A three-dimensional visualization reconstruction method and system for omnidirectional stress in mine boreholes

Abstract

This application provides a three-dimensional visualization reconstruction method and system for omnidirectional stress in mine boreholes, relating to the field of mine monitoring technology. The method includes: firstly, simultaneously acquiring optical images, acoustic signals, and reference image data of the borehole wall during drilling; then, processing the optical images and acoustic signals to generate strain time-series data and event distribution data respectively; next, inputting the reference image data and strain time-series data into a dual-branch convolutional neural network to fuse and generate a dynamic strain field; subsequently, coupling the dynamic strain field and event distribution data to obtain multi-physics field coupled data; furthermore, using a physical information neural network combined with elasticity equations to perform physical constraint optimization on this data, obtaining three-dimensional stress field data of the borehole surrounding rock; finally, simulating and calculating future stress trends based on this stress field data to generate a three-dimensional visualization reconstruction result. This application improves the accuracy and early warning capability of three-dimensional visualization reconstruction of the dynamic stress field of the surrounding rock in deep mine boreholes.

Description

Technical Field

[0001] This application relates to the field of mine monitoring technology, and in particular to a three-dimensional visualization reconstruction method and system for omnidirectional stress in mine boreholes. Background Technology

[0002] Three-dimensional visualization reconstruction of the surrounding rock stress in mine boreholes is of great significance for the safe mining of deep resources. This method can intuitively present the spatial distribution and dynamic evolution of the stress field around the borehole, providing key basis for engineering decisions such as roadway support design and rockburst early warning, and has broad application prospects.

[0003] Existing methods typically rely on a single type of sensor data for stress inversion. For example, some schemes mainly use borehole cameras or optical means to observe borehole wall deformation and then calculate the stress state, while others focus on monitoring acoustic emission events and indirectly assessing the degree of stress concentration through micro-fracture activity. These methods first acquire observational data of a single physical field, then perform calculations using empirical formulas or simplified mechanical models, and finally generate static or quasi-static stress distribution images.

[0004] However, in complex environments with high stress and dynamic disturbances at depth, the stress field of the surrounding rock exhibits strong non-uniformity and time-varying characteristics. Inversion results based on a single data source have limitations in reflecting the absolute value of stress and spatial details. At the same time, they lack the ability to predict the dynamic evolution of the stress field over time, especially future trends. Therefore, existing technologies face the technical challenge of achieving high-precision, high-reliability three-dimensional visualization and advanced prediction of the omnidirectional dynamic stress field of mine boreholes. Summary of the Invention

[0005] This application provides a three-dimensional visualization reconstruction method and system for omnidirectional stress in mine boreholes, which solves the problems of low accuracy and poor early warning capability in the three-dimensional visualization reconstruction of dynamic stress field of surrounding rock in deep mine boreholes in the prior art.

[0006] To address the aforementioned technical problems, in a first aspect, this application provides a three-dimensional visualization reconstruction method for omnidirectional stress in mine boreholes, comprising:

[0007] During the drilling process in the mine, optical image data, acoustic signal data, and reference image data of the borehole wall are acquired.

[0008] Displacement field calculations are performed on the optical image data to generate strain time series data, and the acoustic signal data is extracted and localized to generate event distribution data.

[0009] The reference image data and the strain time series data are input into a dual-branch convolutional neural network for processing to generate a dynamic strain field.

[0010] The dynamic strain field and the event distribution data are coupled and analyzed using a Convolutional Long Short-Term Memory Network (Conv-LSTM) to generate multiphysics coupled data.

[0011] The multiphysics coupling data is input into a physical information neural network, which performs constraint optimization on the multiphysics coupling data based on the elasticity equation to obtain three-dimensional stress field data of the borehole surrounding rock.

[0012] Based on the three-dimensional stress field data, the distribution trend of omnidirectional stress in the borehole surrounding rock over a future set time period is simulated and calculated, generating a visualized reconstruction result.

[0013] Optionally, the step of inputting the reference image data and the strain time-series data into a dual-branch convolutional neural network for processing to generate a dynamic strain field includes:

[0014] The reference image data is input into the first branch of the dual-branch convolutional neural network. The first branch extracts the first stress direction data and stripe spacing data from the reference image data. Based on the first stress direction data and the stripe spacing data, the reference stress distribution is determined.

[0015] The strain time series data is input into the second branch of the dual-branch convolutional neural network, and the second branch performs spatiotemporal convolution operation on the strain time series data to extract condensed data and pattern data.

[0016] The reference stress distribution is transferred to the second branch, and amplitude and spatial corrections are performed on the lumped data and the mode data based on the reference stress distribution;

[0017] The output layer of the dual-branch convolutional neural network fuses and reconstructs the corrected lumped data and corrected pattern data to output a dynamic strain field.

[0018] Optionally, the step of performing amplitude and spatial correction on the lumped data and the mode data based on the reference stress distribution includes:

[0019] Within the second branch, a spatial transformation network is introduced, which pre-establishes a mapping relationship between the reference stress distribution, lumped data, and model data.

[0020] Based on the mapping relationship, according to the stress values ​​at each spatial location in the reference stress distribution, the amplitude correction parameters corresponding to each spatial location are generated using a multilayer perceptron regression model, and the corresponding spatial correction parameters are generated according to the gradient change direction of the reference stress distribution.

[0021] The amplitude correction parameter is multiplied element-wise with the ensemble data and the mode data to achieve amplitude correction.

[0022] Spatial correction is achieved by performing affine transformation operations using the spatial correction parameters and the local features of the ensemble data and the pattern data.

[0023] Optionally, the step of using a Conv-LSTM network to perform coupling analysis on the dynamic strain field and the event distribution data to generate multiphysics coupling data includes:

[0024] The dynamic strain field is divided into multiple strain gauge data in chronological order;

[0025] The event distribution data is divided into multiple time periods according to time sequence, and the kernel density of the event location information in each time period is estimated to generate an event density map.

[0026] The strain gauge data for each time period are combined with the corresponding event density map to form combined data;

[0027] The combined data is input into the Conv-LSTM network, and spatial features are extracted from the combined data through the convolutional layers of the Conv-LSTM network.

[0028] The spatial features are temporally memorized and updated through the recurrent memory unit of the Conv-LSTM network. The recurrent memory unit introduces an attention mechanism during the processing. The attention mechanism calculates spatial weights based on the event density map and uses the spatial weights to adjust the influence of spatial features at different locations on memory.

[0029] The feature generation layer of the Conv-LSTM network generates feature data that integrates temporal and spatial information based on the final memory state output by the recurrent memory unit.

[0030] The feature data is input into the multilayer perceptron of the Conv-LSTM network for fully connected processing to generate multi-physics coupled data.

[0031] Secondly, this application provides a three-dimensional visualization reconstruction system for omnidirectional stress in mine boreholes, comprising:

[0032] The acquisition module is used to acquire optical image data, acoustic signal data, and reference image data of the borehole wall during the drilling process in the mine.

[0033] The calculation module is used to perform displacement field calculation on the optical image data to generate strain time series data, and to extract and locate the acoustic signal data to generate event distribution data.

[0034] The input module is used to input the reference image data and the strain time series data into a dual-branch convolutional neural network for processing to generate a dynamic strain field;

[0035] The analysis module is used to perform coupled analysis of the dynamic strain field and the event distribution data using a Conv-LSTM network to generate multiphysics coupled data.

[0036] An optimization module is used to input the multi-physics coupling data into a physical information neural network. The physical information neural network performs constraint optimization on the multi-physics coupling data based on the elasticity equation to obtain three-dimensional stress field data of the borehole surrounding rock.

[0037] The generation module is used to simulate and calculate the distribution trend of omnidirectional stress in the borehole surrounding rock over a future set time period based on the three-dimensional stress field data, and generate a visualized reconstruction result.

[0038] Thirdly, this application provides an electronic device, comprising:

[0039] Memory, used to store computer programs;

[0040] A processor is configured to execute the computer program to implement the steps of the three-dimensional visualization reconstruction method for omnidirectional stress in mine boreholes as described in the first aspect above.

[0041] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the steps of the three-dimensional visualization reconstruction method for omnidirectional stress in mine boreholes as described in the first aspect above.

[0042] The technical solution provided in this application has the following beneficial effects:

[0043] First, simultaneous perception of the multi-physics field of the surrounding rock condition was achieved by acquiring optical images, acoustic signals, and reference images of the borehole wall. Then, displacement field calculation and event extraction and localization were performed on the optical images and acoustic signals respectively, generating detailed data reflecting the dynamic strain and micro-fracture activity of the borehole wall. Next, a dual-branch convolutional neural network was used to fuse the reference image and strain time series data, effectively combining the absolute stress reference with high-resolution dynamic strain characteristics. Subsequently, a Conv-LSTM network was used to couple the dynamic strain field and event distribution data to capture the spatiotemporal correlation between stress-strain and micro-fracture activity.

[0044] Furthermore, by combining physical information neural networks with elasticity equations for constraint optimization, the final three-dimensional stress field data strictly conforms to physical laws, improving the reliability of the inversion results. Finally, based on the optimized stress field data, future trend simulation calculations are performed, ultimately generating a three-dimensional visualization result that clearly shows the dynamic process of stress redistribution, providing an intuitive basis for engineering early warning.

[0045] Furthermore, when processing with a dual-branch convolutional neural network, the first stress direction and fringe spacing are extracted from the reference image data to determine the reference stress distribution, while lumped data and mode data are extracted from the strain time series data. Then, the reference stress distribution is passed to the second branch, and the amplitude and spatial correction of the lumped data and mode data are performed based on this. Finally, the dynamic strain field is output by fusing and reconstructing the corrected data.

[0046] Furthermore, the effect of this process is that it effectively integrates and calibrates the absolute stress reference provided by the photoelastic method with the high-resolution dynamic strain observations obtained by the digital image correlation method. This results in a dynamic strain field that simultaneously possesses a reliable absolute value reference and fine spatiotemporal evolution details, laying a high-quality intermediate data foundation for subsequent accurate stress field inversion.

[0047] These or other aspects of this application will become more apparent from the description of the following embodiments. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 A flowchart illustrating a three-dimensional visualization reconstruction method for omnidirectional stress in mine boreholes, provided as an embodiment of this application;

[0050] Figure 2 A schematic diagram illustrating a specific implementation of a three-dimensional visualization reconstruction method for omnidirectional stress in a mine borehole, provided in this application embodiment;

[0051] Figure 3 This is a schematic diagram of the structure of a three-dimensional visualization reconstruction system for omnidirectional stress in a mine borehole, provided in an embodiment of this application. Detailed Implementation

[0052] To address the problems of existing technologies, this application proposes a three-dimensional visualization reconstruction method for omnidirectional stress in mine boreholes. The core idea of ​​this method is as follows: First, during borehole construction, optical images of the borehole wall, acoustic emission signals, and a reference stress image obtained through photoelastic freezing technology are simultaneously acquired, constructing a multi-source isomorphic perceptual data foundation. Then, a dual-branch convolutional neural network is used to perform feature alignment and physical correction between the photoelastic fringe image representing the absolute stress reference and strain time-series data reflecting high-resolution dynamic changes, generating a dynamic strain field with both accurate values ​​and fine details. Next, a convolutional long short-term memory network is used to deeply couple the dynamic strain field with the spatiotemporal distribution of acoustic emission events, capturing the intrinsic correlation between stress-strain evolution and micro-fracture activity. Finally, a physical information neural network is used to place the above multi-physics coupled data under the framework of elasticity for constraint optimization, reversing a three-dimensional dynamic stress field that strictly conforms to physical laws, and based on this, a visual simulation of future stress redistribution trends is performed.

[0053] Therefore, this series of collaborative processing steps, for the first time, deeply integrates three types of heterogeneous information—photoelastic absolute reference, speckle dynamic strain, and acoustic emission events—within a unified physical framework. This effectively compensates for the shortcomings of single data sources in terms of spatial integrity, quantitative accuracy, and temporal predictability, thereby achieving high-precision and high-reliability three-dimensional reconstruction and advanced prediction of the omnidirectional dynamic stress field of the surrounding rock in mine boreholes, providing a more effective technical means for the safety of deep engineering.

[0054] To enable those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some embodiments of the present application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0055] The core of this application is to provide a three-dimensional visualization reconstruction method for omnidirectional stress in mine boreholes, and a flowchart of one specific implementation is shown below. Figure 1 As shown, the method includes:

[0056] Step 101: During the drilling process in the mine, acquire optical image data, acoustic signal data, and reference image data of the borehole wall.

[0057] In step 101, optical image data refers to a series of images obtained by photographing speckle patterns pre-made or naturally formed on the surface of the borehole wall, used to record the minute deformation of the borehole wall rock under external force; acoustic signal data refers to the elastic wave signals emitted by micro-fractures generated inside the rock during drilling, which are received by sensors and converted into electrical signal sequences.

[0058] Reference image data refers to a special image obtained through a physical experimental method called "dynamic photoelastic stress freezing". This method projects polarized light onto a special transparent material around the borehole and records the instantaneous stress distribution as alternating bright and dark stripes by rapidly changing the material state. This image provides an absolute reference for the magnitude and direction of stress around the borehole.

[0059] In this embodiment, the process is implemented synchronously by an integrated detection device arranged inside the borehole. This device is equipped with a high-speed camera, an acoustic emission sensor, and a photoelastic imaging module. During drilling or operation, the high-speed camera continuously captures speckle images of the borehole wall to obtain optical image data, and the acoustic emission sensor collects elastic waves generated by micro-fractures in the rock mass in real time to obtain acoustic signal data. During the intervals when the operation is paused, the photoelastic imaging module is activated to project polarized light onto the medium around the borehole and rapidly change its physical state, thereby capturing and recording reference image data characterizing the current instantaneous stress field.

[0060] Step 102: Perform displacement field calculation on the optical image data to generate strain time series data, and extract and locate the acoustic signal data to generate event distribution data.

[0061] Among them, strain time series data refers to a data sequence consisting of multiple strain fields arranged in chronological order, used to describe the continuous evolution of borehole wall strain state over time; event distribution data refers to a structured dataset formed by associating the occurrence time, characteristic parameters and occurrence location information of identified acoustic emission micro-fracture events in three-dimensional space and organizing them in chronological order, used to characterize the spatiotemporal evolution of micro-fracture activity inside the surrounding rock.

[0062] In this embodiment, step 102 includes the following process:

[0063] Step 1021: Based on the optical image data, determine the coordinate changes of the corresponding speckle at adjacent time points, and obtain the displacement data of the borehole wall based on the coordinate changes.

[0064] In step 1021, speckle refers to a small area that is pre-made on the surface of the borehole wall or formed by the natural texture of the rock and appears as a contrast between light and dark in the image; coordinate change refers to the difference in the center position of the same speckle in the pixel coordinate system in the optical image at different times.

[0065] Displacement data refers to the set of data obtained by analyzing and calculating optical images acquired at consecutive time points, which is used to quantify the distance and direction of movement of each tiny region on the borehole wall surface between adjacent time points. This set of data represents the overall deformation and motion of the borehole wall in the form of a vector field.

[0066] In this embodiment, a digital image correlation method is used. First, multiple speckle points are defined as reference points in the optical image at the first moment. Then, in the optical image at the second moment, the best matching position of each reference speckle point is searched and determined by calculating the cross-correlation of the image sub-regions, thereby obtaining the displacement vector of each speckle point. Finally, by spatially interpolating the displacement vectors of all speckle points, a displacement data field covering the entire aperture wall image region is generated.

[0067] In practical applications, assuming that the pixel coordinates of a speckle in the image at time 1 are [50, 100] and the corresponding coordinates in the image at time 2 are [52, 98], then the displacement vector of the speckle is 2 pixels in the horizontal direction and -2 pixels in the vertical direction. By performing bilinear interpolation on the displacement vectors of hundreds or thousands of speckles, a two-dimensional displacement field data describing the movement of each point on the surface of the hole wall is finally obtained.

[0068] Step 1022: Calculate the strain data of the borehole wall based on the displacement data, and combine the strain data corresponding to each time point to form strain time series data.

[0069] In step 1022, strain data refers to a set of physical quantities that describe the degree of local relative deformation of the material, obtained by spatial differential calculation of displacement data. The strain data includes normal strain and shear strain components.

[0070] Strain time series data refers to a three-dimensional data volume formed by arranging and combining strain field data that describe the deformation state of the entire borehole wall surface, calculated at multiple consecutive time points, in chronological order. This data volume represents the distribution of strain on the borehole wall in two spatial dimensions and the continuous evolution of this distribution in the time dimension. It is used to comprehensively record and reflect the dynamic change history of the strain state of the borehole wall surrounding rock over time under construction disturbance.

[0071] In this embodiment, the displacement data obtained in step 1021 is first differentially differentiated in two directions in the image space to calculate the strain tensor components at each pixel location. Then, the strain tensor data of all pixels are arranged according to their original image row and column positions to form a strain field representing the strain state of the entire field at that moment. Finally, the displacement and strain calculations of step 1021 and this step are repeated on the optical image data at multiple consecutive moments, and the strain fields at each moment are arranged in chronological order to form the strain time series data.

[0072] In practical applications, by differentiating the displacement data, the normal strain along the horizontal direction at a certain point can be calculated as 0.002, the normal strain along the vertical direction as 0.001, and the shear strain as 0.0005. Arranging these strain values ​​of all pixels in the current image according to their positions yields the strain field at that moment. Combining the strain fields of multiple consecutive moments such as moment 1, moment 2, and moment 3 in chronological order forms strain time series data describing the dynamic changes of the hole wall strain.

[0073] Step 1023: Identify the pulse signal in the acoustic signal data, and determine the source location data of the pulse signal in the borehole surrounding rock based on the time when the pulse signal arrives at multiple sensors.

[0074] In step 1023, the pulse signal refers to a short-duration, high-amplitude signal segment with specific waveform characteristics that is excited by micro-fractures occurring inside the rock in the acoustic signal data stream; the source location data refers to the coordinates of the micro-fracture event in three-dimensional space calculated using geometric principles by analyzing the time difference between the arrival of the same pulse signal at multiple sensors arranged at different locations in space.

[0075] In this embodiment, firstly, by setting an amplitude threshold and a short-time energy threshold, pulse signal segments exceeding the background noise level are detected and segmented from the continuous acoustic signal data stream. Subsequently, for each identified pulse signal, the absolute time of its arrival at at least four acoustic emission sensors with different spatial coordinates is accurately picked up. Finally, based on the known propagation speed of sound waves in the surrounding rock medium, a system of equations is established with the source position coordinates as unknowns and the coordinates of each sensor and the arrival time difference as known conditions. By solving this system of equations, the source position data of the pulse signal is obtained.

[0076] In practical applications, assuming that the propagation speed of sound waves in the surrounding rock is known to be 4200 meters per second, and that a pulse signal arrives at sensors A, B, C, and D at 1.000 seconds, 1.002 seconds, 1.001 seconds, and 1.005 seconds respectively, the spatial coordinates of the micro-fracture event can be calculated as [1.5, 0.3, -2.1] meters by establishing and solving a hyperbolic equation system based on the propagation speed and time difference.

[0077] Step 1024: Associate the parameters of each pulse signal with the corresponding source location data and organize them in chronological order to generate event distribution data.

[0078] In step 1024, the parameters of the pulse signal refer to the feature values ​​extracted from the identified pulse signal waveform to describe its physical characteristics such as energy and frequency. The parameters of the pulse signal include peak amplitude, dominant frequency and rise time.

[0079] In this embodiment, firstly, for each pulse signal identified in step 1023, the peak amplitude, dominant frequency, and signal rise time are extracted from its waveform as its characteristic parameters; then, a data record is created, and the trigger timestamp of the pulse signal, the extracted characteristic parameters, and the source location data calculated in step 1023 are assigned to the corresponding fields of the record; finally, the data records corresponding to all pulse signals are sorted from earliest to latest according to their trigger timestamps, thereby generating an ordered event distribution data table containing multi-dimensional information of time, parameters, and spatial location.

[0080] In practical applications, the event record corresponding to a micro-fracture event may include: a timestamp of 1024.56 seconds, a peak amplitude of 1.2 volts, a dominant frequency of 120 kHz, a rise time of 8 microseconds, and source location coordinates of [1.5, 0.3, -2.1] meters. By sorting all such event records over the entire monitoring period by time, event distribution data is formed for analyzing the spatiotemporal evolution of micro-fractures.

[0081] This application processes optical images and acoustic signals into structured temporal strain fields and spatiotemporal event datasets through the above process, providing a precise and coordinated input data foundation for subsequent multiphysics fusion and high-precision reconstruction of three-dimensional stress fields.

[0082] Step 103: Input the reference image data and the strain time series data into a dual-branch convolutional neural network for processing to generate a dynamic strain field.

[0083] Among them, the dynamic strain field refers to the strain data field obtained after physical benchmark correction, which contains both high spatiotemporal resolution details and absolute mechanical quantity significance. Its spatial distribution and temporal evolution process are physically consistent with the actual dynamic deformation state of rocks.

[0084] It should be noted that this embodiment does not limit the structural design, parameter design, and training process of the dual-branch convolutional neural network, and can be set accordingly based on the actual situation.

[0085] In this embodiment, step 103 includes the following process, such as... Figure 2 As shown:

[0086] Step 1031: Input the reference image data into the first branch of the dual-branch convolutional neural network. The first branch extracts the first stress direction data and stripe spacing data from the reference image data. Based on the first stress direction data and the stripe spacing data, determine the reference stress distribution.

[0087] In step 1031, the first branch is a network part specifically designed to process static reference image data. Structurally, this branch includes multiple convolutional layers for extracting local features of the image, pooling layers for filtering important features, and fully connected layers for outputting stress direction and fringe spacing parameters. Its core function is to extract quantitative stress information from the photoelastic interference fringe image.

[0088] The first stress direction data refers to the direction of the principal stress at each point on the borehole wall, inferred from the interference fringe pattern in the reference image; the fringe spacing data refers to the average pixel distance between two adjacent bright or dark fringes, which is inversely proportional to the stress magnitude; the reference stress distribution refers to the complete data field of quantitative stress values ​​and their directions at each point on the borehole wall, calculated by integrating the stress direction and fringe spacing information.

[0089] In this embodiment, a first branch of a pre-trained convolutional neural network is used. The network weights of this branch are learned through training on a large number of known stress states and their corresponding photoelastic fringe image samples. During inference, the first branch receives reference image data, and its internal convolutional layers automatically extract the fringe direction and density features. Finally, a fully connected layer outputs a quantitative estimate of the principal stress direction and fringe spacing for each image location. Based on the photoelasticity formula, the fringe spacing estimate for each location is converted into a stress value, and combined with the principal stress direction estimate for that location, a reference stress distribution describing the magnitude and direction of the absolute stress across the entire field is constructed. The expression of the photoelasticity formula can be found in related technologies and will not be elaborated here.

[0090] In practical applications, after the first branch of the network processes the input reference image, it may output that the principal stress direction at a certain location is at an angle of 30 degrees to the horizontal axis and the stripe spacing is 5 pixels. Based on the pre-calibrated conversion coefficient between pixel spacing and stress, for example, each pixel corresponds to 2 MPa, the stress value at that location can be calculated to be approximately 10 MPa, thus determining the stress state of that point in the reference stress distribution.

[0091] Step 1032: Input the strain time series data into the second branch of the dual-branch convolutional neural network, and perform spatiotemporal convolution operation on the strain time series data to extract lumped data and pattern data.

[0092] In step 1032, the second branch is a network part specifically designed to process dynamic strain time series data. This branch introduces a three-dimensional convolution kernel in its structure, enabling its convolutional layer to extract features in both the spatial and temporal dimensions of the image simultaneously, thereby capturing the spatial aggregation phenomenon and the temporal evolution law of strain. Its output is intermediate feature data characterizing the spatiotemporal change pattern of strain.

[0093] Centralized data refers to the characteristic information identified from strain time series data that characterizes areas of high spatial concentration of strain, while pattern data refers to the characteristic information extracted from strain time series data that characterizes the patterns and trends of strain changes over time.

[0094] In this embodiment, the operation is performed through the second branch of a dual-branch convolutional neural network, which is a neural network with a three-dimensional convolutional kernel. The network receives time-stacked strain time-series data, and its three-dimensional convolutional kernel performs sliding calculations simultaneously in two spatial dimensions and one temporal dimension, thereby capturing the spatial aggregation phenomenon of strain and the temporal evolution pattern. After multiple layers of such convolution, pooling, and nonlinear activation operations, the network finally outputs feature maps representing the spatial concentration of strain (i.e., concentrated data) and latent vectors representing the temporal evolution pattern of strain (i.e., pattern data) at the high-level feature layer.

[0095] In practical applications, the strain time-series data input to the second branch may be a three-dimensional data block with a size of 100 pixels × 100 pixels × 10 time steps. Through three-dimensional convolution, the network may identify that the upper left corner of the image consistently exhibits a high strain value during the 5th to 8th time steps, and this feature is encoded as lumped data. At the same time, the network may identify that the strain value of the entire region first increases and then decreases over time, and this feature is encoded as pattern data.

[0096] Step 1033: Transfer the reference stress distribution to the second branch, and perform amplitude correction and spatial correction on the lumped data and the mode data based on the reference stress distribution.

[0097] Step 1033 may specifically include the following steps:

[0098] A1: In the second branch, a spatial transformation network is introduced, which pre-establishes the mapping relationship between the reference stress distribution, lumped data and model data.

[0099] In step A1, the spatial transformation network is a learnable neural network module whose function is to learn the geometric transformation relationship between the spatial coordinate system of the reference stress distribution and the feature spatial coordinate system implied by the lumped data and pattern data.

[0100] In this embodiment of the application, the spatial transformation network takes the reference stress distribution output in step 1031 as input and predicts a set of spatial transformation parameters through a lightweight sub-network. These parameters define a differentiable geometric transformation, such as an affine transformation or a thin plate spline transformation, with the aim of optimally aligning the positions of the lumped data and pattern data in the feature space with the positions of the reference stress distribution in the physical space.

[0101] In practical applications, a spatial transformation network may predict six affine transformation parameters. These parameters together define a transformation matrix that includes translation, rotation, and scaling. This matrix describes how to geometrically deform the lumped data feature map so that the strain concentration area it represents can better match the location of the high-stress area in the reference stress distribution.

[0102] A2: Based on the mapping relationship, according to the stress values ​​at each spatial location in the reference stress distribution, the amplitude correction parameters corresponding to each spatial location are generated using a multilayer perceptron regression model, and the corresponding spatial correction parameters are generated according to the gradient change direction of the reference stress distribution.

[0103] In step A2, both the multilayer perceptron regression model and the spatial transformation network are sub-modules or sub-structures introduced within the second branch to achieve specific correction functions. Specifically, the spatial transformation network, as a whole module, is a component or processing stage within it, which uses the multilayer perceptron regression model to perform the task of generating amplitude correction parameters.

[0104] Therefore, the multilayer perceptron regression model is contained in the spatial transformation network at the functional level, and the spatial transformation network, as the core part of realizing the correction function described in step 1033, is itself a functional sub-module in the second branch network structure.

[0105] Amplitude correction parameters are multiplicative coefficients used to scale the values ​​of lumped data and model data; spatial correction parameters are geometric transformation parameters used to adjust the spatial position of lumped data and model data, such as elements in the aforementioned affine transformation matrix.

[0106] In this embodiment, firstly, the stress value at each location in the reference stress distribution is input into a shared multilayer perceptron, which outputs an amplitude scaling factor proportional to the stress value at that location as an amplitude correction parameter; secondly, the gradient field of the reference stress distribution is calculated to obtain the direction of the fastest stress change at each point, and this direction information is combined with the initial transformation parameters predicted in step A1, and optimized through a regression layer to generate the final spatial correction parameters for fine adjustment.

[0107] In practical applications, for a region with a stress value of 20 MPa in the reference stress distribution, the multilayer sensor may output an amplitude correction parameter of 1.5; for a stress gradient direction of 45 degrees in this region, the regression layer may fine-tune the initial rotation angle parameter based on this to generate a spatial correction parameter that contains a more accurate angle.

[0108] A3: Multiply the amplitude correction parameter element-wise with the ensemble data and the mode data to achieve amplitude correction.

[0109] In this embodiment, the amplitude correction parameter matrix generated in step A2, which corresponds one-to-one with the spatial location, is multiplied element-wise with the lumped data feature map extracted in step 1032; at the same time, the amplitude correction parameter scalar generated based on the global stress level is multiplied with the mode data vector extracted in step 1032; through this operation, the originally dimensionless or relative dimensionless lumped data and mode data are uniformly corrected to an absolute mechanical dimension scale with reference to the reference stress.

[0110] In practical applications, if the eigenvalue of the lumped data at a certain location is 0.8 and the corresponding amplitude correction parameter is 1.5, then the eigenvalue at that location becomes 1.2 after correction. This new value has the same order of magnitude significance as the reference stress distribution.

[0111] A4: Spatial correction is achieved by performing affine transformation operations using the spatial correction parameters and the local features of the lumped data and the pattern data.

[0112] In step A4, local features refer to data units or feature vectors extracted from ensemble data or pattern data that characterize the strain change pattern of a specific small region of an image or its adjacent time period;

[0113] Specifically, in this application, the local features of lumped data refer to the sub-blocks formed by all feature values ​​within a neighborhood window centered on a certain pixel on its two-dimensional feature map. These sub-blocks reflect the strain aggregation characteristics of the central point and its surrounding area. In contrast, the local features of model data refer to the feature segments in its temporal feature sequence associated with a specific time step, reflecting the strain evolution pattern near that moment.

[0114] In this embodiment, the affine transformation matrix constructed using the spatial correction parameters generated in step A2 is used to perform spatial resampling on the lumped data feature map. Specifically, for each target position in the corrected feature map, the source position coordinates corresponding to it in the original lumped data feature map are calculated based on the affine transformation matrix. Then, the feature value of the source position is obtained from the original feature map through bilinear interpolation and assigned to the target position. This process is equivalent to performing an overall rotation, translation, or scaling on the strain spatial distribution represented by the lumped data to precisely align it with the spatial pattern of the reference stress distribution. As a global time feature, the model data is usually not subjected to spatial geometric transformation, but it may be adjusted in relation to the overall results of spatial correction.

[0115] In practical applications, the spatial correction parameter defines a transformation that transforms coordinates [x, y] into [0.9×x-0.1×y+2, 0.1×x+0.9×y-1]. Therefore, the eigenvalue of a point with coordinates [10, 10] in the corrected feature map is taken from the eigenvalue of a point with coordinates approximately [10.9, 9.1] in the original dataset. This mapping achieves spatial deformation and alignment of the feature map.

[0116] Step 1034: The corrected lumped data and the corrected pattern data are fused and reconstructed through the output layer of the dual-branch convolutional neural network to output a dynamic strain field.

[0117] In this embodiment of the application, the corrected lumped data feature map obtained in steps A3 and A4 is concatenated with the corrected pattern data vector; the concatenated joint features are first upsampled through a series of deconvolution layers to gradually restore the spatial size, and different features are fused through convolution layers; finally, the high-dimensional features are mapped to the target number of channels through an output convolution layer to generate the final dynamic strain field, which is a strain data sequence that is continuous in time and has an absolute stress reference in space.

[0118] In practical applications, the corrected ensemble data is a feature map with a size of 25×25×64, and the corrected pattern data is a vector with a length of 128. After concatenation, upsampling and convolutional fusion, the final output is a dynamic strain field with a size of 100×100×N, where N represents the number of time steps. The value of each pixel in the field at each time point represents the strain state after correction by the physical reference.

[0119] In the above process, this application utilizes photoelastic reference to correct high-resolution strain observations to achieve dual accurate calibration of the dynamic strain field in terms of both physical quantities and spatial distribution, providing high-quality input data with a reliable absolute reference for subsequent stress field inversion.

[0120] Step 104: Use a Conv-LSTM network to perform coupling analysis on the dynamic strain field and the event distribution data to generate multiphysics coupling data.

[0121] Among them, multiphysics coupling data refers to a comprehensive data expression that can comprehensively reflect the intrinsic relationship between strain evolution and micro-fracture activity after the dynamic strain field information characterizing rock deformation and the event distribution data information characterizing rock fracture are deeply integrated in the time and space dimensions.

[0122] It should be noted that this embodiment does not limit the structural design, parameter design, and training process of the Conv-LSTM network, and can be set accordingly based on the actual situation.

[0123] In this embodiment, step 104 includes the following process:

[0124] Step 1041: Divide the dynamic strain field into multiple strain gauge data in chronological order.

[0125] In step 1041, strain gauge data refers to a data slice of the dynamic strain field at a specific point in time or within a very short time period. This slice contains the spatial distribution information of the strain of the borehole wall at that moment.

[0126] In this embodiment of the application, the dynamic strain field data is divided into a series of discrete two-dimensional strain distribution maps arranged in chronological order based on the index of the dynamic strain field data in the time dimension, with fixed time intervals or time points, thereby decomposing the continuous dynamic strain field into a series of discrete two-dimensional strain distribution maps arranged in chronological order, where each map represents a strain gauge data.

[0127] Step 1042: Divide the event distribution data into multiple time periods according to time order, and perform kernel density estimation on the event location information in each time period to generate an event density map.

[0128] In step 1042, the event density map refers to a two-dimensional image generated by statistically smoothing the spatial location information of all micro-fracture events that occur within a specific time period. The brightness or value of each pixel in the image represents the density of events occurring within a unit area near that location.

[0129] In this embodiment, the entire monitoring time axis is first divided into multiple time periods that are time-aligned with the strain gauge data in step 1041. Then, for each time period, the source location coordinates of all micro-fracture events occurring within that time period are extracted. Finally, using a kernel density estimation algorithm, the influence of each event point is smoothly diffused within its surrounding spatial region. By superimposing the diffusion contributions of all event points, an event density map with the same spatial resolution as the strain gauge data and numerically representing the spatial aggregation density of events is generated.

[0130] Step 1043: Combine the strain gauge data for each time period with the corresponding event density map to form combined data.

[0131] In step 1043, the combined data refers to the multi-channel two-dimensional data block formed by stitching together strain gauge data and event density maps of the same time period in the channel dimension.

[0132] In this embodiment, the strain gauge data and the event density map are first aligned perfectly on the spatial grid. Then, the strain gauge data is used as the first channel and the event density map is used as the second channel. These two data layers are stacked in the channel direction, so that each spatial location point is simultaneously assigned a strain value and an event density value, forming a dual-channel combined data sample. This operation is repeated for each aligned time period to form a series of time-series combined data.

[0133] Step 1044: Input the combined data into the Conv-LSTM network, and extract spatial features from the combined data through the convolutional layers of the Conv-LSTM network.

[0134] In step 1044, spatial features refer to abstract information extracted from each channel of the combined data that can characterize the strain pattern and event density pattern of the local area.

[0135] In this embodiment, the convolutional layer of the Conv-LSTM network processes the combined data at each time step of the input; the convolutional layer contains multiple two-dimensional convolutional kernels, which slide on the two spatial dimensions of the combined data and perform convolution operations on the strain information of the first channel and the event density information of the second channel, respectively, thereby extracting a spatial feature map containing local strain change patterns and event aggregation patterns.

[0136] Step 1045: The spatial features are processed for temporal memory and update through the recurrent memory unit of the Conv-LSTM network. The recurrent memory unit introduces an attention mechanism during the processing. The attention mechanism calculates spatial weights based on the event density map and uses the spatial weights to adjust the influence of spatial features at different locations on memory.

[0137] In this embodiment of the application, at each time step, the recurrent memory unit receives the current spatial features extracted by step 1044; at the same time, a parallel attention module receives the original event density map of the corresponding time step and calculates a spatial weight matrix based on the value of each pixel position in the event density map, wherein high event density regions are assigned higher weights.

[0138] Then, the spatial weight matrix is ​​used to modulate the spatial features of the current input, so that the features of high event density regions are enhanced when input into the recurrent memory unit. Finally, the recurrent memory unit merges and updates the current spatial features after weight modulation with its own stored memory state of the previous time step to form a new memory state.

[0139] Step 1046: Through the feature generation layer of the Conv-LSTM network, feature data that integrates temporal and spatial information is generated based on the final memory state output by the recurrent memory unit.

[0140] In step 1046, the feature generation layer is one or more subsequent convolutional or fully connected layers in the Conv-LSTM network. Its function is to decode the final memory state output by the recurrent memory unit, which contains the entire temporal information, into a feature representation that is easier for subsequent networks to understand and integrates spatiotemporal information.

[0141] In this embodiment of the application, after processing the combined data of all time steps, the final memory state output by the recurrent memory unit contains the accumulated information of the entire time series; the final memory state is sent to the feature generation layer, which typically contains one or more convolutional layers to further refine and integrate the spatial structure of the memory state, and finally outputs feature data that maintains the structure in space and integrates long-term temporal dependencies in features.

[0142] Step 1047: Input the feature data into the multilayer perceptron of the Conv-LSTM network for fully connected processing to generate multi-physics coupled data.

[0143] In step 1047, the multilayer perceptron is a network structure composed of multiple fully connected layers, used to flatten high-dimensional feature data with spatial structure and map it to the output space of the target dimension.

[0144] In this embodiment, the feature data output by the feature generation layer is first flattened in the spatial dimension and converted into a one-dimensional feature vector. Then, this one-dimensional vector is input into a multilayer perceptron composed of several fully connected layers. These fully connected layers further abstract and compress the feature vector through a series of linear transformations and nonlinear activations, and finally output multi-physics coupling data with fixed dimensions. This data is the direct input to the subsequent physical information neural network.

[0145] This application effectively integrates the spatiotemporal evolution information of dynamic strain field with the spatial distribution pattern of acoustic emission events through the above process, generating multi-physics field coupled data that can comprehensively reflect the intrinsic relationship between rock mass deformation and internal micro-fracture activity, thus laying a solid collaborative information foundation for subsequent accurate physical-driven inversion.

[0146] Step 105: Input the multiphysics coupling data into the physical information neural network. The physical information neural network performs constraint optimization on the multiphysics coupling data based on the elasticity equation to obtain the three-dimensional stress field data of the borehole surrounding rock.

[0147] Among them, three-dimensional stress field data refers to a complete dataset describing the stress state of the rock around the borehole at every point in three-dimensional space, including three normal stress components and three shear stress components.

[0148] It should be noted that this embodiment does not limit the structural design, parameter design, and training process of the physical information neural network, and can be set accordingly based on the actual situation.

[0149] In this embodiment, step 105 includes the following process:

[0150] Step 1051: Input the multi-physics coupling data into the physical information neural network, and extract high-dimensional feature vectors from the multi-physics coupling data through the feature extraction module of the physical information neural network.

[0151] In step 1051, the feature extraction module is the front-end part of the physical information neural network, which consists of several fully connected layers. Its function is to abstract and increase the dimensionality of the input low-dimensional or multi-dimensional coupled data, and extract high-dimensional potential features that can reflect the inherent pattern of the data. The high-dimensional feature vector refers to a one-dimensional array with a dimension much higher than the original input, which is obtained after processing by the feature extraction module.

[0152] In this embodiment, the feature extraction module receives the multiphysics coupling data output in step 104, which is typically a one-dimensional vector of fixed length. The module maps the input vector to a higher-dimensional feature space by applying a nonlinear activation function after a linear transformation of the input through a series of fully connected layers. Finally, the module outputs a high-dimensional feature vector that carries the deep pattern information extracted from the multiphysics coupling data.

[0153] Step 1052: Generate initial stress field data based on the high-dimensional feature vector through the prediction module of the physical information neural network.

[0154] In step 1052, the prediction module is one of the core components of the physical information neural network. Its function is to decode the high-dimensional feature vector and map it into the stress tensor components at each calculation point in the corresponding three-dimensional space. The initial stress field data refers to the three-dimensional stress distribution data that has been initially predicted by the prediction module and has not yet been rigorously verified by physical laws.

[0155] In this embodiment of the application, the input of the prediction module is the high-dimensional feature vector obtained in step 1051; the module first maps the high-dimensional vector to an intermediate dimension through several fully connected layers, and then maps it to the final target dimension through an output layer; the number of neurons in the output layer is equal to the number of all discrete points in the three-dimensional computational domain multiplied by the number of stress components that need to be predicted for each point, thereby directly outputting an initial stress field data vector containing stress information of all points.

[0156] Step 1053: Using the physical constraint module of the physical information neural network, residual data is obtained based on the initial stress field data and the elasticity equation.

[0157] In step 1053, the physical constraint module is a component in the physical information neural network specifically used to calculate the degree of conformity with physical laws; the residual data refers to the difference between the left and right sides of the equation after substituting the initial stress field data output by the prediction module into the basic equation of elasticity. This difference quantifies the degree to which the prediction result violates physical laws.

[0158] In this embodiment, the physical constraint module first reconstructs the stress tensor field of the entire three-dimensional region from the initial stress field data. Then, based on the theory of continuum mechanics, the module automatically calculates whether the stress field satisfies the static equilibrium equation at each internal point, and whether the corresponding strain field satisfies the geometric compatibility equation. For each equation, the module calculates the difference between the left and right sides of the equation, and these differences constitute the residual data. Typically, the residual data is also represented in vector form, with its dimension equal to the number of equations multiplied by the number of calculation points. It should be noted that the relevant interpretations of the theory of continuum mechanics and the static equilibrium equation can be found in related technologies, and will not be elaborated here.

[0159] Step 1054: Input the residual data and the high-dimensional feature vector into the residual optimization layer of the physical information neural network. Iteratively adjust the parameters of the prediction module through the backpropagation algorithm in the residual optimization layer. When the iteration process meets the preset convergence condition, use the output of the adjusted prediction module as the three-dimensional stress field data of the borehole surrounding rock.

[0160] In step 1054, the residual optimization layer is not an independent network layer, but refers to the entire process of updating the network parameters by using the physical residual as part of the loss function and the backpropagation algorithm during the training or optimization phase of the entire network; the convergence condition refers to the criterion used to determine whether the optimization process can be stopped.

[0161] In this embodiment, this step defines the training or online optimization process of the network. First, a total loss function is constructed, which includes both data fitting loss and physical loss. Then, the gradient of the total loss function with respect to all weights and bias parameters in the prediction module is calculated using the backpropagation algorithm. Next, these parameters are updated based on the calculated gradient using gradient descent or a variant thereof, so that the stress field generated by the prediction module better fits the input data features and satisfies the elasticity equations to a greater extent.

[0162] Repeat this iterative process until the value of the total loss function is lower than a preset threshold or the maximum number of iterations is reached. At this point, the parameters of the prediction module have been optimized, and the stress field data output by the module based on the current high-dimensional feature vector is the three-dimensional stress field data of the borehole surrounding rock that conforms to the characteristics of multi-physics data and strictly satisfies physical constraints.

[0163] In the above process, this application utilizes a physical information neural network to combine data learning with the hard constraints of physical laws, ensuring that the final inverted three-dimensional stress field data strictly satisfies the basic principles of elasticity while fitting the multi-source observation characteristics, thereby improving the physical rationality and reliability of the stress field reconstruction results.

[0164] Step 106: Based on the three-dimensional stress field data, simulate and calculate the distribution trend of omnidirectional stress in the borehole surrounding rock within a future set time period, and generate a visual reconstruction result.

[0165] Among them, the visualization reconstruction result refers to the visual output that can intuitively show the spatiotemporal evolution of stress magnitude and direction in the form of dynamic images, by converting the three-dimensional stress field data obtained from simulation calculations at multiple future times into three-dimensional stress field data through graphics technology.

[0166] In this embodiment, step 106 includes the following process:

[0167] Step 1061: The three-dimensional stress field data is used as the initial state input to the time series prediction network. In the time series prediction network, the future time period and time step are set, and iterative calculation is performed according to the time step to obtain the predicted stress field data for multiple future time points.

[0168] In step 1061, the future time period refers to the total duration for which stress trend prediction needs to be performed; the time step refers to the time interval between each forward step when the prediction network performs multi-step prediction; and the predicted stress field data refers to the three-dimensional stress field predicted by the network for a specific future time point.

[0169] In this embodiment, the three-dimensional stress field data obtained in step 105 is first used as the historical sequence endpoint or initial condition of the prediction network. Then, the total future duration to be predicted is set according to the engineering early warning requirements, such as 10 minutes in the future, and the prediction time resolution is set, such as predicting once every 1 minute. Next, the prediction network takes the current stress field as the starting point and, based on the stress evolution dynamics law learned internally, recursively calculates the stress field of the next time step in units of the set time step. Each calculated stress field is used as the input for the next prediction, and so on iteratively until the set future time period is predicted, thereby obtaining a series of predicted stress field data corresponding to different future time points.

[0170] Step 1062: Combine the predicted stress field data at multiple future time points in chronological order to form four-dimensional stress data.

[0171] In step 1062, four-dimensional stress data refers to a data set containing three spatial dimensions and one time dimension, which fully records the information of the stress field continuously changing with space and time over a future period of time.

[0172] In this embodiment of the application, all the predicted stress field data obtained in step 1061 are stacked in the time dimension according to their corresponding time point order; each predicted stress field data is a three-dimensional spatial data volume, and after stacking, a four-dimensional data volume is formed, whose four dimensions represent the length, width, height and time in space, respectively.

[0173] Step 1063: Extract stress magnitude data and second stress direction data from the four-dimensional stress data.

[0174] In step 1063, the stress magnitude data refers to a scalar that characterizes the stress intensity calculated from the stress tensor, usually the maximum principal stress or equivalent stress; the second stress direction data refers to the principal stress direction calculated from the stress tensor, especially the direction of the maximum principal stress. The "second" here is used to distinguish it from the "first stress direction data" extracted from the reference image.

[0175] In this embodiment of the application, tensor operations are performed on each spatial point and each time point in the four-dimensional stress data. First, the eigenvalues ​​and eigenvectors of the stress tensor at that point are calculated. The largest eigenvalue is the magnitude of the maximum principal stress at that point, which constitutes the stress magnitude data. The corresponding eigenvector indicates the direction of the maximum principal stress at that point, which constitutes the second stress direction data.

[0176] Step 1064: Color the stress magnitude data to generate a stress distribution map, and mark the second stress direction data to generate a direction indicator map.

[0177] In step 1064, the stress distribution map is a pseudo-color image in which different colors or shades of color correspond to different stress magnitudes; the direction indicator map is a vector or symbolic map in which arrows, cones or other symbols are used to indicate the direction of stress.

[0178] In this embodiment of the application, when generating the stress distribution map, a color mapping table is first set for the stress magnitude data, such as a gradient from blue to red, corresponding to low stress values ​​to high stress values ​​respectively; then, the stress magnitude data matrix at each time point is converted into a color image according to this mapping table.

[0179] When generating a direction indicator map, representative spatial points are selected on the image, the second stress direction data of these points are read, and directional arrows are drawn at the corresponding image positions. The direction and length of the arrows represent the direction and relative magnitude of the stress, respectively, thus generating a direction indicator map.

[0180] Step 1065: Overlay the stress distribution map and direction indicator map at the same time point to generate a combined image, and arrange the combined images corresponding to all time points in chronological order to form a dynamic image sequence.

[0181] In this embodiment, the stress distribution map and direction indicator map generated at the same time point in step 1064 are first spatially aligned to ensure that their pixel coordinates are completely consistent. Then, through image synthesis technology, such as superimposing the direction indicator map on the stress distribution map in a semi-transparent manner, a combined image is generated that simultaneously contains color-coded stress magnitude and arrows indicating stress direction. Finally, the combined images generated at all future time points are arranged in order from earliest to latest into an image list or image stack to form a dynamic image sequence.

[0182] Step 1066: Play the dynamic image sequence at a predetermined rate to generate a visual reconstruction result.

[0183] In this embodiment of the application, the dynamic image sequence formed in step 1065 is played out in the form of an animation by calling the graphics display interface or using a visualization library; the playback rate can be set as needed, such as 1 frame per second or 10 frames per second, to clearly show the dynamic evolution process of the stress field; the dynamic image finally displayed on the screen is the visualization reconstruction result, which intuitively shows the evolution trend of the future stress magnitude and direction of the borehole surrounding rock.

[0184] This application uses the above process to dynamically predict and visualize the future trends of physically constrained optimized three-dimensional stress field data, ultimately generating a visualization result that can intuitively and dynamically show the evolution of the magnitude and direction of future stress in the borehole surrounding rock. This transforms complex numerical data into intuitive image information that can be directly used for engineering risk assessment and early warning.

[0185] Figure 3 A schematic diagram of a three-dimensional visualization reconstruction system for omnidirectional stress in mine boreholes provided in this application embodiment is shown below. Figure 3 As shown, the system includes:

[0186] The acquisition module 31 is used to acquire optical image data, acoustic signal data and reference image data of the borehole wall during the drilling process in the mine.

[0187] The calculation module 32 is used to perform displacement field calculation on the optical image data to generate strain time series data, and to extract and locate the acoustic signal data to generate event distribution data.

[0188] The input module 33 is used to input the reference image data and the strain time series data into a dual-branch convolutional neural network for processing to generate a dynamic strain field.

[0189] Analysis module 34 is used to perform coupling analysis on the dynamic strain field and the event distribution data using a Conv-LSTM network to generate multiphysics coupling data.

[0190] The optimization module 35 is used to input the multi-physics coupling data into the physical information neural network. The physical information neural network performs constraint optimization on the multi-physics coupling data based on the elasticity equation to obtain the three-dimensional stress field data of the borehole surrounding rock.

[0191] The generation module 36 is used to simulate and calculate the distribution trend of omnidirectional stress in the borehole surrounding rock within a future set time period based on the three-dimensional stress field data, and generate a visualized reconstruction result.

[0192] The three-dimensional visualization reconstruction system for omnidirectional stress in mine boreholes in this application is used to implement the aforementioned three-dimensional visualization reconstruction method for omnidirectional stress in mine boreholes. Therefore, the specific implementation of the three-dimensional visualization reconstruction system for omnidirectional stress in mine boreholes can be found in the embodiment section of the three-dimensional visualization reconstruction method for omnidirectional stress in mine boreholes above. The specific implementation can be referred to the description of the corresponding embodiments, which will not be repeated here.

[0193] This application also provides an electronic device, comprising: a memory for storing a computer program; and a processor for executing the computer program to implement the steps of the three-dimensional visualization reconstruction method for omnidirectional stress in mine boreholes as described above.

[0194] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of any of the above-described methods for three-dimensional visualization reconstruction of omnidirectional stress in mine boreholes.

[0195] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as USB flash drives, read-only memory, random access memory, portable hard drives, magnetic disks, or optical disks.

[0196] The embodiments of this application also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the embodiments of the three-dimensional visualization reconstruction method for omnidirectional stress in mine boreholes described above.

[0197] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0198] The above provides a detailed description of a three-dimensional visualization reconstruction method and system for omnidirectional stress in mine boreholes provided by this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are merely for the purpose of helping to understand the method and its core ideas. It should be noted that those skilled in the art can make various improvements and modifications to this application without departing from its principles, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. A three-dimensional visualization reconstruction method for omnidirectional stress in mine boreholes, characterized in that, include: During the drilling process in the mine, optical image data, acoustic signal data, and reference image data of the borehole wall are acquired. Displacement field calculations are performed on the optical image data to generate strain time series data, and the acoustic signal data is extracted and localized to generate event distribution data. The reference image data and the strain time series data are input into a dual-branch convolutional neural network for processing to generate a dynamic strain field. The dynamic strain field and the event distribution data are coupled and analyzed using a Conv-LSTM network to generate multiphysics coupling data. The multiphysics coupling data is input into a physical information neural network, which performs constraint optimization on the multiphysics coupling data based on the elasticity equation to obtain three-dimensional stress field data of the borehole surrounding rock. Based on the three-dimensional stress field data, the distribution trend of omnidirectional stress in the borehole surrounding rock over a future set time period is simulated and calculated, and a visual reconstruction result is generated. The step of inputting the reference image data and the strain time-series data into a dual-branch convolutional neural network for processing to generate a dynamic strain field includes: The reference image data is input into the first branch of the dual-branch convolutional neural network. The first branch extracts the first stress direction data and stripe spacing data from the reference image data. Based on the first stress direction data and the stripe spacing data, the reference stress distribution is determined. The strain time series data is input into the second branch of the dual-branch convolutional neural network. The second branch performs spatiotemporal convolution on the strain time series data to extract lumped data and pattern data. The lumped data refers to the feature information identified from the strain time series data that characterizes the spatially highly concentrated area of ​​strain. The pattern data refers to the feature information extracted from the strain time series data that characterizes the change pattern and trend of strain in the time dimension. The reference stress distribution is transferred to the second branch, and amplitude and spatial corrections are performed on the lumped data and the mode data based on the reference stress distribution; The output layer of the dual-branch convolutional neural network fuses and reconstructs the corrected lumped data and corrected pattern data to output a dynamic strain field.

2. The method according to claim 1, characterized in that, The amplitude and spatial correction of the lumped data and the model data based on the reference stress distribution includes: Within the second branch, a spatial transformation network is introduced, which pre-establishes a mapping relationship between the reference stress distribution, lumped data, and model data. Based on the mapping relationship, according to the stress values ​​at each spatial location in the reference stress distribution, the amplitude correction parameters corresponding to each spatial location are generated using a multilayer perceptron regression model, and the corresponding spatial correction parameters are generated according to the gradient change direction of the reference stress distribution. The amplitude correction parameter is multiplied element-wise with the ensemble data and the mode data to achieve amplitude correction. Spatial correction is achieved by performing affine transformation operations using the spatial correction parameters and the local features of the ensemble data and the pattern data.

3. The method according to claim 1, characterized in that, The process of using a Conv-LSTM network to couple the dynamic strain field with the event distribution data to generate multiphysics coupled data includes: The dynamic strain field is divided into multiple strain gauge data in chronological order; The event distribution data is divided into multiple time periods according to time sequence, and the kernel density of the event location information in each time period is estimated to generate an event density map. The strain gauge data for each time period are combined with the corresponding event density map to form combined data; The combined data is input into the Conv-LSTM network, and spatial features are extracted from the combined data through the convolutional layers of the Conv-LSTM network. The spatial features are temporally memorized and updated through the recurrent memory unit of the Conv-LSTM network. The recurrent memory unit introduces an attention mechanism during the processing. The attention mechanism calculates spatial weights based on the event density map and uses the spatial weights to adjust the influence of spatial features at different locations on memory. The feature generation layer of the Conv-LSTM network generates feature data that integrates temporal and spatial information based on the final memory state output by the recurrent memory unit. The feature data is input into the multilayer perceptron of the Conv-LSTM network for fully connected processing to generate multi-physics coupled data.

4. The method according to claim 1, characterized in that, The process involves inputting the multiphysics coupling data into a physical information neural network, where the physical information neural network performs constraint optimization on the multiphysics coupling data based on elasticity equations to obtain three-dimensional stress field data of the borehole surrounding rock, including: The multiphysics coupling data is input into a physical information neural network, and a high-dimensional feature vector is extracted from the multiphysics coupling data through the feature extraction module of the physical information neural network. The prediction module of the physical information neural network generates initial stress field data based on the high-dimensional feature vector. The physical constraint module of the physical information neural network obtains residual data based on the initial stress field data using the elasticity equation. The residual data and the high-dimensional feature vector are input together into the residual optimization layer of the physical information neural network. The parameters of the prediction module are iteratively adjusted by the backpropagation algorithm in the residual optimization layer. When the iteration process meets the preset convergence condition, the output result of the adjusted prediction module is used as the three-dimensional stress field data of the borehole surrounding rock.

5. The method according to claim 1, characterized in that, Based on the three-dimensional stress field data, the distribution trend of omnidirectional stress in the borehole surrounding rock over a future set time period is simulated and calculated, generating a visualized reconstruction result, including: The three-dimensional stress field data is used as the initial state input to the time series prediction network. In the time series prediction network, future time periods and time steps are set, and iterative calculations are performed according to the time steps to obtain the predicted stress field data for multiple future time points. The predicted stress field data at multiple future time points are combined in chronological order to form four-dimensional stress data. Extract stress magnitude data and second stress direction data from the four-dimensional stress data; The stress magnitude data is colored to generate a stress distribution map, and the second stress direction data is marked to generate a direction indicator map; The stress distribution map and direction indicator map at the same time point are superimposed to generate a combined image, and the combined images corresponding to all time points are arranged in chronological order to form a dynamic image sequence; The dynamic image sequence is played at a predetermined rate to generate a visual reconstruction result.

6. The method according to claim 1, characterized in that, The process of performing displacement field calculations on the optical image data to generate strain time-series data, and extracting and locating the acoustic signal data to generate event distribution data, includes: Based on the optical image data, determine the coordinate changes of the corresponding speckle at adjacent time points, and obtain the displacement data of the borehole wall based on the coordinate changes; Based on the displacement data, the strain data of the borehole wall is calculated, and the strain data corresponding to each time point is combined to form strain time series data. Identify the pulse signal in the acoustic signal data, and determine the source location data of the pulse signal in the borehole surrounding rock based on the time when the pulse signal arrives at multiple sensors; The parameters of each pulse signal are associated with the corresponding source location data and organized in chronological order to generate event distribution data.

7. A three-dimensional visualization reconstruction system for omnidirectional stress in mine boreholes, characterized in that, include: The acquisition module is used to acquire optical image data, acoustic signal data, and reference image data of the borehole wall during the drilling process in the mine. The calculation module is used to perform displacement field calculation on the optical image data to generate strain time series data, and to extract and locate the acoustic signal data to generate event distribution data. The input module is used to input the reference image data and the strain time series data into a dual-branch convolutional neural network for processing to generate a dynamic strain field; The analysis module is used to perform coupled analysis of the dynamic strain field and the event distribution data using a Conv-LSTM network to generate multiphysics coupled data. An optimization module is used to input the multi-physics coupling data into a physical information neural network. The physical information neural network performs constraint optimization on the multi-physics coupling data based on the elasticity equation to obtain three-dimensional stress field data of the borehole surrounding rock. The generation module is used to simulate and calculate the distribution trend of omnidirectional stress in the borehole surrounding rock over a future set time period based on the three-dimensional stress field data, and generate a visual reconstruction result. The step of inputting the reference image data and the strain time-series data into a dual-branch convolutional neural network for processing to generate a dynamic strain field includes: The reference image data is input into the first branch of the dual-branch convolutional neural network. The first branch extracts the first stress direction data and stripe spacing data from the reference image data. Based on the first stress direction data and the stripe spacing data, the reference stress distribution is determined. The strain time series data is input into the second branch of the dual-branch convolutional neural network. The second branch performs spatiotemporal convolution on the strain time series data to extract lumped data and pattern data. The lumped data refers to the feature information identified from the strain time series data that characterizes the spatially highly concentrated area of ​​strain. The pattern data refers to the feature information extracted from the strain time series data that characterizes the change pattern and trend of strain in the time dimension. The reference stress distribution is transferred to the second branch, and amplitude and spatial corrections are performed on the lumped data and the mode data based on the reference stress distribution; The output layer of the dual-branch convolutional neural network fuses and reconstructs the corrected lumped data and corrected pattern data to output a dynamic strain field.

8. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to execute the computer program to implement the steps of the three-dimensional visualization reconstruction method for omnidirectional stress in mine boreholes as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, enables the three-dimensional visualization reconstruction method for omnidirectional stress in mine boreholes as described in any one of claims 1 to 6.

Images