates a three-dimensional mesh with the spatial coordinates of the current tunneling machine cutting head as the origin, writes the above parameters into the corresponding three-dimensional mesh voxel nodes, and integrates the comprehensive geological anomaly index to perform visualization mapping of color and opacity, finally generating a three-dimensional transparent geological model.

[0043] See attached document Figure 2 This invention provides a method for three-dimensional transparent modeling of fractured rock masses based on deep learning, the method comprising the following steps: S100, the data spatiotemporal synchronization and preprocessing module 10 acquires the mechanical operation status data and heterogeneous physical field waveform data collected by the front-end heterogeneous sensing device cluster. The data spatiotemporal synchronization and preprocessing module 10 performs a time-series alignment operation on the mechanical operation status data and heterogeneous physical field waveform data, and converts the time-series aligned waveform data into a time series matrix.

[0044] S200, the unloading stress time-varying mapping module 20 extracts the tunneling machine operating parameters that have been synchronized in time and space, calculates the unloading stress gradient in the spatial domain in front of the working face according to the rock breaking machinery work relationship, and uses the unloading stress time-varying mapping module 20 to perform order reduction calculation on the initial elastic modulus of the specified spatial region to generate a priori dynamic elastic modulus field.

[0045] The S300 dual-scale decoupled solution module 30 receives the prior dynamic elastic modulus field and time series matrix, and calls the offline pre-trained physical information network base on the explosion-proof edge computing host. The dual-scale decoupled solution module 30 independently solves the low-frequency mechanical branch and the high-frequency electromagnetic branch in the time domain. The low-frequency mechanical branch outputs the displacement response field and the equivalent elastic modulus field, and the high-frequency electromagnetic branch outputs the electric field response field and the relative permittivity field. The displacement response field and the electric field response field are used to construct the observation data residuals and the physical equation residuals. The equivalent elastic modulus field and the relative permittivity field are used as the physical property parameters to be inverted. The dual-scale decoupled solution module 30 performs tensor splicing operation on the static convergence results of the low-frequency mechanical branch and the high-frequency electromagnetic branch in the spatial domain, and outputs a multi-field coupling tensor.

[0046] S400, the drilling data boundary feedback module 40 acquires the drilling measurement data during the anchoring drilling rig construction process, converts the drilling measurement data into physical strength reference values ​​in a specific three-dimensional spatial coordinate system, constructs a Dirichlet boundary loss term based on the physical strength reference values, and inputs the Dirichlet boundary loss term into the network computation graph of the dual-scale decoupling solution module 30. After receiving the Dirichlet boundary loss term, the dual-scale decoupling solution module 30 wakes up the network weight parameters corresponding to the current working face from the memory buffer pool of the edge computing host, applies a forced boundary constraint to the spatial solution set output by the network computation using the Dirichlet boundary loss term, and performs online secondary iteration correction on the initial spatial solution set based on the forced boundary constraint to obtain the corrected multi-field coupling tensor that satisfies the drilling boundary constraint conditions. The corrected multi-field coupling tensor is used as the input data for step S500.

[0047] S500, the dual-parameter mapping and voxel rendering module 50 receives the multi-field coupled tensor after boundary constraint correction, and uses rock physics conversion rules, such as using the Topp empirical formula to convert the dielectric constant and volumetric water content, and using the rock mass elastic modulus empirical relationship model to convert the integrity index, respectively converting the independent parameters contained in the multi-field coupled tensor into the rock mass integrity index and volumetric water content.

[0048] The dual-parameter mapping and voxel rendering module 50 generates a three-dimensional mesh with the current spatial coordinates of the tunnel boring machine cutter head as the origin. The rock mass integrity index and volumetric water content are written into the corresponding three-dimensional mesh voxel nodes. Based on the rock mass integrity index, volumetric water content, or the comprehensive geological anomaly index obtained by fusing the two, the color value and opacity value of the voxel nodes are generated to generate a three-dimensional transparent geological model. Based on the three-dimensional transparent geological model, the mechanical weakening characteristics of the surrounding rock and the hydrological anomaly boundary are extracted, thereby providing a three-dimensional visualization spatial evolution basis for the design of the advanced support and permanent support of the subsequent tunnel working face.

[0049] See attached document Figure 3 The data synchronization module 10 establishes a clock reference for heterogeneous physical field waveform data and mechanical operation status data based on a unified hardware clock protocol. The unified hardware clock protocol can adopt a precise time protocol. The front-end heterogeneous sensing device cluster synchronously starts the underlying data acquisition program based on the hardware clock reference.

[0050] The data synchronization module 10 sends control commands to the high-frequency vibration sensor installed on the cutting arm of the tunneling machine, controlling the high-frequency vibration sensor to record the broadband continuous vibration signal generated by the cutting head's rock-breaking operation. At the same time, the data synchronization module 10 sends pulse trigger commands to the explosion-proof ground radar deployed on the tunnel sidewall, controlling the explosion-proof ground radar to record the high-frequency electromagnetic reflection wave signal. Since the vibration signal is a mechanical elastic wave and the electromagnetic reflection wave signal is a high-frequency electromagnetic wave, the physical propagation speed and time frequency span of the vibration signal and the electromagnetic reflection wave signal are very different. The data synchronization module 10 sets independent sampling frequencies for the broadband continuous vibration signal and the high-frequency electromagnetic reflection wave signal to perform discretized signal acquisition and physical hardware storage.

[0051] The data synchronization module 10 calls the underlying data communication protocol of the tunneling machine monitoring, control and data acquisition system through the industrial Ethernet interface to synchronously acquire the set of mechanical operating statuses of the current cutting and propulsion stage. The set of mechanical operating statuses specifically includes propulsion force, cutting head torque, advance speed and cutting head angular velocity.

[0052] The data synchronization module 10 further extracts the three-dimensional absolute spatial coordinates of the tunneling machine working face at the current working step time point. The acquisition of the three-dimensional absolute spatial coordinates can be achieved using the existing airborne inertial navigation component and total station combined measurement technology. The data synchronization module 10 sets the three-dimensional absolute spatial coordinates as the three-dimensional spatial origin for multi-field collaborative calculation. The data synchronization module 10 provides an absolute position reference for subsequent physical parameter evolution calculation and spatial grid division through the three-dimensional spatial origin.

[0053] The data synchronization module 10 performs time alignment operation on the collected broadband continuous vibration signal and high-frequency electromagnetic reflection wave signal. Using the extracted current step time point as the time anchor point, it maps physical waveform data with different sampling frequencies to the same absolute time reference system, thereby realizing the synchronous response of the front-end heterogeneous sensing device cluster on the physical time scale.

[0054] For wideband continuous vibration signals and high-frequency electromagnetic reflection wave signals with large differences in sampling frequency, the data synchronization module 10 does not force the two to be resampled to the same sampling interval. Instead, it establishes an index association based on the absolute timestamp generated by the unified hardware clock, the current step time point, and the preset time sliding window. The time length of the preset time sliding window ranges from 0.5s to 2.0s.

[0055] Conventional geophysical waveform preprocessing methods convert waveform data into image formats. For conventional processing methods of geophysical waveform image format conversion, existing digital image grayscale mapping methods can be used. However, two-dimensional image preprocessing operations will result in high-frequency feature truncation and spatial dimensionality reduction feature loss.

[0056] The data synchronization module 10 abandons the two-dimensional image preprocessing operation. Instead, it directly maps the broadband continuous vibration signal and the high-frequency electromagnetic reflected wave signal into a one-dimensional time series matrix based on discrete sampling points. The amplitude data is arranged vertically according to time sequence to preserve the transient characteristics of the original physical waveform. The data synchronization module 10 directly stores the one-dimensional time series matrix in the memory buffer pool of the explosion-proof edge computing host. The specific calculation formula for constructing the one-dimensional time series matrix by the data synchronization module 10 is as follows: ; ; In the formula, This is a time series matrix of vibration signals; This is a discrete sampling function for the vibration signal; This refers to the sampling point number of the vibration signal; The sampling time interval for the vibration signal; This represents the total number of sampling points for the vibration signal. It is a time series matrix of electromagnetic wave signals; It is a discrete sampling function for electromagnetic wave signals; This refers to the sampling point number of the electromagnetic wave signal; The sampling time interval for electromagnetic wave signals; This represents the total number of sampling points for the electromagnetic wave signal.

[0057] See attached document Figure 4 The stress mapping module 20 provided by the present invention establishes data communication with the data synchronization module 10 to extract the tunneling machine operating parameters that have been synchronized in time and space. In terms of the implementation of the specific physical mechanism, the stress mapping module 20 uses the rock breaking machinery work relationship to calculate the unloading stress gradient in the spatial domain in front of the working face.

[0058] According to the principle of energy conservation and rock breaking energy conversion, the mechanical energy consumed by the tunneling machine during the rock breaking process is converted into unloading strain energy of the rock mass in front. During the advancement and rotation of the tunneling machine's cutting head, work is done on the rock mass in front. Part of the total energy generated by the work is directly used to break the surface rock, while the other part of the energy is transferred to the deep part of the unbroken rock mass in the form of unloading strain energy, thus forming a non-uniform stress gradient distribution in front of the working face.

[0059] The stress mapping module 20 integrates the work parameters of the tunneling machine in the digging and advancing direction and the work parameters of the tunneling machine in the cutting and rotating direction. First, it calculates the unloading stress intensity index under the current working step, and takes the current three-dimensional absolute spatial coordinates of the tunneling machine cutting head as the origin, takes the tunneling and advancing direction as the main unloading direction, and combines the cutting head attitude angle, effective cutting radius and working face normal to establish the unloading influence space domain.

[0060] The stress mapping module 20 distributes the unloading stress intensity index to different voxel nodes in front of the working face according to the spatial distance attenuation function, specifically a Gaussian attenuation distribution function or an inverse square distance attenuation function. It also determines the unloading stress gradient direction based on the spatial orientation of each voxel node relative to the cutting head, thus forming a three-dimensional unloading stress gradient field in a specific space in front. The specific formula for calculating the three-dimensional unloading stress gradient by the stress mapping module 20 is as follows: ; In the formula, The unloading stress gradient; The conversion factor for work done by propulsion force; For propulsion; This refers to the advance speed; The conversion factor for work done by torque; This refers to the cutting head torque; The cutting head angular velocity; It represents the volume of rock fragmented and cut per unit time.

[0061] Propulsion work conversion coefficient Conversion coefficient with torque for work It is a dimensionless empirical constant, obtained by fitting the factory parameters of the tunneling machine or the field rock cutting test, and its value range is set to 0.6 to 0.9.

[0062] The stress mapping module 20 calculates the volume of rock cut and broken per unit time in real time based on the product of the tunneling advance speed and the effective cross-sectional area of ​​the cutting head.

[0063] After obtaining the unloading stress gradient, the stress mapping module 20 introduces the unloading damage evolution model to perform order reduction calculation on the initial elastic modulus of the specified spatial region. In terms of model data production and parameter determination mechanism, the unloading damage evolution model called by the stress mapping module 20 is pre-constructed based on the rock mechanics experimental data of the mining area.

[0064] The specific construction method is as follows: technicians collect standard rock cores of the target mining strata and conduct multiple sets of triaxial unloading experiments under different confining pressure conditions. They record the stress-strain curves and acoustic emission characteristic data during the triaxial unloading experiments, extract the actual dynamic elastic modulus degradation values ​​of the standard rock cores under different unloading stress gradients as sample labels, establish a nonlinear mapping dataset between unloading stress gradient and dynamic elastic modulus, and then use the least squares method to optimize the nonlinear mapping dataset through numerical analysis software to solve for the rock mass brittle damage evolution parameters with the smallest fitting error. This solidifies the exponential calculation benchmark of the unloading damage evolution model.

[0065] Based on the principles of rock micro-damage mechanics, the expansion of internal micro-cracks in the rock mass under excavation unloading leads to the deterioration of macroscopic mechanical parameters. The initial elastic modulus can be obtained based on existing geological advance prediction logging experience values. The stress mapping module 20 substitutes the calculated unloading stress gradient as the internal damage driving force into the exponential damage evolution equation. The stress mapping module 20 calculates the prior dynamic elastic modulus of the current spatial domain through an exponential decay mechanism. The specific formula for the stress mapping module 20 to perform the order reduction calculation is as follows: ; In the formula, For the a priori dynamic elastic modulus; This is the initial elastic modulus; It is a natural constant; These are parameters for the evolution of brittle damage in rock mass, with values ​​ranging from 0.1 to 0.5. The unloading stress gradient; This is the benchmark value for the uniaxial compressive strength of the rock mass.

[0066] The benchmark value for the uniaxial compressive strength of rock mass can be obtained using existing laboratory rock mechanics testing methods.

[0067] The stress mapping module 20 obtains the dynamic elastic modulus at different spatial coordinate nodes in front of the working face based on the initial elastic modulus reduction calculation process of the specified spatial region. The initial elastic modulus is determined by historical geological exploration reports, well logging data or laboratory rock mechanics test results, and the benchmark value of uniaxial compressive strength of rock mass is determined by core test results or geological exploration sampling reports in the same spatial region.

[0068] The stress mapping module 20 performs three-dimensional meshing interpolation and assembly of the dynamic elastic modulus at each spatial coordinate node in the three-dimensional spatial coordinate system, and organizes the a priori dynamic elastic modulus field into a data structure containing spatial coordinates and dynamic elastic modulus values. The data structure includes at least the x, y, and z coordinates of the voxel nodes and the corresponding dynamic elastic modulus, which are used as the reference coefficient inputs when the decoupled solution module 30 calculates the physical partial differential equations.

[0069] The stress mapping module 20 divides a three-dimensional voxel grid outward based on the three-dimensional absolute spatial coordinate origin set by the system, and calls a three-dimensional inverse distance weighted interpolation algorithm to smoothly map the dynamic elastic modulus values ​​at the discrete spatial coordinate nodes to the unsampled surrounding grid nodes. The stress mapping module 20 generates a priori dynamic elastic modulus field covering the entire area to be detected. The priori dynamic elastic modulus field reflects the non-uniform distribution characteristics of the three-dimensional mechanical parameters inside the rock mass under excavation disturbance. The stress mapping module 20 outputs the priori dynamic elastic modulus field to the system network space.

[0070] The stress mapping module 20 transmits the prior dynamic elastic modulus field to the decoupling solution module 30. The decoupling solution module 30 uses the prior dynamic elastic modulus field as the baseline coefficient input when calculating the physical partial differential equation, thereby providing initial physical field constraints for the subsequent network model. This helps to reduce the convergence dead zone of the decoupling solution module 30 when calculating the time-varying equation due to the lack of physical background constraints.

[0071] See attached document Figure 5 The decoupling solution module 30 provided by this invention calls a physical information network base that is trained offline using historical geological exploration data.

[0072] For the construction of the physical information network base, the decoupling solution module 30 adopts a dual-branch physical information neural network architecture. In one network structure configuration, the physical information network base contains a shallow feature extraction layer at the multi-layer perceptron level, as well as a parallel low-frequency mechanical deep fully connected layer and a high-frequency electromagnetic deep fully connected layer.

[0073] The shallow feature extraction layer contains three hidden layers, each with 256 neurons. The low-frequency mechanical deep fully connected layer and the high-frequency electromagnetic deep fully connected layer each contain four hidden layers, each with 128 neurons. The layers are nonlinearly mapped through the hyperbolic tangent activation function. In the model training data production and preprocessing stage, technicians extract the original vibration waveforms and ground-penetrating radar waveforms obtained from historical geological exploration of the mine. They perform time-domain filtering and amplitude normalization preprocessing on the original vibration waveforms and ground-penetrating radar waveforms to remove high-frequency environmental noise and abnormal baseline drift.

[0074] The processed waveform data is combined with the corresponding spatiotemporal absolute coordinates to transform it into a normalized spatiotemporal input tensor as the model input sample. In terms of the offline training method of the model, the offline training process of the decoupled solution module 30 is deployed on a cloud server cluster. The cloud server cluster uses the normalized spatiotemporal input tensor obtained from the historical geological exploration of the mine as the input sample, and uses the laboratory measured rock mass elastic modulus and relative permittivity corresponding to the historical geological exploration sampling as the label data. The cloud server cluster feeds the normalized spatiotemporal input tensor into the shallow feature extraction layer of the physical information network base for forward propagation calculation.

[0075] After the shallow feature extraction layer and the subsequent parallel low-frequency mechanics deep fully connected layer and high-frequency electromagnetic deep fully connected layer output the predicted equivalent elastic modulus and relative permittivity, the cloud server cluster uses the predicted equivalent elastic modulus and relative permittivity with the same labeled data to calculate the initial mean square error loss. At the same time, the adaptive moment estimation optimizer is used to calculate the gradient matrix of the initial mean square error loss relative to the network nodes according to the chain rule.

[0076] The backpropagation algorithm continuously updates the connection weights and bias parameters of network nodes using the gradient matrix to minimize the mean squared error loss between the predicted value and the label data. The cloud server cluster sets a preset training round for continuous iterative calculation. The preset training round ranges from 5,000 to 10,000 times until the mean squared error loss decreases and stabilizes within a preset convergence interval. The preset convergence interval ranges from 0.001 to 0.005, thereby completing the production and construction of the physical information network base. The decoupled solution module 30 obtains the physical information network base that has converged in the cloud training and distributes the physical information network base to the explosion-proof edge computing host.

[0077] The decoupled solution module 30 executes computing power scheduling rules on the explosion-proof edge computing host to adapt to the limited computing power resources at the edge. The decoupled solution module 30 freezes the network weight parameters of the shallow feature extraction layer of the physical information network base. The decoupled solution module 30 only opens the low-frequency mechanical deep fully connected layer and the high-frequency electromagnetic deep fully connected layer to participate in gradient update fine-tuning.

[0078] The decoupling solution module 30 receives the prior dynamic elastic modulus field output by the stress mapping module 20. The decoupling solution module 30 synchronously reads the one-dimensional time series matrix generated by the data synchronization module 10. The one-dimensional time series matrix contains the vibration signal time series and the electromagnetic wave signal time series. The specific data flow logic is as follows: the decoupling solution module 30 performs tensor concatenation of the vibration signal time series and the electromagnetic wave signal time series with the current three-dimensional spatial coordinates and time nodes, and then inputs them into the shallow feature extraction layer. The prior dynamic elastic modulus field output by the stress mapping module 20 is used as the physical coefficient input or additional feature input of the low-frequency mechanics branch.

[0079] The shallow feature extraction layer jointly encodes the multi-source heterogeneous inputs to extract deep-dimensional spatiotemporal shared latent features. Then, the decoupled solution module 30 separates the spatiotemporal shared latent features and inputs them into the low-frequency mechanics deep fully connected layer and the high-frequency electromagnetic deep fully connected layer for independent forward inference calculation.

[0080] The low-frequency mechanical deep fully connected layer is equipped with a displacement response output head and an equivalent elastic modulus output head, while the high-frequency electromagnetic deep fully connected layer is equipped with an electric field response output head and a relative permittivity output head. The displacement response output head and the electric field response output head are used to establish data residuals with the measured waveform data, while the equivalent elastic modulus output head and the relative permittivity output head are used to form material property parameter channels in the multi-field coupling tensor.

[0081] The internal algorithm logic of the decoupled solution module 30 independently solves the low-frequency mechanics branch and the high-frequency electromagnetic branch in the time domain. In terms of the physical principle of introducing partial differential equation constraints, the physical information neural network calculates the partial derivative of the network output with respect to the spatiotemporal input through an automatic differentiation mechanism, thereby transforming the partial differential equation that controls the evolution of the physical field into a soft constraint penalty term for network training. This makes the prediction results of the neural network not only fit the measured data but also strictly follow the physical laws.

[0082] In the calculation of the low-frequency mechanics branch, the decoupling solution module 30 uses the prior dynamic elastic modulus field output by the stress mapping module 20 as the reference physical coefficients of the partial differential equation of elastic dynamics. The decoupling solution module 30 employs the Latin hypercube sampling algorithm to randomly generate unlabeled mechanical and physical configuration points within a set three-dimensional spatial domain and time window. The decoupling solution module 30 constructs low-frequency mechanical and physical residuals to constrain the prediction space of the neural network. These residuals are established based on the elastic wave propagation equation. The decoupling solution module 30 combines the observation data error and the partial differential equation residuals to calculate the total loss function of the low-frequency mechanics branch. The specific formula for calculating the loss function of the low-frequency mechanics branch by the decoupling solution module 30 is as follows: ; In the formula, This is the low-frequency mechanical branch loss function; Weights for mechanical data residuals; This represents the total number of mechanical observation data points. This refers to the sequence number of the mechanical observation data point; The first prediction for the low-frequency mechanics branch Displacement response at discrete time points; This represents the actual observed displacement amplitude in the vibration signal time series. For mechanical and physical residual weights; The total number of mechanical and physical configuration points; Assign point numbers to mechanics and physics configuration points; This serves as the benchmark value for rock mass density. It is a time variable; For the a priori dynamic elastic modulus; For the low-frequency mechanics branch in the first Predicted displacement response at a mechanical and physical configuration point.

[0083] In the calculation of the high-frequency electromagnetic branch, the decoupling solution module 30 constructs the high-frequency electromagnetic physical residual using the electromagnetic wave propagation partial differential equation obtained by simplifying Maxwell's equations. The decoupling solution module 30 also uses the Latin hypercube sampling algorithm to generate electromagnetic physical placement points in the same spatiotemporal domain. The decoupling solution module 30 independently calculates the total loss function of the high-frequency electromagnetic branch. The specific formula for calculating the high-frequency electromagnetic branch loss function by the decoupling solution module 30 is as follows: ; In the formula, For high-frequency electromagnetic branch loss function; Weights for electromagnetic data residuals; This represents the total number of electromagnetic observation data points. This refers to the serial number of the electromagnetic observation data point; The first prediction for the high-frequency electromagnetic branch Electric field intensity at discrete time points; This represents the actual observed electric field amplitude in the time series of the electromagnetic wave signal. The weights are for electromagnetic physical residuals; The total number of electromagnetic physics configuration points; Assign point numbers to electromagnetic physics configuration points; It is a time variable; For the high-frequency electromagnetic branch in the first Predicted electric field strength at each electromagnetic physical configuration point; The speed of light in a vacuum; The relative permittivity is predicted for the high-frequency electromagnetic branch.

[0084] The decoupled solution module 30 performs iterative network calculations on the low-frequency mechanical branch and the high-frequency electromagnetic branch respectively through an adaptive moment estimation optimization algorithm. The iteration stops when the loss functions of the low-frequency mechanical branch and the high-frequency electromagnetic branch are both less than the convergence threshold. The convergence threshold is set according to the on-site detection accuracy requirements. In the specific engineering implementation environment, the value of the convergence threshold is between 0.0001 and 0.001.

[0085] The decoupled solution module 30 obtains the static convergence results of each independent branch. The static convergence results output by the low-frequency mechanics branch include a displacement response field containing three-dimensional spatial coordinate information and an equivalent elastic modulus characteristic matrix. The displacement response field is used to participate in the calculation of the observation residual and physical residual of the loss function of the low-frequency mechanics branch, and the equivalent elastic modulus characteristic matrix serves as the property inversion result of the low-frequency mechanics branch. The static convergence results output by the high-frequency electromagnetic branch include an electric field response field containing three-dimensional spatial coordinate information and a relative permittivity characteristic matrix.

[0086] The electric field response field is used to participate in the calculation of the observation residual and physical residual of the high-frequency electromagnetic branch loss function, and the relative permittivity characteristic matrix serves as the property inversion result of the high-frequency electromagnetic branch. The decoupling solution module 30 uses a unified three-dimensional absolute space origin as the alignment reference and performs tensor splicing operations in the spatial domain, merging the equivalent elastic modulus characteristic matrix and the relative permittivity characteristic matrix along the characteristic dimension. The decoupling solution module 30 outputs a multi-field coupled tensor containing multi-dimensional property parameters. This multi-field coupled tensor is represented in terms of dimension as a comprehensive high-dimensional matrix containing coordinate information, modulus information, and permittivity information. The splicing calculation formula for the multi-field coupled tensor generated by the decoupling solution module 30 is as follows: ; In the formula, For multi-field coupling tensors; This is a learnable channel weight matrix; This is the characteristic matrix of the equivalent elastic modulus; This is the characteristic matrix of the relative permittivity.

[0087] See attached document Figure 6 The boundary feedback module 40 provided by the present invention establishes data communication with the decoupled solution module 30, and respectively undertakes the tasks of network boundary constraint feedback, spatial quantitative identification of geological anomalies, dynamic updating of boundary conditions of tunneling machine control system, and dynamic optimization of collaborative support design.

[0088] During the network boundary constraint feedback process, the boundary feedback module 40 acquires the drilling measurement data through the data acquisition gateway on the anchored drilling rig, converts the drilling measurement data into physical strength reference values ​​at the borehole trajectory points, and constructs the physical strength reference values ​​as Dirichlet boundary loss terms before returning them to the network calculation graph of the decoupled solution module 30 to correct the spatial solution set output by the decoupled solution module 30.

[0089] During the spatial quantitative identification of geological anomalies, the boundary feedback module 40 acquires a multi-field coupled tensor containing three-dimensional spatial coordinate information. By performing numerical analysis on the equivalent elastic modulus feature matrix and the relative permittivity feature matrix inside the multi-field coupled tensor, the spatial distribution characteristics of hidden disaster-causing geological factors in front of the tunneling machine working face are extracted.

[0090] During the dynamic update of boundary conditions in the tunneling machine control system, the boundary feedback module 40 generates avoidance space constraints, deceleration space constraints, and target advance speed commands based on the three-dimensional geometric boundary, center coordinates, and shortest spatial distance between the continuous geological anomaly and the tunneling machine cutting head.

[0091] For spatial three-dimensional visualization rendering of hidden disaster-causing geological factors, existing three-dimensional volume rendering technology can be used. Based on the principle of multi-source physical field fusion evaluation, the boundary feedback module 40 performs dimensionless processing on the physical parameters in the equivalent elastic modulus feature matrix and the relative permittivity feature matrix by calculating the relative deviation rate with the normal background value, and then weightedly fuses and maps them into a comprehensive geological anomaly index.

[0092] Specifically, for the equivalent elastic modulus feature matrix, the boundary feedback module 40 prioritizes calculating the degree of weakening of the current equivalent elastic modulus relative to the background elastic modulus benchmark value of the geological rock mass without anomalies, in order to characterize the fracture zone, weak interlayer, or collapse risk area; for the relative permittivity feature matrix, the boundary feedback module 40 calculates the degree of increase of the current relative permittivity relative to the background relative permittivity benchmark value, in order to characterize the water-rich anomaly area.

[0093] The boundary feedback module 40 normalizes the mechanical weakening degree and the electromagnetic water-rich degree separately and then performs a weighted fusion to obtain a comprehensive geological anomaly index. This comprehensive geological anomaly index is used to quantitatively evaluate the risk of water inrush or collapse in the rock mass ahead of the tunneling machine's working face, and serves as a core evaluation indicator for adaptively adjusting support design parameters, such as mesh support density and shotcrete thickness. The specific formula for calculating the comprehensive geological anomaly index by the boundary feedback module 40 is as follows: ; In the formula, For comprehensive geological anomaly index; This is the mechanical anomaly weighting coefficient, and its value ranges from 0.4 to 0.6. This is the characteristic matrix of the equivalent elastic modulus; The baseline value for the background elastic modulus of rock masses without abnormal geological features; This is the electromagnetic anomaly weighting coefficient, and its value ranges from 0.4 to 0.6. The characteristic matrix of relative permittivity; The background relative permittivity reference value is the value for rock masses without any abnormal geological features.

[0094] After obtaining the comprehensive geological anomaly index covering the detection area ahead, the boundary feedback module 40 executes spatial clustering logic and early warning logic. The boundary feedback module 40 traverses the comprehensive geological anomaly index in the three-dimensional spatial grid nodes and extracts the set of anomaly nodes that are greater than the preset risk threshold. The preset risk threshold ranges from 0.6 to 0.8. The setting of the preset risk threshold can be based on the relevant experience values ​​of the existing mine safety regulations. The boundary feedback module 40 uses a density-based spatial clustering algorithm to perform spatial connectivity analysis on the set of anomaly nodes in the three-dimensional spatial grid.

[0095] The boundary feedback module 40 employs a density-based spatial clustering method with noise, setting the basic resolution of the system's three-dimensional mesh as the neighborhood search radius and setting the minimum number of continuous voxels required to constitute a geological anomaly as the density threshold. This determines the three-dimensional geometric boundary and center coordinates of the continuous geological anomaly. The boundary feedback module 40 converts the three-dimensional geometric boundary and center coordinates of the continuous geological anomaly into the avoidance space constraints and deceleration space constraints of the tunnel boring machine's automatic cutting control system. The boundary feedback module 40 generates a dynamic advance speed feedback control law based on the relative distance between the continuous geological anomaly and the current three-dimensional absolute spatial coordinates of the tunnel boring machine's cutting head.

[0096] The dynamic advance speed feedback control law is used to reduce the cutting thrust and advance speed of the tunnel boring machine (TBM) when it approaches a continuous geological anomaly to prevent dynamic disasters. The dynamic advance speed feedback control law satisfies the following monotonic control relationship: when the shortest spatial Euclidean distance between the TBM cutting head and the three-dimensional geometric boundary of the continuous geological anomaly increases, the target advance speed command gradually approaches the rated advance speed under normal geological conditions; when the shortest spatial Euclidean distance decreases, the target advance speed command gradually decreases, and triggers speed reduction or shutdown protection when it falls below the minimum safety buffer distance set by the system. The specific formula for calculating the dynamic advance speed feedback control law by the boundary feedback module 40 is as follows: ; In the formula, The feedback is the adjusted target advance speed command; This is the rated advance rate under normal geological conditions; The minimum safe buffer distance set for the system; It represents the shortest spatial Euclidean distance between the tunnel boring machine cutter head and the three-dimensional geometric boundary of the continuous geological anomaly.

[0097] The boundary feedback module 40 sends the target advance speed command to the tunneling machine monitoring, control and data acquisition system to complete the closed-loop control of the physical operating boundary of the tunneling machine. At the same time, the boundary feedback module 40 dynamically adjusts the time sliding window size of the data synchronization module 10 according to the change gradient of the comprehensive geological anomaly index. When the comprehensive geological anomaly index in front of the tunneling machine working face increases and the spatial change gradient along the tunneling advance direction exceeds the preset safety baseline, the safety baseline value ranges from 0.1 to 0.3. The boundary feedback module 40 reduces the time sliding window size according to the inverse proportional mapping logic to improve the spatial sampling resolution of the high-frequency vibration sensor and the explosion-proof ground radar.

[0098] The boundary feedback module 40 feeds back the dynamically adjusted time sliding window size to the data synchronization module 10. The boundary feedback module 40 forms a complete data flow closed-loop feedback system covering the physical perception environment, the edge decoupled computing environment to the mechanical equipment execution environment. The complete closed-loop feedback system helps to overcome the engineering limitations of traditional geological exploration systems that lack equipment linkage control capabilities to a certain extent.

[0099] See attached document Figure 7 The voxel rendering module 50 provided by the present invention establishes data communication connections with the decoupling solution module 30 and the boundary feedback module 40 respectively. The voxel rendering module 50 obtains the multi-field coupling tensor after boundary constraint correction from the decoupling solution module 30, and obtains the comprehensive geological anomaly index and continuous geological anomaly body boundary information from the boundary feedback module 40.

[0100] The voxel rendering module 50 generates discrete three-dimensional mesh data containing three-dimensional absolute spatial coordinates, rock mass integrity index and volumetric water content based on the multi-field coupling tensor, and generates corresponding color values ​​and opacity values ​​based on the comprehensive geological anomaly index. In terms of the physical implementation principle of three-dimensional visualization, the voxel rendering module 50 uses direct volume rendering technology to convert discrete three-dimensional mesh data into images that can be displayed on a two-dimensional display screen.

[0101] Direct volume rendering technology can directly render the internal structure of a 3D data field without extracting surface geometric primitives. This technology helps to fully showcase the gradual changes and spatial connectivity of geological anomalies within a rock mass. The voxel rendering module 50 maps discrete 3D mesh data into a 3D voxel field containing color and opacity information. Based on the numerical distribution range of the comprehensive geological anomaly index, the voxel rendering module 50 constructs a nonlinear transfer function. This transfer function is used to assign corresponding optical properties to each 3D spatial coordinate node and filter out normal geological background. The specific formula for calculating the opacity of voxel nodes in the voxel rendering module 50 is as follows: ; In the formula, Opacity of voxel nodes; This is the opacity scaling factor, which ranges from 5 to 15. For comprehensive geological anomaly index; This is the threshold for anomaly concealment.

[0102] The acquisition of the abnormal hidden threshold for filtering out normal background rock mass no longer relies solely on manual experience. In specific engineering implementation environments, the value range of the abnormal hidden threshold for filtering out normal background rock mass is set to 80% to 100% of the aforementioned preset risk threshold to adaptively filter out safe background data.

[0103] After calculating and obtaining the voxel node opacity, the voxel rendering module 50 converts the comprehensive geological anomaly index into red, green and blue three-channel color values ​​according to the preset pseudo-color mapping table. Areas with high geological anomaly index are mapped to high-saturation warm colors, and areas with low geological anomaly index are mapped to low-saturation cool colors. The voxel rendering module 50 combines the voxel node opacity with the red, green and blue three-channel color values ​​to form a complete four-channel optical property.

[0104] The voxel rendering module 50 initiates a ray casting algorithm to perform projection calculations in the viewpoint screen space. Starting from the viewpoint of the virtual camera, the voxel rendering module 50 emits virtual rays through each pixel of the two-dimensional display screen into the three-dimensional voxel field. The virtual rays are resampled within the three-dimensional voxel field at equal intervals. The interpolation calculation during the resampling process can be performed using existing three-dimensional linear interpolation algorithms. In terms of the physical principle of optical synthesis, the color synthesis accumulation logic simulates the absorption and emission physical process of light propagating in a semi-transparent medium. The multiplication term characterizes the optical attenuation effect caused by the virtual rays being gradually blocked by the voxel nodes in front along the light path.

[0105] The voxel rendering module 50 iteratively accumulates the four-channel optical properties of the light sampling points from front to back along the direction of virtual ray propagation until the virtual ray penetrates the three-dimensional voxel field or the accumulated voxel node opacity reaches the saturation limit. To avoid redundant calculations and algorithm dead zones in the ray casting algorithm, in a specific engineering implementation environment, when the accumulated voxel node opacity reaches 0.95, the system determines that the saturation limit has been reached and immediately triggers the early termination logic of the virtual ray. The specific formula for calculating the final color of the screen pixel by the voxel rendering module 50 is as follows: ; In the formula, The final accumulated color value for each screen pixel; This represents the total number of ray sampling points on the virtual ray; This is the current light sampling point number; For virtual light on the first The color value of each light sampling point; For virtual light on the first Opacity of each light sampling point; This refers to the sequence number of the previous ray sampling point preceding the current ray sampling point; For virtual light on the first The opacity of each light sampling point.

[0106] The voxel rendering module 50 traverses all pixels of the two-dimensional display screen to complete ray projection and color synthesis calculations. The voxel rendering module 50 generates a three-dimensional visualization image containing the three-dimensional geological structure in front of the tunneling machine's working face and the boundaries of continuous geological anomalies. The voxel rendering module 50 outputs the three-dimensional visualization image to the explosion-proof display terminal of the tunneling machine's operating cabin. At the same time, the voxel rendering module 50 responds to the interactive commands input by the operator. The interactive commands include spatial rotation, viewpoint translation, model scaling, and profile cutting operations. The voxel rendering module 50 updates the viewpoint position and projection matrix of the virtual camera in real time according to the interactive commands and re-triggers the ray projection calculation logic.

[0107] The three-dimensional visualization image constructed by the voxel rendering module 50 maps the spatial three-dimensional location and severity of hidden disaster-causing geological factors. The three-dimensional visualization image assists the tunneling machine operator in expanding the perception range of the unknown geological environment ahead. The voxel rendering module 50, together with the dynamic advance speed feedback control law and the recommended parameters for support design output by the boundary feedback module 40, provides human-machine collaborative decision support for the safe and intelligent tunneling and refined support design of deep underground engineering.

[0108] To aid in understanding the technical solution of this invention, a specific application example of three-dimensional transparent modeling of fractured rock mass based on deep learning is provided below.

[0109] This application example uses a deep rock tunnel excavation face with a design depth of -850 meters in a certain mining area as the test object. Historical geological exploration data shows that there is a hidden water-rich fault fracture zone in the area 30 meters in front of the tunnel face. The three-dimensional transparent modeling system for fractured rock mass based on deep learning is equipped with an explosion-proof high-frequency vibration sensor with a sampling frequency of 50kHz at the front end of the tunneling machine's cutting arm, an explosion-proof ground radar with a center frequency of 100MHz is deployed on the tunnel sidewall, and an explosion-proof edge computing host is deployed in the airborne explosion-proof cavity.

[0110] After the tunneling operation starts, the data synchronization module 10 performs timing alignment operation on the broadband vibration signal and electromagnetic reflection signal according to the precise time protocol, extracts a 1.0s time sliding window and converts the physical waveform data within the time sliding window into a one-dimensional time series matrix. The stress mapping module 20 extracts the current tunneling machine's thrust of 456kN and the cutting head torque of 82kN·m, calculates the three-dimensional unloading stress gradient in front of the tunneling working face, and performs a reduction calculation on the initial background elastic modulus of 45.2GPa based on the three-dimensional unloading stress gradient, outputting the a priori dynamic elastic modulus field. The a priori dynamic elastic modulus field value of the excavated disturbance area decreases to 31.4GPa.

[0111] The decoupling solution module 30 calls the physical information network base in the explosion-proof edge computing host and receives the a priori dynamic elastic modulus field as the reference input parameter. The physical information network base solves the low-frequency mechanical branch and the high-frequency electromagnetic branch independently in the time domain and outputs a multi-field coupled tensor containing three-dimensional spatial coordinates, equivalent elastic modulus and relative permittivity. The boundary feedback module 40 obtains the drilling measurement data of the anchoring drill rig and converts it into physical strength reference value. The boundary feedback module 40 uses the physical strength reference value to construct a Dirichlet boundary loss term to perform spatial forced correction on the multi-field coupled tensor.

[0112] The voxel rendering module 50 uses the Topp empirical formula to convert the relative permittivity greater than 24.5 in the multi-field coupling tensor into a volume water content of 32.6%. The voxel rendering module 50 generates a three-dimensional transparent geological model containing a three-dimensional mesh through direct volume rendering technology.

[0113] The three-dimensional transparent geological model generates a highly opaque, red, water-rich fractured zone mapping area 11.3 meters in front of the tunneling working face. The boundary feedback module 40 triggers the dynamic advance speed feedback control law and sends instructions to the tunneling machine monitoring, control and data acquisition system.

[0114] The target advance speed command was automatically adjusted from the rated advance speed of 0.85m / min to 0.18m / min. At the same time, the system automatically generated an early warning for advanced pipe roof grouting based on the spatial boundary of the water-rich fractured zone, and guided the on-site operators to change the support design scheme of this section from conventional anchor cable support to U-shaped steel retractable support superimposed with deep hole grouting reinforcement.

[0115] Table 1: Comparison of Performance Test Data of Physical Property Parameter Inversion and 3D Modeling of Different Detection Algorithms in Deep Rock Tunneling Faces

[0116] According to Table 1 and Figure 8 It can be seen that the root mean square error of the equivalent elastic modulus of the full waveform inversion algorithm is 1.137 GPa, but the time taken for a single spatial three-dimensional modeling iteration is 47.38 s, which cannot meet the real-time control frequency requirements of the edge end of the tunneling equipment. The pure data-driven convolutional neural network model reduces the time taken for a single spatial three-dimensional modeling iteration to 2.14 s, but the lack of physical mechanism constraints causes the root mean square error of the equivalent elastic modulus to rise to 2.684 GPa.

[0117] The deep learning-based 3D transparent modeling method for fractured rock masses introduces residual constraints from elastic dynamics and Maxwell's partial differential equations during the model training phase, keeping the root mean square error of the equivalent elastic modulus at 1.216 GPa. The accuracy of physical parameter inversion is close to that of the full waveform inversion algorithm. The decoupled solution module 30 freezes the shallow feature extraction layer of the physical information network base and only performs fine-tuning calculations on the weights of deep network nodes, further reducing the time consumption of a single spatial 3D modeling iteration to 1.37s.

[0118] In addition, the boundary feedback module 40 constructs a Dirichlet boundary loss term using measurement-while-drilling data and performs boundary constraint correction on the network prediction results. After introducing measurement-while-drilling data, the spatial positioning deviation of the anomaly center coordinates is reduced from 1.17m in the convolutional neural network model to 0.42m, and the geometric boundary identification deviation of the anomaly is reduced from 1.25m to 0.58m. The three-dimensional transparent modeling method of fractured rock mass based on deep learning meets the timeliness requirements of edge end calculation while ensuring spatial modeling accuracy. It also outputs a dynamic advance speed feedback control law based on the spatial geometric calculation results to realize closed-loop control of the physical operating boundary of the tunneling equipment.

[0119] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0120] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A three-dimensional transparent modeling system for fractured rock masses based on deep learning, characterized in that, The system includes: The data synchronization module is used to receive mechanical operating status data and heterogeneous physical field waveform data, perform time alignment operation on the heterogeneous physical field waveform data and convert it into a one-dimensional time series matrix; The stress mapping module is used to acquire the mechanical operating state data and calculate the unloading stress gradient, and use the unloading stress gradient to perform order reduction calculation on the initial elastic modulus to generate a priori dynamic elastic modulus field. The decoupled solution module is used to call the physical information network base containing mechanical and electromagnetic branches, receive the prior dynamic elastic modulus field and the one-dimensional time series matrix, solve the mechanical and electromagnetic branches respectively, and splice the solution convergence results to output a multi-field coupling tensor. The boundary feedback module is used to convert the drilling measurement data into boundary constraint terms and feed them back to the decoupled solution module to obtain the corrected multi-field coupling tensor. Based on the corrected multi-field coupling tensor, the module calculates the comprehensive geological anomaly index and generates control commands based on the comprehensive geological anomaly index. The voxel rendering module is used to receive the corrected multi-field coupling tensor and the comprehensive geological anomaly index, convert the corrected multi-field coupling tensor into rock physics parameters and write them into voxel nodes, and fuse the comprehensive geological anomaly index to generate a three-dimensional transparent geological model.

2. The deep learning-based three-dimensional transparent modeling system for fractured rock masses according to claim 1, characterized in that, The heterogeneous physical field waveform data includes continuous vibration signals and electromagnetic reflection wave signals; The data synchronization module is specifically used to: obtain the time point of the work step as the time anchor point, establish an index association based on the absolute timestamp, the time point of the work step and the preset time sliding window, and arrange the continuous vibration signal and the electromagnetic reflection wave signal vertically according to the amplitude data of discrete sampling points to construct the one-dimensional time series matrix.

3. The deep learning-based three-dimensional transparent modeling system for fractured rock masses according to claim 1, characterized in that, The stress mapping module calculates the unloading stress gradient, specifically including: Extract the propulsion force and advance speed from the mechanical operating status data, and use the product of propulsion force and advance speed as the work parameter in the propulsion direction by combining the propulsion force work conversion coefficient; extract the cutting head torque and angular velocity from the mechanical operating status data, and use the product of cutting head torque and angular velocity as the work parameter in the cutting rotation direction by combining the torque work conversion coefficient; add the work parameter in the propulsion direction and the work parameter in the cutting rotation direction, and divide by the volume of rock broken by cutting per unit time to calculate the unloading stress gradient.

4. The deep learning-based three-dimensional transparent modeling system for fractured rock masses according to claim 1, characterized in that, The stress mapping module uses the unloading stress gradient to perform a reduction-order calculation on the initial elastic modulus to generate a priori dynamic elastic modulus field, specifically including: The ratio of the unloading stress gradient to the benchmark value of the uniaxial compressive strength of the rock mass is used as the internal damage driving force. It is substituted into the exponential damage evolution equation constructed based on the brittle damage evolution parameters of the rock mass to calculate the dynamic elastic modulus at the spatial coordinate nodes. The three-dimensional inverse distance weighted interpolation algorithm is called to smoothly map the dynamic elastic modulus values ​​at the spatial coordinate nodes to the unsampled surrounding grid nodes to generate the a priori dynamic elastic modulus field.

5. The deep learning-based three-dimensional transparent modeling system for fractured rock masses according to claim 1, characterized in that, The physical information network base includes a shallow feature extraction layer, as well as a parallel low-frequency mechanical deep fully connected layer and a high-frequency electromagnetic deep fully connected layer. The mechanical branch outputs a displacement response field and an equivalent elastic modulus characteristic matrix through the low-frequency mechanical deep fully connected layer; the electromagnetic branch outputs an electric field response field and a relative permittivity characteristic matrix through the high-frequency electromagnetic deep fully connected layer; the multi-field coupling tensor is generated by merging the equivalent elastic modulus characteristic matrix and the relative permittivity characteristic matrix along the characteristic dimension.

6. The deep learning-based three-dimensional transparent modeling system for fractured rock masses according to claim 5, characterized in that, When the decoupled solution module solves the mechanical branch and the electromagnetic branch, it constructs branch loss functions containing observation data residuals and physical equation residuals respectively for iterative optimization. The physical equation residuals of the mechanics branch are constructed based on the partial differential equations of elastic dynamics, and the a priori dynamic elastic modulus field is used as the reference physical coefficient of the partial differential equations of elastic dynamics; the physical equation residuals of the electromagnetic branch are constructed based on the partial differential equations of electromagnetic wave propagation.

7. The deep learning-based three-dimensional transparent modeling system for fractured rock masses according to claim 5, characterized in that, The boundary feedback module calculates the comprehensive geological anomaly index, specifically including: The degree of weakening of the equivalent elastic modulus feature matrix relative to the background elastic modulus benchmark value of the geological rock mass without anomalies, and the degree of increase of the relative permittivity feature matrix relative to the background relative permittivity benchmark value are calculated respectively. After normalizing the degree of weakening and the degree of increase, a weighted fusion is performed according to the preset mechanical anomaly weight coefficient and electromagnetic anomaly weight coefficient to obtain the comprehensive geological anomaly index.

8. The deep learning-based three-dimensional transparent modeling system for fractured rock masses according to claim 1, characterized in that, The boundary feedback module generates control commands, specifically including: The comprehensive geological anomaly index exceeding the preset risk threshold is extracted for spatial clustering, and the three-dimensional geometric boundary of the continuous geological anomaly body is extracted. Based on the shortest spatial Euclidean distance between the three-dimensional geometric boundary and the spatial coordinates of the tunnel boring machine cutter head, combined with the set minimum safety buffer distance, the dynamic advance speed feedback control law is calculated based on the monotonic control relationship, and then the target advance speed command is generated.

9. The deep learning-based three-dimensional transparent modeling system for fractured rock masses according to claim 1, characterized in that, The voxel rendering module generates a three-dimensional transparent geological model, specifically including: The comprehensive geological anomaly index is converted into the opacity of voxel nodes using a nonlinear transfer function and background geological data is filtered out. Combined with a pseudo-color mapping table, the comprehensive geological anomaly index is converted into red, green and blue three-channel color values, which are combined to form the four-channel optical properties of the light sampling points. Virtual light is emitted from a virtual camera into the three-dimensional voxel field, and the four-channel optical properties of the light sampling points are iteratively synthesized and accumulated along the light propagation direction. The accumulation calculation stops when the accumulated opacity reaches the saturation limit, and a three-dimensional visualization image is output.

10. The deep learning-based three-dimensional transparent modeling system for fractured rock masses according to claim 2, characterized in that, The boundary feedback module is also used to monitor the spatial variation gradient of the comprehensive geological anomaly index along the advancement direction. When the spatial variation gradient exceeds the preset safety baseline, the feedback control module reduces the preset time sliding window size to improve the spatial sampling resolution of the heterogeneous physical field waveform data.

Images