Tunnel face peripheral hidden joint distribution prediction method and system based on spatiotemporal deep learning

Abstract

This invention relates to the field of tunnel engineering geological exploration and support design technology. Specifically, it relates to a method and system for predicting the distribution of hidden joints on the periphery of a tunnel face based on spatiotemporal deep learning. First, a training dataset is constructed based on a DFN model. Then, a dual-stream spatiotemporal deep learning model is built to extract the temporal features of historical cross-sections and the spatial constraint features of the exposed areas of the current cross-section. Finally, the features are fused and the joint distribution of the hidden areas of the current cross-section is reconstructed. This invention can achieve accurate prediction of joints in unexcavated areas. The prediction results can be directly imported into numerical simulation software to establish a refined geological model, and verified through subsequent construction using the bench method, providing a reliable basis for tunnel support design and safe construction.

Description

Technical Field

[0001] This invention relates to the field of tunnel engineering geological exploration and support design technology, specifically to a method and system for predicting the distribution of hidden joints on the periphery of the tunnel face based on spatiotemporal deep learning. Background Technology

[0002] During tunnel excavation, accurately understanding the joint distribution characteristics within the surrounding rock is crucial for assessing rock stability and optimizing support design. This is especially true when using the bench excavation method, where there is a strong spatial correlation between the geological information revealed by the upper bench excavation and the geological conditions of the lower bench and deeper surrounding rock.

[0003] Currently, joint prediction mainly relies on time-series prediction models, which predict the joint distribution at future excavation sections based on joint data from historically excavated sections. While this "predicting the future along the axial direction" method can provide a reference for construction planning, it has the following significant shortcomings in refined support design and numerical simulation analysis: 1. Lack of characterization of the integrity of the current cross section: Existing methods often ignore the strong constraint effect of the exposed part of the current cross section (such as the upper step) on the unexposed part (such as the lower step), resulting in insufficient prediction accuracy for the deep or unexcavated areas of the current cross section.

[0004] II. Lack of Numerical Modeling Basis: When conducting two-dimensional numerical simulations of tunnels (such as UDEC and PFC), it is usually necessary to establish a complete joint network model that includes the deep parts of the surrounding rock. Existing methods can only obtain the joint traces on the surface of the tunnel wall. Hidden joints outside the contour lines are usually randomly generated, which cannot truly reflect the mechanical anisotropy of the current cross-section.

[0005] Third, verification is difficult: conventional prediction methods predict the geological conditions a few meters deep in the surrounding rock, which are often difficult to verify directly through subsequent construction, leading to doubts about the reliability of the prediction model.

[0006] Therefore, there is an urgent need for a method that can combine historical evolution patterns and current known local information to accurately deduce the joint distribution in the current cross-section hidden area, so as to serve refined numerical modeling and targeted support design. Summary of the Invention

[0007] The present invention provides a method and system for predicting the distribution of hidden joints on the periphery of a tunnel face based on spatiotemporal deep learning. It makes full use of the full information of historical cross sections and the local known information of the current cross section to achieve accurate reconstruction and prediction of the distribution of joints in the unexposed area of ​​the current cross section and the surrounding rock mass. The method is verified through subsequent construction and establishes definite geological boundary conditions for the numerical model.

[0008] According to the present invention, a method for predicting the distribution of hidden joints around the tunnel face based on spatiotemporal deep learning includes the following steps: S1. Training data generation: 1.1) Establish a three-dimensional discrete fracture network (DFN) geological model based on the regional geological data of the target tunnel; 1.2) The DFN geological model is serialized and cut into sections along the tunnel axis to generate a two-dimensional joint network dataset of multiple continuous sections; 1.3) For each training sample cross-section, construct input-output data pairs; 1.4) Encode the joint data to form a feature vector representation; S2. Spatiotemporal deep learning model construction: 2.1) Construct a spatiotemporal deep learning prediction model. The model adopts a two-stream input structure, including: Temporal feature stream: used to receive historical tunnel face joint sequences and extract the spatial evolution features of geological structures along the tunnel axis using recurrent neural networks or 3D convolution; Spatial constraint flow: used to receive joint data of the exposed area of ​​the current cross section and extract the local geometric constraint features of the current cross section using a convolutional neural network; 2.2) Feature fusion decoding; S3, Model Training: 3.1) Divide the dataset generated in step S1 into a training set, a validation set, and a test set, and train the model constructed in step S2; 3.2) Define a composite loss function to measure the consistency between the predicted hidden region results and the actual DFN slices; 3.3) Use the gradient descent algorithm to optimize the model parameters until the model's prediction accuracy on the validation set meets the requirements; S4. Engineering Application Prediction.

[0009] Preferably, in step 1.1), the three-dimensional discrete fracture network (DFN) geological model includes the statistical distribution of joint dip, dip angle, trace length, density, and spacing.

[0010] Preferably, in step 1.3), the input-output data pair includes: Input data: includes complete joint distribution data of the first N historical cross sections, and joint distribution data of the exposed area of ​​the current cross section; Output data includes the distribution data of hidden joints in the unexposed area of ​​the current cross section and the surrounding set range.

[0011] As a preferred option, in step 1.4), the encoding processing method includes raster encoding, graph structure encoding, and point cloud encoding; during raster encoding, multi-source data feature alignment is performed, specifically: the raster features of geophysical data and joint data after raster encoding are fused through a cross-attention mechanism to highlight the geophysical signal features that are strongly correlated with the joint distribution.

[0012] Preferably, in step 2.2), the temporal evolution features are fused with the current spatial constraint features; and the complete joint network distribution of the current cross section is reconstructed through a decoder or a generator of a generative adversarial network.

[0013] Preferably, in step 3.2), the composite loss function includes: Reconstruction loss: used to predict pixel-level or feature-level differences between joint images and the true distribution; Topological loss: used to ensure that the predicted joint network is consistent with the actual geological statistical laws in terms of connectivity and cross-cutting relationships.

[0014] Preferably, step S4 specifically includes: 4.1) Data Acquisition: At the tunnel construction site, collect sketches of the previous N completed excavation sections, as well as joint sketch data of the step face on the current section. 4.2) Hidden Joint Prediction: The measured data is encoded and input into the trained model to predict and output the distribution of joint networks in the area to be excavated in the lower step of the current cross section and in the deep part of the surrounding rock. 4.3) Numerical modeling and analysis: Import the predicted complete joint network into numerical simulation software, establish a refined geological model, and perform surrounding rock stability analysis and support stress calculation; 4.4) Construction verification and feedback: After the lower bench is excavated, the measured joint data of the lower bench is collected and compared with the prediction results of step 4.2) for verification. The error is then fed back to the model for online fine-tuning. 4.5) Support optimization: Based on the numerical simulation results and the predicted location of weak structural surfaces, optimize the support parameters of the lower step and invert arch in advance.

[0015] Preferred numerical simulation software includes UDEC, PFC, and FLAC3D.

[0016] This invention provides a system for predicting the distribution of hidden joints around a tunnel face based on spatiotemporal deep learning. It employs the aforementioned method for predicting the distribution of hidden joints around a tunnel face based on spatiotemporal deep learning and includes: The data processing module is used to build the DFN model, generate slice data, and encode the measured data; The model building and training module is used to build a dual-stream spatiotemporal deep learning network and optimize its parameters. The prediction module is used to infer the joint distribution in the hidden area based on the input historical data and currently known data; The numerical modeling interface module is used to convert the predicted joint network into a geometric model file that can be recognized by numerical simulation software.

[0017] The beneficial effects of this invention are as follows: (1) The prediction logic is rigorous and the verifiability is strong: This invention utilizes the time difference of the step method construction to predict the next step by the upper step. The prediction results can be directly verified in the subsequent construction, which solves the problem that traditional prediction of deep surrounding rock cannot be verified. (2) Serving refined numerical simulation: It can reconstruct the hidden joint network outside the cave wall outline, which solves the problem that the deep joints of the surrounding rock can only be generated randomly in numerical simulation, resulting in the model boundary not matching the actual measurement, and significantly improves the accuracy of numerical analysis; (3) Make full use of local geometric constraints: Compared with pure time series prediction, the present invention adds constraint information of the exposed part of the current section and utilizes the spatial continuity of joint distribution within the same section, which significantly improves the prediction accuracy; (4) Pre-control of geological risks: The distribution of joints can be accurately grasped before the excavation of the lower bench, and the interconnected joints and potential rockfall areas can be identified in advance, effectively guiding safe construction. Attached Figure Description

[0018] Figure 1 This is a flowchart of a method for predicting the distribution of hidden joints around the tunnel face based on spatiotemporal deep learning, as described in Example 1. Figure 2 This is a schematic diagram of the prediction of hidden joints in the tunnel cross section in Example 1; Figure 3 This is a schematic diagram of the joint coding method in Example 1; Figure 4 This is a diagram of the spatiotemporal deep learning model architecture in the embodiment. Detailed Implementation

[0019] To further understand the content of this invention, a detailed description of the invention will be provided in conjunction with the accompanying drawings and embodiments. It should be understood that the embodiments are merely illustrative and not limiting of the invention.

[0020] Example 1 like Figure 1 As shown in the figure, this embodiment provides a method for predicting the distribution of hidden joints on the periphery of a tunnel face based on spatiotemporal deep learning, which includes the following steps: S1. Training data generation: 1.1) A three-dimensional discrete fracture network (DFN) geological model was established based on the regional geological data of the target tunnel. The three-dimensional discrete fracture network (DFN) geological model includes the statistical distribution of joint dip, dip angle, trace length, density, and spacing.

[0021] 1.2) The DFN geological model is serialized and cut into sections along the tunnel axis to generate a two-dimensional joint network dataset of multiple continuous sections.

[0022] 1.3) For each training sample cross-section, construct input-output data pairs; the input-output data pairs include: Input data: Includes complete joint distribution data of the first N historical cross sections, as well as joint distribution data of the exposed areas of the current cross section (such as the upper step or the outline of the tunnel wall); Output data includes the distribution data of hidden joints in the currently unexposed area of ​​the tunnel cross-section (such as the step below) and within a defined outer range (such as 3-5 meters outside the outline). A schematic diagram of hidden joint prediction for a tunnel cross-section is shown below. Figure 2 As shown.

[0023] 1.4) The joint data is encoded to form a feature vector representation; encoding methods include rasterization encoding, graph structure encoding, and point cloud encoding. Joint encoding methods include... Figure 3 As shown, (a) is rasterized encoding; (b) is graph structure encoding; and (c) is point cloud encoding.

[0024] When rasterizing (e.g., ground-penetrating radar profiles), multi-source data feature alignment is performed. Specifically, the raster features of geophysical data and joint data after rasterization are fused through a cross-attention mechanism to highlight geophysical signal features that are strongly correlated with joint distribution.

[0025] S2. Spatiotemporal deep learning model construction: 2.1) As Figure 4 As shown, a spatiotemporal deep learning prediction model is constructed. The model adopts a two-stream input structure, including: Temporal feature stream (temporal feature extraction): used to receive historical tunnel face joint sequences and extract the spatial evolution features of geological structures along the tunnel axis using recurrent neural networks (RNN / LSTM) or three-dimensional convolution (3D-CNN); Spatial constraint flow (spatial feature extraction): used to receive joint data of the exposed area of ​​the current cross section and use a convolutional neural network (CNN) to extract the local geometric constraint features of the current cross section.

[0026] 2.2) Feature fusion decoding; specifically: The temporal evolution features are fused with the current spatial constraint features (Concatenation / Attention); and the complete joint network distribution (including hidden regions) of the current cross section is reconstructed through a decoder or a generator of a generative adversarial network (GAN).

[0027] S3, Model Training: 3.1) Divide the dataset generated in step S1 into a training set, a validation set, and a test set, and train the model constructed in step S2.

[0028] 3.2) Define a composite loss function to measure the consistency between the predicted hidden region results and the actual DFN slices; Composite loss functions include: Reconstruction loss: used to predict pixel-level or feature-level differences between joint images and the true distribution; Topological loss: used to ensure that the predicted joint network is consistent with the actual geological statistical laws in terms of connectivity and cross-cutting relationships.

[0029] 3.3) Use the gradient descent algorithm to optimize the model parameters until the model's prediction accuracy on the validation set meets the requirements.

[0030] S4. Engineering Application Prediction.

[0031] In step S4, specifically: 4.1) Data Acquisition: At the tunnel construction site (such as bench method construction), collect sketches of the previous N completed excavation sections, as well as joint sketch data of the working face of the bench on the current section. 4.2) Hidden Joint Prediction: The measured data is encoded and input into the trained model to predict and output the distribution of joint networks in the area to be excavated in the lower step of the current cross section and in the deep part of the surrounding rock. 4.3) Numerical Modeling and Analysis: The predicted complete joint network is imported into numerical simulation software to establish a refined geological model, and the surrounding rock stability analysis and support stress calculation are performed; the numerical simulation software includes UDEC, PFC, and FLAC3D; 4.4) Construction verification and feedback: After the lower bench is excavated, the measured joint data of the lower bench is collected and compared with the prediction results of step 4.2) for verification. The error is then fed back to the model for online fine-tuning. 4.5) Support optimization: Based on the numerical simulation results and the predicted location of weak structural surfaces, optimize the support parameters of the lower step and invert arch in advance.

[0032] This embodiment provides a system for predicting the distribution of hidden joints around a tunnel face based on spatiotemporal deep learning. It employs the aforementioned method for predicting the distribution of hidden joints around a tunnel face based on spatiotemporal deep learning and includes: The data processing module is used to build the DFN model, generate slice data, and encode the measured data; The model building and training module is used to build a dual-stream spatiotemporal deep learning network and optimize its parameters. The prediction module is used to infer the joint distribution in the hidden area based on the input historical data and currently known data; The numerical modeling interface module is used to convert the predicted joint network into a geometric model file that can be recognized by numerical simulation software.

[0033] This embodiment makes full use of the full information of historical cross sections and the local known information of the current cross section (such as the upper step joint) to achieve accurate reconstruction and prediction of the distribution of joints in the unexposed area of ​​the current cross section (such as the lower step) and the surrounding rock mass. The reconstruction is verified through subsequent construction to establish the defined geological boundary conditions for the numerical model.

[0034] Example 2 This embodiment uses ConvLSTM to perform full-section hidden joint simulation and prediction, as detailed below: This embodiment uses a virtual simulation experiment based on the DFN geological model to verify the effectiveness of the method in predicting internal and external hidden joints based on joint outcrops on the tunnel wall outline in a full-section excavation scenario.

[0035] Data preparation and coding: (1) Use FracMan software or other similar open source programs to generate a three-dimensional DFN model to simulate the joint distribution in the granite area, including 3 sets of dominant joints. (2) Cut a two-dimensional cross section every 1 meter to generate a total of 1000 cross section samples. (3) Input data: Extract the joint intersection positions and attitudes on the tunnel design outline (horseshoe shape) of each cross section to simulate the tunnel wall sketch after full-section excavation; at the same time, extract the complete joint network of the previous 5 cross sections. (4) Output labels: The actual joint network distribution within the outline of the cross section (i.e., the tunnel cavity area, used for reverse verification) and within 5 meters outside the outline. (5) Use raster coding to divide the space into 10cm×10cm grids and record the joint attributes in each grid.

[0036] Model construction: (1) ConvLSTM (Convolutional Long Short-Term Memory Network) was selected as the prediction model, which is good at processing raster images of spatiotemporal sequence data. (2) An encoder-decoder structure was designed. The encoder extracts the features of historical cross-section sequences and splices and fuses them with the contour constraint features of the current cross-section. The decoder reconstructs the global joint distribution probability map of the current cross-section.

[0037] Prediction and Validation: (1) Input the cave wall contour data from the test set into the trained model. (2) The model outputs a prediction map of the hidden joints in the current section. (3) Validation results: Compare the prediction map with the actual slice data in the DFN model. The results show that the overlap (IoU) between the predicted joint traces and the actual traces reaches 78%, and the hidden intersecting joints that are not completely connected on the cave wall are successfully restored. This proves that the method can infer the internal structure based on the boundary conditions and can provide a high-precision geological model for numerical simulation.

[0038] Example 3 This embodiment uses the ST-GCN method for on-site prediction via the step method, as detailed below: This embodiment is applied to an actual mountain tunnel project constructed using the step-by-step method.

[0039] Training data generation: (1) Establish an engineering-scale DFN model based on geological survey data and generate slice data. (2) Input data: full-section joint map of historically excavated sections + measured joint map of the steps on the current section. (3) Output data: joint map of the area to be excavated on the step below the current section. (4) Use graph structure encoding to convert joint traces into graph nodes and edges, and use graph neural networks to process non-Euclidean data.

[0040] Model construction and training: (1) Construct a prediction model based on ST-GCN (Spatiotemporal Graph Convolutional Network). (2) Introduce an attention mechanism to enhance the spatial association weights of the upper step nodes to the lower step nodes. (3) Loss function: Combine node position regression loss and edge connection topology loss to optimize the connectivity prediction of joints.

[0041] Practical Applications: (1) On-site Data Collection: When construction reached chainage K10+050, the upper bench had been excavated and supported, while the lower bench had not been excavated. Geological engineers used photogrammetry to obtain joint data of the upper bench face and retrieved historical data from K10+045 to K10+049. (2) Prediction and Modeling: The data was input into the model, predicting that there was a set of continuous joints J2 in the lower bench area of ​​section K10+050, which extended into the depth of the surrounding rock. The prediction results were imported into UDEC software to establish a two-dimensional model. The calculation showed that the joints might cause the right sidewall of the lower bench to collapse. (3) Support Decision: Based on the simulation results, it was decided to add anchor pipes at the right arch foot of the upper bench before excavating the lower bench, and to reserve core soil on the right side of the lower bench. (4) Verification: Two days later, the lower bench was excavated, and the measured joint orientation and location matched the prediction results by 85%, proving the effectiveness of the prediction method.

[0042] The present invention and its embodiments have been described above illustratively. This description is not restrictive, and the figures shown are only one embodiment of the present invention; the actual structure is not limited thereto. Therefore, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the present invention, such designs should fall within the protection scope of the present invention.

Claims

1. A method for predicting distribution of concealed joints in the periphery of a tunnel face based on spatiotemporal deep learning, characterized in that: Includes the following steps: S1. Training data generation: 1.1) Establish a three-dimensional discrete fracture network (DFN) geological model based on the regional geological data of the target tunnel; 1.2) The DFN geological model is serialized and cut into sections along the tunnel axis to generate a two-dimensional joint network dataset of multiple continuous sections; 1.3) For each training sample cross-section, construct input-output data pairs; 1.4) Encode the joint data to form a feature vector representation; S2. Spatiotemporal deep learning model construction: 2.1) Construct a spatiotemporal deep learning prediction model. The model adopts a two-stream input structure, including: Temporal feature stream: used to receive historical tunnel face joint sequences and extract the spatial evolution features of geological structures along the tunnel axis using recurrent neural networks or 3D convolution; Spatial constraint flow: used to receive joint data of the exposed area of ​​the current cross section and extract the local geometric constraint features of the current cross section using a convolutional neural network; 2.2) Feature fusion decoding; S3, Model Training: 3.1) Divide the dataset generated in step S1 into a training set, a validation set, and a test set, and train the model constructed in step S2; 3.2) Define a composite loss function to measure the consistency between the predicted hidden region results and the actual DFN slices; 3.3) Use the gradient descent algorithm to optimize the model parameters until the model's prediction accuracy on the validation set meets the requirements; S4. Engineering Application Prediction.

2. The spatiotemporal deep learning-based tunnel face periphery hidden joint distribution prediction method according to claim 1, characterized in that: In step 1.1), the three-dimensional discrete fracture network (DFN) geological model includes the statistical distribution of joint dip, dip angle, trace length, density, and spacing.

3. The spatiotemporal deep learning-based tunnel face periphery hidden joint distribution prediction method according to claim 2, characterized in that: In step 1.3), the input-output data pair includes: Input data: Includes complete joint distribution data of the first N historical cross sections, and joint distribution data of the exposed area of ​​the current cross section; Output data includes the distribution data of hidden joints in the unexposed area of ​​the current cross section and the surrounding set range.

4. The method for predicting the distribution of hidden joints around the tunnel face based on spatiotemporal deep learning according to claim 3, characterized in that: In step 1.4), the encoding processing methods include rasterization encoding, graph structure encoding, and point cloud encoding. During rasterization encoding, multi-source data feature alignment is performed. Specifically, the raster features of geophysical data and joint data after rasterization encoding are fused through a cross-attention mechanism to highlight geophysical signal features that are strongly correlated with joint distribution.

5. The spatiotemporal deep learning-based tunnel face peripheral hidden joint distribution prediction method according to claim 4, characterized in that: In step 2.2), the temporal evolution features are fused with the current spatial constraint features; and the complete joint network distribution of the current cross section is reconstructed through a decoder or a generator of a generative adversarial network.

6. The spatiotemporal deep learning-based tunnel face periphery blind joint distribution prediction method according to claim 5, characterized in that: In step 3.2), the composite loss function includes: Reconstruction loss: used to predict pixel-level or feature-level differences between joint images and the true distribution; Topological loss: used to ensure that the predicted joint network is consistent with the actual geological statistical laws in terms of connectivity and cross-cutting relationships.

7. The spatiotemporal deep learning-based tunnel face periphery hidden joint distribution prediction method according to claim 6, characterized in that: In step S4, specifically: 4.1) Data Acquisition: At the tunnel construction site, collect sketches of the previous N completed excavation sections, as well as joint sketch data of the step face on the current section. 4.2) Hidden Joint Prediction: The measured data is encoded and input into the trained model to predict and output the distribution of joint networks in the area to be excavated in the lower step of the current cross section and in the deep part of the surrounding rock. 4.3) Numerical modeling and analysis: Import the predicted complete joint network into numerical simulation software, establish a refined geological model, and perform surrounding rock stability analysis and support stress calculation; 4.4) Construction verification and feedback: After the lower bench is excavated, the measured joint data of the lower bench is collected and compared with the prediction results of step 4.2) for verification. The error is then fed back to the model for online fine-tuning. 4.5) Support optimization: Based on the numerical simulation results and the predicted location of weak structural surfaces, optimize the support parameters of the lower step and invert arch in advance.

8. The spatiotemporal deep learning-based tunnel face periphery hidden joint distribution prediction method according to claim 7, characterized in that: Numerical simulation software includes UDEC, PFC, and FLAC3D.

9. A system for predicting distribution of concealed joints in the periphery of a tunnel face based on spatiotemporal deep learning, characterized by: It employs the method for predicting the distribution of hidden joints around the tunnel face based on spatiotemporal deep learning as described in any one of claims 1-8, and includes: The data processing module is used to build the DFN model, generate slice data, and encode the measured data; The model building and training module is used to build a dual-stream spatiotemporal deep learning network and optimize its parameters. The prediction module is used to infer the joint distribution in the hidden area based on the input historical data and currently known data; The numerical modeling interface module is used to convert the predicted joint network into a geometric model file that can be recognized by numerical simulation software.

Images