A tunnel unfavorable geological condition advanced detection and early warning analysis method

Abstract

The application discloses a tunnel unfavorable geology advanced detection and early warning analysis method, relates to the technical field of geophysical exploration, and is characterized in that a multi-source geophysical detection array is arranged behind a working face, microsecond-level synchronization is realized through Beidou timing, seismic reflection waves, geological radar electromagnetic waves and advanced drilling geological recording data are collected, multi-dimensional features are extracted after differential preprocessing, a multi-source feature fusion model is constructed to generate a fusion data set, a pre-training model is combined to identify unfavorable geological bodies and calculate a construction risk grade, corresponding early warnings are triggered, and the identification result and early warning content are dynamically updated during construction. The application improves the unfavorable geology identification accuracy through multi-source geophysical data fusion, quantitatively evaluates the risk judgment objectivity, guarantees the early warning timeliness through a dynamic updating mechanism, and provides reliable technical support for tunnel safe construction under complex geological conditions.

Description

Technical Field

[0001] This invention relates to the field of geophysical exploration technology, and in particular to a method for advanced detection and early warning analysis of adverse geological conditions in tunnels. Background Technology

[0002] Tunnels are key infrastructure projects in transportation, water conservancy, and mining, traversing areas with complex geological conditions, including widespread distribution of unfavorable geological bodies such as faults, fracture zones, karst caves, water-rich zones, and weak interlayers. These unfavorable geological bodies are a major cause of engineering disasters such as tunnel collapses, water and mud inrushes, and rock bursts, directly threatening the lives of construction workers and the progress of the project. Traditional advanced geological exploration of tunnels mainly relies on manual geological logging and single geophysical methods. Manual logging can only reflect the geological information exposed at the tunnel face and cannot detect unfavorable geological bodies hidden ahead of the tunnel face. Single geophysical methods have significant technical limitations. Seismic methods are effective in identifying geological bodies with significant differences in elastic wave impedance, but have low sensitivity to low-resistivity, water-rich zones. Ground-penetrating radar has high resolution for geological bodies with large differences in dielectric constant, but its detection depth is limited and it is severely affected by metal interference. Advanced drilling methods have high accuracy, but their detection range is small, construction costs are high, and operational efficiency is low, making it difficult to achieve large-scale continuous exploration. The problem of multiple solutions with a single method is prominent, and it is easy to miss or misjudge, which cannot meet the safety requirements of tunnel construction under complex geological conditions.

[0003] While existing multi-source data fusion detection technologies have to some extent compensated for the shortcomings of single methods, they still have many deficiencies. Most schemes fail to achieve high-precision time synchronization of multi-source data, resulting in significant discrepancies in the time dimension between different data sources and leading to spatial misalignment in the fusion results. Data preprocessing lacks specificity; using a uniform preprocessing method for different types of geophysical data fails to effectively remove noise interference specific to each data source, reducing the signal-to-noise ratio and usability of the data. Feature fusion often employs a fixed-weighted average method, failing to consider the adaptability differences of different data sources under varying geological conditions and construction environments. Low-quality data can significantly negatively impact the fusion results, further amplifying errors in identifying adverse geological conditions. Risk assessment relies heavily on qualitative judgments based on engineering experience, lacking unified quantitative assessment indicators and methods. The judgments of different assessors vary considerably, failing to provide an objective and accurate basis for construction decisions.

[0004] Existing tunnel geological early warning systems lack dynamic updating capabilities. Most systems can only achieve one-time detection and static early warning, failing to dynamically correct warning results by incorporating newly added detection data during construction and geological information revealed during excavation at the tunnel face. When actual geological conditions deviate from previous detection results, the early warning system cannot adjust in time, easily leading to delayed or false warnings. Furthermore, existing systems provide insufficient early warning information, offering only simple risk level indications, lacking detailed spatial location, distribution range, and geometric scale information of adverse geological bodies, and failing to provide targeted construction and treatment suggestions, making it difficult to directly guide on-site construction. As tunnel engineering develops towards deeper burial, longer distances, and more complex geological conditions, higher demands are placed on the accuracy, reliability, and real-time performance of advanced detection, urgently requiring an advanced method capable of accurate fusion of multi-source data, intelligent identification of adverse geological conditions, and dynamic risk early warning. Summary of the Invention

[0005] This invention proposes a method for advanced detection and early warning analysis of adverse geological conditions in tunnels, in order to solve the problems mentioned in the prior art.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a method for advanced detection and early warning analysis of adverse geological conditions in tunnels, comprising the following steps: A multi-source geophysical detection array is set up at a predetermined distance behind the tunnel face. The array includes a three-component seismic detector array, a multi-frequency ground-penetrating radar antenna array, and a borehole parameter acquisition unit. The microsecond-level time synchronization of all detection equipment is achieved through the BeiDou time synchronization system, and the seismic reflection wave data, ground-penetrating radar electromagnetic wave data, and geological logging data of the advanced borehole within the predetermined detection range in front of the tunnel face are collected simultaneously. Differentiated preprocessing procedures were performed on the collected multi-source raw data to complete random noise removal, surface wave separation, automatic first arrival picking and velocity analysis of seismic wave data, background clutter removal, gain compensation, migration imaging and data interpolation of ground-penetrating radar data, and standardization and digitization of borehole geological logging data. Extract multi-dimensional feature parameters from the preprocessed multi-source data, including the time-domain amplitude, frequency-domain dominant frequency, phase shift, P-wave velocity, and S-wave velocity characteristics of seismic waves, and the reflection intensity, dielectric constant distribution, electromagnetic wave attenuation coefficient, and reflection interface depth characteristics of ground-penetrating radar. A multi-source feature fusion model is constructed to perform feature layer fusion processing on feature parameters of different types and dimensions to generate a fused feature dataset containing geological structure and physical property information. Based on the fused feature dataset and the pre-trained intelligent identification model for adverse geological conditions, common adverse geological bodies in front of the tunnel face are identified, and the spatial location, distribution range and geometric scale of the adverse geological bodies are determined. By combining the design parameters of the tunnel project, the regional geological background conditions, and the parameters of the identified adverse geological bodies, the risk level of the adverse geological bodies to the tunnel construction is calculated; based on the calculated risk level, the corresponding level of early warning is triggered, and a standardized early warning report containing detailed information on the adverse geological bodies, risk assessment results, and targeted treatment suggestions is generated. During tunnel excavation, newly added detection data and geological information revealed by the excavation face are accessed in real time, and the results of adverse geological identification and risk assessment models are dynamically updated.

[0007] Furthermore, it also includes a step for quantitative assessment of adverse geological risks, which combines the scale parameters, physical property parameters, and engineering impact parameters of the adverse geological body to complete the quantitative calculation of the risk level. The calculation formula is as follows: ;in This is a dimensionless adverse geological risk value. For the dimensionless weighting coefficients of the scale parameter, These are dimensionless weighting coefficients for physical property parameters. These are the dimensionless weighting coefficients for engineering influence parameters. The normalized size value of the adverse geological body. The normalized physical property values ​​of the adverse geological body. This represents the normalized impact value of adverse geological bodies on the project.

[0008] Furthermore, it also includes a dynamic weight allocation step for multi-source data fusion, which combines quality parameters, environmental impact parameters, and geological adaptability parameters from different data sources to dynamically calculate the fusion weights. The calculation formula is as follows: ;in For the first Dimensionless fusion weights of multiple data sources For the first Normalized data quality values ​​for each data source. For the first Normalized environmental adaptability values ​​of each data source. For the first Normalized geological type adaptability values ​​from each data source, dimensionless. The total number of data sources participating in the integration.

[0009] Furthermore, the multi-source geophysical exploration array adopts a layered arrangement. The three-component seismic detector array is symmetrically arranged along the two side walls of the tunnel, with the detector spacing being a preset fixed value. The arrangement length covers a preset range behind the tunnel face, completing seismic wave detection at a depth of 100 to 150 meters in front of the tunnel face. The multi-frequency ground-penetrating radar antenna array is installed on a movable support, including antennas of three different frequencies: low frequency, medium frequency, and high frequency. The borehole parameter acquisition unit is integrated on the advanced drilling rig, which collects drilling speed, torque, drilling pressure, and rotational speed parameters in real time during the drilling process. All detection equipment obtains a unified time reference through the Beidou time synchronization module.

[0010] Furthermore, the multi-source data differential preprocessing adopts a multi-stage processing architecture. Seismic wave data preprocessing first removes 50Hz power frequency interference and high-frequency random noise through bandpass filtering, then separates surface waves and effective reflected waves through FK filtering, and automatically picks up the first arrival wave using a combination of energy ratio method and polarization analysis. The velocity distribution model of P-waves and S-waves in the formation is obtained through velocity scanning method. Ground penetrating radar data preprocessing first eliminates direct waves and ground clutter through average background removal method, then compensates for the attenuation of electromagnetic waves with propagation distance through automatic gain compensation, corrects the spatial position of the reflection interface through Kirchhoff migration imaging algorithm, and completes the missing data areas through linear interpolation. Borehole geological logging data preprocessing converts qualitative descriptions into standardized quantitative parameters and generates digital borehole geological profiles. After preprocessing, the quality of all data is assessed.

[0011] Furthermore, the multi-dimensional feature extraction adopts a hierarchical feature extraction architecture. For seismic wave data, it extracts features such as root mean square amplitude, peak amplitude, energy attenuation rate, and duration in the time domain; dominant frequency, bandwidth, energy distribution entropy, and harmonic ratio in the frequency domain; and physical property features. For ground-penetrating radar data, it extracts features such as mean single-channel reflection intensity, reflection layer thickness, dielectric constant profile, attenuation coefficient profile, and number of reflection interfaces. For borehole data, it extracts features such as lithology index, rock mass integrity index, water content index, and joint density index. The mutual information method is used to select features from all extracted features, eliminating redundant and irrelevant features and retaining features that contribute highly to the identification of adverse geological conditions.

[0012] Furthermore, the multi-source feature fusion model adopts a cross-modal fusion network architecture based on an attention mechanism. The network contains three independent feature extraction branches and a global fusion attention module. The three feature extraction branches process seismic wave features, ground-penetrating radar features, and borehole features, respectively. Each branch consists of multiple convolutional layers and pooling layers to extract deep abstract features from each data source. The global fusion attention module automatically learns the importance weights of different features under different geological conditions, performs weighted fusion of the deep features output by each branch, updates parameters through an adaptive moment estimation algorithm, and introduces a regularization term during training to reduce the model's overfitting.

[0013] Furthermore, the intelligent identification model for adverse geological conditions adopts a hybrid architecture combining convolutional neural networks and long short-term memory networks. The convolutional neural network extracts the spatial structural features of the fusion features to capture the geometric shape and distribution patterns of adverse geological bodies, while the long short-term memory network extracts the temporal variation features of the fusion features to capture the continuous variation trend of stratigraphic parameters. The model output layer adopts a multi-task learning structure, simultaneously outputting the type, spatial location, distribution range, and geometric scale parameters of adverse geological bodies. The model training dataset contains thousands of sets of tunnel advance detection measured data and corresponding excavation verification data under different geological conditions. During training, data augmentation techniques such as random pruning and adding noise are used to expand the dataset size, and the five-fold cross-validation method is used to evaluate the model performance.

[0014] Furthermore, adverse geological risks are classified into four levels: Level 1 corresponds to extremely high risk, Level 2 to high risk, Level 3 to medium risk, and Level 4 to low risk. The corresponding risk level is determined based on the calculated comprehensive risk value. Different warnings for different levels are sent out using different methods and response procedures. Level 1 warnings are sent directly to the project manager and construction command center through multiple means such as audible and visual alarms, SMS, and telephone, activating the emergency response mechanism, suspending construction, and formulating an emergency response plan. Level 2 warnings are sent to the technical manager and construction teams, formulating specific construction response plans and strengthening on-site monitoring. Level 3 and Level 4 warnings are sent to on-site technical personnel, increasing the frequency of construction monitoring and preparing for construction.

[0015] Furthermore, the dynamic update mechanism adopts a real-time data-driven incremental update method. During the tunnel excavation process, after each preset advance is completed, a multi-source data acquisition and analysis is automatically triggered. The newly added detection data is input into the identification model and risk assessment model to update the identification results and risk levels of adverse geology in front of the tunnel face. At the same time, the actual geological information exposed by the excavation of the tunnel face is entered into the system as label data to fine-tune the identification model online and correct the model parameters. The system automatically stores all historical detection data and excavation verification data to build a local geological database.

[0016] Compared with existing technologies, the beneficial effects of this invention are: This invention employs the BeiDou time synchronization system to achieve microsecond-level time synchronization of multi-source geophysical exploration equipment, providing a unified time reference for multi-source data fusion, eliminating time deviations between different data sources, and improving the accuracy of initial alignment of multi-source data. The multi-source geophysical exploration array integrates seismic, ground-penetrating radar, and borehole drilling detection methods, fully leveraging the technical advantages of each method to achieve comprehensive detection at different depths and resolutions, covering a large geological space ahead of the tunnel face, and overcoming the limitations of single detection methods.

[0017] This invention employs a differentiated preprocessing workflow, designing specific processing steps for different types of geophysical data to effectively remove noise interference unique to each data source and improve the quality of the raw data. A multi-dimensional feature extraction architecture comprehensively extracts the temporal, frequency, and physical property features of seismic waves, ground-penetrating radar, and borehole data. Redundant and irrelevant features are eliminated through feature selection, retaining features that contribute significantly to the identification of adverse geological conditions, thus providing a high-quality feature foundation for subsequent fusion and identification.

[0018] This invention employs an attention-based cross-modal fusion network, combined with a dynamic weight allocation method, to automatically learn the importance weights of different data sources and features under different geological conditions. This enhances the contribution of high-quality data in the fusion process and strengthens the adaptability and robustness of multi-source data fusion. The intelligent identification model for adverse geological formations uses a hybrid architecture combining convolutional neural networks and long short-term memory networks. It simultaneously captures the spatial structural features of adverse geological bodies and the temporal variation characteristics of stratigraphic parameters, improving the accuracy and generalization ability for identifying different types of adverse geological bodies.

[0019] This invention calculates the comprehensive risk value of adverse geological conditions through multi-dimensional parameter fusion, providing a quantitative basis for risk level classification and improving the objectivity and consistency of risk assessment. The tiered early warning mechanism employs differentiated push methods and response processes based on different risk levels. Early warning reports include detailed spatial information of adverse geological bodies and targeted treatment suggestions, directly guiding on-site construction decisions. A real-time data-driven incremental update mechanism, combined with geological information revealed during excavation, fine-tunes the model online, dynamically updating the adverse geological condition identification results and risk levels, ensuring the accuracy and timeliness of early warning results and providing reliable technical support for safe tunnel construction. Attached Figure Description

[0020] Figure 1 This is a schematic flowchart of the overall process for advanced detection and early warning analysis of adverse geological conditions in tunnels proposed in this invention; Figure 2 This is a schematic block diagram of the multi-source geophysical detection array layout and synchronization architecture proposed in this invention; Figure 3 A flowchart illustrating the hierarchical processing of data preprocessing and feature extraction; Figure 4 A flowchart for intelligent identification of adverse geological conditions based on hybrid neural networks; Figure 5 This is a feedback loop diagram for risk quantification assessment and dynamic updates. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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.

[0022] Reference Figures 1 to 5 A method for advanced detection and early warning analysis of adverse geological conditions in tunnels, comprising the following steps: A multi-source geophysical detection array is set up at a predetermined distance behind the tunnel face. The array includes a three-component seismic detector array, a multi-frequency ground-penetrating radar antenna array, and a borehole parameter acquisition unit. The microsecond-level time synchronization of all detection equipment is achieved through the BeiDou time synchronization system, and the seismic reflection wave data, ground-penetrating radar electromagnetic wave data, and geological logging data of the advanced borehole within the predetermined detection range in front of the tunnel face are collected simultaneously. Differentiated preprocessing procedures were performed on the collected multi-source raw data to complete random noise removal, surface wave separation, automatic first arrival picking and velocity analysis of seismic wave data, background clutter removal, gain compensation, migration imaging and data interpolation of ground-penetrating radar data, and standardization and digitization of borehole geological logging data. Multi-dimensional feature parameters of preprocessed multi-source data are extracted, including the time-domain amplitude, frequency-domain dominant frequency, phase shift, P-wave velocity, and S-wave velocity characteristics of seismic waves; the reflection intensity, dielectric constant distribution, electromagnetic wave attenuation coefficient, and reflection interface depth characteristics of ground-penetrating radar; and the lithology, integrity, and water content characteristics of borehole data. A multi-source feature fusion model is constructed to perform feature layer fusion processing on feature parameters of different types and dimensions to generate a fused feature dataset containing geological structure and physical property information. Based on the fused feature dataset and the pre-trained intelligent identification model for adverse geology, common adverse geological bodies such as faults, fracture zones, karst caves, water-rich zones, and weak interlayers in front of the tunnel face are identified, and the spatial location, distribution range and geometric scale of adverse geological bodies are determined. By combining the design parameters of the tunnel project, the regional geological background conditions, and the parameters of the identified adverse geological bodies, the risk level of the adverse geological bodies to the tunnel construction is calculated; based on the calculated risk level, the corresponding level of early warning is triggered, and a standardized early warning report containing detailed information on the adverse geological bodies, risk assessment results, and targeted treatment suggestions is generated. During tunnel excavation, newly added detection data and geological information revealed by the excavation face are accessed in real time, and the results of adverse geological identification and risk assessment models are dynamically updated to correct early warning content.

[0023] This invention also includes a step for quantitative assessment of adverse geological risks, which combines the scale parameters, physical property parameters, and engineering impact parameters of the adverse geological body to complete the quantitative calculation of the risk level. The calculation formula is as follows: ;in This is a dimensionless adverse geological risk value. For the dimensionless weighting coefficients of the scale parameter, These are dimensionless weighting coefficients for physical property parameters. These are the dimensionless weighting coefficients for engineering influence parameters. The normalized size value of the adverse geological body. These are the normalized physical properties of the adverse geological body, dimensionless. The normalized impact value of adverse geological bodies on the project is dimensionless. The comprehensive risk value is obtained by fusion calculation of multi-dimensional parameters, which provides a quantitative basis for risk level classification and improves the objectivity and consistency of risk assessment. At the same time, the weight coefficients of each parameter can be dynamically adjusted according to the characteristics of different tunnel projects to improve the pertinence and accuracy of risk assessment.

[0024] This invention also includes a step for dynamically allocating multi-source data fusion weights. This involves dynamically calculating the fusion weights by combining quality parameters, environmental impact parameters, and geological adaptability parameters from different data sources. The calculation formula is as follows: ;in For the first Dimensionless fusion weights of multiple data sources For the first Normalized data quality values ​​from each data source, dimensionless. For the first Normalized environmental adaptability values ​​of each data source, dimensionless. For the first Normalized geological type adaptability values ​​from each data source, dimensionless. The total number of data sources participating in the fusion is dimensionless. By dynamically allocating the fusion weights of different data sources, the contribution of high-quality data in the fusion process is enhanced, thereby strengthening the adaptability and robustness of multi-source data fusion.

[0025] In this invention, the multi-source geophysical detection array adopts a layered arrangement. The three-component seismic detector array is symmetrically arranged along the two sidewalls of the tunnel, with the detector spacing being a preset fixed value. The arrangement length covers a preset range behind the tunnel face, enabling seismic wave detection at depths of 100 to 150 meters in front of the tunnel face. The multi-frequency ground-penetrating radar antenna array is mounted on a movable support and includes antennas of three different frequencies: low frequency, medium frequency, and high frequency. The low-frequency antenna is used for deep, large-area detection, the medium-frequency antenna is used for medium-depth, medium-resolution detection, and the high-frequency antenna is used for shallow, high-resolution detection. The borehole parameter acquisition unit is integrated on the advanced drilling rig, which collects drilling speed, torque, drilling pressure, and rotational speed parameters in real time during the drilling process. All detection equipment obtains a unified time reference through the BeiDou time synchronization module, adding precise microsecond-level timestamps to each set of acquired data, achieving accurate alignment of multi-source data in the time dimension.

[0026] In this invention, the multi-source data differential preprocessing adopts a multi-stage processing architecture. Seismic wave data preprocessing first removes 50Hz power frequency interference and high-frequency random noise through bandpass filtering, then separates surface waves and effective reflected waves through FK filtering, and automatically picks up the first arrival wave using a combination of energy ratio method and polarization analysis with millisecond-level picking accuracy. The velocity distribution model of P-waves and S-waves in the formation is obtained through velocity scanning method. Ground penetrating radar data preprocessing first eliminates direct waves and surface clutter through average background removal method, then compensates for the attenuation of electromagnetic waves with propagation distance through automatic gain compensation, corrects the spatial position of the reflection interface using Kirchhoff migration imaging algorithm, and completes missing data areas through linear interpolation. Borehole geological logging data preprocessing converts qualitative descriptions such as lithological description, structural surface information, water content, and joint development degree into standardized quantitative parameters to generate digital borehole geological profiles. After preprocessing, all data are quality evaluated, and data segments that do not meet the quality requirements are removed.

[0027] In this invention, multi-dimensional feature extraction adopts a hierarchical feature extraction architecture. For seismic wave data, it extracts features such as root mean square amplitude, peak amplitude, energy attenuation rate, and duration in the time domain; dominant frequency, bandwidth, energy distribution entropy, and harmonic ratio in the frequency domain; and physical properties such as P-wave velocity, S-wave velocity, wave velocity ratio, and Poisson's ratio. For ground-penetrating radar data, it extracts features such as the mean single-channel reflection intensity, reflection layer thickness, dielectric constant profile, attenuation coefficient profile, and number of reflection interfaces. For borehole data, it extracts features such as lithology index, rock mass integrity index, water content index, and joint density index. The mutual information method is used to select features from all extracted features, eliminating redundant and irrelevant features and retaining features that contribute significantly to identifying adverse geological conditions. All retained features are normalized to eliminate the influence of different dimensions and orders of magnitude, generating standardized feature vectors.

[0028] In this invention, the multi-source feature fusion model adopts a cross-modal fusion network architecture based on an attention mechanism. The network includes three independent feature extraction branches and a global fusion attention module. The three feature extraction branches process seismic wave features, ground-penetrating radar features, and borehole features, respectively. Each branch consists of multiple convolutional layers and pooling layers to extract deep abstract features from each data source. The global fusion attention module automatically learns the importance weights of different features under different geological conditions and performs weighted fusion of the deep features output by each branch to generate a fusion feature vector containing complementary information from multiple sources. The model adopts an end-to-end training method with the accuracy of identifying adverse geological conditions as the optimization objective. The parameters are updated through an adaptive moment estimation algorithm. During the training process, a regularization term is introduced to reduce the overfitting degree of the model and improve the discriminative and representational capabilities of the fusion features.

[0029] In this invention, the intelligent identification model for adverse geological features adopts a hybrid architecture combining convolutional neural networks and long short-term memory networks. The convolutional neural network extracts the spatial structural features of the fusion features to capture the geometric shape and distribution patterns of adverse geological bodies, while the long short-term memory network extracts the temporal variation features of the fusion features to capture the continuous variation trend of stratigraphic parameters. The model output layer adopts a multi-task learning structure, simultaneously outputting the type, spatial location, distribution range, and geometric scale parameters of adverse geological bodies. The model training dataset contains thousands of sets of tunnel advance detection measured data and corresponding excavation verification data under different geological conditions. During training, data augmentation techniques such as random pruning and adding noise are used to expand the dataset size. The five-fold cross-validation method is used to evaluate the model performance, thereby improving the model's recognition accuracy and generalization ability for different types of adverse geological bodies.

[0030] In this invention, adverse geological risks are classified into four levels: Level 1 corresponds to extremely high risk, Level 2 to high risk, Level 3 to medium risk, and Level 4 to low risk. The corresponding risk level is determined based on the calculated comprehensive risk value. Different warnings for different levels are sent through different methods and response procedures. Level 1 warnings are sent directly to the project manager and construction command center via audible and visual alarms, SMS, and telephone, activating the emergency response mechanism, suspending construction, and developing an emergency response plan. Level 2 warnings are sent to the technical manager and construction team, developing a specific construction response plan and strengthening on-site monitoring. Level 3 and Level 4 warnings are sent to on-site technical personnel, increasing the frequency of construction monitoring and preparing for construction. The warning report includes the three-dimensional spatial location, distribution range, risk level, possible engineering disaster types, and targeted construction response suggestions for the adverse geological body.

[0031] In this invention, the dynamic update mechanism adopts a real-time data-driven incremental update method. During tunnel excavation, after each preset advance is completed, a multi-source data acquisition and analysis is automatically triggered. The newly added detection data is input into the identification model and risk assessment model to update the identification results and risk levels of adverse geology ahead of the tunnel face. At the same time, the actual geological information exposed by the excavation at the tunnel face is entered into the system as tag data to fine-tune the identification model online, correct the model parameters, and improve the accuracy of subsequent detection. The system automatically stores all historical detection data and excavation verification data to build a local geological database. When there is a large deviation between the identification results and the excavation information, a full-range re-detection and analysis is automatically triggered. The model is corrected by combining historical data to maintain the accuracy and timeliness of the early warning results.

[0032] The following two examples further illustrate the specific implementation of this system: This invention relates to a method for advanced detection and early warning analysis of adverse geological conditions in tunnels. It integrates multi-source geophysical detection technology, cross-modal feature fusion technology, and artificial intelligence recognition technology to achieve accurate identification, risk quantification assessment, and dynamic early warning of adverse geological bodies ahead of the tunnel face. The system comprises a data acquisition layer, a data processing layer, an intelligent analysis layer, and an early warning output layer. The data acquisition layer synchronously acquires multi-source geophysical data; the data processing layer performs differentiated preprocessing and multi-dimensional feature extraction; the intelligent analysis layer achieves multi-source feature fusion and intelligent identification of adverse geological conditions; and the early warning output layer completes risk quantification assessment and graded early warning. The method is adaptable to various complex geological conditions, such as deep-buried tunnels in mountainous areas and tunnels in karst regions, providing technical support for safe tunnel construction. The invention will be further described in detail below with reference to two specific embodiments.

[0033] Example 1 This embodiment is applied to the advanced detection scenario of a deep-buried tunnel on a mountain highway. The tunnel is 5.2 kilometers long with a maximum burial depth of 680 meters. The geological conditions of the area it traverses are complex, with multiple faults and fracture zones, and local water-rich zones. During construction, engineering disasters such as collapse, water inrush, and mudslide are likely to occur. The requirement is to achieve adverse geological detection within 150 meters in front of the tunnel face, control the positioning error within a reasonable range, and issue early warning information in advance to guide construction.

[0034] During the system deployment phase, a multi-source geophysical detection array is deployed within a range of 20 to 50 meters behind the tunnel face. A three-component seismic detector array is symmetrically arranged along both sides of the tunnel, with a detector spacing of 2 meters and 16 detectors on each side, for a total length of 30 meters. This allows for seismic wave detection at depths of 100 to 150 meters in front of the tunnel face. A multi-frequency ground-penetrating radar antenna array is mounted on a movable support, including low-frequency, medium-frequency, and high-frequency antennas. The low-frequency antenna has a detection depth of up to 50 meters, the medium-frequency antenna 30 meters, and the high-frequency antenna 15 meters, enabling layered detection at different depths. A borehole parameter acquisition unit is integrated into an advanced drilling rig, with three boreholes drilled to a depth of 60 meters, arranged in a triangular pattern at the tunnel face. All detection equipment achieves microsecond-level time synchronization via a BeiDou time synchronization module, adding precise timestamps to each set of acquired data to ensure accurate alignment of multi-source data in the time dimension.

[0035] During the data acquisition phase, a hammer-driven seismic source was used to generate seismic waves, and a seismic detector array simultaneously received the reflected wave signals, with a sampling frequency set to 1000Hz. The ground-penetrating radar antenna moved horizontally along the tunnel face, scanning in 0.1-meter increments to collect electromagnetic wave data covering the entire face. During the advanced drilling operation, the borehole parameter acquisition unit collected drilling speed, torque, drilling pressure, and rotational speed parameters in real time, while simultaneously completing borehole geological logging, recording information such as lithology, structural planes, water content, and joint development.

[0036] In the data preprocessing stage, differentiated processing procedures are implemented for different types of data. Seismic wave data first undergoes bandpass filtering to remove 50Hz power frequency interference and high-frequency random noise, then FK filtering separates surface waves from effective reflected waves. First arrival waves are automatically acquired using a combination of energy ratio method and polarization analysis, and velocity distribution models for P-waves and S-waves are obtained through velocity scanning. Ground penetrating radar data first eliminates direct waves and surface clutter using average background removal, then compensates for electromagnetic wave attenuation with propagation distance using automatic gain compensation, corrects the spatial location of the reflection interface using Kirchhoff migration imaging algorithm, and completes missing data areas using linear interpolation. Borehole geological logging data converts qualitative descriptions into standardized quantitative parameters, generating digital borehole geological profiles. After preprocessing, all data undergoes quality assessment, and data segments that fail to meet quality standards are removed.

[0037] In the feature extraction stage, a hierarchical feature extraction architecture is adopted to extract multi-dimensional feature parameters. For seismic wave data, time-domain features include root mean square amplitude, peak amplitude, energy attenuation rate, and duration; frequency-domain features include dominant frequency, bandwidth, energy distribution entropy, and harmonic ratio; and physical properties such as P-wave velocity, S-wave velocity, wave velocity ratio, and Poisson's ratio are extracted. For ground-penetrating radar data, features such as mean single-channel reflection intensity, reflector layer thickness, dielectric constant profile, attenuation coefficient profile, and number of reflector interfaces are extracted. For borehole data, features such as lithology index, rock mass integrity index, water content index, and joint density index are extracted. Mutual information is used for feature selection, eliminating redundant and irrelevant features, retaining features that contribute significantly to the identification of adverse geological conditions, and normalizing all retained features to generate standardized feature vectors.

[0038] In the feature fusion and recognition stage, standardized feature vectors are input into a cross-modal fusion network based on an attention mechanism. The network automatically learns the importance weights of different features to generate a fused feature vector. The fused feature vector is then input into an intelligent geological adverse identification model. The model outputs that there is a fault fracture zone with a width of 8 meters ahead of the tunnel face, 42 meters away from the tunnel face. The rock mass within the fracture zone is fractured and has a high water content.

[0039] During the risk assessment and early warning phase, the comprehensive risk value is calculated by combining the parameters of the adverse geological body and the tunnel engineering design parameters, and the risk level is determined to be Level II. The system generates a standardized early warning report, which includes the three-dimensional spatial location, distribution range, risk level, types of engineering disasters that may be caused by the fault fracture zone, and targeted construction and disposal suggestions. The early warning information is pushed to the technical supervisor and construction team. During construction, multi-source data acquisition and analysis are automatically triggered every 10 meters of advance, dynamically updating the adverse geological identification results and risk level. When the excavation face reaches a position 5 meters away from the fault fracture zone, the system raises the risk level to Level I, activates the emergency response mechanism, suspends construction, and formulates an advanced support plan.

[0040] Table 1 Comparison of Detection Results of Deep-Buried Tunnels in Mountainous Areas Table 1 presents the application effects of this invention in deep-buried tunnel scenarios in mountainous areas through five core indicators. The improved accuracy of adverse geological feature identification and spatial positioning stems from the complementary fusion of multi-source detection technologies and the application of a cross-modal feature fusion network, fully leveraging the technical advantages of different detection methods and compensating for the ambiguity of single methods. The increased identification rate of water-rich zones is attributed to the joint analysis of seismic wave velocity characteristics and ground-penetrating radar dielectric constant characteristics, effectively distinguishing between water-rich fractured zones and dry fractured zones. The improved accuracy of risk assessment relies on a quantitative assessment method based on multi-dimensional parameter fusion, eliminating the subjective differences of human experience-based judgment. The extended early warning time comes from the combination of large-scale deep detection and a dynamic update mechanism, enabling earlier detection of hidden adverse geological bodies and allowing sufficient time for construction and treatment.

[0041] Example 2 This embodiment is applied to the advanced detection scenario of a water conveyance tunnel in a karst area. The tunnel is 8.7 kilometers long and traverses a typical karst landform with strong karst development and widespread distribution of karst caves and weak interlayers of varying sizes. There are also some filled karst caves. During construction, disasters such as mudslides and collapses are likely to occur. The requirement is to accurately identify karst caves and weak interlayers within 100 meters in front of the tunnel face to provide a basis for adjusting the construction plan.

[0042] During the system deployment phase, a multi-source geophysical detection array was deployed within a range of 15 to 40 meters behind the tunnel face. A three-component seismic detector array was symmetrically arranged along both sides of the tunnel, with a detector spacing of 1.5 meters and 18 detectors on each side, for a total length of 25.5 meters. The multi-frequency ground-penetrating radar antenna array used a combination of intermediate-frequency and high-frequency antennas; the intermediate-frequency antenna was used to detect deep karst caves, while the high-frequency antenna was used to detect shallow small caves and weak interlayers. The borehole parameter acquisition unit was integrated onto an advanced drilling rig, with an advanced drilling depth of 80 meters and four boreholes, arranged in the four quadrants of the tunnel face. All detection equipment achieved microsecond-level time synchronization via a BeiDou time synchronization module, ensuring the time consistency of multi-source data.

[0043] During the data acquisition phase, an electric spark source was used to excite seismic waves, and a seismic detector array synchronously received the reflected wave signals, with the sampling frequency set to 2000Hz. The ground-penetrating radar antenna scanned along the tunnel face in a grid pattern, with a horizontal and vertical scanning step of 0.1 meters to achieve full coverage of the tunnel face. During the advanced drilling operation, drilling parameters and geological logging data were collected in real time, with a focus on recording the location, size, infill type, and thickness and distribution of weak interlayers in karst caves.

[0044] In the data preprocessing stage, the preprocessing workflow was optimized to suit the characteristics of karst geology. For seismic wave data, the focus was on removing scattering noise specific to karst areas, improving the signal-to-noise ratio of the effective signal through multiple superpositions, and employing an improved first-arrival picking algorithm to enhance the picking accuracy of karst interfaces. For ground-penetrating radar data, the focus was on processing multiple wave interference from cave reflection signals, improving vertical resolution through deconvolution processing, and correcting the spatial morphology of caves using migration imaging algorithms. For borehole geological logging data, the focus was on quantifying the degree of karst development and the properties of infill materials, generating a digital profile of karst development.

[0045] In the feature extraction stage, the content of feature extraction is optimized based on the characteristics of karst geology. For seismic wave data, the focus is on extracting features such as abrupt changes in wave velocity, abrupt changes in energy attenuation, and scattering characteristics to reflect the differences in physical properties between karst caves and weak interlayers. For ground-penetrating radar data, the focus is on extracting features such as strong reflection interfaces, reflection signal attenuation, and interruption of in-phase axes to identify karst caves of different sizes and filling types. For borehole data, the focus is on extracting features such as abrupt changes in drilling parameters and lithological changes to pinpoint the accurate location of karst caves and weak interlayers. Mutual information is used for feature selection, retaining features that contribute significantly to karst identification and generating standardized feature vectors.

[0046] In the feature fusion and recognition stage, standardized feature vectors are input into a cross-modal fusion network to generate fused feature vectors. The fused feature vectors are then input into an intelligent identification model for adverse geological conditions. The model outputs that there are three karst caves and one weak interlayer ahead of the tunnel face. The largest cave has a diameter of 6 meters and is 35 meters from the tunnel face, filled with clay. The weak interlayer is 2.5 meters thick and is 58 meters from the tunnel face.

[0047] During the risk assessment and early warning phase, a comprehensive risk value is calculated, determining the risk level of the largest karst cave to be Level II and the risk level of the weak interlayer to be Level III. The system generates an early warning report containing detailed information on each adverse geological body and targeted construction and treatment suggestions, and the early warning information is pushed to relevant personnel. During construction, data collection and analysis are automatically triggered every 8 meters of excavation, dynamically updating the identification results and risk levels. When the excavation face reaches a position 10 meters away from the karst cave, the system increases the detection frequency, monitors changes in the karst cave boundary in real time, and guides the construction unit to take advanced pre-grouting reinforcement measures.

[0048] Table 2 Comparison of Tunnel Detection Results in Karst Areas Table 2 illustrates the advantages of this invention in karst tunnel scenarios through five core indicators. The improved identification rate of karst caves and weak interlayers stems from the multi-dimensional characterization of karst geological features using multi-source data. Seismic waves can detect large, deep karst caves, ground-penetrating radar can identify small, shallow karst caves and thin, weak interlayers, and borehole data provides accurate calibration. The improved spatial positioning accuracy benefits from the cross-validation and fusion positioning of multi-source data, eliminating the positioning bias of single methods. The improved risk assessment accuracy relies on a quantitative risk assessment model, combining the size, filling material type, and distribution location of karst caves for comprehensive judgment. The shortened dynamic update response time comes from automated data acquisition and real-time analysis processes, enabling rapid response to geological changes during construction, timely adjustment of early warning information, and ensuring construction safety.

[0049] Reference Figure 2 This diagram focuses on the physical implementation and spatiotemporal alignment logic of the detection end. The hardware layer adopts a "layered, symmetrical, and integrated" layout strategy: three-component seismic detectors are symmetrically arranged along the sidewalls to capture large-scale deep reflected signals; the ground-penetrating radar uses movable supports to mount antennas of different frequencies, achieving coverage from shallow high resolution to deep panoramic views; the borehole acquisition unit is integrated into the drilling rig to capture mechanical dynamic parameters in real time. The technical highlight of this architecture lies in the introduction of the "BeiDou Time Service Center," which provides a unified microsecond-level time reference for all heterogeneous sensors. This high-precision synchronization eliminates the time drift caused by the difference in sampling frequencies between seismic waves and electromagnetic waves, ensuring that during subsequent feature fusion, each set of seismic features, radar features, and borehole features points to the exact same geological point in space. This is a prerequisite for achieving "spatial mapping" and "feature complementarity" of multi-source data.

[0050] Reference Figure 3 This figure details the purification process from "noisy data" to "standard feature vectors." Different preprocessing pipelines are designed for different physical media: seismic waves focus on wavefield separation and velocity scanning, using the energy ratio method to pinpoint the first arrival wave; radar waves focus on background suppression and migration imaging to correct the geometric position of the reflecting interface; and borehole data undergoes a digital transformation from qualitative description to quantitative indicators. In the feature extraction stage, the system not only captures static features in the time and frequency domains (such as amplitude and dominant frequency) but also delves deeper into dynamic features (such as wave velocity ratio and attenuation coefficient). To prevent the "curse of dimensionality" caused by feature redundancy, this process specifically introduces mutual information for feature selection, eliminating irrelevant variables and ultimately outputting a set of normalized, high-contribution feature vectors. This process is logically rigorous, ensuring that the data input to the recognition model is both physically representative and computationally efficient.

[0051] Reference Figure 4This diagram reveals the system's "intelligent brain" architecture, employing a design philosophy of cross-modal fusion and hybrid deep learning. The recognition process is divided into two core dimensions: spatial structure and temporal trend. First, through a global fusion attention module, the system can automatically adjust the weights of each detection source according to the current geological environment, such as increasing the weight of ground-penetrating radar features in water-rich zones. Subsequently, a convolutional neural network (CNN) is responsible for extracting the spatial geometry of adverse geological bodies, capturing abrupt spatial changes in features such as faults and karst caves; while a long short-term memory network (LSTM) is responsible for analyzing the continuous evolution trend of surrounding rock parameters as the detection depth increases. This dual perception model of "spatial + temporal," combined with a multi-task learning output layer, can simultaneously determine the type, location, extent, and scale of geological bodies. The model, trained through closed-loop training on historical excavation data, possesses extremely strong generalization capabilities.

[0052] Reference Figure 5 This diagram illustrates the logical generation of early warning decisions and the self-evolution mechanism of the model. The system transforms the identified geological parameters into engineering risk language, calculates a quantified risk score by comprehensively considering three dimensions: scale, physical properties, and engineering impact, and classifies four early warning levels accordingly. The early warning system adopts a differentiated response strategy, from "work stoppage" for level one risk to "intensified monitoring" for level four risk, achieving precise policy implementation. More importantly, this process establishes a "real-time data-driven" incremental update mechanism. In actual construction, with each step forward at the tunnel face, the actual geological conditions revealed by the excavation (labeled data) are collected and compared in real time with the previous identification results. This cycle of "measurement-early warning-verification-correction" enables the system to quickly correct model parameters when geological conditions change abruptly, effectively solving the common problem of "accurate detection, difficult prediction" in geological exploration, and significantly improving the timeliness and reliability of early warnings.

[0053] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for advanced detection and early warning analysis of adverse geological conditions in tunnels, characterized in that, Includes the following steps: A multi-source geophysical detection array is set up at a predetermined distance behind the tunnel face. The array includes a three-component seismic detector array, a multi-frequency ground-penetrating radar antenna array, and a borehole parameter acquisition unit. The microsecond-level time synchronization of all detection equipment is achieved through the BeiDou time synchronization system, and the seismic reflection wave data, ground-penetrating radar electromagnetic wave data, and geological logging data of the advanced borehole within the predetermined detection range in front of the tunnel face are collected simultaneously. Differentiated preprocessing procedures were performed on the collected multi-source raw data to complete random noise removal, surface wave separation, automatic first arrival picking and velocity analysis of seismic wave data, background clutter removal, gain compensation, migration imaging and data interpolation of ground-penetrating radar data, and standardization and digitization of borehole geological logging data. Extract multi-dimensional feature parameters from the preprocessed multi-source data, including the time-domain amplitude, frequency-domain dominant frequency, phase shift, P-wave velocity, and S-wave velocity characteristics of seismic waves, and the reflection intensity, dielectric constant distribution, electromagnetic wave attenuation coefficient, and reflection interface depth characteristics of ground-penetrating radar. A multi-source feature fusion model is constructed to perform feature layer fusion processing on feature parameters of different types and dimensions to generate a fused feature dataset containing geological structure and physical property information. Based on the fused feature dataset and the pre-trained intelligent identification model for adverse geological conditions, common adverse geological bodies in front of the tunnel face are identified, and the spatial location, distribution range and geometric scale of the adverse geological bodies are determined. By combining the design parameters of the tunnel project, the regional geological background conditions, and the parameters of the identified adverse geological bodies, the risk level of the adverse geological bodies to the tunnel construction is calculated; based on the calculated risk level, the corresponding level of early warning is triggered, and a standardized early warning report containing detailed information on the adverse geological bodies, risk assessment results, and targeted treatment suggestions is generated. During tunnel excavation, newly added detection data and geological information revealed by the excavation face are accessed in real time, and the results of adverse geological identification and risk assessment models are dynamically updated.

2. The method for advanced detection and early warning analysis of adverse geological conditions in tunnels according to claim 1, characterized in that, It also includes a step for quantitative assessment of adverse geological risks, which combines the scale parameters, physical property parameters, and engineering impact parameters of the adverse geological body to complete the quantitative calculation of the risk level. The calculation formula is as follows: ;in This is a dimensionless adverse geological risk value. For the dimensionless weighting coefficients of the scale parameter, These are dimensionless weighting coefficients for physical property parameters. These are the dimensionless weighting coefficients for engineering influence parameters. The normalized size value of the adverse geological body. The normalized physical property values ​​of the adverse geological body. This represents the normalized impact value of adverse geological bodies on the project.

3. The method for advanced detection and early warning analysis of adverse geological conditions in tunnels according to claim 1, characterized in that, It also includes a dynamic weight allocation step for multi-source data fusion, which combines quality parameters, environmental impact parameters, and geological adaptability parameters from different data sources to dynamically calculate the fusion weights. The calculation formula is as follows: ;in For the first Dimensionless fusion weights of multiple data sources For the first Normalized data quality values ​​for each data source. For the first Normalized environmental adaptability values ​​of each data source. For the first Normalized geological type adaptability values ​​from each data source, dimensionless. The total number of data sources participating in the integration.

4. The method for advanced detection and early warning analysis of adverse geological conditions in tunnels according to claim 1, characterized in that, The multi-source geophysical exploration array adopts a layered arrangement. The three-component seismic detector array is symmetrically arranged along the two side walls of the tunnel, with the detector spacing set to a preset fixed value. The arrangement length covers a preset range behind the tunnel face, completing seismic wave detection at a depth of 100 to 150 meters in front of the tunnel face. The multi-frequency ground-penetrating radar antenna array is installed on a movable support and includes antennas of three different frequencies: low frequency, medium frequency, and high frequency. The borehole parameter acquisition unit is integrated on the advanced drilling rig to collect drilling speed, torque, drilling pressure, and rotational speed parameters in real time during the drilling process. All detection equipment obtains a unified time reference through the Beidou time synchronization module.

5. The method for advanced detection and early warning analysis of adverse geological conditions in tunnels according to claim 1, characterized in that, The multi-source data differential preprocessing adopts a multi-stage processing architecture. Seismic wave data preprocessing first removes 50Hz power frequency interference and high-frequency random noise through bandpass filtering, then separates surface waves and effective reflected waves through FK filtering, and automatically picks up the first arrival wave by combining the energy ratio method and polarization analysis. The velocity distribution model of P-waves and S-waves in the formation is obtained by velocity scanning. Ground penetrating radar data preprocessing first eliminates direct waves and ground clutter by average background removal, then compensates for the attenuation of electromagnetic waves with propagation distance through automatic gain compensation, corrects the spatial position of the reflection interface by Kirchhoff migration imaging algorithm, and completes the missing data areas by linear interpolation. Borehole geological logging data preprocessing converts qualitative descriptions into standardized quantitative parameters and generates digital borehole geological profiles. After preprocessing, the quality of all data is assessed.

6. The method for advanced detection and early warning analysis of adverse geological conditions in tunnels according to claim 1, characterized in that, Multi-dimensional feature extraction employs a hierarchical feature extraction architecture. For seismic wave data, it extracts features such as root mean square amplitude, peak amplitude, energy attenuation rate, and duration in the time domain; dominant frequency, bandwidth, energy distribution entropy, and harmonic ratio in the frequency domain; and physical property features. For ground-penetrating radar data, it extracts features such as mean single-channel reflection intensity, reflector layer thickness, dielectric constant profile, attenuation coefficient profile, and number of reflector interfaces. For borehole data, it extracts features such as lithology index, rock mass integrity index, water content index, and joint density index. The mutual information method is used to select features from all extracted features, eliminating redundant and irrelevant features and retaining features that contribute significantly to the identification of adverse geological conditions.

7. The method for advanced detection and early warning analysis of adverse geological conditions in tunnels according to claim 1, characterized in that, The multi-source feature fusion model adopts a cross-modal fusion network architecture based on an attention mechanism. The network contains three independent feature extraction branches and a global fusion attention module. The three feature extraction branches process seismic wave features, ground-penetrating radar features, and borehole features, respectively. Each branch consists of multiple convolutional layers and pooling layers to extract deep abstract features from each data source. The global fusion attention module automatically learns the importance weights of different features under different geological conditions and performs weighted fusion of the deep features output by each branch. The parameters are updated through an adaptive moment estimation algorithm. During training, a regularization term is introduced to reduce the overfitting of the model.

8. The method for advanced detection and early warning analysis of adverse geological conditions in tunnels according to claim 1, characterized in that, The intelligent identification model for adverse geological conditions adopts a hybrid architecture combining convolutional neural networks and long short-term memory networks. The convolutional neural network extracts the spatial structural features of the fusion features to capture the geometric shape and distribution patterns of adverse geological bodies, while the long short-term memory network extracts the temporal variation features of the fusion features to capture the continuous variation trend of stratigraphic parameters. The model output layer adopts a multi-task learning structure, simultaneously outputting the type, spatial location, distribution range, and geometric scale parameters of adverse geological bodies. The model training dataset contains thousands of sets of measured data from tunnel advance detection under different geological conditions and corresponding excavation verification data. During training, data augmentation techniques such as random pruning and adding noise are used to expand the dataset size, and the five-fold cross-validation method is used to evaluate the model performance.

9. The method for advanced detection and early warning analysis of adverse geological conditions in tunnels according to claim 1, characterized in that, Adverse geological risks are classified into four levels: Level 1 corresponds to extremely high risk, Level 2 to high risk, Level 3 to medium risk, and Level 4 to low risk. The corresponding risk level is determined based on the calculated comprehensive risk value. Different warnings for different levels are sent out using different methods and response procedures. Level 1 warnings are sent directly to the project manager and construction command center through multiple means, including audible and visual alarms, SMS, and telephone, to activate the emergency response mechanism, suspend construction, and formulate an emergency response plan. Level 2 warnings are sent to the technical manager and construction teams to formulate specific construction response plans and strengthen on-site monitoring. Level 3 and Level 4 warnings are sent to on-site technical personnel to increase the frequency of construction monitoring and prepare for construction.

10. The method for advanced detection and early warning analysis of adverse geological conditions in tunnels according to claim 1, characterized in that, The dynamic update mechanism adopts a real-time data-driven incremental update method. During the tunnel excavation process, after each preset advance is completed, a multi-source data acquisition and analysis is automatically triggered. The newly added detection data is input into the identification model and risk assessment model to update the identification results and risk levels of adverse geology in front of the tunnel face. At the same time, the actual geological information exposed by the excavation of the tunnel face is entered into the system as label data to fine-tune the identification model online and correct the model parameters. The system automatically stores all historical detection data and excavation verification data to build a local geological database.

Images