CN-121561597-B - High-ground-stress soft rock tunnel large deformation prediction method based on deep learning

Abstract

The invention discloses a high-ground-stress soft rock tunnel large-deformation prediction method based on deep learning, and belongs to the field of tunnel engineering geological disaster prediction. By constructing a double-branch network architecture of geological mode judgment-deep learning prediction, the accurate prediction of large tunnel deformation is realized. Firstly, 6 qualitative indexes such as rock mass structure, weathering degree, groundwater condition and the like, 3 quantitative indexes such as strength stress ratio, supporting rigidity, equivalent hole diameter and the like are collected, after coding and normalization pretreatment, an upper layer and a lower layer of networks are respectively input, wherein the upper layer adopts a convolutional neural network to judge the geological mode of surrounding rock, the probability distribution of 6 easily-developed large deformation modes such as weak homogeneity type and block fracture structure type and the like is output, and the lower layer predicts the deformation quantity and divides the risk level through a fully-connected neural network based on the judging result and the quantitative indexes of the geological mode. The method can fully reflect the complex coupling mechanism of the large deformation of the high-ground-stress soft rock tunnel, and can realize adaptability prediction aiming at different geological modes.

Inventors

• CHEN KEPU
• YANG ZULIN
• SONG YU
• LIU ZONGJIAN
• GAO FENG
• LI HUAILIANG
• LI YOUGEN
• HE YUEBAO
• SHI YUQIANG
• SHUAI YONGJIAN
• SHANG MIAOMIAO
• Tian Yuanxun

Assignees

• 中国水利水电第七工程局有限公司
• 成都理工大学

Dates

Publication Date

20260508

Application Date

20260123

Claims (1)

1. 1. A high-ground-stress soft rock tunnel large deformation prediction method based on deep learning is characterized by comprising the following steps: The parameter acquisition and preprocessing stage comprises integrating geological survey data, field test data and construction period dynamic monitoring data before construction, constructing a large deformation influence parameter database comprising a plurality of qualitative indexes and quantitative indexes, carrying out hierarchical coding on the qualitative indexes, carrying out min-max normalized mapping on the quantitative indexes to a [0,1] interval, removing abnormal values through a3 sigma rule, and supplementing missing values through K neighbor interpolation, so as to form standardized input data; Inputting the preprocessed qualitative index coding vector into a geological mode judging network to obtain probability distribution of multiple types of geological modes, judging that the geological mode is corresponding to the geological mode when the probability of the geological mode of a certain type is more than or equal to a preset value, and identifying the geological mode as 'easy-to-develop large deformation surrounding rock', otherwise identifying the geological mode as 'difficult-to-develop large deformation surrounding rock'; the deformation prediction stage comprises the steps of splicing the geological mode coding and the preprocessed quantitative index normalization vector into a multidimensional fusion vector if the judgment result is 'easy to generate large deformation surrounding rock', inputting the multidimensional fusion vector into a deep learning prediction network, outputting a continuous deformation prediction value through a linear activation function, and dividing discrete large deformation grades based on a deformation interval; outputting a deformation predicted value and a large deformation grade corresponding to the deformation predicted value, and providing a geological mode judgment result as a risk early warning basis; The qualitative indexes comprise rock mass structure, weathering degree, groundwater condition, main control structure surface and hole axis relation, construction influence degree and rock mass hardness degree, wherein the quantitative indexes comprise correction strength stress ratio, support rigidity and equivalent hole diameter; In the parameter acquisition and pretreatment stage, the qualitative index coding rule is that the rock mass structure is divided into 1-5 levels according to the discrete shape, the fragmentation shape, the block fragmentation shape, the lamellar shape and the integral shape, the weathering degree is divided into 1-5 levels according to the total weathering, the strong weathering, the middle weathering and the micro weathering, and the non-weathering, and the underground water condition, the relation between the main control structure surface and the hole axis, the structural influence degree and the rock mass softness degree are all divided into 1-5 levels according to the severity degree; The geological mode identification network is based on a convolutional neural network architecture, extracts characteristics through a convolutional layer, and stores key information through pooling layer downsampling, and outputs probability distribution of 6 types of geological modes, wherein the probability distribution comprises weak homogeneity type, block fracture structure type, weak lamellar type, lamellar error type, weak mutual lamellar type and lamellar buckling type; the concrete architecture of the geological mode judgment network comprises the following steps: An input layer for receiving a 6-dimensional qualitative index encoding vector; the two convolution layers extract single-index associated features through convolution kernels, mine multi-index coupling features and capture geological mode related differences; the pooling layer is used for carrying out maximum pooling operation, reducing the calculated amount and retaining key characteristics; A flattening layer for converting the multi-dimensional feature map into a one-dimensional vector; The full-connection layer is used for outputting probability distribution of 6 types of geological modes; The deep learning prediction network is based on a fully-connected neural network architecture and comprises: the input layer is used for receiving the 4-dimensional fusion vector and is formed by splicing geological mode codes and 3 quantitative index normalization vectors; full connection layer 1 and full connection layer 2, activation function is ReLU, dropout rate is 0; An output layer 1 for outputting a continuous deformation prediction value through a linear activation function; And an output layer 2 for dividing discrete large deformation grades based on deformation intervals.

Description

High-ground-stress soft rock tunnel large deformation prediction method based on deep learning Technical Field The invention relates to the field of tunnel engineering geological disaster prediction, in particular to a high-ground-stress soft rock tunnel large-deformation prediction method based on deep learning. Background In high-ground-stress soft rock tunnel engineering, surrounding rock large deformation is a main hidden danger causing construction delay, support system failure and even safety accidents. Conventional tunnel deformation prediction methods often predict based on a single geologic parameter (e.g., rock strength, ground stress, etc.), empirical formulas, and conventional functions. The prediction of arch crown deformation of a highway tunnel based on regression analysis [ C ], 2022:566-571.DOI:10.26914 ] is based on actual measurement data of arch crown settlement of a section of a highway tunnel in different levels of surrounding rocks, and the fitting and prediction accuracy of 5 regression function models in the current specifications and related researches on the arch crown settlement deformation of the tunnel under the different levels of surrounding rocks are compared and analyzed. The polynomial function has the highest fitting precision to the measured data, but is not suitable for deformation prediction of tunnel vault. Along with the gradual development of tunnel construction in China to deep burial, long tunnels and large diameters, tunnels with high ground stress, soft rock, faults and other complex geological conditions are traversed. The influences of the buried depth, span, surrounding rock strength stress ratio, groundwater, rock structural plane and the like of the tunnel need to be comprehensively considered during prediction. For example Chen Weizhong, tian Yun, wang Xuehai et al (soft rock tunnel extrusion deformation prediction based on corrected [ BQ ] values [ J ]. Geotechnical mechanics, 2019, 40 (08): 3125-3134.DOI: 10.16285) propose that extrusion deformation amount can be directly and rapidly predicted based on tunnel burial depth, span and rock mass corrected [ BQ ] index parameters. The method is mainly suitable for deformation prediction when the tunnel is not excavated, and real-time adjustment is difficult to realize. In the tunnel construction process, various factors are required to be realized to effectively predict large deformation disasters in various stages of tunnel investigation, design, construction and the like. And further, the management and the optimization of the supporting scheme are realized, reworking is avoided, and the construction safety is ensured. In Huang Hongjian (the research of the large deformation numerical simulation prediction research of the construction of the high-ground stress weak surrounding rock section of the burger tunnel [ J ]. Railway standard design, 2009, (03): 93-95.DOI: 10.13238), the three-dimensional numerical simulation prediction deformation is performed on the large deformation section construction of the burger tunnel of the Yiwanrailway by adopting a large finite difference program. However, as the distance between the detection section and the tunnel surface increases, the prediction error also increases gradually. The conventional tunnel deformation prediction method relies on an empirical formula, numerical simulation and on-site monitoring, and often has the limitations of difficult adaptation to complex geological conditions, high numerical simulation calculation cost, poor real-time performance, low on-site monitoring data utilization efficiency, incapability of realizing advanced prediction, dynamic adjustment and the like. In recent years, along with the rapid development of artificial intelligence technology, the application of the intelligent prediction method in the field of tunnel engineering is gradually deepened, and a technical breakthrough direction is provided for large deformation prevention and control under complex geological conditions such as high ground stress, soft rock and the like. The deep learning technology effectively solves the problem of insufficient suitability of a traditional empirical formula (such as Hooke-Brownian criterion) and a linear regression model in a complex scene by virtue of the capability of the deep learning technology in automatic feature extraction and complex mapping modeling of high-dimensional nonlinear data, and becomes one of core technical paths of current tunnel deformation prediction. Xing Pengtao, and the like, optimizes a XGBOOST regression model by adopting a ZOA algorithm (CN202510682278. X 'a multi-risk source tunnel construction early warning method'), expands tunnel face images and construction site monitoring data into important data sources, establishes a corresponding database, builds a tunnel face joint extraction model by utilizing algorithms such as image recognition, neural network, machine learning and the like, builds a tunnel def