A deep learning-based high-stress soft rock tunnel large deformation prediction method

Abstract

The application discloses a high-stress soft rock tunnel large deformation prediction method based on deep learning, and belongs to the field of tunnel engineering geological disaster prediction. Through the construction of a double-branch network architecture of "geological pattern identification-deep learning prediction", the accurate prediction of tunnel large deformation is realized. First, six qualitative indexes such as rock mass structure, weathering degree and underground water condition and three quantitative indexes such as strength stress ratio, support stiffness and equivalent hole diameter are collected, and after coding and normalization preprocessing, they are respectively input into the upper and lower networks: the upper layer adopts a convolutional neural network to identify the surrounding rock geological pattern, and outputs the probability distribution of six types of large deformation modes such as soft homogeneous type and block structure type; the lower layer predicts the deformation amount and divides the risk grade through a fully connected neural network based on the geological pattern identification result and the quantitative index. The application can fully reflect the complex coupling mechanism of high-stress soft rock tunnel large deformation; and adaptive prediction is realized for different geological patterns.

Description

Technical Field

[0001] This invention relates to the field of geological hazard prediction in tunnel engineering, and in particular to a method for predicting large deformations in high-stress soft rock tunnels based on deep learning. Background Technology

[0002] In high-stress soft rock tunnel engineering, large deformation of the surrounding rock is a major hidden danger leading to construction delays, support system failures, and even safety accidents. Traditional tunnel deformation prediction methods are often based on single geological parameters (such as rock strength, ground stress, etc.), empirical formulas, and traditional functions. Gou Chao, Long Xin, Liu Zhi, et al. (Prediction of Arch Deformation of Highway Tunnel Based on Regression Analysis [C], 2022:566-571.DOI:10.26914) Based on the measured data of arch settlement of a highway tunnel in different levels of surrounding rock, this paper compares and analyzes the fitting and prediction accuracy of five regression function models in current specifications and related studies on the arch settlement deformation of tunnels under different levels of surrounding rock. Among them, the polynomial function has the highest fitting accuracy for the measured data, but it is not suitable for the deformation prediction of tunnel arches. With the gradual development of tunnel construction in my country towards deep burial, long tunnels, and large diameters, tunnels traversing complex geological conditions such as high ground stress, soft rock, and faults are becoming increasingly common. When making predictions, it is necessary to comprehensively consider the effects of tunnel burial depth, span, surrounding rock strength-stress ratio, groundwater, and rock mass structure.

[0003] For example, Chen Weizhong, Tian Yun, Wang Xuehai, et al. (Prediction of Compression Deformation of Soft Rock Tunnels Based on Corrected [BQ] Value [J]. Rock and Soil Mechanics, 2019, 40(08):3125-3134.DOI:10.16285) proposed that the amount of compression deformation can be predicted directly and quickly based on the tunnel burial depth, span, and rock mass corrected [BQ] index parameters. This method is mainly applicable to deformation prediction before tunnel excavation and is difficult to adjust in real time. During tunnel construction, multiple factors should be considered in various stages such as tunnel exploration, design, and construction to effectively predict large deformation disasters. This will help in management, optimization of support schemes, avoidance of rework, and ensuring construction safety.

[0004] In Huang Hongjian's study (Research on Numerical Simulation Prediction of Large Deformation in the Construction of Baozhen Tunnel in High Ground Stress and Weak Surrounding Rock Section [J]. Railway Standard Design, 2009, (03): 93-95. DOI: 10.13238), a large-scale finite difference program was used to conduct three-dimensional numerical simulation prediction of the large deformation section of Baozhen Tunnel on the Yichang-Wanzhou Railway. However, the prediction error gradually increases as the distance between the detection section and the tunnel face increases. It can be seen that traditional tunnel deformation prediction methods rely on empirical formulas, numerical simulation and field monitoring, which often have limitations such as difficulty in adapting to complex geological conditions, high computational cost and poor real-time performance of numerical simulation, low efficiency of field monitoring data utilization, and inability to achieve advanced prediction and dynamic adjustment.

[0005] In recent years, with the rapid development of artificial intelligence technology, the application of intelligent prediction methods in the field of tunnel engineering has gradually deepened, especially providing a technological breakthrough for the prevention and control of large deformations under complex geological conditions such as high ground stress and soft rock. Among them, deep learning technology, with its ability to automatically extract features and model complex mappings from high-dimensional nonlinear data, effectively makes up for the insufficient adaptability of traditional empirical formulas (such as the Hawke-Brown criterion) and linear regression models in complex scenarios, and has become one of the core technical paths for tunnel deformation prediction. Xing Pengtao et al. optimized the XGBOOST regression model using the ZOA algorithm (CN202510682278.X "A Multi-Risk Source Tunnel Construction Early Warning Method"). By expanding tunnel face images and construction site monitoring data into important data sources, a corresponding database was established. Using image recognition, neural networks, machine learning algorithms, and other algorithms, a tunnel face joint extraction model was built, and a tunnel deformation prediction model was constructed, which can effectively predict deformation. Wang Lan et al. (CN202411772417.X "A Multi-Factor Tunnel Surrounding Rock Deformation Prediction Method Based on Deep Learning") used the ISAI-MHA-LSTM model to dynamically focus on key time-series data for deformation prediction. The model is built based on seven information factors specific to the project, but its adaptability to different geological conditions and tunnel types is low. Su Xulin et al. (CN202110639207.3 "Tunnel Construction Deformation Prediction Method and System Based on Composite Neural Network") integrated CNN, LSTM, and MLP networks. CNN extracts features from construction status images, LSTM captures the temporal patterns of deformation data, and MLP corrects the prediction results, with clear division of labor and complementarity. However, it requires simultaneous acquisition of high-quality construction status images and long-term deformation monitoring data. Image acquisition is affected by tunnel lighting and occlusion, and missing data directly reduces the prediction effect. Furthermore, multiple models place high demands on computing power. The essence of large tunnel deformation is the external manifestation of stress redistribution and damage evolution in the rock mass after the mechanical equilibrium of the surrounding rock is broken by excavation disturbance. Its prediction accuracy directly depends on the ability to characterize the coupling relationship between construction disturbance, geological characteristics, and mechanical response.

[0006] In practical application, the industry has conducted targeted explorations focusing on different construction methods and deformation types. Wang Mingnian and Yi Wenhao, in their patent (CN202310491104.6 "Method, Device, and Medium for Predicting Large Deformation Levels in Drill-and-Blast Tunnel Construction"), focused on the dynamic disturbance characteristics of drill-and-blast construction. They innovatively used drilling speed, drilling pressure, and blast hole layout parameters during drilling, along with parameters such as charge quantity and vibration energy during blasting, as input features to a deep learning model, constructing a technical closed loop of "real-time parameter acquisition - online model prediction - risk classification and early warning." This method effectively solves the pain points of traditional manual monitoring relying on experience-based interpretation and having a delayed response, reducing manual input by more than 40%. However, limited by the application scenario, its core parameter system focuses more on the disturbance intensity of drill-and-blast construction and does not fully incorporate the unique geological constitutive parameters of high-stress soft rock (such as rock damage threshold and long-term creep characteristics). This results in insufficient accuracy in predicting slow deformation induced by low disturbance but high stress in deeply buried soft rock tunnels.

[0007] To address the asymmetric deformation problem more prone to occur in soft rock tunnels, Xue Yiguo et al., in their patent "A Method and Device for Grading and Intelligent Prediction of Asymmetric Deformation in Soft Rock Tunnels" (CN202211313735.0), proposed a technical approach combining data dimensionality reduction and nonlinear modeling. This method uses principal component analysis (PCA) to screen core features with a cumulative contribution rate exceeding 85% from multi-source monitoring data such as surrounding rock stress difference, cross-sectional convergence rate, and groundwater seepage flow, effectively eliminating data redundancy. Then, combined with an improved support vector machine, a graded prediction model is constructed, controlling the relative error of asymmetric deformation prediction to within 10%. However, its technical limitation lies in the fact that the feature selection process focuses more on the deformation monitoring data itself, failing to systematically correlate the coupling relationship between the magnitude and direction of ground stress, surrounding rock lithology, and RQD value under high ground stress conditions. This makes it difficult to explain the geological roots of asymmetric deformation (such as stress concentration caused by weak interlayers in local rock masses), resulting in a decrease in the model's generalization ability in areas of abrupt geological changes (such as fault fracture zones).

[0008] From the perspective of geomechanical mechanisms, Wang Xue et al., in patent CN202210801985.2 "A Method for Constructing a Geomechanical Model of Large Deformation Due to Compression in Deeply Buried Tunnels," further revealed the essential laws of deformation. This patent innovatively introduces a dynamic plastic potential function to characterize the continuous evolution of the surrounding rock from damage expansion to fracture and dilatation after the excavation of a deep-buried tunnel. When the damage variable (fracture density) is small, the plastic potential function is dominated by damage expansion; as damage accumulates, the function linearly transitions to a fracture and dilatation mode, accurately reproducing the time effect of slow initial deformation and subsequent rapid increase in deformation of soft rock under high ground stress. This research confirms that the geomechanical model is the core for improving the physical interpretability of prediction models. However, this model focuses more on theoretical mechanism derivation and is not combined with the real-time data processing capabilities of deep learning, making it difficult to achieve dynamic prediction during construction. Furthermore, it does not cover the regulatory role of engineering design parameters such as support stiffness and equivalent tunnel diameter on deformation, resulting in a certain disconnect from practical engineering applications.

[0009] Regarding the integration of geological models and prediction models, Tao Zhiping and Zhou Depei, in their paper "Deformation Law and Disaster Prediction of Creep-Type Landslide Tunnels" (Journal of Southwest Jiaotong University, 2007, 02: 163-168), first constructed a technical framework of "geomechanical model-prediction model-disaster correlation". This study, focusing on tunnels within landslide bodies, established three types of geomechanical models reflecting the coupling relationship between landslide thrust and tunnel deformation. Combining a grey isodimensional information model with finite element numerical simulation, they achieved the correlation prediction of "tunnel deformation-landslide stress displacement," providing an early approach to embedding geological models into prediction models. However, limited by technical means, the grey model used has a weak ability to fit nonlinear data, and the finite element simulation relies on numerous parameter assumptions, making it difficult to adapt to the characteristics of "variable geological conditions and real-time updates of construction data" in high-stress soft rock tunnels. The prediction accuracy and real-time performance can no longer meet the current needs of intelligent construction.

[0010] In summary, while existing tunnel deformation prediction technologies have made progress in data-driven approaches, mechanism analysis, and scenario adaptation, they still suffer from three core shortcomings when applied to the specific scenario of high-stress soft rock tunnels: First, the integration of geological models and deep learning is insufficient. Existing deep learning models largely rely on construction parameters or monitoring data, failing to systematically embed key geological and engineering parameters such as burial depth, RQD value, ground stress magnitude, and support stiffness, resulting in a lack of geological specificity. Second, there is a certain disconnect between mechanism-driven and data-driven approaches, with an emphasis on theoretical mechanisms or data fitting, failing to address the complex nonlinearities of high-stress soft rock deformation. Third, scenario adaptability is insufficient. Existing methods do not fully consider the characteristics of high-stress soft rock tunnels, making it difficult to dynamically adjust the prediction model during construction, leading to insufficient early warning windows. These technological gaps mean that the accuracy, generalization ability, and engineering practicality of current prediction methods in high-stress soft rock tunnels need improvement. There is an urgent need to construct an intelligent prediction model that deeply integrates geological models and deep learning to meet the requirements of safe construction. Therefore, there is an urgent need for an intelligent prediction method that can deeply integrate geological model characteristics with real-time monitoring data. Summary of the Invention

[0011] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for predicting large deformation of soft rock tunnels under high ground stress based on deep learning.

[0012] The objective of this invention is achieved through the following technical solution: a method for predicting large deformation of high-stress soft rock tunnels based on deep learning, comprising the following steps:

[0013] Parameter acquisition and preprocessing stage: Integrate geological survey data before construction, field test data and dynamic monitoring data during construction to construct a database of large deformation influence parameters including multiple qualitative and quantitative indicators; classify and encode the qualitative indicators, normalize the quantitative indicators to the [0,1] interval using min-max normalization, and remove outliers using the 3σ criterion and supplement missing values ​​using K-nearest neighbor interpolation to form standardized input data;

[0014] Geological model identification stage: The preprocessed qualitative index encoding vector is input into the geological model identification network to obtain the probability distribution of multiple geological models. When the probability of a certain geological model is greater than or equal to the preset value, it is determined to be the corresponding geological model and marked as "surrounding rock prone to large deformation". Otherwise, it is marked as "surrounding rock not prone to large deformation".

[0015] Deformation prediction stage: If the identification result is "surrounding rock prone to large deformation", the geological model code and the preprocessed quantitative index normalized vector are concatenated into a multi-dimensional fusion vector, which is then input into the deep learning prediction network; the continuous deformation prediction value is output through the linear activation function, and the discrete large deformation level is divided based on the deformation range.

[0016] Results output stage: Output the predicted deformation value and its corresponding large deformation level, and provide geological model identification results as a basis for risk warning.

[0017] Preferably, the qualitative indicators include rock mass structure, weathering degree, groundwater conditions, relationship between the main controlling structural plane and the tunnel axis, degree of tectonic influence, and rock mass hardness; the quantitative indicators include corrected strength-stress ratio, support stiffness, and equivalent tunnel diameter.

[0018] Preferably, in the parameter acquisition and preprocessing stage, the qualitative index coding rules are as follows: the rock mass structure is divided into 1-5 levels according to the loose, fragmented, blocky, layered, and integral state; the weathering degree is divided into 1-5 levels according to the fully weathered, strongly weathered, moderately weathered, slightly weathered, and unweathered state; the groundwater conditions, the relationship between the main control structural plane and the tunnel axis, the degree of tectonic influence, and the rock mass hardness are all divided into 1-5 levels according to the severity.

[0019] Preferably, the geological pattern identification network is based on a convolutional neural network architecture. It extracts features through convolutional layers and retains key information through pooling layers, outputting the probability distribution of six types of geological patterns, including weak homogeneous type, block-fracture structure type, weak layered type, layered fault type, weak interlayered type, and layered fold type.

[0020] Preferably, the specific architecture of the geological model identification network includes:

[0021] The input layer is used to receive a 6-dimensional qualitative index encoding vector;

[0022] Two convolutional layers extract single-indicator correlation features and mine multi-indicator coupling features through convolutional kernels to capture geological model-related differences;

[0023] The pooling layer is used to perform max pooling operations, reducing computation while preserving key features;

[0024] Flattening layers are used to convert multidimensional feature maps into one-dimensional vectors;

[0025] A fully connected layer is used to output the probability distribution of six geological models.

[0026] Preferably, the deep learning prediction network is based on a fully connected neural network architecture, comprising:

[0027] The input layer is used to receive a 4-dimensional fusion vector, which is composed of geological model encoding and normalized vectors of three quantitative indicators.

[0028] Fully connected layer 1 and fully connected layer 2, with ReLU activation function and 0 dropout rate;

[0029] Output layer 1 is used to output the predicted value of continuous deformation through a linear activation function;

[0030] Output layer 2 is used to classify discrete large deformation levels based on the deformation range.

[0031] The beneficial effects of this invention are:

[0032] 1) This invention breaks through the limitations of traditional prediction methods that rely on single parameters or linear models. It integrates nine key engineering and geological parameters, including rock hardness, weathering intensity, groundwater conditions, and rock structure. It combines "qualitative identification of geological patterns prone to deformation of surrounding rock" with "quantitative prediction of tunnel deformation" to form a dual-task output system. Through multi-layer neural networks, it automatically mines the nonlinear correlation between parameters, which can fully reflect the complex coupling mechanism of large deformation of soft rock tunnels under high ground stress. It achieves adaptive prediction for the differentiated deformation patterns of different geological patterns, effectively reducing the prediction error caused by insufficient adaptability of geological conditions and greatly improving the accuracy of deformation prediction. Attached Figure Description

[0033] Figure 1 This is a flowchart of the method of the present invention;

[0034] Figure 2 Network diagram for geological model discrimination and deformation prediction;

[0035] Figure 3 This is a graph showing the changes in model training loss and accuracy. Detailed Implementation

[0036] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] See Figures 1-3 This invention provides a technical solution: a method for predicting large deformation of high-stress soft rock tunnels based on deep learning, comprising the following steps:

[0038] Parameter acquisition and preprocessing stage: Integrate geological survey data before construction, field test data and dynamic monitoring data during construction to construct a database of large deformation influence parameters including multiple qualitative and quantitative indicators; classify and encode the qualitative indicators, normalize the quantitative indicators to the [0,1] interval using min-max normalization, and remove outliers using the 3σ criterion and supplement missing values ​​using K-nearest neighbor interpolation to form standardized input data;

[0039] Geological model identification stage: The preprocessed qualitative index encoding vector is input into the geological model identification network to obtain the probability distribution of multiple geological models. When the probability of a certain geological model is greater than or equal to the preset value, it is determined to be the corresponding geological model and marked as "surrounding rock prone to large deformation". Otherwise, it is marked as "surrounding rock not prone to large deformation".

[0040] Deformation prediction stage: If the identification result is "surrounding rock prone to large deformation", the geological model code and the preprocessed quantitative index normalized vector are concatenated into a multi-dimensional fusion vector, which is then input into the deep learning prediction network; the continuous deformation prediction value is output through the linear activation function, and the discrete large deformation level is divided based on the deformation range.

[0041] Results output stage: Output the predicted deformation value and its corresponding large deformation level, and provide geological model identification results as a basis for risk warning.

[0042] In this embodiment, a database of surrounding rock geological models is constructed, and combined with a deep learning model, to predict tunnel deformation in real time and with high accuracy. This not only draws on the advantages of existing intelligent prediction technologies but also strengthens the correlation analysis between geological models and deformation responses, thereby improving the generalization ability and engineering applicability of the prediction model.

[0043] First, parameter acquisition and preprocessing are performed, integrating pre-construction geological survey data, field test data, and dynamic monitoring data during construction to construct a database of large deformation influence parameters containing 6 qualitative indicators and 3 quantitative indicators. These include rock mass structure (divided into loose, fragmented, blocky, layered, and integral types) and weathering degree (fully weathered, strongly weathered, moderately weathered, slightly weathered, and unweathered). The 3 quantitative indicators include corrected strength-stress ratio, support stiffness, and equivalent tunnel diameter. Qualitative indicators are coded at levels 1-5, and quantitative indicators are normalized to the [0,1] interval using the min-max method. Outliers are removed using the 3σ criterion, and missing values ​​are supplemented using K-nearest neighbor interpolation. Subsequently, phased prediction is performed, using a geological model identification network with upper branches to achieve qualitative identification. The preprocessed 6 qualitative indicator encoding vectors are input into the identification network. The system outputs the probability distributions of six geological models (weak homogeneous, blocky fracture, weak layered, layered fault, weak interlayered, and layered fold). When the probability of a certain geological model is ≥0.6, it is identified as the corresponding geological model and is classified as "surrounding rock prone to large deformation." When the probabilities of all models are <0.6 and the overall identification score is high, it is classified as "surrounding rock not prone to large deformation." A deep learning prediction network with lower branches is used to achieve quantitative prediction. The geological model codes (1-6) are concatenated with the normalized vectors of the three preprocessed quantitative indicators to form a 4-dimensional fusion vector, which is then used to output the predicted deformation amount through a linear activation function. Finally, large deformation amount prediction is performed, and the level is determined by the predicted deformation amount. Simultaneously, multi-task learning is used to optimize two related but different prediction tasks, improving the model's generalization ability and achieving the prediction of surrounding rock deformation level.

[0044] like Figure 2 As shown, in order to ensure that the key features of relevant indicators can be effectively combined, the model adopts a "dual-branch cascade" architecture, which is divided into a geological model identification branch (upper layer) and a deformation prediction branch (lower layer). The specific parameter settings are as follows.

[0045] The geological pattern identification branch receives preprocessed qualitative index data (six qualitative index encoding vectors) from the input layer. It then extracts correlated features and biochemical features through two convolutional layers to capture differences in geological patterns. Pooling layers reduce the number of parameters to avoid overfitting. A flattened format is used to provide the input format for the fully connected layers, ultimately outputting a geological pattern classification.

[0046] Deformation prediction branch: This branch receives preprocessed quantitative index data (normalized vector of quantitative index) from the input layer and flattens it into a one-dimensional feature vector, which is convenient for splicing with the feature vector of the geological model.

[0047] Finally, the feature indicators of the two branches are concatenated and used for final prediction. A connecting layer is used to fuse the geological model feature vector and the quantitative deformation feature vector into a "qualitative-quantitative" feature, improving the comprehensiveness of the prediction. During the prediction process, a fully connected layer connects to the output layer. The fully connected layer uses ReLU as the activation function and has a dropout rate of 0, deeply fusing the features from both branches.

[0048] The main parameters used in model training are as follows: Adam is used as the optimizer, the initial learning rate is set to 0.001, and the learning rate is decayed every 5 epochs at a rate of 0.8 times the original learning rate. The training consists of 50 epochs with a batch size of 16.

[0049] See Figure 3 This visually demonstrates the model's excellent convergence characteristics and generalization ability during training, validating the effectiveness of the dual-task output under the "dual-branch cascade" architecture. From the loss curve ( Figure 3 (a) and Figure 3 (c) From this perspective, the loss values ​​of y1 and y2 both show a steep downward trend in the early stages of iteration (approximately 0-10 rounds), and then quickly stabilize. Furthermore, the training and validation set curves are highly correlated, with no rebound in validation loss, indicating that the model effectively suppresses overfitting while rapidly extracting rock mass features. Regarding accuracy ( Figure 3 (b) and Figure 3 (d) Both curves show a step-like increase, with the final accuracy of the y2 output (corresponding to the lower deformation prediction branch) climbing to a high level of around 98%, better than the approximately 86% level of the y1 output (corresponding to the upper geological model identification branch). This not only confirms the description in the manual regarding the model's "rapid learning of data features" and "good generalization ability," but also confirms that the model can effectively establish a nonlinear mapping between geological parameters and deformation levels through multi-layer neural networks, meeting the accurate prediction requirements of high-stress soft rock tunnel engineering. From the curves showing the changes in training set and validation set loss and accuracy during model training, it can be seen that in the rapid convergence phase of 0-10 rounds, the training set loss decreased from 2.1 to 0.6, and the validation set loss decreased from 2.0 to 0.7; the training set accuracy increased from 35% to 78%, and the validation set accuracy increased from 32% to 75%, indicating that the model quickly learns data features without significant overfitting.

[0050] From rounds 11-30, during the stable optimization phase: the training set loss decreased from 0.6 to 0.3, and the validation set loss decreased from 0.7 to 0.4; the training set accuracy increased from 78% to 92%, and the validation set accuracy increased from 75% to 88%. The model gradually optimized its parameters and fitted complex nonlinear correlations. From rounds 31-50, during the convergence and stabilization phase: the training set loss stabilized between 0.25 and 0.3, and the validation set loss stabilized between 0.35 and 0.4; the training set accuracy stabilized between 92% and 93%, and the validation set accuracy stabilized between 88% and 89%. The model converged and demonstrated good generalization ability.

[0051] See Figure 1 and Figure 2 Taking the Kangding No. 1 Tunnel as an example, qualitative indicators such as rock characteristics and groundwater conditions of each tunnel are first obtained through reasonable methods. Then, quantitative indicators from the tunnel excavation process, such as support stiffness and equivalent tunnel diameter, are used as inputs x1 (qualitative) and x2 (quantitative) to the model, respectively. Finally, large deformation level identification and prediction are performed. The predicted and identified large deformation levels are the deformation risk levels (Level 1, Level 2, and Level 3, with higher levels indicating greater risk) obtained from the model prediction and manual comprehensive assessment, respectively. The model validation set contains 161 data points, with 133 correct identifications, resulting in a validation set accuracy of 82%. This intelligent prediction model for large tunnel deformation, integrating multiple qualitative and quantitative indicators, can effectively and accurately predict large deformations and identify their levels.

[0052] In some embodiments, the qualitative indicators include rock mass structure, degree of weathering, groundwater conditions, relationship between the main control structural plane and the tunnel axis, degree of tectonic influence, and rock mass hardness; the quantitative indicators include the corrected strength-stress ratio, support stiffness, and equivalent tunnel diameter.

[0053] In this embodiment, based on the "Engineering Rock Mass Classification Standard" and statistical analysis of multiple high-stress soft rock tunnel engineering cases, nine key parameters were selected and divided into qualitative indicators (for geological model identification) and quantitative indicators (for deformation prediction). The rock mass structure, weathering degree, and tectonic influence were obtained through engineering geological surveys; the maximum principal stress was obtained through in-situ stress testing, and the corrected strength-stress ratio was preliminarily calculated; the tunnel face was observed every 5m of excavation to update the groundwater conditions and the relationship between the main control structural plane and the tunnel axis; the shotcrete thickness and anchor bolt parameters were recorded to update the support stiffness and correct the equivalent tunnel diameter.

[0054] Parameter standardization and qualitative index coding: The six qualitative indicators are converted into numerical codes of 1-5 according to the above classification criteria, forming a coding vector of dimension 6×1 (denoted as Input x1); Quantitative index normalization: Min-max normalization is used to map the three quantitative indicators to the [0,1] interval; Outlier and missing value handling: The 3σ criterion is used to remove outliers (data that exceed the range of [x_mean-3σ, x_mean+3σ]) in the quantitative indicators, and missing values ​​are filled by K-nearest neighbor interpolation (K=5, based on the five most similar samples).

[0055] In some embodiments, during the parameter acquisition and preprocessing stage, the qualitative index coding rules are as follows: the rock mass structure is classified into 1-5 levels according to its loose, fragmented, blocky, layered, and integral state; the weathering degree is classified into 1-5 levels according to its complete weathering, strong weathering, moderate weathering, slight weathering, and unweathered state; and the groundwater conditions, the relationship between the main control structural plane and the tunnel axis, the degree of tectonic influence, and the rock mass hardness are all classified into 1-5 levels according to their severity.

[0056] In some embodiments, the geological pattern identification network is based on a convolutional neural network architecture. It extracts features through convolutional layers and retains key information through pooling layers, outputting the probability distribution of six types of geological patterns, including weak homogeneous type, block-fracture structure type, weak layered type, layered fault type, weak interlayered type, and layered fold type.

[0057] In this embodiment, the upper-level branch qualitative identification stage identifies the geological model of the surrounding rock. By constructing a geological model identification network of "convolution-pooling-fully connected layers," it achieves qualitative identification of "whether large deformation is likely" and "specific geological model type," providing geological mechanism constraints for the prediction of lower-level deformation. First, qualitative indicators determine whether large deformation is likely. The collected parameters are input into the upper-level branch surrounding rock geological model identification module. The core of this module is the model identification network. The identification indicators are extracted from the input data through two convolutional layers to mine geological model-related feature information. Then, a pooling layer is used for downsampling to retain key features and reduce computational load. A flattening layer is used to prepare for the fully connected layer. Finally, preliminary classification results are output.

[0058] In some embodiments, the specific architecture of the geological pattern identification network includes:

[0059] The input layer is used to receive a 6-dimensional qualitative index encoding vector;

[0060] Two convolutional layers extract single-indicator correlation features and mine multi-indicator coupling features through convolutional kernels to capture geological model-related differences;

[0061] The pooling layer is used to perform max pooling operations, reducing computation while preserving key features;

[0062] Flattening layers are used to convert multidimensional feature maps into one-dimensional vectors;

[0063] A fully connected layer is used to output the probability distribution of six geological models.

[0064] In this embodiment, the first stage involves constructing a geological model discrimination network for surrounding rocks prone to large deformation based on the qualitative index system, performing preliminary geological model discrimination. The input to the geological model discrimination network is a preprocessed 6-dimensional qualitative index encoding vector. The convolutional layer extracts single-index correlation features and mines multi-index coupling features through convolutional kernels. The output is downsampled to reduce computation by 30% while retaining key features. Various qualitative indicators, such as the relationship between structural plane attitude and axis, rock mass structure, and groundwater conditions, are used as model input. The upper-layer network determines whether the surrounding rock is prone to large deformation, specifically classifying it as either easily deformable or not. This achieves preliminary early warning of surrounding rock condition and large deformation risk. The "easily deformable / not easily deformable" discrimination result serves as a preliminary warning signal: if the surrounding rock is easily deformable, deformation prediction is triggered; if it is not easily deformable, a low-risk result is directly output.

[0065] In some embodiments, the deep learning prediction network is based on a fully connected neural network architecture, including:

[0066] The input layer is used to receive a 4-dimensional fusion vector, which is composed of geological model encoding and normalized vectors of three quantitative indicators.

[0067] Fully connected layer 1 and fully connected layer 2, with ReLU activation function and 0 dropout rate;

[0068] Output layer 1 is used to output the predicted value of continuous deformation through a linear activation function;

[0069] Output layer 2 is used to classify discrete large deformation levels based on the deformation range.

[0070] In this embodiment, the lower-level branch quantitative prediction stage performs deformation prediction and grade determination. Based on the upper-level geological model identification results, a deep learning prediction network is constructed in conjunction with quantitative indicators to achieve "accurate deformation prediction - grade classification," forming a dual-task output. The network structure is determined as follows: the deep learning prediction network (lower-level branch) adopts a fully connected neural network architecture, consisting of an input layer, fully connected layer 1, fully connected layer 2, output layer 1 (deformation prediction), and output layer 2 (grade determination). The output layer is based on the actual deformation labels of the training dataset and simultaneously grades the predicted deformation.

[0071] If the surrounding rock is determined to be prone to large deformation, quantitative indicators and geological models are input into the lower-level branch deep learning prediction module. The lower-level branch outputs predicted deformation based on the geological model identification results and newly added deformation prediction indicators. The core is to introduce indicators directly related to deformation. Finally, the outputs of the upper-level branch and the flattened outputs of the lower-level branch are concatenated, and the deformation is predicted by combining the geological model and deformation-related characteristics. Finally, the deformation level is determined based on the deformation and standards.

[0072] The second stage, targeting a qualitative + quantitative system, constructs a surrounding rock deformation prediction network for prediction and early warning by combining surrounding rock geological model types and field survey data. The specific implementation of the deformation prediction stage includes: a 4-dimensional fusion vector is composed of "three quantitative index normalized vectors (corrected strength-stress ratio, support stiffness, and equivalent tunnel diameter) + one geological model code (1-6 corresponding to six models)," where the geological model code adopts a unique thermal coding method;

[0073] The concatenation order of the multidimensional comprehensive feature vector is "64-dimensional qualitative feature vector (flattened layer output) + 4-dimensional fusion vector". After concatenation, no secondary normalization is required, and it is directly input into the fully connected layer of the prediction network. The output of the prediction network includes two dimensions: one is the continuous deformation amount output by the linear activation function, and the other is the discrete large deformation level based on the deformation amount interval, realizing the dual task output of "quantitative + qualitative".

[0074] The cascading of the two-stage neural network is achieved through "feature vector concatenation": the flattened layer of the first stage (geological pattern identification network) outputs a 64-dimensional feature vector, which is directly used as one of the feature inputs of the second stage (deep learning prediction network) without the need for additional data transformation; the model training uses 1200 samples divided into training set, validation set and test set in a 7:2:1 ratio, and the samples are from typical tunnel projects.

[0075] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A method for predicting large deformation of high-stress soft rock tunnels based on deep learning, characterized in that: Includes the following steps: Parameter acquisition and preprocessing stage: Integrate geological survey data before construction, field test data and dynamic monitoring data during construction to construct a database of large deformation influence parameters including multiple qualitative and quantitative indicators; classify and encode the qualitative indicators, normalize the quantitative indicators to the [0,1] interval using min-max normalization, and remove outliers using the 3σ criterion and supplement missing values ​​using K-nearest neighbor interpolation to form standardized input data; Geological model identification stage: The preprocessed qualitative index encoding vector is input into the geological model identification network to obtain the probability distribution of multiple geological models. When the probability of a certain geological model is greater than or equal to the preset value, it is determined to be the corresponding geological model and marked as "surrounding rock prone to large deformation". Otherwise, it is marked as "surrounding rock not prone to large deformation". Deformation prediction stage: If the identification result is "surrounding rock prone to large deformation", the geological model code and the preprocessed quantitative index normalized vector are concatenated into a multi-dimensional fusion vector, which is then input into the deep learning prediction network. The continuous deformation prediction value is output through a linear activation function, and the discrete large deformation level is divided based on the deformation range. Results output stage: Output the predicted deformation value and its corresponding large deformation level, and provide geological model identification results as a basis for risk warning; The qualitative indicators include rock mass structure, weathering degree, groundwater conditions, relationship between the main controlling structural plane and the tunnel axis, degree of tectonic influence, and rock mass hardness; the quantitative indicators include corrected strength-stress ratio, support stiffness, and equivalent tunnel diameter. In the parameter acquisition and preprocessing stage, the qualitative index coding rules are as follows: the rock mass structure is divided into 1-5 levels according to its loose, fragmented, blocky, layered, and integral state; the weathering degree is divided into 1-5 levels according to its completely weathered, strongly weathered, moderately weathered, slightly weathered, and unweathered state; the groundwater conditions, the relationship between the main control structural plane and the tunnel axis, the degree of tectonic influence, and the rock mass hardness are all divided into 1-5 levels according to their severity. The geological pattern identification network is based on a convolutional neural network architecture. It extracts features through convolutional layers and retains key information through pooling layers, outputting the probability distribution of six types of geological patterns, including weak homogeneous type, block-fracture structure type, weak layered type, layered fault type, weak interlayered type, and layered fold type. The specific architecture of the geological model identification network includes: The input layer is used to receive a 6-dimensional qualitative index encoding vector; Two convolutional layers extract single-indicator correlation features and mine multi-indicator coupling features through convolutional kernels to capture geological model-related differences; The pooling layer is used to perform max pooling operations, reducing computation while preserving key features; Flattening layers are used to convert multidimensional feature maps into one-dimensional vectors; A fully connected layer is used to output the probability distribution of six geological models; The deep learning prediction network is based on a fully connected neural network architecture and includes: The input layer is used to receive a 4-dimensional fusion vector, which is composed of geological model encoding and normalized vectors of three quantitative indicators. Fully connected layer 1 and fully connected layer 2, with ReLU activation function and 0 dropout rate; Output layer 1 is used to output the predicted value of continuous deformation through a linear activation function; Output layer 2 is used to classify discrete large deformation levels based on the deformation range.

Images