Tunnel boring machine rock mass grade classification method based on weak supervised learning

Abstract

The invention discloses a tunnel boring machine rock mass grade classification method based on weak supervised learning, and the method comprises the steps: collecting TBM multi-channel sensor time sequence data and geological record labels, and constructing a data set with noise labels; carrying out preprocessing and sequence alignment on the data; dividing a transition section and a stable section based on the geological label change points, and generating a probabilistic soft label guided by domain information; a teacher-student network architecture is constructed, feature learning and label correction are performed through a double-graph self-distillation mechanism of a semantic graph and a category graph, and model parameters are optimized in combination with prototype learning and a composite loss function; and finally, reasoning the real-time TBM data by using the trained student network, and outputting a rock mass grade classification result. According to the method, robust rock mass feature representation can be automatically learned from the engineering labels containing noise, correction and stable classification of the noise labels are achieved, and the self-adaptive tunneling capability and construction reliability of the TBM under the complex geological condition are improved.

Description

Technical Field

[0001] This invention relates to tunnel construction and intelligent geological identification technologies and methods, and in particular to a dynamic rock mass classification method based on tunnel boring machine (TBM) operation data and a weakly supervised learning framework. Background Technology

[0002] Tunnel boring machines (TBMs) are core equipment in modern underground engineering, and their tunneling efficiency and safety highly depend on the accurate identification of surrounding rock geological conditions. This invention provides an innovative solution to the key technical challenges in TBM rock mass classification. Traditional rock mass classification methods (such as RMR, Q system, BQ index, etc.) are mostly designed based on the drill-and-blast method construction background, focusing on the evaluation of rock mass self-stability. They lack adaptability to the high-speed tunneling of TBMs and the changing geological conditions, and are difficult to reflect the real-time mechanical behavior of the rock mass.

[0003] In recent years, artificial intelligence-based rock mass identification methods have gradually emerged, automatically classifying rock masses using massive amounts of sensor data (such as thrust, torque, and tunneling rate) generated during TBM operation. However, most of these methods employ a supervised learning paradigm, where training labels rely on manual geological logging or experience-based judgment. This results in problems such as strong subjectivity, poor consistency, and significant noise, hindering the reliable application of these models in practical engineering.

[0004] Furthermore, while existing unsupervised or semi-supervised methods reduce reliance on labels, their inherent mechanisms remain sensitive to label noise, resulting in limited robustness under complex geological conditions. Of particular concern is the observation from engineering practice that, although engineering-corrected rock mass classification results are relatively accurate in most excavated sections, rock mass label updates are often delayed or incorrectly assigned in discontinuous areas or sections influenced by subjective factors, leading to prominent local label noise issues. Summary of the Invention

[0005] Purpose of the invention: To address the shortcomings of existing methods, this invention proposes a rock mass classification method based on weakly supervised learning.

[0006] Technical solution: The rock mass classification method for tunnel boring machines based on weakly supervised learning described in this invention includes the following steps: (1) Data acquisition and construction: Acquire multi-channel sensor time-series data of tunnel boring machine (TBM) and corresponding geological logging rock mass grade labels, and construct a dataset with noise labels; (2) Data preprocessing: missing value filling, outlier handling, and shutdown section removal are performed on sensor data; stable tunneling sections are extracted by change point identification and sequence alignment and normalization are performed. (3) Domain information-guided soft tag generation: Based on the geological tag change points, transitional and stable sections are divided, and then probabilistic soft tags are generated; (4) Weakly supervised model training: Construct a teacher-student network architecture, perform feature learning and label correction through a dual-graph self-distillation mechanism of semantic graph and category graph, and optimize model parameters by combining prototype learning and composite loss function; (5) Real-time classification of rock mass grade: The trained student network is used to perform forward inference on the real-time collected TBM sensor data and output the rock mass grade classification results.

[0007] Further, step (2) is performed as follows: S2.1, for missing TBM runtime data, impute the missing values ​​using the mean of the data column containing the missing values; for outliers, use... The criteria combine the capping method for identification and processing; S2.2 identifies and eliminates data from the shutdown section based on parameters such as thrust, torque, and tunneling rate; S2.3, using the cumulative and change point identification method, identify and extract the time series of stable change segments of each variable within each tunneling cycle of the TBM; S2.4, using interpolation methods, the data length of each tunneling cycle segment is standardized to 1024 data points to construct a standardized TBM operation parameter dataset. T ; S2.5, use the nearest neighbor interpolation method to complete the geological exploration labels to ensure the integrity of the labels and data; S2.6 uses 0-1 standardization for continuous variables and one-hot encoding for discrete variables.

[0008] Further, step (3) is performed as follows: S3.1, Identify the dataset T Geological label change points in the data.

[0009] Traverse the preprocessed tunnel excavation dataset T Identify the locations in the tunneling cycle where the engineering label for the rock mass grade changes. Define these locations as geological label change points. } and use it as an engineering indicator signal for potential rock layer interfaces.

[0010] S3.2 defines the transition section and the stable section.

[0011] For each geological label change point Expanding forward and backward from the center L One tunneling cycle ( L Based on the hyperparameters preset according to the tunnel diameter and geological conditions, this interval is defined as the "transition section". T1. This section is a physical region where rock mass mechanical parameters may change. The dataset contains all... T The remaining portion outside of section 1 is defined as the "stable section". T 2. The rock mass properties are relatively uniform in this type of section, and the engineering labels are highly reliable.

[0012] S3.3, Confidence allocation of soft labels for partitioned segments.

[0013] Based on different assessments of the reliability of engineering tags in different sections, this invention adopts a differentiated confidence level allocation strategy: S3.3.1, Soft label generation for transition sections: Confidence allocation based on structured attenuation For the transition section T 1. This invention proposes a structured attenuation scheme based on domain knowledge to simulate the real variations in rock mass properties near geological interfaces. The core of this scheme lies in: at the point of change where label reliability is most uncertain... The lowest confidence level is assigned to the relevant dominant category; as the tunneling location moves further away from the point, its confidence level will monotonically increase according to the Gaussian weighting function to reflect the process of geological judgment certainty gradually recovering with distance.

[0014] S3.3.2, Soft label generation for stable segments: Confidence allocation based on dominant trust and uniform regularization For stable sections T 2. The allocation strategy that combines dominant trust and uniform regularization is adopted. The core of this strategy is based on the following engineering understanding: In sections with relatively uniform lithology, manually interpreted engineering labels have high reliability and should be the main source of supervision information; however, in order to enhance the model's generalization ability and suppress its overfitting to potential labeling errors, a fixed and uniformly distributed small uncertainty needs to be artificially introduced.

[0015] Further, step (4) is performed as follows: In the teacher-student architecture, the teacher network receives LSL as a supervision signal and uses LSL and feature embeddings to generate high-confidence category prototypes. The student network receives high-confidence targets from the teacher network, performs feature learning and label correction. S4.1 uses fully convolutional networks such as U-Net as the backbone network to extract multi-channel time series features, which are then processed by the feature embedding head and the classification head to extract high-dimensional features and output class probability distributions, respectively. S4.2, feature embedding and class probability are fused and iterated through a self-distillation module; The design scheme of the semantic-category dual-graph self-distillation module is as follows: S4.2.1, Semantic-Level Graph (SLG) Construction; A graph structure is constructed based on feature embedding, preserving the features of each node. k The nearest neighbor connections are used, and the affinity matrix is ​​calculated as follows: ; In the formula, Representing the eigenvector and The Euclidean distance between them S4.2.2, Class-Level Graph (CLG) Construction, A graph structure is constructed based on the category probability distribution, connecting nodes predicted to be of the same category. The edge weights are determined by the probability distribution distance, and the adjacency matrix is ​​calculated as follows: ; In the formula, p i and p j It is the first i and j The class probability of each sample, cdist( ) represents the Euclidean distance. Define the index of the most likely class for each sample. S4.2.3, Self-Distillation process. By dynamic threshold Distinguishing between high and low confidence nodes: ; In the formula, Indicates the time step t For category c Confidence level, This indicates the indicator function that selects the prediction class with the highest confidence at each time step. σ It is a fixed threshold hyperparameter. Define high and low confidence masks based on dynamic thresholds. M high and M low : ; ; Implementing SLG-CLG dual-graph collaborative optimization using an alternating update strategy: ; ; ; In the formula, It is the Gaussian error linear unit activation function. k This indicates the number of distillation iterations performed. α These are hyperparameters that control the balance before and after the update. S4.3, Design of Composite Loss Function The overall loss function of model S4.3.1 is a weighted sum of the four component losses: ; The four loss components are defined as follows: ; In the formula, n This is the number of samples included in a single batch operation. c It is the number of categories. ; The superscript "source" indicates whether the calculation belongs to the teacher model, LSL, or student model. Indicates sample i At time step l Category c The probability, It is the corresponding feature embedding. ; In the formula, The teacher model is in the first i The first sample and the first l Prediction accuracy at each time step The student model is in the first i Pseudo-labeling of teachers at individual sample locations The prediction accuracy ; In the formula, ,in Represents a set of similar sample pairs. Represents a set of samples from different classes, based on their pseudo-labels. ; In the formula, It is the first i The feature vector of each sample This indicates the corresponding prediction category. global prototype , S4.3.2, Global Prototype Updated via moving average ; In the formula, This is the update rate hyperparameter.

[0016] Further, step (5) is performed as follows: S5.1, Model Training: First, as shown in S1-S2, a dataset containing data from the complete stable tunneling phase of the TBM and corresponding geological labels is established. T Then, the data augmentation strategy shown in S3 is used to solve the problems of long-tail distribution and class imbalance of actual field data, obtain balanced data suitable for model training, input the data into the weakly supervised learning TBM rock mass classification algorithm established in S4, train the model, optimize the composite loss function until convergence. S5.2, Model Deployment: During the deployment and inference phase, only the student network branch of the model established in S4 is used. The real-time collected TBM sensor data is input into the student network branch, and the corresponding rock mass classification result can be output, supporting dynamic TBM participation in intelligent tunneling decision-making.

[0017] Compared with the prior art, the present invention has the following beneficial effects: By constructing a teacher-student architecture, introducing domain information and a dual-graph self-distillation mechanism, and combining prototype learning to establish representative feature vectors for each type of rock mass as the category origin, automatic correction and stable classification of noisy labels are achieved. This method does not rely on clean labels and can learn geologically meaningful rock mass category representations from noisy TBM operation data, making it suitable for real-time, dynamic TBM intelligent tunneling scenarios.

[0018] This invention establishes a weakly supervised learning framework guided by domain information. By introducing domain knowledge-guided soft label generation and a dual-image self-distillation mechanism, it achieves automatic and accurate classification of rock mass grades based on noisy geological labels. This effectively solves the problems of poor reliability and insufficient adaptability of rock mass identification caused by subjective, untimely, and inconsistent labeling under complex geological conditions, and improves the reliability and adaptability of TBM construction to complex geological conditions. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the workflow of the present invention; Figure 2 This is a schematic diagram of the overall teacher-student structure of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be further described below.

[0021] The rock mass classification method for tunnel boring machines based on weakly supervised learning described in this embodiment is as follows: Figure 1 As shown, the steps are as follows: S1, Dataset Construction.

[0022] Collect actual TBM tunneling data, including multi-channel time-series sensor data such as thrust, torque, and tunneling rate, as well as corresponding geological logging rock mass grade labels. The constructed dataset should contain a sufficient sample size to cover different geological conditions.

[0023] S2, Data Preprocessing.

[0024] Data preprocessing includes the following steps: S2.1, for missing TBM runtime data, impute the missing values ​​using the mean of the data column containing the missing values; for outliers, use... The criteria are combined with the capping method for identification and processing.

[0025] S2.2 identifies and eliminates data from the shutdown section based on parameters such as thrust, torque, and tunneling rate.

[0026] S2.3 uses the cumulative and change point identification method to identify and extract the time series of stable change segments of each variable within each tunneling cycle of the TBM.

[0027] S2.4, using interpolation methods, the data length of each tunneling cycle segment is standardized to 1024 data points to construct a standardized TBM operation parameter dataset. T .

[0028] S2.5, the nearest neighbor interpolation method is used to complete the geological exploration labels to ensure the integrity of the labels and data.

[0029] S2.6 uses 0-1 standardization for continuous variables and one-hot encoding for discrete variables.

[0030] S3, Leader Soft Label (LSL) generation guided by domain information.

[0031] Based on engineering practice observations, this invention proposes a method for generating guiding soft tags: S3.1, Identify the dataset T Geological label change points in the data.

[0032] Traverse the preprocessed tunnel excavation dataset T Identify the locations in the tunneling cycle where the engineering label for the rock mass grade changes. Define these locations as geological label change points. } and use it as an engineering indicator signal for potential rock layer interfaces.

[0033] S3.2 defines the transition section and the stable section.

[0034] For each geological label change point Expanding forward and backward from the center LOne tunneling cycle ( L Based on the hyperparameters preset according to the tunnel diameter and geological conditions, this interval is defined as the "transition section". T 1. This section is a physical region where rock mass mechanical parameters may change. The dataset contains all... T The remaining portion outside of section 1 is defined as the "stable section". T 2. The rock mass properties are relatively uniform in this type of section, and the engineering labels are highly reliable.

[0035] S3.3, Confidence allocation of soft labels for partitioned segments.

[0036] Based on different assessments of the reliability of engineering tags in different sections, this invention adopts a differentiated confidence level allocation strategy: S3.3.1, Soft label generation for transition sections: Confidence allocation based on structured attenuation For the transition section T 1. This invention proposes a structured attenuation scheme based on domain knowledge to simulate the real variations in rock mass properties near geological interfaces. The core of this scheme lies in: at the point of change where label reliability is most uncertain... The lowest confidence level is assigned to the relevant dominant category; as the tunneling location moves further away from the point, its confidence level will monotonically increase according to the Gaussian weighting function to reflect the process of geological judgment certainty gradually recovering with distance.

[0037] The specific implementation steps are as follows: 1. At the change point The location defines the emerging dominant category. With the original dominant category that is about to be withdrawn .

[0038] 2. For emerging categories Its confidence level increases from the point of change in the positive tunneling direction. Specifically, at the location... ( The confidence level at point () is given by the following Gaussian weighting function: ; in, Representative at the point of change Minimum basic confidence level at; This represents the peak confidence level reached after moving far from the point of change; The scaling parameter used to control the confidence growth rate.

[0039] 3. Regarding exit categories Its confidence level increases from the point of change towards the negative tunneling direction, and its functional form is symmetrical to the above: ; 4. At the change point There is This reflects the most ambiguous understanding of the rock mass properties and the least reliable engineering labels at this location.

[0040] 5. To ensure the probabilistic interpretation of soft labels, in the transition section... T For any point within 1, it is stipulated that The above function design naturally satisfies this constraint through its symmetric form. The confidence levels for all other non-dominant categories are set to 0 within this segment.

[0041] S3.3.2, Soft label generation for stable segments: Confidence allocation based on dominant trust and uniform regularization For stable sections T 2. This invention employs an allocation strategy that combines dominant trust with uniform regularization. This strategy is based on the following engineering understanding: in sections with relatively homogeneous lithology, manually interpreted engineering labels have high reliability and should be the primary source of supervisory information; however, to enhance the model's generalization ability and suppress its overfitting to potential labeling errors, a fixed, uniformly distributed small uncertainty needs to be artificially introduced.

[0042] The specific implementation steps are as follows: 1. For T For any tunneling cycle sample within section 2, the category corresponding to its original engineering hard label is determined as the dominant category for that sample. .

[0043] 2. As the dominant category Assign a higher base confidence level This demonstrates a basic level of trust in the engineering label. Its confidence level is: ; In a preferred embodiment, take .

[0044] Use the remaining confidence level The residual uncertainty is treated as a uniformly distributed uncertainty and distributed equally among the rest. Each non-dominant category. The obtained confidence level is: (For example At that time, each got ); 3. Finally, the soft label of this sample satisfy: Furthermore, the dominant category component is significantly higher than the other components.

[0045] S4, Model Architecture Design.

[0046] See Figure 2 This invention employs a teacher-student learning paradigm, comprising a teacher network and a student network that share the same feature extraction backbone network. In this teacher-student architecture, the teacher network receives LSL (Leadership Level Selection) as a supervision signal and utilizes LSL and feature embeddings to generate high-confidence category prototypes. The student network receives high-confidence targets from the teacher network and performs feature learning and label correction.

[0047] S4.1 uses fully convolutional networks such as U-Net as the backbone network to extract multi-channel time series features, which are then processed by the feature embedding head and classification head to extract high-dimensional features and output class probability distributions, respectively.

[0048] S4.2, feature embeddings and class probabilities are fused iteratively through a self-distillation module.

[0049] See Figure 2 This invention proposes a semantic-category dual-graph self-distillation module design scheme: S4.2.1 Semantic-Level Graph (SLG) Construction.

[0050] A graph structure is constructed based on feature embedding, preserving the features of each node. k The nearest neighbor connections are used, and the affinity matrix is ​​calculated as follows: ; In the formula, Representing the eigenvector and The Euclidean distance between them.

[0051] S4.2.2, Class-Level Graph (CLG) Construction.

[0052] A graph structure is constructed based on the category probability distribution, connecting nodes predicted to be of the same category. The edge weights are determined by the probability distribution distance, and the adjacency matrix is ​​calculated as follows: ; In the formula, p i and p j It is the first i and j The class probability of each sample, cdist( ) represents the Euclidean distance. Define the index of the most likely category for each sample.

[0053] S4.2.3, Self-Distillation process.

[0054] By dynamic threshold Distinguishing between high and low confidence nodes: ; In the formula, Indicates the time step t For category c Confidence level, This indicates the indicator function that selects the prediction class with the highest confidence at each time step. σ It is a fixed threshold hyperparameter.

[0055] Define high and low confidence masks based on dynamic thresholds. M high and M low : ; ; Implementing SLG-CLG dual-graph collaborative optimization using an alternating update strategy: ; ; ; In the formula, It is the Gaussian error linear unit activation function. k This indicates the number of distillation iterations performed. α It is a hyperparameter that controls the balance before and after the update.

[0056] S4.3, Design of composite loss function.

[0057] The overall loss function of model S4.3.1 is a weighted sum of the four component losses: ; The four loss components are defined as follows: ; In the formula, n This is the number of samples included in a single batch operation. c It is the number of categories. ; The superscript "source" indicates whether the calculation belongs to the teacher model, LSL, or student model. Indicates sample i At time step l Category c The probability, This is the corresponding feature embedding.

[0058] ; In the formula, The teacher model is in the first i The first sample and the first l Prediction accuracy at each time step The student model is in the first i Pseudo-labeling of teachers at individual sample locations The accuracy of predictions.

[0059] ; In the formula, ,in Represents a set of similar sample pairs. This represents a set of different class sample pairs, based on their pseudo-labels.

[0060] ; In the formula, It is the first i The feature vector of each sample This indicates the corresponding prediction category. global prototype .

[0061] S4.3.2, Global Prototype Updated using a moving average.

[0062] ; In the formula, This is the update rate hyperparameter.

[0063] S5, Training and Deployment of Weakly Supervised Learning TBM Rock Mass Grading Classification Model.

[0064] S5.1, Model Training. During model training in this invention, a dataset containing data from the complete stable tunneling phase of a TBM and corresponding geological labels is first established as shown in S1-S2. T Then, the data augmentation strategy shown in S3 is used to address the long-tail distribution and class imbalance issues of the actual field data, obtaining balanced data suitable for model training. The data is then input into the weakly supervised learning TBM rock mass classification algorithm established in S4 for model training, optimizing the composite loss function until convergence.

[0065] S5.2, Model Deployment. During the deployment inference phase, only the student network branch of the model established in S4 is used. Real-time collected TBM sensor data is input into the student network branch, which outputs the corresponding rock mass classification results, supporting dynamic TBM participation in intelligent tunneling decision-making.

[0066] This invention establishes a weakly supervised learning framework guided by domain information. By introducing domain knowledge-guided soft label generation and a dual-image self-distillation mechanism, it achieves automatic and accurate classification of rock mass grades based on noisy geological labels. This effectively solves the problems of poor reliability and insufficient adaptability of rock mass identification caused by subjective, untimely, and inconsistent labeling under complex geological conditions, and improves the reliability and adaptability of TBM construction to complex geological conditions.

[0067] The above are merely preferred embodiments of the present invention and do not constitute any limitation on the present invention. Any equivalent substitutions or modifications made by those skilled in the art to the technical solutions and content disclosed in the present invention without departing from the scope of the present invention shall be deemed to have remained within the protection scope of the present invention.

Claims

1. A rock mass classification method for tunnel boring machines based on weakly supervised learning, characterized in that, Includes the following steps: (1) Data acquisition and construction: acquire multi-channel sensor time-series data of tunnel boring machine (TBM) and corresponding geological logging rock mass grade labels, and construct a dataset with noise labels; (2) Data preprocessing: Filling in missing values, handling outliers, and removing downtime segments from the sensor data; Stable tunneling sections are extracted using change point identification, followed by sequence alignment and normalization. (3) Domain information-guided soft tag generation: Based on the geological tag change points, transitional and stable sections are divided, and then probabilistic soft tags are generated; (4) Weakly supervised model training: Construct a teacher-student network architecture, perform feature learning and label correction through a dual-graph self-distillation mechanism of semantic graph and category graph, and optimize model parameters by combining prototype learning and composite loss function; (5) Real-time classification of rock mass grade: The trained student network is used to perform forward inference on the real-time collected TBM sensor data and output the rock mass grade classification results.

2. The method for classifying rock mass grades for tunnel boring machines based on weakly supervised learning according to claim 1, characterized in that, The method for step (2) is as follows: S2.1, for missing TBM runtime data, impute the missing values ​​using the mean of the data column containing the missing values; for outliers, use... The criteria combine the capping method for identification and processing; S2.2 identifies and eliminates data from the shutdown section based on parameters such as thrust, torque, and tunneling rate; S2.3, using the cumulative and change point identification method, identify and extract the time series of stable change segments of each variable within each tunneling cycle of the TBM; S2.4, using interpolation methods, the data length of each tunneling cycle segment is standardized to 1024 data points to construct a standardized TBM operation parameter dataset. T ; S2.5, use the nearest neighbor interpolation method to complete the geological exploration labels to ensure the integrity of the labels and data; S2.6 uses 0-1 standardization for continuous variables and one-hot encoding for discrete variables.

3. The method for classifying rock mass grades for tunnel boring machines based on weakly supervised learning according to claim 1, characterized in that, The method for step (3) is as follows: S3.1, Identify the dataset T Geological label change points in: Traverse the preprocessed tunnel excavation dataset T Identify the tunneling cycle locations where the rock mass grade engineering label changes, and define these locations as geological label change points. }, and use it as an engineering indicator signal for potential rock layer interfaces, S3.2, Define the transition section and the stable section: For each geological label change point Expanding forward and backward from the center L One tunneling cycle, L Based on the preset hyperparameters for tunnel diameter and geological conditions, this section is defined as a transition zone. T 1. This section is a physical region where rock mass mechanical parameters may change. The dataset contains all... T The remaining portion outside of section 1 is defined as a stable section. T 2. The rock mass properties within this type of section are relatively homogeneous, making the engineering labels highly reliable. S3.3, Confidence allocation of soft labels for different partitions: Based on different assessments of the reliability of engineering labels in different sections, a differentiated confidence level allocation strategy is adopted: S3.3.1, Soft label generation for transition sections: Confidence allocation based on structured attenuation: For the transition section T 1. A structured attenuation scheme based on domain knowledge to simulate the real variation of rock mass properties near geological interfaces; S3.3.2, Soft label generation for stable segments: Confidence allocation based on dominant trust and uniform regularization: For stable sections T 2. An allocation strategy combining dominant trust and uniform regularization is adopted.

4. The method for classifying rock mass grades for tunnel boring machines based on weakly supervised learning according to claim 1, characterized in that, The method for step (4) is as follows: In the teacher-student architecture, the teacher network receives LSL as a supervision signal and uses LSL and feature embeddings to generate high-confidence category prototypes. The student network receives high-confidence targets from the teacher network, performs feature learning and label correction. S4.1 uses fully convolutional networks such as U-Net as the backbone network to extract multi-channel time series features, which are then processed by the feature embedding head and the classification head to extract high-dimensional features and output class probability distributions, respectively. S4.2, feature embedding and class probability are fused and iterated through a self-distillation module; The design scheme of the semantic-category dual-graph self-distillation module is as follows: S4.2.1, Semantic-Level Graph (SLG) Construction; A graph structure is constructed based on feature embedding, preserving the features of each node. k The nearest neighbor connections are calculated, and the affinity matrix is ​​as follows: ； In the formula, Representing the eigenvector and The Euclidean distance between them S4.2.2, Class-Level Graph (CLG) Construction, A graph structure is constructed based on the category probability distribution, connecting nodes predicted to be of the same category. The edge weights are determined by the probability distribution distance, and the adjacency matrix is ​​calculated as follows: ； In the formula, p i and p j It is the first i and j The class probability of each sample, cdist( ) represents the Euclidean distance. Define the index of the most likely class for each sample. S4.2.3, Self-Distillation process. By dynamic threshold Distinguishing between high and low confidence nodes: ； In the formula, Indicates the time step t For category c Confidence level, This indicates the indicator function that selects the prediction class with the highest confidence at each time step. σ It is a fixed threshold hyperparameter. Define high and low confidence masks based on dynamic thresholds. M high and M low : ； ； Implementing SLG-CLG dual-graph collaborative optimization using an alternating update strategy: ； ； ； In the formula, It is the Gaussian error linear unit activation function. k This indicates the number of distillation iterations performed. α These are hyperparameters that control the balance before and after the update. S4.3, Design of Composite Loss Function The overall loss function of model S4.3.1 is a weighted sum of the four component losses: ； The four loss components are defined as follows: ； In the formula, n This is the number of samples included in a single batch operation. c It is the number of categories. ; The superscript "source" indicates whether the calculation belongs to the teacher model, LSL, or student model. Indicates sample i At time step l Category c The probability, It is the corresponding feature embedding. ； In the formula, The teacher model is in the first i The first sample and the first l Prediction accuracy at each time step The student model is in the first i Pseudo-labeling of teachers at individual sample locations The prediction accuracy ； In the formula, ,in Represents a set of similar sample pairs. Represents a set of samples from different classes, based on their pseudo-labels. ； In the formula, It is the first i The feature vector of each sample This indicates the corresponding prediction category. global prototype , S4.3.2, Global Prototype Updated via moving average ； In the formula, This is the update rate hyperparameter.

5. The method for classifying rock mass grades for tunnel boring machines based on weakly supervised learning according to claim 1, characterized in that, The method for step (5) is as follows: S5.1, Model Training: First, as shown in S1-S2, a dataset containing data and corresponding geological labels for the complete stable tunneling phase of the TBM is established. T Then, the data augmentation strategy shown in S3 is used to solve the problems of long-tail distribution and class imbalance of actual field data, obtain balanced data suitable for model training, input the data into the weakly supervised learning TBM rock mass classification algorithm established in S4, train the model, optimize the composite loss function until convergence. S5.2, Model Deployment: During the deployment and inference phase, only the student network branch of the model established in S4 is used. The real-time collected TBM sensor data is input into the student network branch, and the corresponding rock mass classification result can be output, supporting dynamic TBM participation in intelligent tunneling decision-making.

Images