A method and system for dynamic classification of surrounding rock in TBM tunnels based on cutterhead scratches

Abstract

This invention relates to the field of data processing technology and discloses a dynamic classification method and system for TBM tunnel surrounding rock based on cutterhead scratches. It aims to solve the problems of poor real-time performance and accuracy in existing methods. The main scheme includes: after the TBM completes one ring of tunneling, point cloud and image data are simultaneously collected behind the shield tail, along with short-term deformation time-series data of the surrounding rock and cutterhead operating condition data; the point cloud data is preprocessed and divided into grid cells; scratch morphology parameters are extracted within each grid cell, and scratch feature indices are further calculated; the deformation time-series data is analyzed to extract surrounding rock deformation response parameters; the scratch morphology parameters, scratch feature indices, and surrounding rock deformation response parameters are standardized and fused with the cutterhead operating condition data to form a unified fused feature vector, which is then input into a pre-trained surrounding rock classification model to obtain the surrounding rock grade for each grid cell. This invention improves the real-time performance and accuracy of surrounding rock classification.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and specifically to a dynamic classification method and system for TBM tunnel surrounding rock based on cutterhead scratches. Background Technology

[0002] Tunnel boring machines (TBMs), as efficient and continuous full-face mechanized tunneling equipment, have been widely used in long tunnel projects in railways, highways, water conservancy, and mining. With engineering construction trending towards deeper, longer, and more complex geological conditions, the stability of the surrounding rock has become an increasingly prominent issue. Accurately and in real-time determining the quality grade of the surrounding rock during tunneling is crucial for achieving safe and efficient construction and dynamically optimizing support decisions.

[0003] Traditional methods for classifying surrounding rock mainly rely on geological surveys before construction and manual logging and experience-based judgment during construction. These methods suffer from significant lag and discontinuity, making it difficult to provide real-time, dynamic geological information feedback for continuous TBM tunneling. Especially when encountering adverse geological formations such as fault fracture zones, weak interlayers, and abrupt lithological changes, traditional methods are unable to provide timely warnings, easily leading to construction risks.

[0004] To enhance the geological perception capabilities of TBM construction, existing technologies have explored various online identification or inversion methods, primarily including: The first is empirical inversion based on tunneling parameters, which involves indirectly inferring the strength and integrity of the surrounding rock by real-time monitoring of operating parameters such as cutterhead torque, thrust, and rotational speed, combined with historical data statistics or empirical formulas, thereby achieving surrounding rock classification; the second is indirect analysis based on rock cuttings or vibration signals, which involves statistically analyzing the particle size distribution of tunneling debris or performing spectral analysis of machine vibration signals to determine lithological changes, thereby achieving surrounding rock classification; the third is based on advanced geological prediction technologies, such as the Transit Seismic Wave Spectroscopy (TSP) or ground-penetrating radar, which involves physical detection ahead of the tunnel to identify large-scale geological anomalies, thereby achieving surrounding rock classification; and the fourth is based on manual or automated logging behind the excavation face or shield tail, which involves identifying cracks, spalling, and other conditions through manual observation or laser scanning after the surrounding rock is exposed, and then classifying it accordingly.

[0005] While the aforementioned methods have some application, they all suffer from fundamental limitations. First, information acquisition is generally delayed or non-real-time, with long interpretation cycles for advance forecasts, and observation of the exposed face lags significantly behind the excavation face, failing to meet the urgent need for real-time, dynamic perception of the surrounding rock condition in continuous TBM tunneling. Second, these methods often rely on a single data source, using only macroscopic tunneling parameters or rear-view images, making their judgments susceptible to interference from equipment conditions, tool wear, construction dust, and human experience, resulting in poor noise resistance and a high misjudgment rate. Furthermore, existing technologies almost completely ignore the direct and rich physical information carrier—the scratches left on the tunnel wall after the interaction between the cutterhead and the surrounding rock—failing to reflect the essence of cutting behavior and failing to systematically correlate and quantify their geometric morphology and texture distribution with the mechanical behavior and failure modes of the surrounding rock. They lack a standardized framework capable of deeply integrating multi-source heterogeneous data (such as geometry, texture, temporal deformation, and operating conditions), leading to poor accuracy in surrounding rock classification. Summary of the Invention

[0006] This invention aims to solve the problems of poor real-time performance and accuracy of existing tunnel surrounding rock classification methods, and proposes a dynamic classification method and system for TBM tunnel surrounding rock based on cutterhead scratches.

[0007] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:

[0008] In a first aspect, the present invention provides a dynamic classification method for surrounding rock in TBM tunnels based on cutterhead scraping marks, the method comprising:

[0009] After the TBM completes one ring of tunneling, three-dimensional laser scanning and surface image acquisition of the tunnel wall are performed simultaneously in the area behind the shield tail. Point cloud data for characterizing the three-dimensional geometry of the cutterhead scratches and image data for characterizing its surface texture features are obtained respectively. At the same time, short-term deformation time series data of the surrounding rock and cutterhead working condition data are collected.

[0010] The point cloud data is preprocessed and divided into grid cells. Based on the point cloud data and image data in each grid cell, the scratch morphology parameters of the blade scratch are extracted, and the scratch feature index is calculated according to the scratch morphology parameters. The scratch feature index includes scratch clarity index, scratch uniformity index and scratch depth index.

[0011] The short-term deformation time series data are processed to extract the surrounding rock deformation response parameters, which include the surrounding rock convergence rate and creep model parameters.

[0012] The cutterhead working condition data, scratch morphology parameters, scratch characteristic indicators, and surrounding rock deformation response parameters are standardized and fused to form a unified fused feature vector.

[0013] The fused feature vector is input into a pre-trained surrounding rock classification model to obtain the surrounding rock grade of each grid cell. Based on the surrounding rock grades of all grid cells, a spatial distribution cloud map of the surrounding rock grades is generated.

[0014] Furthermore, the three-dimensional laser scanning is performed using a 3D laser scanner, the image acquisition is performed using an industrial camera, and the area behind the shield tail is a range of 3-10 meters behind the shield tail.

[0015] Furthermore, the preprocessing of the point cloud data includes: denoising, registration, and removal of non-surrounding rock surfaces;

[0016] The process of dividing the grid unit includes: mapping the point cloud on the cave wall surface to a two-dimensional unfolded coordinate system, and then dividing the grid unit according to a preset scale.

[0017] Furthermore, the scratch morphology parameters for extracting the blade disc scratches include:

[0018] Depth maps are generated from point cloud data based on grid cells, and edge detection and connected component analysis are performed in combination with image data to identify valid scratches;

[0019] For each valid scratch within a grid cell, extract the scratch length, average depth, maximum depth, depth variance, orientation angle, scratch edge sharpness, and scratch coverage area. Calculate the average scratch length, average average depth, average average maximum depth, average depth variance, average orientation angle, and average scratch edge sharpness for all valid scratches within the grid cell. Then, calculate the scratch coverage rate of the grid cell based on the scratch coverage area.

[0020] The average value of the scratch length, average value of the average depth, average value of the maximum depth, average value of the depth variance, average value of the direction angle, average value of the scratch edge sharpness, and scratch coverage of all valid scratches within the grid cell are used as the scratch morphology parameters.

[0021] Furthermore, the formula for calculating the scratch sharpness index for each grid cell is as follows:

[0022] ;

[0023] in, This indicates the scratch sharpness index of the grid cells. This represents the average edge sharpness of all valid scratches within the grid cell. This represents the average depth of all valid scratches within the grid cell. This represents the average depth variance of all valid scratches within the given grid cell. , and These represent calibration constants.

[0024] Furthermore, the formula for calculating the scratch uniformity index of each grid cell is as follows:

[0025] ;

[0026] in, The scratch uniformity index represents the grid cell. This indicates the scratch coverage of the grid cell. This represents the variance of the scratch coverage of all graded grid cells. This represents the average scratch coverage of all graded grid cells. This represents a small constant.

[0027] Furthermore, the formula for calculating the scratch depth index for each grid cell is as follows:

[0028] ;

[0029] in, The scratch depth index represents the grid cell. This represents the data set consisting of the average depth of all valid scratches within the grid cell. Indicates reference depth. This indicates taking the median.

[0030] Furthermore, the surrounding rock convergence rate is the initial convergence rate obtained by linearly fitting the deformation-time data during the initial period after excavation, and its calculation formula is as follows:

[0031] ;

[0032] in, This represents the initial convergence rate of the surrounding rock. Indicates the first The time corresponding to each monitoring point This represents the amount of convergence deformation of the surrounding rock at that moment. This represents the average value across all monitoring times. This represents the average value of all monitored deformations.

[0033] Furthermore, the creep model parameters are obtained by fitting a logarithmic creep model used to characterize the creep properties of the surrounding rock. The time span of the deformation-time data used for fitting is greater than the initial time period. The logarithmic creep model is as follows:

[0034] ;

[0035] in, Indicates time The cumulative convergence deformation of the surrounding rock. and Indicates the parameters of the creep model. Characterizing the amplitude of creep deformation, Characterizing the decay of creep rate over time, It represents the natural logarithm.

[0036] In a second aspect, the present invention provides a dynamic grading system for TBM tunnel surrounding rock based on cutterhead scratches, used to implement the dynamic grading method for TBM tunnel surrounding rock based on cutterhead scratches as described in the first aspect, the system comprising:

[0037] The data acquisition module is used to simultaneously perform three-dimensional laser scanning and surface image acquisition of the tunnel wall in the area behind the shield tail after the TBM completes one ring of tunneling. It obtains point cloud data to characterize the three-dimensional geometric morphology of the cutterhead scratches and image data to characterize their surface texture features. At the same time, it acquires short-term deformation time series data of the surrounding rock and cutterhead working condition data.

[0038] The data processing module is used to preprocess the point cloud data and divide it into grid cells. Based on the point cloud data and image data in each grid cell, it extracts the scratch morphology parameters of the cutterhead scratch and calculates scratch feature indices according to the scratch morphology parameters. The scratch feature indices include scratch clarity index, scratch uniformity index, and scratch depth index. The module also processes the short-term deformation time series data to extract the surrounding rock deformation response parameters, which include the surrounding rock convergence rate and creep model parameters.

[0039] The feature fusion module is used to standardize and fuse the cutterhead working condition data, scratch morphology parameters, scratch feature indicators, and surrounding rock deformation response parameters to form a unified fused feature vector.

[0040] The surrounding rock classification module is used to input the fused feature vector into the pre-trained surrounding rock classification model to obtain the surrounding rock level of each grid cell, and generate a spatial distribution cloud map of the surrounding rock level based on the surrounding rock levels of all grid cells.

[0041] The beneficial effects of this invention are as follows: The TBM tunnel surrounding rock dynamic classification method and system based on cutterhead scratches provided by this invention synchronously collects three-dimensional point cloud and surface image data of cutterhead scratches, short-term deformation time series data of surrounding rock, and cutterhead working condition data. Based on grid cells, it extracts scratch morphological parameters and comprehensive indicators, and constructs a complete technical system from data acquisition, feature extraction, multi-source fusion to intelligent inference. This achieves real-time, gridded, and physically interpretable dynamic classification of surrounding rock quality, overcoming the shortcomings of traditional methods such as lag, susceptibility to interference from single data sources, and inability to reflect the essence of cutting behavior. It significantly improves the accuracy, robustness, and engineering applicability of surrounding rock classification results, providing reliable technical support for real-time support decision-making and risk warning in TBM construction. Attached Figure Description

[0042] Figure 1 A flowchart illustrating the dynamic classification method for TBM tunnel surrounding rock based on cutterhead scratches provided in this embodiment;

[0043] Figure 2 This is a schematic diagram of the structure of a dynamic grading system for TBM tunnel surrounding rock based on cutterhead scratches, provided as an example. Detailed Implementation

[0044] Because existing technologies rely heavily on a single data source for surrounding rock classification and lack effective means to integrate multi-source sensor data, the results of surrounding rock classification are severely lagging, sensitive to single data sources and working condition interference, and difficult to achieve real-time dynamic updates and high-resolution spatial representation. Therefore, they cannot meet the urgent need of TBM intelligent construction for real-time and continuous perception and evaluation of surrounding rock conditions.

[0045] Based on this, the technical solution of the present invention is proposed. In the present invention, firstly, after the TBM completes one ring of tunneling, three-dimensional laser point cloud data, surface image data, short-term deformation time series data of the surrounding rock, and cutterhead working condition data of the tunnel wall are simultaneously collected in the stable area behind the shield tail. Next, the point cloud and image data are preprocessed and divided into grid cells. Within each grid cell, the scratch morphology parameters of the cutterhead scratches are extracted, and scratch feature indices such as scratch clarity index, scratch uniformity index, and scratch depth index are further calculated. At the same time, the deformation time series data is analyzed to extract the surrounding rock convergence rate and creep model parameters. Then, the above-mentioned scratch morphology parameters, scratch feature indices, deformation response parameters, and cutterhead working condition data are standardized and feature fused to form a unified fused feature vector, which is input into a pre-trained surrounding rock classification model to obtain the surrounding rock grade of each grid cell. Finally, a spatial distribution cloud map of the surrounding rock grade is generated based on the classification results of all grid cells.

[0046] The technical solutions in this embodiment 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.

[0047] Figure 1 A flowchart illustrating a dynamic classification method for TBM tunnel surrounding rock based on cutterhead scratches is shown. Please refer to [link / reference]. Figure 1 The method includes the following steps:

[0048] Step 1: After the TBM completes one ring of tunneling, three-dimensional laser scanning and surface image acquisition of the tunnel wall are performed simultaneously in the area behind the shield tail to obtain point cloud data for characterizing the three-dimensional geometric morphology of the cutterhead scratches and image data for characterizing its surface texture features. At the same time, short-term deformation time series data of the surrounding rock and cutterhead working condition data are collected.

[0049] In this embodiment, the three-dimensional laser scanning is performed using a 3D laser scanner, the image acquisition is performed using an industrial camera, and the area behind the shield tail is a range of 3-10 meters behind the shield tail.

[0050] In practical applications, after the TBM completes one ring of tunneling, three data acquisition operations are simultaneously carried out in a relatively stable area approximately 3-10 meters behind the tail shield and from the excavation face. Specifically, firstly, a 3D laser scanner with a ranging accuracy of no less than ±5mm is used to perform a full-section scan of the exposed tunnel wall to obtain high-density point cloud data for characterizing the three-dimensional geometry of the cutterhead scratches; simultaneously, an industrial camera equipped with a dustproof housing and auxiliary lighting module is used to photograph the same area to obtain two-dimensional image data for characterizing the surface texture features of the scratches; in addition, high-precision displacement sensors (such as fiber Bragg gratings (FBG) or LVDT displacement gauges) pre-embedded in the cross-section continuously collect short-term deformation time-series data of the surrounding rock, and sensors installed on the cutterhead drive system collect cutterhead operating parameters such as torque, thrust, and rotational speed in real time. All acquisition devices are synchronized via a unified clock source to ensure that the multi-source data have a consistent spatiotemporal reference, providing high-quality input for subsequent feature extraction and fusion.

[0051] Step 2: Preprocess the point cloud data and divide it into grid cells. Based on the point cloud data and image data in each grid cell, extract the scratch morphology parameters of the blade scratch and calculate the scratch feature index according to the scratch morphology parameters. The scratch feature index includes scratch clarity index, scratch uniformity index and scratch depth index.

[0052] In this embodiment, the preprocessing of the point cloud data includes: denoising, registration, and removal of non-surrounding rock surfaces; the division of grid units includes: mapping the point cloud of the cave wall surface to a two-dimensional unfolded coordinate system, and then dividing it into grid units according to a preset scale.

[0053] In practical applications, the three-dimensional point cloud data obtained in step 1 is first preprocessed. This includes: using statistical outlier filtering to remove noise points caused by dust and water droplets; using initial registration based on FPFH features and the ICP algorithm for iterative fine matching to fuse the point clouds from multiple scans into a high-precision unified model; subsequently, applying algorithms such as RANSAC plane fitting to automatically identify and remove non-surrounding rock surfaces such as tunnel lining segments and construction equipment, retaining only the point cloud of exposed surrounding rock areas. Next, the processed surrounding rock surface point cloud is mapped to a two-dimensional unfolded coordinate system (such as a circumferential-axial coordinate system) based on the tunnel axis, and meshed according to a preset engineering scale (usually 0.5m×0.5m to 1.0m×1.0m), discretizing the continuous cross-section into a series of independent analysis units.

[0054] For each grid cell, the cutter head scratches are detected and parameters are extracted by combining the point cloud data within it with the corresponding image data. In this embodiment, the extraction of the scratch morphology parameters of the cutter head scratches includes steps 21 to 23:

[0055] Step 21: Generate a depth map based on the point cloud data of the grid cells, and combine it with the image data to perform edge detection and connected component analysis to identify valid scratches;

[0056] Step 22: For each valid scratch within a grid cell, extract the scratch length, average depth, maximum depth, depth variance, orientation angle, scratch edge sharpness, and scratch coverage area. Calculate the average scratch length, average average depth, average maximum depth, average depth variance, average orientation angle, and average scratch edge sharpness of all valid scratches within the grid cell. Calculate the scratch coverage rate of the grid cell based on the scratch coverage area.

[0057] Step 23: The average value of the scratch length, the average value of the average depth, the average value of the maximum depth, the average value of the depth variance, the average value of the direction angle, the average value of the scratch edge sharpness, and the scratch coverage rate of all valid scratches in the grid cell are used as the scratch morphology parameters.

[0058] In practical applications, the point cloud is first subjected to local curvature calculation and depth map generation to enhance the geometric features of the scratches. Then, image edge detection algorithms (such as the Canny operator) are used to identify potential scratch contours. Through connected component analysis, valid scratches with lengths and depths exceeding engineering-significant thresholds are selected. For each valid scratch, its morphological parameters are quantified and extracted, including: the average scratch length, average average depth, average maximum depth, average depth variance, average orientation angle, average scratch edge sharpness, and scratch coverage rate for all valid scratches within the grid cell.

[0059] Based on the extracted scratch morphology parameters, scratch feature indices are further calculated, including scratch clarity index, scratch uniformity index, and scratch depth index.

[0060] In this embodiment, the formula for calculating the scratch sharpness index of each grid cell is as follows:

[0061] ;

[0062] in, This indicates the scratch sharpness index of the grid cells. This represents the average edge sharpness of all valid scratches within the grid cell. This represents the average depth of all valid scratches within the grid cell. This represents the average depth variance of all valid scratches within the given grid cell. , and They represent the calibration constants, where and These are calibration constants used to adjust the sensitivity and offset of logic functions. To prevent calibration constants with a denominator of zero.

[0063] The formula for calculating the scratch uniformity index of each grid cell is as follows:

[0064] ;

[0065] in, The scratch uniformity index represents the grid cell. This indicates the scratch coverage of the grid cell. This represents the variance of the scratch coverage of all graded grid cells. This represents the average scratch coverage of all graded grid cells. This is a small constant used to prevent the denominator from being zero.

[0066] The formula for calculating the scratch depth index for each grid cell is as follows:

[0067] ;

[0068] in, The scratch depth index represents the grid cell. This represents the data set consisting of the average depth of all valid scratches within the grid cell. This refers to a reference depth used for normalization, pre-defined based on engineering experience or lithology. This indicates taking the median.

[0069] The scratch sharpness index combines edge sharpness and depth stability; a higher value indicates a sharper scratch outline and more uniform depth, typically corresponding to a more intact rock mass and a more brittle shear failure mode. The scratch uniformity index characterizes the spatial distribution of scratches, reflecting the homogeneity of the surrounding rock structure and the distribution characteristics of structural surfaces. The scratch depth index, through normalization of the average depth of grid cells, reflects the cutting resistance and local failure depth of the surrounding rock. These three indices together constitute the key feature vector for inverting surrounding rock properties from scratch morphology.

[0070] Step 3: Process the short-term deformation time series data to extract the surrounding rock deformation response parameters, which include the surrounding rock convergence rate and creep model parameters.

[0071] This step aims to extract quantitative parameters characterizing the time-dependent deformation behavior of the surrounding rock from short-term deformation monitoring data. These deformation response parameters mainly include the surrounding rock convergence rate, which reflects the rapid deformation trend of the surrounding rock, and creep model parameters, which characterize its medium- and long-term rheological properties.

[0072] In this embodiment, the surrounding rock convergence rate is the initial convergence rate obtained by linearly fitting the deformation-time data during the initial period after excavation, and its calculation formula is as follows:

[0073] ;

[0074] in, This represents the initial convergence rate of the surrounding rock. Indicates the first The time corresponding to each monitoring point This represents the amount of convergence deformation of the surrounding rock at that moment. This represents the average value across all monitoring times. This represents the average value of all monitored deformations.

[0075] In practical applications, for newly excavated sections, displacement sensors are deployed at key measuring points such as the arch crown and arch waist to collect surrounding rock convergence deformation-time series data in the initial stage after excavation (e.g., 0 to 24 hours). Linear regression analysis is performed on the deformation-time data of this initial period, and the initial convergence rate is obtained by fitting with the least squares method, which is then used as the surrounding rock convergence rate.

[0076] In this embodiment, the creep model parameters are obtained by fitting a logarithmic creep model used to characterize the creep properties of the surrounding rock. The time span of the deformation-time data used for fitting is greater than the initial time period. The logarithmic creep model is as follows:

[0077] ;

[0078] in, Indicates time The cumulative convergence deformation of the surrounding rock. and Indicates the parameters of the creep model. Characterizing the amplitude of creep deformation, Characterizing the decay of creep rate over time, It represents the natural logarithm.

[0079] Specifically, this embodiment employs a physically meaningful logarithmic creep model to describe the gradual convergence of deformation of the surrounding rock over time. To evaluate the time-dependent deformation characteristics of the surrounding rock, it is necessary to analyze deformation monitoring data over a longer period (significantly longer than the initial period, e.g., several days to several weeks). Therefore, the time span of the deformation-time data used for fitting is longer than the initial period. The creep model parameters obtained from the fitting are... and Together, they constitute a quantitative description of the long-term rheological capacity and stability trend of the surrounding rock.

[0080] Step 4: Standardize and fuse the cutterhead working condition data, scratch morphology parameters, scratch characteristic indicators, and surrounding rock deformation response parameters to form a unified fused feature vector.

[0081] This step aims to standardize and merge diverse heterogeneous features from different sensors, with different physical meanings and dimensions, into a unified digital feature vector that can be used for model inference.

[0082] In practical applications, firstly, all features corresponding to the cutterhead working condition data, scratch morphology parameters, scratch characteristic indices, and surrounding rock deformation response parameters are standardized to eliminate dimensional differences and ensure numerical comparability of each feature. For example, a z-score standardization method based on training data statistics can be used, or a normalization interval can be determined based on feature engineering experience.

[0083] Then, based on the physical importance and empirical weights of the features in the surrounding rock classification task, the standardized features are weighted and combined to construct a unified fusion feature vector. For example, initial empirical weights (e.g., weights of 0.3, 0.25, 0.15, 0.2, and 0.1 respectively) are assigned to core features such as scratch clarity index, scratch depth index, scratch uniformity index, initial convergence rate, and cutterhead operating parameters (e.g., normalized torque), with a sum of 1. These weights can be automatically optimized and adjusted through the model's cross-validation process. Finally, all weighted features are concatenated into a unified fusion feature vector.

[0084] This embodiment constructs a fused feature vector that combines physical interpretability and data robustness by integrating cutterhead working condition data, scratch morphology parameters, scratch characteristic indicators, and surrounding rock deformation response parameters. This fused feature vector cross-validates the surrounding rock state from multiple dimensions such as spatial morphology, temporal evolution, and construction load, enabling the surrounding rock classification model to learn the complex mapping relationship between geological conditions and multi-source sensing information more comprehensively and accurately. This effectively overcomes the shortcomings of traditional single-data source methods, such as susceptibility to interference and biased judgment, and significantly improves the accuracy, reliability, and engineering applicability of dynamic classification results.

[0085] Step 5: Input the fused feature vector into the pre-trained surrounding rock classification model to obtain the surrounding rock grade of each grid cell. Based on the surrounding rock grades of all grid cells, generate a spatial distribution cloud map of the surrounding rock grades.

[0086] Specifically, the fused feature vector generated in step 4 is input into a pre-trained surrounding rock classification model. This model is preferably a hybrid model architecture to balance accuracy and interpretability. Specifically, the surrounding rock classification model typically comprises two main parts: a static tree model module (e.g., using LightGBM or Random Forest algorithms) for efficiently processing the fused static feature vector and outputting the preliminary probability distribution of the surrounding rock grade for each grid cell; and a temporal enhancement module (e.g., using LSTM or temporal convolutional networks) for further processing the original displacement time-series data associated with the grid cells, performing deformation dynamics-based correction and uncertainty quantification on the preliminary results of the static model. During the training phase, the model utilizes a large amount of labeled historical engineering data, with labels derived from a combination of traditional geological logging (e.g., RMR or Q-system classification), indoor mechanical test results, and expert judgment.

[0087] After receiving the fused feature vector from each grid cell, the surrounding rock classification model performs online inference and outputs the surrounding rock grade of that grid cell (usually represented as grades I to V or corresponding numerical scores) and a confidence score to quantify the reliability of the grade determination. Simultaneously, the model provides dominant feature contribution analysis (e.g., through SHAP values), visually demonstrating the key features influencing the classification result of that cell and their direction of influence, enhancing the transparency and interpretability of the model's decision-making.

[0088] Finally, based on the output results (grade, confidence level) of all grid cells, spatial interpolation and rendering are performed on the 2D unfolded diagram or 3D model of the tunnel cross-section to generate a spatial distribution cloud map of the surrounding rock grade. This cloud map clearly displays the spatial distribution of different grades of surrounding rock using continuous color bands, and can highlight low-confidence areas or high-risk weak zones, forming an intuitive digital portrait of the surrounding rock quality. This provides a direct and visual basis for subsequent support decisions and risk management.

[0089] In this embodiment, support recommendations based on the surrounding rock grade are also included. Specifically, the system has a built-in support recommendation knowledge base, which presets corresponding support measure types and key parameter ranges according to different surrounding rock grades. After obtaining the surrounding rock grade of each grid cell or the entire cross section, the system automatically matches and outputs the corresponding support recommendations from the support recommendation knowledge base.

[0090] For example, for Class I surrounding rock (generally intact), conventional lining is recommended, with shotcrete of 30-40 mm thickness and anchor spacing greater than 2.5 meters. For Class II surrounding rock (locally jointed), standard support is recommended, with shotcrete thickness of 40-60 mm and anchor length of 2-4 meters. For Class III surrounding rock (moderately fractured), reinforced support is required, with shotcrete thickness of 60-80 mm, anchor spacing of 1.5-2.0 meters, and local reinforcement mesh or steel arches may be added. For Class IV surrounding rock (highly fractured or weak zone), significantly reinforced support should be implemented, with closely spaced anchors (spacing less than 1.5 meters), combined with casing grouting and pre-reinforcement methods. For Class V surrounding rock (extremely fractured or strongly weak strata), it is recommended to suspend excavation and implement comprehensive measures such as temporary scaffolding, high-strength support, and large-scale grouting reinforcement, while initiating an expert review process.

[0091] In summary, the TBM tunnel surrounding rock dynamic classification method based on cutterhead scratches provided in this embodiment integrates multi-source information such as the morphological characteristics of cutterhead scratch marks, the temporal response of surrounding rock deformation, and cutterhead working conditions, and performs intelligent inversion based on grid cells. This achieves real-time, dynamic, and high-resolution classification of TBM tunnel surrounding rock, overcoming the shortcomings of traditional methods such as reliance on experience, strong lag, and weak anti-interference ability. It significantly improves the accuracy, robustness, and engineering interpretability of the classification results, providing reliable technical support for intelligent decision-making and safety control in tunnel construction.

[0092] Based on the above technical solutions, this embodiment also proposes a dynamic rock grading system for TBM tunnels based on cutterhead scratches, used to implement the dynamic rock grading method for TBM tunnels based on cutterhead scratches as described in the embodiment. Please refer to [link to relevant documentation]. Figure 2 The system includes:

[0093] The data acquisition module is used to simultaneously perform three-dimensional laser scanning and surface image acquisition of the tunnel wall in the area behind the shield tail after the TBM completes one ring of tunneling. It obtains point cloud data to characterize the three-dimensional geometric morphology of the cutterhead scratches and image data to characterize their surface texture features. At the same time, it acquires short-term deformation time series data of the surrounding rock and cutterhead working condition data.

[0094] The data processing module is used to preprocess the point cloud data and divide it into grid cells. Based on the point cloud data and image data in each grid cell, it extracts the scratch morphology parameters of the cutterhead scratch and calculates scratch feature indices according to the scratch morphology parameters. The scratch feature indices include scratch clarity index, scratch uniformity index, and scratch depth index. The module also processes the short-term deformation time series data to extract the surrounding rock deformation response parameters, which include the surrounding rock convergence rate and creep model parameters.

[0095] The feature fusion module is used to standardize and fuse the cutterhead working condition data, scratch morphology parameters, scratch feature indicators, and surrounding rock deformation response parameters to form a unified fused feature vector.

[0096] The surrounding rock classification module is used to input the fused feature vector into the pre-trained surrounding rock classification model to obtain the surrounding rock level of each grid cell, and generate a spatial distribution cloud map of the surrounding rock level based on the surrounding rock levels of all grid cells.

[0097] It is understood that the TBM tunnel surrounding rock dynamic classification system based on cutterhead scratches described in this embodiment is a system used to implement the TBM tunnel surrounding rock dynamic classification method based on cutterhead scratches described in the embodiment. As the system disclosed in the embodiment corresponds to the method disclosed in the embodiment, the description is relatively simple. For relevant parts, please refer to the description of the method. It will not be repeated here.

Claims

1. A dynamic classification method for surrounding rock in TBM tunnels based on cutterhead scratches, characterized in that, The method includes: After the TBM completes one ring of tunneling, three-dimensional laser scanning and surface image acquisition of the tunnel wall are performed simultaneously in the area behind the shield tail. Point cloud data for characterizing the three-dimensional geometry of the cutterhead scratches and image data for characterizing its surface texture features are obtained respectively. At the same time, short-term deformation time series data of the surrounding rock and cutterhead working condition data are collected. The point cloud data is preprocessed and divided into grid cells. Based on the point cloud data and image data in each grid cell, the scratch morphology parameters of the blade scratch are extracted, and the scratch feature index is calculated according to the scratch morphology parameters. The scratch feature index includes scratch clarity index, scratch uniformity index and scratch depth index. The scratch morphology parameters for extracting the blade disc scratches include: Depth maps are generated from point cloud data based on grid cells, and edge detection and connected component analysis are performed in combination with image data to identify valid scratches; For each valid scratch within a grid cell, extract the scratch length, average depth, maximum depth, depth variance, orientation angle, scratch edge sharpness, and scratch coverage area. Calculate the average scratch length, average average depth, average average maximum depth, average depth variance, average orientation angle, and average scratch edge sharpness for all valid scratches within the grid cell. Then, calculate the scratch coverage rate of the grid cell based on the scratch coverage area. The average value of the scratch length, the average value of the average depth, the average value of the maximum depth, the average value of the depth variance, the average value of the direction angle, the average value of the scratch edge sharpness, and the scratch coverage rate of all valid scratches within the grid cell are used as the scratch morphology parameters. The short-term deformation time series data are processed to extract the surrounding rock deformation response parameters, which include the surrounding rock convergence rate and creep model parameters. The cutterhead working condition data, scratch morphology parameters, scratch characteristic indicators, and surrounding rock deformation response parameters are standardized and fused to form a unified fused feature vector. The fused feature vector is input into a pre-trained surrounding rock classification model to obtain the surrounding rock grade of each grid cell. Based on the surrounding rock grades of all grid cells, a spatial distribution cloud map of the surrounding rock grades is generated.

2. The dynamic classification method for TBM tunnel surrounding rock based on cutterhead scratches according to claim 1, characterized in that, The three-dimensional laser scanning is performed using a 3D laser scanner, the image acquisition is performed using an industrial camera, and the area behind the shield tail is a range of 3-10 meters behind the shield tail.

3. The dynamic classification method for TBM tunnel surrounding rock based on cutterhead scratches according to claim 1, characterized in that, The preprocessing of the point cloud data includes: denoising, registration, and removal of non-surrounding rock surfaces; The process of dividing the grid unit includes: mapping the point cloud on the cave wall surface to a two-dimensional unfolded coordinate system, and then dividing the grid unit according to a preset scale.

4. The dynamic classification method for TBM tunnel surrounding rock based on cutterhead scratches according to claim 1, characterized in that, The formula for calculating the scratch sharpness index for each grid cell is as follows: ； in, This indicates the scratch sharpness index of the grid cells. This represents the average edge sharpness of all valid scratches within the grid cell. This represents the average depth of all valid scratches within the grid cell. This represents the average depth variance of all valid scratches within the given grid cell. , and These represent calibration constants.

5. The dynamic classification method for TBM tunnel surrounding rock based on cutterhead scratches according to claim 1, characterized in that, The formula for calculating the scratch uniformity index of each grid cell is as follows: ； in, The scratch uniformity index represents the grid cell. This indicates the scratch coverage of the grid cell. This represents the variance of the scratch coverage of all graded grid cells. This represents the average scratch coverage of all graded grid cells. This represents a small constant.

6. The dynamic classification method for TBM tunnel surrounding rock based on cutterhead scratches according to claim 1, characterized in that, The formula for calculating the scratch depth index for each grid cell is as follows: ； in, The scratch depth index represents the grid cell. This represents the data set consisting of the average depth of all valid scratches within the grid cell. Indicates reference depth. This indicates taking the median.

7. The dynamic classification method for TBM tunnel surrounding rock based on cutterhead scratches according to claim 1, characterized in that, The surrounding rock convergence rate is the initial convergence rate obtained by linearly fitting the deformation-time data during the initial period after excavation, and its calculation formula is as follows: ； in, This represents the initial convergence rate of the surrounding rock. Indicates the first The time corresponding to each monitoring point This represents the amount of convergence deformation of the surrounding rock at that moment. This represents the average value across all monitoring times. This represents the average value of all monitored deformations.

8. The dynamic classification method for TBM tunnel surrounding rock based on cutterhead scratches according to claim 7, characterized in that, The creep model parameters are obtained by fitting a logarithmic creep model used to characterize the creep properties of the surrounding rock. The time span of the deformation-time data used for fitting is longer than the initial time period. The logarithmic creep model is as follows: ； in, Indicates time The cumulative convergence deformation of the surrounding rock. and Indicates the parameters of the creep model. Characterizing the amplitude of creep deformation, Characterizing the decay of creep rate over time, It represents the natural logarithm.

9. A dynamic grading system for surrounding rock in TBM tunnels based on cutterhead scratches, characterized in that, For implementing the dynamic classification method for TBM tunnel surrounding rock based on cutterhead scraping as described in any one of claims 1 to 8, the system comprises: The data acquisition module is used to simultaneously perform three-dimensional laser scanning and surface image acquisition of the tunnel wall in the area behind the shield tail after the TBM completes one ring of tunneling. It obtains point cloud data to characterize the three-dimensional geometric morphology of the cutterhead scratches and image data to characterize their surface texture features. At the same time, it acquires short-term deformation time series data of the surrounding rock and cutterhead working condition data. The data processing module is used to preprocess the point cloud data and divide it into grid cells. Based on the point cloud data and image data in each grid cell, it extracts the scratch morphology parameters of the cutterhead scratch and calculates scratch feature indices according to the scratch morphology parameters. The scratch feature indices include scratch clarity index, scratch uniformity index, and scratch depth index. The module also processes the short-term deformation time series data to extract the surrounding rock deformation response parameters, which include the surrounding rock convergence rate and creep model parameters. The feature fusion module is used to standardize and fuse the cutterhead working condition data, scratch morphology parameters, scratch feature indicators, and surrounding rock deformation response parameters to form a unified fused feature vector. The surrounding rock classification module is used to input the fused feature vector into the pre-trained surrounding rock classification model to obtain the surrounding rock level of each grid cell, and generate a spatial distribution cloud map of the surrounding rock level based on the surrounding rock levels of all grid cells.

Images