【作者】 李超；

【导师】 刘征宇；

【作者基本信息】 山东大学 ， 岩土工程， 2025， 硕士

【摘要】 随着我国隧道工程向西南复杂地质区域推进,隧道建设面临断层、溶洞、含水构造等不良地质体的严重威胁,极易引发突水突泥、塌方等灾害,威胁施工安全。传统单一物探方法存在多解性强、探测角度有限等问题,而现有综合解译方法依赖人工经验,效率低且主观性强。深度学习技术因其强大的特征提取与融合能力,为多源探测数据的智能解译提供了新思路。然而,隧道多源探测数据的智能解译仍面临地球物理属性与地质资料关系不清、多探测方法数据融合困难、地质信息利用不足等挑战。为此,本文提出基于深度学习的隧道多源探测结果综合解译方法,旨在实现不良地质体的精准识别与分类。本文通过理论分析、数值模拟和现场试验,系统研究了地质建模、特征融合与分类方法,形成了完整的智能解译技术体系。首先,基于断层、溶洞、水体等典型不良地质与地质信息的关系,提出了多物性参数三维建模方法,构建了包含地质信息、物性参数及探测数据的综合数据集,为深度学习训练提供了数据基础。其次,针对多源数据特征差异问题,设计了“先分别提取特征后统一融合”的策略,采用Gabor滤波器提取地震成像的纹理特征、可变卷积提取电阻率成像的扩散特征,结合多层注意力机制实现多源数据的高效融合,显著提升了不良地质位置与形态的识别精度。进一步,基于集成学习框架,融合物探结果与地质信息,通过随机森林树多角度分析数据关系,实现了不良地质类别与风险等级的准确预测,并引入迁移学习优化实际数据应用效果。具体工作内容如下:(1)基于地质信息与异常体关系的复杂地质建模方法。针对隧道地球物理属性与地质资料关系不清,模拟数据难建立的问题,本文分析了典型不良地质与地质信息、三维物性建模的相关关系,优选了地质信息并将其数字化,进而提出了隧道典型不良地质多物性参数三维建模方法,在此基础上,提出了地质-物性-数据的多参数批量建模方法,最终构建了一套适用于隧道智能判识研究的隧道多源综合探测数据集。(2)基于多源滤波器改进的不良地质位置与形态识别方法。针对多探测方法探距、特征不同、难以有效融合的问题,本文通过分析各类超前预报成像结果特征,提出了多探测方法成像特征融合策略,采用不同滤波器对空间对齐的各类预报结果进行特征提取,进而以多层注意力机制进行特征融合,最终建立了隧道不良地质识别网络,实现从多源探测结果中获取异常体的位置和形态。(3)基于集成学习与地质信息辅助的不良地质分类方法。针对地质信息多源异构、难以有效利用,不良地质难解译的问题,通过采用集成学习融入地质信息和探测结果,对不良地质类别进行预测分析,同时开展迁移学习实现模拟数据训练网络在实际数据上的应用,最终实现了对各类超前预报结果中不良地质的有效分类。基于上述研究成果,在某铁路工程隧道现场开展了探测试验,探测结果较准确的揭示了超前预报探测区域的断层破碎带等不良地质以及探测区域的分段的地质风险,验证了本文方法的可行性和有效性。更多还原

【Abstract】 With the advancement of tunnel engineering in China towards complex geological areas in the southwest,tunnel construction is facing serious threats from adverse geological bodies such as faults,caves,and water bearing structures,which can easily cause disasters such as water and mud bursts,landslides,and pose a threat to construction safety.Traditional single geophysical methods have problems such as strong ambiguity and limited detection angles,while existing comprehensive interpretation methods rely on manual experience,which is inefficient and subjective.Deep learning technology provides new ideas for intelligent interpretation of multi-source detection data due to its powerful feature extraction and fusion capabilities.However,the intelligent interpretation of multi-source detection data in tunnels still faces challenges such as unclear relationship between geophysical properties and geological data,difficulty in integrating data from multiple detection methods.and insufficient utilization of geological information.Therefore,this article proposes a comprehensive interpretation method for multi-source detection results of tunnels based on deep learning.aiming to achieve accurate identification and classification of adverse geological bodies.This article systematically studies geological modeling,feature fusion,and classification methods through theoretical analysis,numerical simulation,and field experiments.forming a complete intelligent interpretation technology system.Firstly,based on the relationship between typical adverse geological conditions such as faults,caves,and water bodies and geological information,a multi property parameter 3D modeling method was proposed,and a comprehensive dataset containing geological information,property parameters,and detection data was constructed,providing a data foundation for deep learning training.Secondly,in response to the problem of feature differences in multi-source data,a strategy of "extracting features separately before unified fusion" was designed.Gabor filters were used to extract texture features from seismic imaging,and variable convolutions were used to extract diffusion features from resistivity imaging.Combined with a multi-layer attention mechanism,efficient fusion of multi-source data was achieved,significantly improving the recognition accuracy of adverse geological locations and shapes.Furthermore,based on the ensemble learning framework,the integration of geophysical exploration results and geological information,and the multi angle analysis of data relationships through random forest trees,accurate prediction of adverse geological categories and risk levels was achieved,and transfer learning was introduced to optimize the actual data application effect.The specific job responsibilities are as follows:(1)A complex geological modeling method based on the relationship between geological information and anomalous bodies.In response to the problem of unclear relationship between tunnel geophysical properties and geological data,and difficulty in establishing simulated data,this paper analyzes the correlation between typical adverse geology and geological information,as well as 3D physical property modeling.Geological information is optimized and digitized,and a multi parameter 3D modeling method for typical adverse geology in tunnels is proposed.Based on this,a multi parameter batch modeling method for geological physical property data is proposed,and a tunnel multi-dimensional comprehensive detection dataset suitable for intelligent identification research is finally constructed.(2)A method for identifying the location and morphology of unfavorable geological conditions based on multi-source filter improvement.In response to the problems of different detection methods,different features,and difficulty in effective fusion,this paper proposes a multi detection method imaging feature fusion strategy by analyzing the characteristics of various advanced prediction imaging results.Different filters are used to extract features from various spatial prediction results,and then multi-layer attention mechanism is used for feature fusion.Finally,a tunnel adverse geological identification network is established to obtain the position and shape of abnormal bodies from multiple detection results.(3)A method for classifying unfavorable geology based on ensemble learning and geological information assistance.In response to the problem of heterogeneous and difficult to effectively utilize geological information,as well as the difficulty in translating adverse geological conditions,ensemble learning was adopted to integrate geological information and detection results for predicting and analyzing adverse geological categories.At the same time,transfer learning was carried out to simulate the application of data training networks on actual data,ultimately achieving effective classification of adverse geological conditions in various overdue prediction results.Based on the above research results,detection experiments were carried out at the tunnel site of the Western Plateau Railway Project.The detection results accurately revealed the adverse geological conditions such as fault fracture zones in the advanced prediction detection area and the geological risks of segmented detection areas,verifying the feasibility and effectiveness of the method proposed in this paper.更多还原