A multi-source information fusion rock three-dimensional reconstruction method and system

Abstract

The present application relates to the technical field of rock mechanics, and discloses a rock three-dimensional reconstruction method and system based on multi-source information fusion, which acquires data and carries out pretreatment, learns spatial features of a fusion feature vector through a 3D-CNN network, and constructs a three-dimensional voxel model of rock mesoscopic damage; a three-dimensional reconstruction algorithm based on point clouds is used to convert the fused image data into a point cloud model of the surrounding rock structure of an underground cavern, and based on an implicit surface reconstruction algorithm, a digital twin framework of the surrounding rock structure of the underground cavern is constructed; the feature parameters output by the three-dimensional voxel model and the digital twin framework are taken as inputs, and an LSTM-CNN fusion model is used to output optimal support timing and support parameters; in the process of underground engineering construction, surrounding rock deformation data are collected in real time, and a deep deterministic policy gradient algorithm is used to adjust the support scheme in real time; the present application can improve the scientificity and timeliness of support design under complex geological conditions.

Description

Technical Field

[0001] This invention relates to the field of rock mechanics technology, specifically to a method and system for three-dimensional reconstruction of rocks using multi-source information fusion. Background Technology

[0002] In rock mechanics research and underground engineering construction, accurately obtaining the microstructure of rock damage and the surrounding rock structure of underground caverns is crucial for support design. Traditional rock mechanics tests suffer from data dispersion, resulting in an inaccurate description of the internal damage state of the rock. At the same time, support design often relies on engineering experience and lacks a real-time dynamic adjustment mechanism, making it difficult to adapt to engineering needs under complex geological conditions. Although some existing technologies use single or limited data sources to reconstruct rock structures, they do not fully utilize the fusion of multi-source data, making it impossible to construct high-precision three-dimensional models and achieve intelligent optimization of support schemes. Summary of the Invention

[0003] To address the aforementioned problems, this invention presents a method and system for three-dimensional reconstruction of rocks by fusing multi-source information.

[0004] The first aspect of this invention provides a method for three-dimensional reconstruction of rocks by fusing multi-source information, the method comprising the following steps:

[0005] Acquire CT scan data, acoustic emission signals, DIC deformation field data, borehole imaging images, and ground-penetrating radar data of the rock, and perform preprocessing.

[0006] Preprocessed CT scan data, acoustic emission signals and DIC deformation field data are fused together, and spatial features are learned from the fused feature vectors through a 3D-CNN network to construct a three-dimensional voxel model of rock microstructure damage.

[0007] The preprocessed borehole imaging images and ground-penetrating radar data are fused together. A point cloud-based 3D reconstruction algorithm is used to convert the fused image data into a point cloud model of the surrounding rock structure of the underground cavern. Based on the implicit surface reconstruction algorithm, a digital twin framework of the surrounding rock structure of the underground cavern is constructed.

[0008] The feature parameters output from the 3D voxel model and the digital twin framework are used as inputs, and the optimal support timing and support parameters are output using the LSTM-CNN fusion model.

[0009] During the construction of underground engineering projects, real-time data on the deformation of the surrounding rock is collected, and the support scheme is adjusted in real time using a depth-deterministic gradient algorithm.

[0010] Optionally, in a first implementation of the first aspect of the present invention, the acquisition of CT scan data, acoustic emission signals, DIC deformation field data, borehole imaging images, and ground-penetrating radar data of the rock, and the preprocessing thereof, includes:

[0011] For CT scan data, nonlocal mean filtering is used to remove metal artifacts;

[0012] For acoustic emission signals, noise is filtered out based on wavelet packet transform;

[0013] For DIC deformation field data, median filtering is used to remove random noise from speckle images, and subpixel interpolation is used to enhance pixels;

[0014] For borehole imaging images, histogram equalization combined with the Otsu algorithm is used to enhance fracture contrast and remove borehole water film reflection.

[0015] For ground-penetrating radar data, the moving average method is used to remove background noise, and the KS transform is used to suppress multiple waves.

[0016] Optionally, in a second implementation of the first aspect of the present invention, the fusion of the preprocessed CT scan data, acoustic emission signal, and DIC deformation field data includes:

[0017] The preprocessed CT scan data, acoustic emission signals, and DIC deformation field data are converted into data feature vectors, and position codes are added before being input into the Transformer encoder.

[0018] The Transformer encoder uses 8 self-attention heads. Each attention head independently calculates the QKV matrix of three modal data and obtains the correlation between pairs of data through dot product operation.

[0019] After self-attention computation, the output features of self-attention are layer normalized, and the input of the Transformer encoder is added to the output features through residual connections.

[0020] After the fusion features are obtained through self-attention and layer normalization, the high-order features after cross-modal fusion are extracted through two fully connected layers. After multi-layer processing by the Transformer encoder, the fused feature vector is output.

[0021] Optionally, in a third implementation of the first aspect of the present invention, the step of constructing a three-dimensional voxel model of rock microstructural damage by learning spatial features from the fused feature vectors using a 3D-CNN network includes:

[0022] Convert the raw spatial resolution of the CT scan data output by the Transformer into a three-dimensional tensor;

[0023] A three-dimensional tensor is input into a 3D-CNN network, and multi-scale spatial features are extracted hierarchically through the 3D-CNN network to output a three-dimensional voxel probability matrix.

[0024] Mineral grain nodes are obtained by extracting high-value regions from CT scan data. Based on the three-dimensional voxel probability matrix, voxel clusters with a probability greater than or equal to 0.7 are extracted to obtain microcrack nodes.

[0025] When the minimum distance between mineral grains and microcracks is less than 0.05 mm, an undirected edge is established; if the distance between the tips of microcracks is less than 0.1 mm, a directed edge is established, resulting in a graph structure.

[0026] Based on the graph structure, physical interactions are captured through a GNN network to obtain corrected voxel features. The GNN network is then mapped back to a 3D-CNN network for iterative optimization.

[0027] A three-dimensional voxel model of rock microstructure damage was constructed by alternating processing with 3D-CNN and GNN networks.

[0028] Optionally, in a fourth implementation of the first aspect of the present invention, the step of inputting a three-dimensional tensor into a 3D-CNN network, extracting multi-scale spatial features hierarchically through the 3D-CNN network, and outputting a three-dimensional voxel probability matrix includes:

[0029] The bottom convolutional layer slides with a stride of 1, performing convolution operations on each voxel and its 26 neighboring voxels to extract basic spatial features;

[0030] The middle convolutional layers use downsampling with a stride of 2 to expand the receptive field and integrate the spatial relationships between local features;

[0031] High-level convolutional layers integrate full-space features through 1×1×1 convolutional kernels, ultimately outputting a three-dimensional voxel probability matrix.

[0032] Optionally, in a fifth implementation of the first aspect of the present invention, the step of using the feature parameters output by the three-dimensional voxel model and the digital twin framework as input, and using the LSTM-CNN fusion model to output the optimal support timing and support parameters, includes:

[0033] The feature parameters output from the 3D voxel model and the digital twin framework are preprocessed and concatenated to obtain a hybrid feature vector, which is then used as the input to the LSTM-CNN fusion model.

[0034] The CNN branch scans the fracture distribution tensor with a 3×3 convolution kernel to identify typical spatial patterns in the surrounding rock structure and outputs a 64-channel spatial feature map. The local features are aggregated into a global structural description through a max pooling layer, and a 64×64 spatial feature matrix is ​​output.

[0035] The LSTM branch first performs time series encoding, and then captures the elastic modulus drop event, identifies the gradual decay trend of cohesion and detects the second derivative of crack propagation rate through forget gate, input gate and output gate. The fully connected layer transforms the hidden state of LSTM into a 64-dimensional temporal feature vector.

[0036] The spatial feature matrix output by the CNN branch is flattened into a one-dimensional vector and concatenated with the temporal feature vector output by the LSTM branch to form a comprehensive feature vector that includes spatial distribution patterns and temporal evolution trends.

[0037] The output layer generates the final result through linear transformation, obtaining the optimal support timing and support parameters.

[0038] Optionally, in the sixth implementation of the first aspect of the present invention, the step of collecting surrounding rock deformation data in real time during the underground engineering construction process and adjusting the support scheme in real time using a depth-deterministic gradient algorithm includes:

[0039] After removing noise by Kalman filtering, the surrounding rock deformation data is input into the depth deterministic strategy gradient algorithm.

[0040] The Actor network outputs specific parameters based on the current state, generates support adjustment instructions after verification, and transmits them to the corresponding execution device.

[0041] The adjusted support scheme is fed back into the digital twin framework to simulate and predict the evolution of the surrounding rock structure.

[0042] The second aspect of the present invention provides a multi-source information fusion rock three-dimensional reconstruction system, a data acquisition module for acquiring CT scan data, acoustic emission signals, DIC deformation field data, borehole imaging images and ground-penetrating radar data of rocks, and performing preprocessing.

[0043] The first construction module is used to fuse preprocessed CT scan data, acoustic emission signals and DIC deformation field data, and to learn spatial features of the fused feature vectors through a 3D-CNN network to construct a three-dimensional voxel model of rock microstructure damage.

[0044] The second construction module is used to fuse the preprocessed borehole imaging images and ground-penetrating radar data. It adopts a point cloud-based 3D reconstruction algorithm to convert the fused image data into a point cloud model of the surrounding rock structure of the underground cavern. Based on the implicit surface reconstruction algorithm, it constructs a digital twin framework of the surrounding rock structure of the underground cavern.

[0045] The output module takes the feature parameters output by the 3D voxel model and the digital twin framework as input, and uses the LSTM-CNN fusion model to output the optimal support timing and support parameters.

[0046] The real-time adjustment module is used to collect surrounding rock deformation data in real time during underground engineering construction and to adjust the support scheme in real time using a depth-deterministic gradient algorithm.

[0047] A third aspect of the present invention provides a multi-source information fusion rock three-dimensional reconstruction device, the multi-source information fusion rock three-dimensional reconstruction device including a memory and at least one processor, the memory storing instructions; the at least one processor calling the instructions in the memory to cause the multi-source information fusion rock three-dimensional reconstruction device to perform the various steps of the multi-source information fusion rock three-dimensional reconstruction method as described in any of the preceding claims.

[0048] A fourth aspect of the present invention provides a computer-readable storage medium storing instructions that, when executed by a processor, implement the steps of the multi-source information fusion rock three-dimensional reconstruction method as described in any of the preceding claims.

[0049] The technical solution provided by this invention involves acquiring CT scan data, acoustic emission signals, DIC deformation field data, borehole imaging images, and ground-penetrating radar data of rocks, and performing preprocessing. The preprocessed CT scan data, acoustic emission signals, and DIC deformation field data are then fused, and spatial feature learning is performed on the fused feature vectors using a 3D-CNN network to construct a three-dimensional voxel model of the rock's microstructural damage. The preprocessed borehole imaging images and ground-penetrating radar data are then fused, and a point cloud-based three-dimensional reconstruction algorithm is used to convert the fused image data into a point cloud model of the surrounding rock structure of an underground cavern. Based on an implicit surface reconstruction algorithm, a digital twin framework of the surrounding rock structure of the underground cavern is constructed. Finally, the three-dimensional voxel model and the data... The feature parameters output by the digital twin framework are used as input, and the optimal support timing and support parameters are output using an LSTM-CNN fusion model. During the construction of underground engineering, the surrounding rock deformation data is collected in real time, and the support scheme is adjusted in real time using a depth deterministic gradient algorithm. This invention realizes the high-precision construction of a three-dimensional model of rock microscopic damage and the establishment of a digital twin framework for the surrounding rock structure of underground caverns. By dynamically associating the evolution law of rock mechanical parameters with the response characteristics of the support structure, an optimization model for support timing and parameters is constructed, realizing real-time feedback adjustment of the support scheme. This can improve the scientificity and timeliness of support design under complex geological conditions, reduce the risk of support lag in deep underground engineering, and has broad application prospects. Attached Figure Description

[0050] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention.

[0051] Figure 1A flowchart of a multi-source information fusion method for three-dimensional reconstruction of rocks provided in an embodiment of the present invention;

[0052] Figure 2 This is a schematic diagram of the structure of a multi-source information fusion rock three-dimensional reconstruction system provided in an embodiment of the present invention;

[0053] Figure 3 This is a schematic diagram of the structure of a rock three-dimensional reconstruction device for multi-source information fusion provided in an embodiment of the present invention. Detailed Implementation

[0054] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a particular order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” or “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, apparatus, product, or device that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0055] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1 A flowchart of a multi-source information fusion method for three-dimensional reconstruction of rocks provided in this invention embodiment, the method specifically includes the following steps:

[0056] Step 101: Acquire CT scan data, acoustic emission signals, DIC deformation field data, borehole imaging images, and ground-penetrating radar data of the rock, and perform preprocessing.

[0057] In this embodiment, CT scan data is obtained using a 128-slice spiral CT scanner to acquire DICOM format data of the density distribution inside the rock, covering a three-dimensional density matrix of the entire rock sample volume; acoustic emission signals are obtained using an 8-channel acoustic emission sensor array with a sensor resonant frequency of 150-450kHz, simultaneously acquiring load-time curves and spatiotemporal location data of acoustic emission events; DIC deformation field data is obtained using a binocular vision DIC system with a 2-megapixel high-speed camera, combined with a 3D Digital ImageCorrelation algorithm, to acquire spatiotemporal sequence data of surface displacement and strain fields; borehole imaging images are obtained using a panoramic borehole camera system to acquire 360° unfolded images of the borehole inner wall; and ground-penetrating radar data is obtained using a 200MHz high-frequency ground-penetrating radar to acquire dielectric constant distribution data within a depth range of 0-30m underground, generating a radar profile grayscale map.

[0058] In this embodiment, nonlocal mean filtering is used to remove metal artifacts from CT scan data; wavelet packet transform is used to filter out noise from acoustic emission signals; median filtering is used to remove random noise from speckle images from DIC deformation field data, and subpixel interpolation is used to enhance pixels; histogram equalization combined with the Otsu algorithm is used to enhance fracture contrast and remove borehole water film reflections from borehole images; and moving average method is used to remove background noise from ground-penetrating radar data, and KS transform is used to suppress multiple waves.

[0059] Step 102: The preprocessed CT scan data, acoustic emission signal and DIC deformation field data are fused together, and the spatial features of the fused feature vector are learned by a 3D-CNN network to construct a three-dimensional voxel model of rock microstructure damage.

[0060] In this embodiment, the preprocessed CT scan data, acoustic emission signal, and DIC deformation field data are converted into data feature vectors, and position encoding is added before being input into the Transformer encoder. The Transformer encoder employs eight sets of self-attention, with each attention head independently calculating the QKV matrix of the three modal data. The correlation between pairs of data is obtained through dot product operations. After the self-attention calculation, the output features of the self-attention are layer-normalized, and the input of the Transformer encoder is added to the output features through residual connections. The fused features after self-attention and layer normalization are extracted through two fully connected layers to obtain high-order features after cross-modal fusion. After multi-layer processing by the Transformer encoder, the fused feature vector is output.

[0061] In this embodiment, a Transformer encoder layer is formed by self-attention, layer normalization, and feedforward network. Typically, 12-24 layers are stacked. The output of each encoder layer serves as the input of the next layer, enabling multi-round cross-modal information interaction. For example, the first layer captures the local correlation between CT and acoustic emission, the second layer integrates the global strain distribution of the DIC deformation field, and the third layer correlates the results of the first two layers to form a multi-scale fusion feature of local damage and global deformation. As the number of layers increases, the model gradually transitions from intramodal features to shallow intermodal correlations, eventually forming deep fusion features that can simultaneously describe micro-crack morphology, surface deformation trends, and energy release patterns.

[0062] In this embodiment, the original spatial resolution of the CT scan data output by the Transformer is converted into a three-dimensional tensor. The three-dimensional tensor is input into a 3D-CNN network, which extracts multi-scale spatial features layer by layer and outputs a three-dimensional voxel probability matrix. High-value regions in the CT scan data are extracted to obtain mineral grain nodes. Based on the three-dimensional voxel probability matrix, voxel clusters with a probability greater than or equal to 0.7 are extracted to obtain microcrack nodes. When the minimum distance between mineral grains and microcracks is less than 0.05 mm, undirected edges are established. If the distance between the tips of microcracks is less than 0.1 mm, directed edges are established to obtain a graph structure. Based on the graph structure, a GNN network is used to capture physical interactions to obtain corrected voxel features. The GNN network is mapped back to the 3D-CNN network for iterative optimization. After alternating processing by the 3D-CNN network and the GNN network, a three-dimensional voxel model of rock microstructural damage is constructed.

[0063] In this embodiment, the bottom convolutional layer slides with a stride of 1, performing convolution operations on each voxel and its 26 neighboring voxels to extract basic spatial features, such as: identifying density anomalies in individual voxels, capturing strain gradients between adjacent voxels, detecting local peaks in acoustic emission signal energy in space, introducing nonlinearity into the activation function to enhance sensitivity to damage features such as density decrease-strain increase-acoustic signal activity, and outputting a low-level feature map to characterize the basic geometric units of rock microstructure, such as mineral grain boundaries and micropore contours.

[0064] The middle convolutional layer uses downsampling with a stride of 2 to expand the receptive field and integrate the spatial relationships between local features. For example, it connects adjacent microcrack voxels to identify the direction of crack segments, associates the density distribution of mineral particles with the surrounding strain field to determine the hindering effect of hard minerals on crack propagation, and combines the spatial localization of acoustic emission signals to determine the main path of crack propagation. By skipping connections, it preserves the details of the lower layers and avoids the loss of edge information caused by downsampling. It outputs a middle layer feature map to characterize millimeter-level damage structures, such as microcracks with a length of 0.2 mm and mineral particles with a diameter of 0.1 mm.

[0065] High-level convolutional layers integrate full-space features through 1×1×1 convolutional kernels, ultimately outputting a three-dimensional voxel probability matrix. For example, they can identify the main crack network that runs through multiple mineral particles to determine the connectivity of the damage, calculate the damage degree index of different regions to generate a damage heat map, predict the future direction of crack propagation, and finally output a three-dimensional voxel probability matrix, with the probability of each voxel belonging to a microcrack, mineral particle, or pore, serving as the initial structure of the mesoscopic damage model.

[0066] In this embodiment, physical interactions include information flow and message aggregation between nodes; mineral particle nodes transmit hindering effects to adjacent crack nodes, and high-density minerals suppress crack propagation rates through boundary features. For example, minerals with high elastic modulus reduce the stress concentration at the crack tip; crack nodes transmit damage feedback to adjacent mineral nodes, and the stress waves generated by crack propagation may weaken the mineral particle boundaries.

[0067] A neighbor feature weighted average strategy is adopted, with the weights determined by edge features such as contact area ratio and stress transfer efficiency. For example, when updating the propagation rate of a microcrack, the hindering effect of the three surrounding mineral particles and the guiding effect of the two adjacent cracks are combined to form a propagation prediction that is more in line with the actual physical laws.

[0068] In this embodiment, the three-dimensional voxel model of rock microstructural damage includes:

[0069] Geometric details: clearly showing micron-scale pores and nanoscale grain boundary slip;

[0070] Physical properties: mineral composition, degree of damage, and strain state of each voxel;

[0071] Evolution process: Supports dynamic simulation of crack propagation path during loading, such as progressive fracturing along mineral weak surfaces and transgranular fracture across grains, providing microscale mechanical response basis for support design, such as prediction of sudden drop in surrounding rock strength caused by crack connectivity.

[0072] Step 103: The preprocessed borehole imaging images and ground-penetrating radar data are fused together. A point cloud-based 3D reconstruction algorithm is used to convert the fused image data into a point cloud model of the surrounding rock structure of the underground cavern. Based on the implicit surface reconstruction algorithm, a digital twin framework of the surrounding rock structure of the underground cavern is constructed.

[0073] In this embodiment, image enhancement technology is used to improve the uneven illumination of the borehole inner wall and highlight the edge contours of structural surfaces such as fractures and bedding in borehole imaging images. An image segmentation algorithm is used to automatically identify rock types and generate a two-dimensional unfolded image with semantic labels, with each pixel labeled with the rock type and fracture development status.

[0074] To remove interference signals such as those directly emitted by the radar antenna from ground-penetrating radar data, and background noise such as clutter reflected from fixed objects, effective signals such as strong reflection wave groups at fault interfaces and water-bearing areas are enhanced through time-frequency analysis. The depth coordinates of the radar profile are converted into actual geographic coordinates, and a spatial location index of the radar reflection signal is established by combining the dip angle and azimuth angle parameters of the borehole.

[0075] Using the underground cavern design coordinate system as a reference, the borehole's three-dimensional trajectory data, borehole depth, dip angle, and azimuth angle obtained by the inclinometer are used to map the two-dimensional unfolded image of the borehole imaging into three-dimensional space, forming a cylindrical unfolded image along the borehole path; coordinate transformation is performed on the ground-penetrating radar data to align the spatial position of the radar reflection point with the rock structure position of the borehole imaging, ensuring that the spatial position of the same geological body is consistent in the two types of data.

[0076] The fracture orientation extracted from the borehole images is cross-validated with the corresponding low dielectric constant anomaly band in the radar data to confirm the existence of real structural surfaces; rock type labels are associated with radar wave velocity values ​​to correct the identification error of single data, such as avoiding misjudging rock fractures as faults; the rock density derived from borehole imaging is weighted and averaged with radar dielectric constant data to generate fused physical property feature values.

[0077] Edge detection is performed on the fused image data to extract structural boundary points and rock type abrupt change points. Each feature point is assigned spatial coordinates, structural type, and initial values ​​of rock mechanical parameters.

[0078] A triangulation algorithm is used to extend two-dimensional feature points to three-dimensional space. Based on the depth information of borehole trajectory and radar reflection, a sparse point cloud is generated. Feature points of the same geological body in different boreholes and radar profiles are clustered and merged to remove duplicate points and improve the purity of the point cloud. In areas with complex structural surfaces, such as fault zones, missing points are supplemented by interpolation algorithm, and Gaussian smoothing filter is applied to reduce point cloud noise, making the attribute changes of adjacent points tend to be continuous. With each point cloud as the center, the weighted average of its neighboring points is calculated to generate continuous local tangent planes. Finally, they are stitched together to form a global NURBS surface. By adjusting the fitting parameters, the detail preservation and computational efficiency of the surface are balanced to ensure that the surface can accurately reflect the concave and convex morphology of the surrounding rock.

[0079] Geometric Model: Import the curved surface model into the digital twin platform to create a three-dimensional geometric entity of the surrounding rock of the underground cavern, including the cavern outline, surrounding rock structure surface, fault zone, etc., and render and query in real time, such as measuring the crack length between any two points;

[0080] The rock mechanical parameters of the point cloud are mapped to each grid cell on the surface through an interpolation algorithm to form a mechanical property field. Real-time monitoring data is associated with the corresponding grid cells to achieve a three-in-one modeling of geometric structure, mechanical properties and real-time status. The stress distribution and deformation trend of the surrounding rock under different support schemes are simulated through the finite element analysis module, and the visualization results are output.

[0081] Step 104: Using the feature parameters output by the 3D voxel model and the digital twin framework as input, the LSTM-CNN fusion model is used to output the optimal support timing and support parameters.

[0082] In this embodiment, the feature parameters output from the 3D voxel model and the digital twin framework are preprocessed and concatenated to obtain a hybrid feature vector, which serves as the input to the LSTM-CNN fusion model. The CNN branch scans the fracture distribution tensor with 3×3 convolutional kernels to identify typical spatial patterns in the surrounding rock structure, outputting a 64-channel spatial feature map. The local features are aggregated into a global structural description through a max pooling layer, outputting a 64×64 spatial feature matrix. The LSTM branch first performs time-series encoding, and then uses forget gates, input gates, and output gates to capture events of sudden drops in elastic modulus, identify the gradual decay trend of cohesion, and detect the second derivative of crack propagation rate. The fully connected layer converts the hidden state of the LSTM into a 64-dimensional temporal feature vector. The spatial feature matrix output from the CNN branch is flattened into a one-dimensional vector and concatenated with the temporal feature vector output from the LSTM branch to form a comprehensive feature vector containing spatial distribution patterns and temporal evolution trends. The output layer generates the final result through linear transformation, obtaining the optimal support timing and support parameters.

[0083] Step 105: During the construction of underground engineering, collect surrounding rock deformation data in real time and use a depth deterministic strategy gradient algorithm to adjust the support scheme in real time.

[0084] In this embodiment, the surrounding rock deformation data is filtered by Kalman filtering to remove noise and then input into the depth deterministic strategy gradient algorithm. The Actor network outputs specific parameters based on the current state, which are verified to generate support adjustment instructions and transmitted to the corresponding execution device. The adjusted support scheme is fed back into the digital twin framework to simulate and predict the evolution process of the surrounding rock structure.

[0085] In this embodiment, the input to the Actor network is the real-time surrounding rock condition. The hidden layer includes two fully connected layers with 256 neurons in each layer. The ReLU activation function is used to extract nonlinear correlations in the state features, such as risk patterns corresponding to high deformation rate and low support density. The output is a 4-dimensional continuous action vector, including anchor bolt spacing, anchor cable preload, support density, and construction delay time. Specific parameters are directly output through a deterministic strategy.

[0086] The Critic network takes a concatenation of a state vector and an action vector as input. The hidden layers consist of three fully connected layers, each with 512 neurons. The activation function is LeakyReLU. The network evaluates the long-term value of taking an action in the current state, such as the duration for which the support scheme can maintain the stability of the surrounding rock. The output is a single value scalar used to guide the Actor network's optimization strategy.

[0087] Please see Figure 2 A schematic diagram of the structure of a multi-source information fusion rock three-dimensional reconstruction system provided in this embodiment of the invention. The system includes:

[0088] The data acquisition module is used to acquire CT scan data, acoustic emission signals, DIC deformation field data, borehole imaging images, and ground-penetrating radar data of rocks, and to perform preprocessing.

[0089] The first construction module is used to fuse preprocessed CT scan data, acoustic emission signals and DIC deformation field data, and to learn spatial features of the fused feature vectors through a 3D-CNN network to construct a three-dimensional voxel model of rock microstructure damage.

[0090] The second construction module is used to fuse the preprocessed borehole imaging images and ground-penetrating radar data. It adopts a point cloud-based 3D reconstruction algorithm to convert the fused image data into a point cloud model of the surrounding rock structure of the underground cavern. Based on the implicit surface reconstruction algorithm, it constructs a digital twin framework of the surrounding rock structure of the underground cavern.

[0091] The output module takes the feature parameters output by the 3D voxel model and the digital twin framework as input, and uses the LSTM-CNN fusion model to output the optimal support timing and support parameters.

[0092] The real-time adjustment module is used to collect surrounding rock deformation data in real time during underground engineering construction and to adjust the support scheme in real time using a depth-deterministic gradient algorithm.

[0093] Figure 3 This is a schematic diagram of a multi-source information fusion rock 3D reconstruction device 300 provided in an embodiment of the present invention. The multi-source information fusion rock 3D reconstruction device 300 can vary significantly due to different configurations or performance. It may include one or more central processing units (CPUs) 310 (e.g., one or more processors) and a memory 320, and one or more storage media 330 (e.g., one or more mass storage devices) for storing application programs 333 or data 332. The memory 320 and storage media 330 can be temporary or persistent storage. The program stored in the storage media 330 may include one or more modules (not shown in the diagram), each module may include a series of instruction operations on the multi-source information fusion rock 3D reconstruction device 300. Furthermore, the processor 310 may be configured to communicate with the storage media 330 and execute the series of instruction operations in the storage media 330 on the multi-source information fusion rock 3D reconstruction device 300 to implement the method provided in the above embodiment.

[0094] The multi-source information fusion rock 3D reconstruction device 300 may also include one or more power supplies 340, one or more wired or wireless network interfaces 350, one or more input / output interfaces 360, and / or one or more operating devices 331, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. Those skilled in the art will understand that... Figure 3 The multi-source information fusion rock three-dimensional reconstruction device structure shown does not constitute a limitation on the computer device provided by the present invention. It may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0095] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when the instructions are executed on a computer, cause the computer to perform the various steps of the multi-source information fusion rock three-dimensional reconstruction method provided in the above embodiments.

[0096] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the above-described equipment or apparatus / unit can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0097] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0098] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A rock three-dimensional reconstruction method based on multi-source information fusion, characterized in that, The method comprises the following steps: CT scanning data, acoustic emission signals, DIC deformation field data, borehole imaging images and geological radar data of the rock are acquired and preprocessed; The preprocessed CT scanning data, acoustic emission signals and DIC deformation field data are fused, spatial feature learning of the fused feature vectors is performed through a 3D-CNN network, and a three-dimensional voxel model of rock meso-damage is constructed; The preprocessed borehole imaging images and geological radar data are fused, a point cloud model of the underground cavern surrounding rock structure is converted from the fused image data through a point cloud-based three-dimensional reconstruction algorithm, and a digital twin framework of the underground cavern surrounding rock structure is constructed based on an implicit surface reconstruction algorithm; Feature parameters output by the three-dimensional voxel model and the digital twin framework are taken as inputs, and an optimal supporting time and supporting parameters are output by an LSTM-CNN fusion model; In the construction process of the underground engineering, surrounding rock deformation data are collected in real time, and a deep deterministic policy gradient algorithm is used to adjust the supporting scheme in real time; The fusion of the preprocessed CT scanning data, acoustic emission signals and DIC deformation field data comprises the following steps: The preprocessed CT scanning data, acoustic emission signals and DIC deformation field data are converted into data feature vectors, and position encoding is added, and then input into a Transformer encoder; In the Transformer encoder, 8 groups of self-attention are adopted, each group of attention head independently calculates Q-K-V matrices of the three kinds of modal data, and the correlation between each two data is obtained through dot product operation; After the self-attention calculation, the output features of the self-attention are subjected to layer normalization, and the input of the Transformer encoder is added to the output features through residual connection; The fusion features subjected to self-attention and layer normalization are subjected to two fully connected layers to extract high-order features after cross-modal fusion, and the fusion feature vectors are output after multi-layer processing of the Transformer encoder; The fusion of the preprocessed CT scanning data, acoustic emission signals and DIC deformation field data comprises the following steps: The feature parameters output by the three-dimensional voxel model and the digital twin framework are preprocessed and spliced to obtain a mixed feature vector, which is taken as the input of the LSTM-CNN fusion model; The CNN branch scans the fracture distribution tensor through a 3×3 convolution kernel, identifies the typical spatial mode in the surrounding rock structure, outputs a 64-channel spatial feature map, aggregates the local features into a global structure description through a max-pooling layer, and outputs a 64×64 spatial feature matrix; The LSTM branch first performs time series encoding, and then captures the second derivative of the elastic modulus sudden drop event, identifies the progressive attenuation trend of the cohesion, and detects the crack propagation rate through the forget gate, input gate and output gate, and a fully connected layer converts the hidden state of the LSTM into a 64-dimensional time series feature vector; The spatial feature matrix output by the CNN branch is flattened into a one-dimensional vector, which is concatenated with the time series feature vector output by the LSTM branch to form a comprehensive feature vector containing spatial distribution rules and time evolution trends. The output layer generates the final result through linear transformation to obtain the optimal supporting time and supporting parameters.

2. The method of claim 1, wherein the method further comprises: The CT scan data, acoustic emission signals, DIC deformation field data, borehole imaging images and geological radar data of the rock are acquired and preprocessed, including: For the CT scan data, non-local mean filtering is used to remove metal artifacts; For the acoustic emission signals, wavelet packet transform is used to filter out noise; For the DIC deformation field data, median filtering is used to remove speckle image random noise, and sub-pixel interpolation is used to enhance pixels; For the borehole imaging image, the histogram equalization combined with the Otsu algorithm is used to enhance the contrast of the fissure and remove the drilling water film reflection; For the geological radar data, the sliding average method is used to remove background noise, and the K-S transform is used to suppress multiple waves.

3. The method for three-dimensional reconstruction of rocks by multi-source information fusion as described in claim 1, characterized in that, The 3D-CNN network is used to learn the spatial features of the fusion feature vector, and a three-dimensional voxel model of the rock micro-damage is constructed, including: The original spatial resolution based on the CT scan data output by the Transformer is converted into a three-dimensional tensor; The three-dimensional tensor is input into the 3D-CNN network, and the 3D-CNN network is used to extract multi-scale spatial features in layers, and output a three-dimensional voxel probability matrix; The high-value area in the CT scan data is extracted to obtain a mineral particle node, and based on the three-dimensional voxel probability matrix, a voxel cluster with a probability greater than or equal to 0.7 is extracted to obtain a micro-crack node; When the minimum distance between the mineral particles and the micro-cracks is less than 0.05mm, an undirected edge is established, and if the tip distance between the micro-cracks is less than 0.1mm, a directed edge is established, to obtain a graph structure; Based on the graph structure, the GNN network is used to capture the physical interaction to obtain a corrected voxel feature, and the GNN network is mapped back to the 3D-CNN network for iterative optimization; After the alternating processing of the 3D-CNN network and the GNN network, a three-dimensional voxel model of the rock micro-damage is constructed.

4. The multi-source information fusion rock three-dimensional reconstruction method of claim 3, wherein, The three-dimensional tensor is input into the 3D-CNN network, and the 3D-CNN network is used to extract multi-scale spatial features in layers, and output a three-dimensional voxel probability matrix, including: The bottom convolution layer slides with a step of 1, and performs convolution operation on each voxel and its 26 adjacent neighbors to extract basic spatial features; The middle convolution layer uses a step of 2 for downsampling to expand the receptive field and integrate the spatial relationship between local features; The high-level convolution layer integrates the full spatial features through a 1x1x1 convolution kernel, and finally outputs a three-dimensional voxel probability matrix.

5. The method for three-dimensional reconstruction of rocks by multi-source information fusion as described in claim 1, characterized in that, During the underground engineering construction process, the surrounding rock deformation data is collected in real time, and the deep deterministic policy gradient algorithm is used to adjust the supporting scheme in real time, including: The surrounding rock deformation data is input into the deep deterministic policy gradient algorithm after being filtered by Kalman filtering to remove noise; The Actor network outputs specific parameters according to the current state, generates supporting adjustment instructions after verification, and transmits them to the corresponding execution equipment; The adjusted supporting scheme is fed back to the digital twin framework to simulate and predict the evolution process of the surrounding rock structure.

6. A multi-source information fusion rock three-dimensional reconstruction system, characterized in that, The system comprises: a data acquisition module for acquiring CT scan data, acoustic emission signals, DIC deformation field data, borehole imaging images and geological radar data of the rock, and performing preprocessing; The first construction module is used for fusing the preprocessed CT scan data, acoustic emission signals and DIC deformation field data, performing spatial feature learning on the fused feature vectors through a 3D-CNN network, and constructing a three-dimensional voxel model of rock micro-damage; the preprocessed CT scan data, acoustic emission signals and DIC deformation field data are converted into data feature vectors, and position encoding is added, and then input into a Transformer encoder; eight groups of self-attention are adopted in the Transformer encoder, each group of attention head independently calculates the Q-K-V matrix of the three modal data, and the correlation between the data is obtained through dot product operation; after the self-attention calculation, the output features of the self-attention are normalized by layer, and the input of the Transformer encoder is added to the output features through residual connection; the fused features after self-attention and layer normalization are extracted through two fully connected layers to obtain high-order features after cross-modal fusion, and the fused feature vectors are output after multi-layer processing of the Transformer encoder; The second construction module is used for fusing the preprocessed borehole imaging image and the geological radar data, adopting a point cloud-based three-dimensional reconstruction algorithm to convert the fused image data into a point cloud model of the underground cavern surrounding rock structure, and based on an implicit surface reconstruction algorithm, a digital twin framework of the underground cavern surrounding rock structure is constructed; The output module is used for outputting the feature parameters of the three-dimensional voxel model and the digital twin framework as inputs, and outputting the optimal support timing and support parameters by using an LSTM-CNN fusion model; the feature parameters output by the three-dimensional voxel model and the digital twin framework are preprocessed and spliced to obtain a mixed feature vector, which is used as the input of the LSTM-CNN fusion model; the CNN branch scans the fracture distribution tensor through a 3×3 convolution kernel, identifies the typical spatial mode in the surrounding rock structure, outputs a 64-channel spatial feature map, aggregates the local features into a global structure description through a max-pooling layer, and outputs a 64×64 spatial feature matrix; the LSTM branch first performs time series encoding, and then captures the second derivative of the elastic modulus sudden drop event, identifies the progressive attenuation trend of cohesion, and detects the crack propagation rate through the forget gate, input gate and output gate, and a full connection layer converts the hidden state of the LSTM into a 64-dimensional time series feature vector; the spatial feature matrix output by the CNN branch is flattened into a one-dimensional vector, which is concatenated with the time series feature vector output by the LSTM branch to form a comprehensive feature vector containing spatial distribution rules and time evolution trends; the output layer generates the final result through linear transformation to obtain the optimal support timing and support parameters; The real-time adjustment module is used for real-time collection of surrounding rock deformation data during the construction of underground engineering, and real-time adjustment of the support scheme by using a deep deterministic policy gradient algorithm.

7. A rock three-dimensional reconstruction device of multi-source information fusion, characterized in that, The multi-source information fusion rock three-dimensional reconstruction device comprises a memory and at least one processor, and the memory stores instructions; the at least one processor invokes the instructions in the memory, so that the multi-source information fusion rock three-dimensional reconstruction device executes the steps of the multi-source information fusion rock three-dimensional reconstruction method according to any one of claims 1-5.

8. A computer-readable storage medium having stored thereon instructions, the computer-readable storage medium comprising: The instructions are executed by the processor to implement the steps of the multi-source information fusion rock three-dimensional reconstruction method according to any one of claims 1-5.

Images