CN-121167476-B - Shield tunnel surrounding rock classification method based on multitask learning and countermeasure training

CN121167476BCN 121167476 BCN121167476 BCN 121167476BCN-121167476-B

Abstract

The invention relates to a shield tunnel surrounding rock classification method based on multi-task learning and countermeasure training, which comprises the steps of collecting different tunnel engineering data, carrying out signal decomposition and structuring treatment, constructing a database, constructing a transition learning model by utilizing a multi-task learning frame and a countermeasure training mechanism, carrying out training and verification on the transition learning model based on the database, obtaining target tunnel engineering data, inputting the target tunnel engineering data into the final transition learning model, and outputting predicted surrounding rock types and corresponding corrected BQ values, wherein the surrounding rock categories and corresponding corrected BQ values of different tunnel segments and the time sequence data of shield machine operation parameters recorded by a shield machine PLC system during construction of each ring pipe segment. According to the invention, the common characteristics among different engineering data are extracted through the migration learning model framework combining the multi-task learning framework and the countermeasure training, so that the generalization capability of the model is improved, and the problem of surrounding rock classification in a scarce scene of target engineering data is solved.

Inventors

FANG QIAN
MENG YUXIANG
HE WEIGUO
GUO JING
ZHENG GUOLI
LI QIMING

Assignees

中铁第六勘察设计院集团有限公司
北京交通大学

Dates

Publication Date: 20260508
Application Date: 20250909

Claims (5)

1. The shield tunnel surrounding rock classification method based on multitask learning and countermeasure training is characterized by comprising the following steps: Collecting different tunnel engineering data, carrying out signal decomposition and structuring treatment, and constructing a database, wherein the data comprise surrounding rock types at different pipe pieces of a tunnel, corresponding correction BQ values and time sequence data of shield machine operation parameters recorded by a shield machine PLC system during construction of each ring pipe piece; Constructing a transfer learning model by utilizing a multi-task learning framework and an countermeasure training mechanism, training and verifying the transfer learning model based on the database to obtain a final transfer learning model, wherein the transfer learning model comprises a coding layer, an attention layer, a multi-classification layer, a regression layer, a gradient inversion layer and a data reconstruction layer, input data is firstly converted into deep features through the coding layer, the deep features are converted into weighted features through the attention layer, the weighted features are then flowed to three different tasks, the regression layer executes a first task, the multi-classification layer executes a second task, and the gradient inversion layer and the data reconstruction layer execute a third task: the first task is a regression task, the regression layer is adopted to predict and correct BQ values according to the weighting characteristics, and surrounding rock categories are divided according to the corrected BQ value range to serve as prediction results, wherein the regression layer adopts a full-connection layer and a RELU function is adopted as an activation function; the second task is a multi-classification task, the probability of belonging to different surrounding rock categories is predicted by adopting the multi-classification layer according to the weighting characteristics, wherein the multi-classification layer adopts a full-connection layer and adopts an RELU function as an activation function; the third task is a data reconstruction task, the weighted feature firstly flows through the gradient inversion layer, then the data reconstruction layer is input, and reconstruction data of the input data is output, wherein the data reconstruction layer adopts a full-connection layer and adopts RELU functions as activation functions; And acquiring target tunnel engineering data, inputting the target tunnel engineering data into the final transfer learning model, and outputting predicted surrounding rock types and corresponding corrected BQ values.
2. The method for classifying surrounding rock of a shield tunnel based on multitasking learning and countermeasure training as recited in claim 1, wherein performing the signal decomposition comprises: Performing outlier removal and normalization processing on the time sequence data of the shield machine operation parameters in the construction of each ring pipe slice to obtain effective time sequence data; carrying out signal decomposition on the effective time sequence data by adopting a VMD-DFA method to obtain a plurality of sub-signals, and calculating the scale index of each sub-signal; Dividing the range of the scale index into a trend signal, a fluctuation signal and a noise signal based on the range of the scale index, and eliminating the noise signal.
3. The method for classifying surrounding rocks of a shield tunnel based on multitasking learning and countermeasure training as claimed in claim 2, wherein performing the structuring process includes: And uniformly measuring all trend signals and fluctuation signals, and constructing a structured database of different projects by taking each ring as a data sample, wherein the data sample comprises a representative value of shield machine operation data, surrounding rock types of positions of the rings and corresponding correction BQ values in each ring pipe piece construction, the representative value of the trend signals is represented by an average value, and the representative value of the fluctuation signals is represented by a variation coefficient.
4. The method for classifying surrounding rocks of a shield tunnel based on multi-task learning and countermeasure training according to claim 1, wherein the coding layer adopts a multi-layer fully-connected neural network and adopts RELU functions as activation functions.
5. The shield tunnel surrounding rock classification method based on multi-task learning and countermeasure training according to claim 1, wherein the attention layer is composed of a dense layer, a Softmax layer and a weighting layer, wherein the dense layer is used for calculating attention scores of deep features and adopting RELU functions as activation functions, the Softmax layer is used for normalizing the attention scores, and the weighting layer is used for carrying out weighted evaluation on the deep features based on the normalized attention scores and outputting weighted features.

Description

Shield tunnel surrounding rock classification method based on multitask learning and countermeasure training Technical Field The invention relates to the technical field of shield tunnel construction, in particular to a shield tunnel surrounding rock classification method based on multi-task learning and countermeasure training. Background The shield machine is widely used in tunnel construction, and the operation parameters of the shield machine are required to be matched with geological conditions, so that a driver of the shield machine is required to adjust the operation parameters of the shield machine in real time according to the geological conditions. In the prior art, a plurality of machine learning models, such as Random Forest (RF), convolutional Neural Network (CNN), deep Neural Network (DNN) and the like, have been used for classifying surrounding rock in real time, but the models have the problems that 1, the generalization ability of the models is poor, the training data of the classification models come from a certain designated project, and therefore, the optimal parameter combination of the models and the learned characteristics have strong correlation with the distribution characteristics of the training data. When the model is applied to another engineer, the data distribution characteristics of the target data set and the training data set are deviated due to the difference of the shield tunneling machine configuration and the geological conditions, so that the model has poor performance, and 2, the retraining cost is high, because the conventional model has poor generalization capability, different projects often need to train a single model, and training one model often needs to consume a large amount of computing resources and needs a large amount of data as a support. When the data volume of the new project is sparse in the initial construction stage, the model cannot be effectively trained, so that the model cannot guide the construction of the project in the initial construction stage. In the prior art, although the precision is attempted to be improved through a physical model or a hybrid network, the problem of knowledge migration across projects and across shield machines is not solved. Therefore, there is a need for a classification method of surrounding rock of a shield tunnel, which can efficiently utilize historical data and adapt to different engineering scenes. Disclosure of Invention The invention aims to provide a shield tunnel surrounding rock classification method based on multi-task learning and countermeasure training, common characteristics among different engineering data are extracted through a transition learning model framework combining a multi-task learning framework and the countermeasure training, so that model generalization capability is improved, and the surrounding rock classification problem under a scarce scene of target engineering data is solved. In order to achieve the above object, the present invention provides the following solutions: a shield tunnel surrounding rock classification method based on multitask learning and countermeasure training comprises the following steps: Collecting different tunnel engineering data, carrying out signal decomposition and structuring treatment, and constructing a database, wherein the data comprise surrounding rock types at different pipe pieces of a tunnel, corresponding correction BQ values and time sequence data of shield machine operation parameters recorded by a shield machine PLC system during construction of each ring pipe piece; Constructing a transfer learning model by utilizing a multi-task learning framework and an countermeasure training mechanism, and training and verifying the transfer learning model based on the database to obtain a final transfer learning model; And acquiring target tunnel engineering data, inputting the target tunnel engineering data into the final transfer learning model, and outputting predicted surrounding rock types and corresponding corrected BQ values. Optionally, performing the signal decomposition includes: Performing outlier removal and normalization processing on the time sequence data of the shield machine operation parameters in the construction of each ring pipe slice to obtain effective time sequence data; carrying out signal decomposition on the effective time sequence data by adopting a VMD-DFA method to obtain a plurality of sub-signals, and calculating the scale index of each sub-signal; Dividing the range of the scale index into a trend signal, a fluctuation signal and a noise signal based on the range of the scale index, and eliminating the noise signal. Optionally, performing the structuring process, building the database includes: And uniformly measuring all trend signals and fluctuation signals, and constructing a structured database of different projects by taking each ring as a data sample, wherein the data sample comprises a representative value of shie