A Method for Predicting Surface Settlement During Tunnel Excavation Based on Deep Learning and Feature Transfer Learning

Abstract

This invention discloses a method for predicting surface settlement during tunnel excavation based on deep learning and feature transfer learning, relating to the field of tunnel boring machine (TBM) technology. The method includes: collecting surface settlement parameters and constructing a surface settlement database; performing data augmentation on the surface settlement database to obtain a source domain dataset; preprocessing the source domain dataset and using the preprocessed dataset to establish a surface settlement prediction model; performing domain-adaptive transfer learning on the surface settlement prediction model to obtain a surface settlement prediction model for the target project; and inputting the collected target domain dataset into the surface settlement prediction model for the target project to obtain a predicted surface settlement for the target project. The surface settlement prediction model established by this invention can address the problem of inapplicability across different projects, enabling the established prediction model to be truly applied.

Description

Technical Field

[0001] This invention relates to the field of tunnel boring machine (TBM) technology, and in particular to a method for predicting surface settlement during tunnel excavation based on deep learning and feature transfer learning. Background Technology

[0002] With the rapid development of urban infrastructure construction, underground projects such as tunnels and subways are increasing. The shield tunneling method, with its advantages of high safety and weather-independent construction, has been widely used. However, shield tunnel construction inevitably impacts the surrounding geological environment. The degree of ground deformation caused during shield tunnel excavation is an important basis for evaluating project safety. Therefore, establishing practical analytical methods to determine the ground deformation caused during shield tunnel excavation is of great significance.

[0003] The current mainstream approach is to use machine learning or deep learning methods combined with parameters such as geological parameters, shield tunneling parameters, and tunnel geometry from the actual construction process to establish a surface settlement prediction model for predicting surface settlement. However, the surface settlement prediction models established in this way are often only applicable to the projects that provide training data and cannot be used across projects, making it difficult for the established prediction models to be truly applied.

[0004] Therefore, proposing a method for predicting surface settlement during tunnel excavation based on deep learning and feature transfer learning to address the difficulties in existing technologies is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] In view of this, the present invention provides a method for predicting surface settlement during tunnel excavation based on deep learning and feature transfer learning. The surface settlement prediction model established can address the problem of not being applicable across projects, thus enabling the established prediction model to be truly applied.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A method for predicting surface settlement during tunnel excavation based on deep learning and feature transfer learning includes the following steps:

[0008] S1: Collect surface subsidence parameters and construct a surface subsidence database;

[0009] S2: Perform data augmentation on the surface subsidence database to obtain the source domain dataset;

[0010] S3: Preprocess the source domain dataset and use the preprocessed source domain dataset to establish a land subsidence prediction model;

[0011] S4: Perform domain-adaptive transfer learning on the surface subsidence prediction model to obtain the surface subsidence prediction model for the target project;

[0012] S5: Input the collected target domain dataset into the surface subsidence prediction model of the target project to obtain the surface subsidence prediction for the target project.

[0013] In the above method, the surface settlement parameters in S1 may include, but are not limited to: geological parameters, shield parameters, characteristic value combinations of tunnel geometry, and surface settlement label values.

[0014] Optionally, in the above method, Gaussian noise is used to augment the surface subsidence database in S2.

[0015] Optionally, in the above method, the specific preprocessing of the source domain dataset in S3 is as follows: The data in the source domain dataset is normalized using the following formula:

[0016]

[0017] Where x' is the normalized feature or label value, and x is the original feature or label value. min x max These are the minimum and maximum values ​​for a given feature or label.

[0018] Optionally, in the above method, the specific content of establishing a land subsidence prediction model using the preprocessed source domain dataset in S3 is as follows: the processed source domain dataset is divided into a training set, a validation set, and a test set in a ratio of 8:1:1.

[0019] The training set is used to train the LSTM model in deep learning algorithms;

[0020] The validation set is used to tune the hyperparameters of the LSTM model;

[0021] The test set is used to evaluate the model's generalization ability.

[0022] Optionally, the training set used to train the LSTM model in the deep learning algorithm includes the following steps: selecting the optimal hyperparameter combination through the results of the differential evolution algorithm and the validation set; training the LSTM model using the training set under the optimal hyperparameter combination, and evaluating the predictive performance of the LSTM model using the test set; and using the trained LSTM model as the source model for the land subsidence prediction model.

[0023] Optionally, the specific content of domain-adaptive transfer learning in S4 of the above method is as follows: mapping the source domain dataset and the target domain dataset to a high-dimensional feature space, and calculating the distance between their MMD values.

[0024]

[0025] Where X and Y represent the data distributions of the source and target domains, n and m are the number of samples in the source and target domains, φ is the mapping function that maps the samples to the high-dimensional feature space, and x i and y j Let H be the samples from the source domain and the samples from the target domain, and let H be the distance measured by mapping the data to a high-dimensional feature space using φ().

[0026] As can be seen from the above technical solution, compared with the prior art, the present invention provides a method for predicting surface settlement during tunnel excavation based on deep learning and feature transfer learning, which has the following beneficial effects: using geological parameters, shield parameters, tunnel geometry and shield attitude as feature values, and surface settlement as a label, an artificial intelligence algorithm is trained to establish a source model for surface settlement prediction. Then, through domain adaptive transfer learning, the dataset of the source model and the dataset of the target domain to be predicted are mapped to a high-dimensional feature space. The mean difference between these two distributions in the high-dimensional feature space is calculated, and the mapping that minimizes the mean difference is found, so that the distributions of the source domain and the target domain in the high-dimensional feature space are as close as possible. This solves the problem that the surface settlement prediction model is difficult to use in different projects. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0028] Figure 1 The flowchart of the surface settlement prediction method for tunnel excavation based on deep learning and feature transfer learning provided by this invention is shown below.

[0029] Figure 2 A diagram illustrating the domain-adaptive transfer learning process provided in this embodiment of the invention. Detailed Implementation

[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] Reference Figure 1 As shown, this invention discloses a method for predicting surface settlement during tunnel excavation based on deep learning and feature transfer learning, comprising the following steps:

[0032] S1: Collect surface subsidence parameters and construct a surface subsidence database;

[0033] S2: Perform data augmentation on the surface subsidence database to obtain the source domain dataset;

[0034] S3: Preprocess the source domain dataset and use the preprocessed source domain dataset to establish a land subsidence prediction model;

[0035] S4: Perform domain-adaptive transfer learning on the surface subsidence prediction model to obtain the surface subsidence prediction model for the target project;

[0036] S5: Input the collected target domain dataset into the surface subsidence prediction model of the target project to obtain the surface subsidence prediction for the target project.

[0037] Furthermore, the surface settlement parameters in S1 include, but are not limited to: geological parameters, shield tunneling parameters, characteristic value combinations of tunnel geometry, and surface settlement label values.

[0038] Specifically, geological parameters include: weighted average elastic modulus of soil, weighted internal friction angle of soil, and weighted average Poisson's ratio of soil; shield tunneling parameters include: tunneling thrust, grouting pressure, and soil chamber pressure; and tunnel geometry includes: tunnel diameter and tunnel depth.

[0039] Furthermore, Gaussian noise is used in S2 to augment the surface subsidence database.

[0040] Specifically, the process of using Gaussian noise for data augmentation is as follows:

[0041] T1 loads surface subsidence database data.

[0042] T2 defines the noise level: select an appropriate standard deviation (sigma) for Gaussian noise.

[0043] T3 generates noisy copies: For each data point, we generate multiple copies with different noise levels.

[0044] T4 Merge Data: Merges all noisy copies into a large dataset, building the source domain dataset.

[0045] Furthermore, the specific preprocessing steps for the source domain dataset in S3 are as follows: The data in the source domain dataset is normalized using the following formula:

[0046]

[0047] Where x' is the normalized feature or label value, and x is the original feature or label value. min x max These are the minimum and maximum values ​​for a given feature or label.

[0048] Specifically, the source domain dataset is preprocessed by normalizing the data, which involves scaling each feature value and label to the range of [-1, 1] according to a certain ratio. This ensures good data symmetry, avoids gradient vanishing or gradient exploding problems, and eliminates the influence of data scale on model training.

[0049] Furthermore, the specific content of S3 for establishing a land subsidence prediction model using the preprocessed source domain dataset is as follows: the processed source domain dataset is divided into a training set, a validation set, and a test set in a ratio of 8:1:1.

[0050] The training set is used to train the LSTM model in deep learning algorithms;

[0051] The validation set is used to tune the hyperparameters of the LSTM model;

[0052] The test set is used to evaluate the model's generalization ability.

[0053] Furthermore, the specific content of the training set for training the LSTM model in the deep learning algorithm is as follows: the optimal hyperparameter combination is selected through the results of the differential evolution algorithm and the validation set; under the optimal hyperparameter combination, the LSTM model is trained using the training set, and the predictive performance of the LSTM model is evaluated using the test set; the trained LSTM model is used as the source model for the land subsidence prediction model.

[0054] Specifically, the processed source domain dataset is divided into a training set, a validation set, and a test set in an 8:1:1 ratio. The training set is used to train the Long Short-Term Memory (LSTM) network model in the deep learning algorithm; the validation set is used to tune the hyperparameters of the LSTM model; and the test set is used to evaluate the model's generalization ability.

[0055] The LSTM model is trained, and the optimal hyperparameter combination is selected using the Differential Evolution (DE) algorithm and the results of the validation set. Under the optimal hyperparameter combination, the LSTM model is trained using the training set, and its predictive performance is evaluated using the test set. The trained LSTM model serves as the source model for the land subsidence prediction model. The DE optimization process for the LSTM model is as follows:

[0056] T1 defines the objective function MSE, which is used to evaluate the value of the objective function after each DE iteration. The specific formula for MSE is as follows:

[0057]

[0058] Among them, y i y represents the true value of the label. p The predicted value for the label.

[0059] T2 initialization: This involves initializing a set of candidate solutions for selecting hyperparameters. In LSTM, the candidate set of hyperparameters includes learning rate, batch size, number of hidden layer units, Dropout ratio, etc.

[0060] T3 mutation: For each candidate solution, the DE algorithm randomly selects three other different candidate solutions from the set, calculates their differences, and then adds this difference to the original candidate solution to generate a new candidate solution.

[0061] T4 crossover: The DE algorithm crossovers the mutated candidate solutions with the original candidate solutions (i.e., swaps some parameters) to generate the final experimental solution.

[0062] T5 Selection: Finally, the DE algorithm compares the experimental solution with the original candidate solution and selects the one with the smaller objective function value as the next generation candidate solution.

[0063] T6 iteration: Repeat steps T3-T5 until the set number of iterations is met.

[0064] Furthermore, the specific content of domain-adaptive transfer learning in S4 is as follows: mapping the source domain dataset and the target domain dataset to a high-dimensional feature space, and calculating the distance between their MMD values:

[0065]

[0066] Where X and Y represent the data distributions of the source and target domains, n and m are the number of samples in the source and target domains, φ is the mapping function that maps the samples to the high-dimensional feature space, and x i and y j Let H be the samples from the source domain and the samples from the target domain, and let H be the distance measured by mapping the data to a high-dimensional feature space using φ().

[0067] Specifically, the domain-adaptive transfer learning process is as follows: Figure 2 As shown, by minimizing the loss function, the distributions of the source and target domains in the high-dimensional feature space are made as close as possible, thereby improving the model's generalization performance in the target domain.

[0068] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for predicting surface settlement during tunnel excavation based on deep learning and feature transfer learning, characterized in that, Includes the following steps: S1: Collect surface subsidence parameters and construct a surface subsidence database; S2: Perform data augmentation on the surface subsidence database to obtain the source domain dataset; S3: Preprocess the source domain dataset and use the preprocessed source domain dataset to establish a land subsidence prediction model; S4: Perform domain-adaptive transfer learning on the surface subsidence prediction model to obtain the surface subsidence prediction model for the target project; S5: Input the collected target domain dataset into the surface subsidence prediction model of the target project to obtain the surface subsidence prediction for the target project; Gaussian noise was used to augment the surface subsidence database in S2. Specifically, the process of using Gaussian noise for data augmentation is as follows: T1 loads surface subsidence database data; T2 defines the noise level: selects the standard deviation of Gaussian noise, sigma; T3 generates noisy copies: For each data point, multiple copies with different noise levels are generated; T4 Merge Data: Merge all noisy copies into a single dataset to build the source domain dataset; The specific content of S3 in establishing a land subsidence prediction model using the preprocessed source domain dataset is as follows: the processed source domain dataset is divided into training set, validation set and test set in a ratio of 8:1:

1. The training set is used to train the LSTM model in deep learning algorithms; The validation set is used to tune the hyperparameters of the LSTM model; The test set is used to evaluate the model's generalization ability; The training set is used to train the LSTM model in the deep learning algorithm. The specific content is as follows: the optimal hyperparameter combination is selected through the differential evolution algorithm and the results of the validation set; under the optimal hyperparameter combination, the LSTM model is trained using the training set, and the prediction performance of the LSTM model is evaluated using the test set; the trained LSTM model is used as the source model for the land subsidence prediction model.

2. The method for predicting surface settlement during tunnel excavation based on deep learning and feature transfer learning according to claim 1, characterized in that, The surface settlement parameters in S1 include, but are not limited to: geological parameters, shield tunneling parameters, characteristic value combinations of tunnel geometry, and surface settlement label values.

3. The method for predicting surface settlement during tunnel excavation based on deep learning and feature transfer learning according to claim 1, characterized in that, The specific preprocessing steps for the source domain dataset in S3 are as follows: Normalizing the data in the source domain dataset using the following formula: Where, x ' x represents the normalized feature or label value, and x represents the original feature or label value. min x max These are the minimum and maximum values ​​for a given feature or label.

4. The method for predicting surface settlement during tunnel excavation based on deep learning and feature transfer learning according to any one of claims 1-3, characterized in that, The specific content of domain-adaptive transfer learning in S4 is as follows: mapping the source domain dataset and the target domain dataset to a high-dimensional feature space, and calculating the distance between their MMD (Match Size Distance). Where X and Y represent the data distributions of the source and target domains, n and m are the number of samples in the source and target domains, φ is the mapping function that maps the samples to the high-dimensional feature space, and x i and y j Let H be the samples from the source domain and the samples from the target domain, and let H be the distance measured by mapping the data to a high-dimensional feature space using φ().

Images