Prediction Method for the Distribution of Media behind the Shield Tunnel Wall Based on Ground Penetrating Radar and GAN

Abstract

The present invention is a method for predicting the distribution of the medium behind the shield tunnel wall based on ground penetrating radar and GAN. Install and use ground penetrating radar in the shield tunnel to collect signals and images; conduct a model test on the post-grouting of the ground penetrating radar, collect the ground penetrating radar images of the shield tunnel with known grouting thickness, and construct the corresponding distribution map of the medium behind the wall; after preprocessing the data, obtain a data sample set; establish a prediction model for the distribution of the medium behind the shield tunnel wall based on the generative adversarial network GAN of the ground penetrating radar image; optimize the calculation parameters of the generative adversarial network GAN; the prediction model of the generative adversarial network GAN after optimizing the parameters realizes the detection of the post-grouting behind the tunnel wall. Use the data set of the preprocessed ground penetrating radar time-domain signal to construct a generative adversarial network. Finally, predict and identify the ground penetrating radar images collected by the ground penetrating radar for the post-grouting behind the shield tunnel wall.

Description

Technical Field

[0001] The present invention relates to a non-destructive detection method for tunnels. Technical Background

[0002] The existing detection of post-grouting behind the shield tunnel wall in China mainly relies on ground penetrating radar detection. The thickness and density uniformity of the grouting are detected by a ground penetrating radar device. However, the image of the ground penetrating radar is not a direct image of the underground structure. Its returned signal is an image formed by the signal of electromagnetic waves, which needs to be analyzed and explained. The inspectors conduct inspections at the tunnel site, record and identify the ground penetrating radar images of the post-grouting behind the shield tunnel wall. This method not only has a slow detection speed but also has a strong dependence on the experience of the inspectors and has great subjectivity.

[0003] Note: Generative Adversarial Networks (GAN) is a deep learning model that can generate images through confrontation.

[0004] Compared with the existing technology, if it is possible to generate an image of the medium distribution behind the wall using the ground penetrating radar image of the post-grouting behind the shield tunnel wall, it is possible to predict potential dangers in advance and provide valuable time for timely supplementary grouting. Summary of the Invention

[0005] Aiming at the problems and requirements mentioned in the background technology, the purpose of the present invention is to provide a method for generating the medium distribution behind the wall in the field of post-grouting detection of shield tunnel walls. Using the principle of generative adversarial networks, simulate the post-grouting behind the shield tunnel wall in model tests or numerical simulations, collect data and construct a data set, and use digital signal processing methods for preprocessing. Use the data set of the ground penetrating radar time-domain signal after processing to construct a generative adversarial network model for generation and confrontation. Finally, input the measured ground penetrating radar data of the tunnel into the model formed by the confrontation to obtain the corresponding medium image, so as to obtain the situation map of the medium distribution behind the wall.

[0006] Aiming at the engineering difficulty that the ground penetrating radar images of the post-grouting behind the shield tunnel wall are difficult to identify, the present invention uses the principle of generative adversarial network GAN, trains the generative adversarial network after preprocessing according to the data set of the model test, establishes the mapping relationship between the two-dimensional ground penetrating radar image and the medium distribution of the shield tunnel, and inputs the image collected from the ground penetrating radar model test into the prediction model, so as to realize the prediction of the post-grouting diseases behind the shield tunnel wall.

[0007] To achieve the above object, a machine learning prediction method applied to the field of post-grouting detection of shield tunnel walls provided by the present invention includes:

[0008] A method for predicting post-grouting diseases of shield tunnel walls, characterized by including the following steps:

[0009] S1. Install and use a ground penetrating radar in a shield tunnel to collect signals and images.

[0010] S2. Conduct a model test on the post - grouting of the ground penetrating radar, collect the ground penetrating radar images of the shield tunnel with known grouting thickness, and construct the corresponding post - grouting medium distribution map.

[0011] S3. After pre - processing the data, obtain a data sample set; establish a prediction model for the post - grouting medium distribution of the shield tunnel based on the generative adversarial network GAN of the ground penetrating radar image.

[0012] S4. Optimize the calculation parameters of the generative adversarial network GAN.

[0013] S5. Use the prediction model of the generative adversarial network GAN after optimizing the parameters to realize the detection of post - grouting behind the tunnel wall.

[0014] Beneficial effects brought by the technical solution

[0015] The present invention provides a machine - learning prediction method applied to the field of post - grouting detection behind the shield tunnel wall. Using the principle of the generative adversarial network, simulate the post - grouting behind the shield tunnel wall in a model test, collect data to construct a data set, and use digital signal processing methods for pre - processing. Use the data set of the time - domain signal of the pre - processed ground penetrating radar to construct a generative adversarial network. Finally, predict and identify the ground penetrating radar images collected by the ground penetrating radar for post - grouting behind the shield tunnel wall. It can effectively solve the problem that the data of the ground penetrating radar for detecting post - grouting is difficult to interpret, and provide real - time feedback and guidance for the grouting construction. Brief description of the drawings

[0016] Figure 1 For the operation process of the generative adversarial network in S3 (in the figure: discriminator = discriminator)

[0017] Figure 2 Original data of the input ground penetrating radar image (example)

[0018] Figure 3 Input image of the post - grouting medium distribution behind the shield tunnel wall (schematic examples of multiple working conditions)

[0019] Figure 4 For the operation process in S33 Specific implementation manners

[0020] The following will further illustrate the technical solution of the present invention in combination with specific embodiments and their accompanying drawings. With the following description, the advantages and features of the present invention will be clearer.

[0021] It should be noted that the embodiments of the present invention have good implementability and are not any form of limitation to the present invention. The technical features described in the embodiments of the present invention or the combination of technical features should not be considered isolated, and they can be combined with each other to achieve better technical effects. The scope of the preferred implementation manner of the present invention may also include other implementations, and this should be understood by those skilled in the art to which the embodiments of the present invention belong.

[0022] For technologies, methods, and devices known to those of ordinary skill in the relevant art, they may not be discussed in detail, but where appropriate, the said technologies, methods, and devices should be regarded as part of the authorization specification. In all examples shown and discussed here, any specific value should be interpreted as merely exemplary and not as a limitation. Therefore, other examples of the exemplary embodiments may have different values.

[0023] A prediction method for post - grouting diseases in the shield tunnel wall, characterized by comprising the following steps:

[0024] S1. Install and use a ground - penetrating radar in the shield tunnel to collect signals and images.

[0025] S2. Conduct a model test on the post - grouting of the ground - penetrating radar, collect the ground - penetrating radar images of the shield tunnel with known grouting thickness (such as Figure 2 sample) and construct a corresponding post - grouting medium distribution map (such as Figure 3 sample).

[0026] S3. After pre - processing the data, obtain a data sample set; establish a prediction model for the post - grouting medium distribution of the shield tunnel wall based on the generative adversarial network GAN of the ground - penetrating radar image;

[0027] S4. Optimize the calculation parameters of the generative adversarial network GAN.

[0028] S5. Input the real - time collected ground - penetrating radar image of the post - grouting in the shield tunnel wall into the prediction model of the generative adversarial network GAN with optimized parameters in S4 to realize the detection of the post - grouting in the tunnel wall.

[0029] The specific steps of step S1 include the following steps:

[0030] S11. Use a ground - penetrating radar to scan and detect the lining and post - grouting behind the shield tail of the shield tunnel under construction. A commercial radar can be selected for the ground - penetrating radar. The frequency and time window can be selected according to the on - site situation. The position of the ground - penetrating radar should be as close as possible to the segment, and the bolts and grouting holes of the shield segment should be avoided as much as possible during scanning.

[0031] S12. Record the position and scanning path of the radar scan, collect the ground - penetrating radar signals on the computer, and mark the position corresponding to each signal on the scanning path.

[0032] S13. Number each radar signal collected starting from 0.

[0033] S14. Correlate the radar signals with the scanning positions to form a data set.

[0034] The specific steps of step S2 are as follows:

[0035] S21. Select segment lining of shield tunnel and grouting slurry for model test, make a model according to the distribution ratio of 1:1 of segment lining and grouting, place the segment lining in the same soil and inject the same slurry.

[0036] S22. According to the engineering site conditions and engineering experience, divide the medium behind the lining (grouting layer, soil) into voids, uneven grouting, etc. in the model test.

[0037] S23. Use the same ground penetrating radar for model test and conduct radar scanning on the grouting layer with known thickness classification.

[0038] S24. Record the position and scanning path of the radar scanning, collect the ground penetrating radar signals on the computer and mark the medium distribution corresponding to each signal on the scanning path.

[0039] The specific steps of step S3 are as follows:

[0040] S31. Preprocess the radar data samples (excluding medium distribution) in S24, including removing drift, Butterworth band-pass filtering, moving average, F-K migration and normalization, etc., and finally obtain the radar signals as data between [-1, 1].

[0041] S32. After preprocessing the radar data collected from multiple model tests as in S31, form a sample set; the sample set includes data sets from the ground penetrating radar image set {z 1 , z 2 , z 3 ,..., z m} and from the distribution map of the medium behind the shield tunnel lining {x 1 , x 2 , x 3 ,..., x m}.

[0042] S33. Initialize the generation function G: R d →R n and the discriminant function D(x): R n →[0, 1], where the generation function G: R d →R n is a convolutional neural network CNN, and the discriminant function D(x): R n→ [0, 1] is a multi-layer neural network DNN;

[0043] The discriminant function D(x): R n → [0, 1] can be a binary classification neural network (denoted as D 2 ) or a multi-classification neural network (denoted as D 1 ). The training process of the entire model is as Figure 4 shown.

[0044] S34. The adversarial game of GAN is mathematically represented by the min-max value of the objective function between the discriminant function D(x): R n → [0, 1] and the generator function G: R d → R n . The iteration starts with t = 1. Select sample points {x 1 , x 2 , x 3 ,..., x m} from the distribution map of the medium behind the shield tunnel wall. Select the corresponding m images from the ground penetrating radar image set as the vector {z 1 , z 2 , z 3 ,..., z m}. Where m is a hyperparameter and needs to be adjusted according to the training situation.

[0045] S35. Take z in S3.4 as the input to obtain m generated data Where

[0046] S36. According to the following formula 1, for each sample {x 1 , x 2 , x 3 ,..., x m} optimize the neural network variable θ d , update the parameters of the generator D(x) to maximize the loss function of the discriminant function Where η is also a hyperparameter and needs to be tried and adjusted.

[0047]

[0048] S37. According to the following formula 2, optimize the neural network variable θ g , update the parameters of the generator D(x) to minimize the loss function of the generator function Where η is also a hyperparameter and needs to be tried and adjusted.

[0049]

[0050] S38. Through iterations with the number of rounds \(t = 1, 2,\cdots,T\), an optimized generative adversarial network model is established.

[0051] The said step S4 includes:

[0052] S41. Preprocess the ground penetrating radar data collected in S1. The preprocessing steps are the same as S31;

[0053] S42. Input it into the optimized generative adversarial network model established in S38 to obtain the image of the medium distribution behind the wall.

Claims

1. A prediction method for post - grouting diseases in the shield tunnel wall, characterized in that, it includes the following steps: S1. Install and use a ground - penetrating radar in the shield tunnel to collect signals and images; S2. Conduct a model test on post - grouting of the ground - penetrating radar, collect the ground - penetrating radar images of the shield tunnel with known grouting thickness, and construct the corresponding post - wall medium distribution map; S3. After pre - processing the data, obtain a data sample set; establish a prediction model for the post - wall medium distribution of the shield tunnel based on the generative adversarial network GAN of the ground - penetrating radar image; S4. Optimize the calculation parameters of the generative adversarial network GAN; S5. Use the prediction model of the generative adversarial network GAN with optimized parameters to detect the post - grouting behind the tunnel wall; The specific steps of step S3 include the following steps: S31. Pre - process the radar data samples in S24, and finally obtain data with the radar signal between [-1, 1]; S32. After preprocessing the radar data collected from multiple model tests as in S31, a sample set is formed. The sample set includes the ground penetrating radar image set and the dataset of the medium distribution map behind the shield tunnel wall dataset; S33. Initialize the generation function and the discrimination function , where the generation function is a convolutional neural network CNN, and the discrimination function is a deep neural network DNN; S34. The adversarial game of GAN is mathematically represented by the minimax value of the objective function between the discriminant function and the generation function ; At the beginning of iteration, t = 1, select sample points from the distribution map of the medium behind the shield tunnel wall dataset; Select m corresponding images from the ground penetrating radar image set as vectors ; where m is a hyperparameter S35. Use z in S3.4 as the input to obtain m generated data , where ; S36. For each sample, according to the following formula 1 Optimize the neural network variables , update the generator parameters to maximize the loss function of the discriminant function ; where is also a hyperparameter; Formula 1 S37. Optimize the neural network variables according to the following formula 2 , and update the parameters of the generator to minimize the loss function of the generation function ; where is also a hyperparameter; Formula 2 S38. Through iterations with the number of rounds t = 1, 2,... T, establish an optimized generative adversarial network model.

2. A prediction method for post - grouting diseases in the shield tunnel wall according to claim 1, characterized in that, the specific steps of step S1 include the following steps: S11. Use a ground - penetrating radar to scan and detect the lining and post - grouting behind the shield tail of the shield tunnel under construction; S12. Record the position and scanning path of the radar scan, collect the ground - penetrating radar signals on the computer, and mark the position corresponding to each signal on the scanning path; S13. Number each collected radar signal starting from 0; S14. Correlate the radar signals with the scanning positions to form a data set.

3. A prediction method for post - grouting diseases in the shield tunnel wall according to claim 1, characterized in that, the specific steps of step S2 include the following steps: S21. Select the shield tunnel segment and grouting slurry for model testing, make a model according to the distribution of the segment and grouting 1:1, place the segment in the same soil and inject the same slurry; S22. According to the engineering site conditions and engineering experience, divide the post - wall medium into voids and uneven grouting in the model test; S23. Use the same ground - penetrating radar for model testing and scan the grouting layer with known thickness classification by radar; S24. Record the position and scanning path of the radar scan, collect the ground - penetrating radar signals on the computer, and mark the medium distribution corresponding to each signal on the scanning path.

4. A prediction method for post - grouting diseases in the shield tunnel wall according to claim 1, characterized in that, The discrimination function is a binary classification neural network, denoted as D 2 or a multi-classification neural network, denoted as D 1 .

5. A prediction method for post - grouting diseases in the shield tunnel wall according to claim 1, characterized in that, the steps of step S4 include: S41. Pre - process the ground - penetrating radar data collected in S1; the pre - processing steps are the same as S31; S42. Input it into the optimized generative adversarial network model established in S38 to obtain the post - wall medium distribution image.

Images