Search

CN-122020397-A - Tunnel boring machine rock mass grade classification method based on weak supervision learning

CN122020397ACN 122020397 ACN122020397 ACN 122020397ACN-122020397-A

Abstract

The invention discloses a tunneling machine rock mass grade classification method based on weak supervision learning, which comprises the steps of collecting time sequence data and geological record labels of a TBM multichannel sensor, constructing a label data set with noise, preprocessing and aligning the data, dividing a transition section and a stable section based on geological label change points, regenerating a probability soft label guided by field information, constructing a teacher-student network architecture, carrying out feature learning and label correction through a double-graph self-distillation mechanism of a semantic graph and a class graph, optimizing model parameters by combining prototype learning and a composite loss function, and finally reasoning real-time TBM data by using a trained student network, and outputting a rock mass grade classification result. The method can automatically learn the robust rock mass characteristic representation from the noisy engineering label, realize the correction and stable classification of the noisy label, and improve the self-adaptive tunneling capability and construction reliability of the TBM under complex geological conditions.

Inventors

  • ZHU MENGQI
  • ZHU HEHUA
  • RUI YI

Assignees

  • 同济大学

Dates

Publication Date
20260512
Application Date
20251225

Claims (5)

  1. 1. The rock mass grade classification method of the tunnel boring machine based on weak supervision learning is characterized by comprising the following steps of: (1) Acquiring time sequence data of a multichannel sensor of a tunnel boring machine TBM and corresponding geological logging rock mass grade labels, and constructing a data set with noise labels; (2) The data preprocessing comprises the steps of filling a missing value, processing an abnormal value, eliminating a shutdown section of sensor data, extracting a stable tunneling section by adopting variable point identification, and performing sequence alignment and normalization processing; (3) The soft label generation guided by the field information comprises the steps of dividing a transition section and a stable section based on geological label change points, and further generating a probability soft label; (4) Constructing a teacher-student network architecture, performing feature learning and label correction through a double-graph self-distillation mechanism of a semantic graph and a class graph, and optimizing model parameters by combining prototype learning and a composite loss function; (5) And (3) classifying rock mass grades in real time, namely performing forward reasoning on TBM sensor data acquired in real time by using a trained student network, and outputting rock mass grade classification results.
  2. 2. The method for classifying rock mass grades of a tunnel boring machine based on weak supervised learning as set forth in claim 1, wherein the step (2) is as follows: S2.1, filling the missing TBM operation data by using the average value of the data column where the missing value is located, and filling abnormal values by using The criterion is combined with a capping method to carry out identification and treatment; s2.2, identifying and eliminating shutdown segment data based on parameters such as thrust, torque, tunneling speed and the like; S2.3, identifying and extracting time sequences of stable variable change segments of each TBM in each tunneling cycle by using an accumulation and change point identification method; S2.4, unifying the data length of each tunneling cycle segment into 1024 data points by using an interpolation method, and constructing a standard TBM operation parameter data set T; S2.5, complementing the geological survey label by adopting a nearest neighbor interpolation method, and ensuring the integrity of the label and the data; S2.6, normalizing the continuous variable by using 0-1, and performing one-heat coding on the discrete variable.
  3. 3. The method for classifying rock mass grades of a tunnel boring machine based on weak supervised learning as set forth in claim 1, wherein the step (3) is as follows: s3.1, identifying geological tag change points in the data set T: Traversing the preprocessed tunneling data set T, identifying tunneling circulation positions in which the rock mass grade engineering labels change, and defining the positions as geological label change points { And uses it as engineering indication signal of potential rock stratum interface, S3.2, defining a transition section and a stable section: with each geological tag change point L tunneling cycles are extended forwards and backwards by taking the center as the center, L is a super parameter preset according to the tunnel diameter and geological conditions, the interval is defined as a transition section T 1 which is a physical area where the mechanical parameters of the rock mass can change, the rest part except all the sections T 1 in the data set is defined as a stable section T 2 , the rock mass property in the sections is relatively uniform, the engineering label reliability is higher, S3.3, partitioning soft label confidence assignment: based on different evaluations of the reliability of engineering labels in different sections, a differentiated confidence allocation strategy is adopted: s3.3.1, transition segment soft label generation: confidence assignment based on structured decay: For the transition section T 1 , a structured attenuation scheme based on domain knowledge is used for simulating the real variation condition of the rock mass attribute near the geological interface; S3.3.2, stable segment soft label generation, confidence assignment based on dominant trust and uniform regularization: For the stable section T 2 , an allocation strategy that combines dominant trust with uniform regularization is employed.
  4. 4. The method for classifying rock mass grades of a tunnel boring machine based on weak supervised learning as set forth in claim 1, wherein the step (4) is as follows: in a teacher-student architecture, a teacher network receives LSL as a supervision signal, and generates a category prototype with high confidence by utilizing LSL and feature embedding The student network receives the high confidence target from the teacher network, performs feature learning and label correction, S4.1, using a full convolution network such as U-Net as a backbone network, extracting multi-channel time sequence features, and then processing by a feature embedding head and a classification head to respectively extract high-dimensional features and output class probability distribution; s4.2, feature embedding and class probability are fused and iterated through a self-distillation module; The semantic-category double-graph self-distillation module design scheme is as follows: s4.2.1, semantic graph (Semantic-LEVEL GRAPH, SLG) construction; based on the feature embedding construction graph structure, k nearest neighbor connections of each node are reserved, and an affinity matrix is calculated as follows: ; In the formula, Representing feature vectors And The euclidean distance between the two, S4.2.2, class diagram (Class-LEVEL GRAPH, CLG) construction, Based on the class probability distribution, constructing a graph structure, connecting nodes predicted to be in the same class, determining edge weights by probability distribution distances, and calculating an adjacency matrix as follows: ; Wherein, p i and p j are class probabilities of the ith and j samples, cdist% ) The euclidean distance is represented as, Defining the most probable class index for each sample, S4.2.3, a Self-distillation (Self-Distillation) process, By dynamic threshold Distinguishing high and low confidence nodes: ; In the formula, Indicating the confidence in category c at time step t, An indication Fu Hanshu indicating that the prediction class with the highest confidence is selected at each time step, σ is a fixed threshold hyper-parameter, Based on the dynamic threshold, high and low confidence masks M high and M low are defined: ; ; the SLG-CLG double-graph collaborative optimization is realized by adopting an alternate updating strategy: ; ; ; In the formula, Is a Gaussian error linear unit activation function, k represents the number of distillation iterations passed, alpha is a super parameter controlling the balance before and after updating, S4.3, designing a composite loss function, The overall loss function of the S4.3.1 model is a weighted sum of four quantile losses: ; wherein the four loss terms are defined as follows: ; Where n is the number of samples included in a batch operation, c is the number of categories, ; Wherein the superscript "source" indicates whether the calculation belongs to a teacher model, an LSL, or a student model, Representing the probability that sample i belongs to category c at time step l, Is the embedding of the corresponding feature, ; In the formula, Is the predicted accuracy of the teacher model at the ith sample and the ith time step, Is pseudo-label of student model to teacher at ith sample Is used for the prediction accuracy of the (c) in the (c), ; In the formula, Wherein Representing a collection of pairs of homogeneous samples, Representing a collection of different classes of sample pairs, based on their pseudo tags, ; In the formula, Is the eigenvector of the i-th sample, Representation corresponds to prediction category Global prototypes of (a) , S4.3.2 Global prototype By means of the updating of the moving average, ; In the formula, Is an update rate super parameter.
  5. 5. The method for classifying rock mass grades of a tunnel boring machine based on weak supervised learning as set forth in claim 1, wherein the step (5) is as follows: S5.1, model training, namely firstly establishing a data set T containing data of a complete TBM stable tunneling stage and corresponding geological labels as shown in S1-S2, then adopting a data enhancement strategy as shown in S3 to solve the problem of unbalance between long tail distribution and classes of actual field data, acquiring balanced data suitable for model training, inputting the data into a weak supervision learning TBM rock mass class classification algorithm established in S4, performing model training, and optimizing a composite loss function until convergence; And S5.2, model deployment, namely in a deployment reasoning stage, only using the student network branch of the model established in the step S4 to input TBM sensor data acquired in real time into the student network branch, and outputting a corresponding rock mass grade classification result to support TBM dynamic adjustment to participate in intelligent tunneling decision.

Description

Tunnel boring machine rock mass grade classification method based on weak supervision learning Technical Field The invention relates to a method for dynamically classifying rock mass grades based on Tunnel Boring Machine (TBM) operation data and a weak supervision learning framework, belonging to the technical methods of tunnel construction and intelligent geological recognition. Background Tunnel Boring Machines (TBMs) are core equipment in modern underground engineering, and the boring efficiency and safety of which are highly dependent on accurate identification of surrounding rock geological conditions. The invention provides an innovative solution to the key technical problem in TBM rock mass grade classification. The traditional rock mass classification method (such as RMR, Q system, BQ index and the like) is mostly based on the construction background design of the drilling and blasting method, focuses on the self-stability evaluation of the rock mass, has insufficient adaptability under the condition of variable geological conditions when TBM is tunneled at high speed, and is difficult to reflect the real rock mass mechanical behavior in real time. In recent years, a rock mass identification method based on artificial intelligence is gradually rising, and automatic classification is performed through massive sensor data (such as thrust, torque, tunneling speed and the like) generated during TBM operation. However, most of the methods adopt a supervised learning paradigm, the training labels of the methods depend on manual geological logging or experience judgment, and the problems of strong subjectivity, poor consistency, obvious noise and the like exist, so that the reliable application of the model in actual engineering is restricted. In addition, although the existing unsupervised or semi-supervised method reduces the dependence on the label, the inherent mechanism of the existing unsupervised or semi-supervised method is still sensitive to label noise, and the robustness under complex geological conditions is limited. Of particular concern, it has been found from engineering practice that despite the relatively accurate classification of the engineering corrected rock mass in most excavated sections, there is often delay or misdistribution of rock mass label updates in discrete areas or zones affected by subjective factors, resulting in significant local label noise problems. Disclosure of Invention Aiming at the defects existing in the prior art, the invention provides a rock mass grade classification method based on weak supervision learning. The technical scheme is that the rock mass grade classification method of the tunnel boring machine based on weak supervision learning comprises the following steps: (1) Acquiring time sequence data of a multichannel sensor of a tunnel boring machine TBM and corresponding geological logging rock mass grade labels, and constructing a data set with noise labels; (2) The data preprocessing comprises the steps of filling a missing value, processing an abnormal value, eliminating a shutdown section of sensor data, extracting a stable tunneling section by adopting variable point identification, and performing sequence alignment and normalization processing; (3) The soft label generation guided by the field information comprises the steps of dividing a transition section and a stable section based on geological label change points, and further generating a probability soft label; (4) Constructing a teacher-student network architecture, performing feature learning and label correction through a double-graph self-distillation mechanism of a semantic graph and a class graph, and optimizing model parameters by combining prototype learning and a composite loss function; (5) And (3) classifying rock mass grades in real time, namely performing forward reasoning on TBM sensor data acquired in real time by using a trained student network, and outputting rock mass grade classification results. Further, the method in the step (2) is as follows: S2.1, filling the missing TBM operation data by using the average value of the data column where the missing value is located, and filling abnormal values by using The criterion is combined with a capping method to carry out identification and treatment; s2.2, identifying and eliminating shutdown segment data based on parameters such as thrust, torque, tunneling speed and the like; S2.3, identifying and extracting time sequences of stable variable change segments of each TBM in each tunneling cycle by using an accumulation and change point identification method; S2.4, unifying the data length of each tunneling cycle segment into 1024 data points by using an interpolation method, and constructing a standard TBM operation parameter data set T; S2.5, complementing the geological survey label by adopting a nearest neighbor interpolation method, and ensuring the integrity of the label and the data; S2.6, normalizing the continuous variab