A method for denoising steel bars in a tunnel lining

Abstract

This invention discloses a method for denoising steel reinforcement in tunnel lining, comprising the following steps: Step S1, designing a defect model of the steel reinforcement mesh covering the tunnel lining to conform to practical significance, obtaining a forward simulation map through finite-difference time-domain forward modeling (FDTD), and simultaneously designing a real model map corresponding to the defect; Step S2, training and iterating the parameters of a GPR-invNet network structure using the forward simulation map as input and the real model map as output, to obtain the optimal GPR-invNet network structure; Step S3, outputting a denoised prediction result map based on the optimal GPR-invNet network structure after training. This invention has the advantages of fast processing time, simple operation, and high accuracy.

Description

Technical Field

[0001] This invention relates to a method for noise reduction of reinforcing steel bars in tunnel lining, belonging to the technical field of ground penetrating radar. Background Technology

[0002] Ground-penetrating radar (GPR) is a non-destructive testing tool based on high-frequency electromagnetic wave propagation. It features rapid, non-destructive, and continuous detection, and can image underground structural profiles in real time. Therefore, it is widely used in engineering and environmental geophysical exploration, and is an excellent tool for detecting hidden defects under tunnel linings. However, due to the coverage of radar signals by the shallow steel reinforcement network in the tunnel lining, GPR cannot accurately identify defects beneath the reinforcement network. Some studies use GprMax forward modeling for manual image interpretation of defects under the reinforcement network, but this method is time-consuming and has low accuracy. Therefore, how to accurately remove the influence of steel reinforcement noise has become a major challenge for GPR in identifying defects under tunnel linings.

[0003] Currently, the most widely used methods for denoising steel bars are manual interpretation and deep learning.

[0004] In 2004, Xu Maohui et al. studied radar images of reinforcing bars through extensive forward modeling and subsequently used a backpropagation (BP) artificial neural network to identify the diameter of the reinforcing bars. However, this method failed to identify defects under the reinforcing bar mesh, and forward modeling was time-consuming, costly, and required an insufficient number of specimens.

[0005] In 2020, Ye Yuhui et al. studied various defects under tunnel lining using deep learning. However, this method did not provide a method for identifying defects under strong signal interference from steel bars.

[0006] In 2020, Li Shufan et al. studied the identification and denoising of voids and delamination defects in concrete under steel mesh using deep learning. However, the defect model studied in this method is relatively simple and has not been applied to tunnel lining inspection.

[0007] In summary, although various solutions have emerged for the identification and noise reduction of defects under the steel reinforcement mesh in tunnel lining, some are time-consuming and have a small number of specimens; some have studied different types of defects in tunnel lining but have not provided specific identification methods for defects under the steel reinforcement mesh; and some have studied methods for identifying defects such as voids and cavities under the steel reinforcement mesh, but the defect models studied are relatively simple and have not been applied to the tunnel lining environment. Summary of the Invention

[0008] To overcome the problems in the prior art, the present invention provides a method for noise reduction of reinforcing steel bars in tunnel lining.

[0009] The technical solution provided by this invention to solve the above-mentioned technical problems is: a method for noise reduction of reinforcing steel in tunnel lining, comprising the following steps:

[0010] Step S1: Design a defect model for the tunnel lining structure under the steel mesh covering the tunnel lining to make it realistic. Obtain the forward simulation diagram through finite-difference time-domain forward modeling (FDTD) and design the real model diagram corresponding to the defect.

[0011] Step S2: Use the forward simulation graph as input and the real model graph as output in the GPR-inv Net network structure for training and parameter iteration to obtain the optimal GPR-inv Net network structure;

[0012] Step S3: Output the denoised prediction result image based on the optimal GPR-inv Net network structure after training.

[0013] A further technical solution is that the specific training process in step S2 is as follows:

[0014] Step S21: First, put the data x i Input into the first layer;

[0015] Step S22: The output layer is finally reached through the calculations of the intermediate layer;

[0016] Step S23: Calculate data x i mean μ β and variance Then, for the data x i Standardize the data; finally, set the learnable parameters γ and β to restore the data characteristics.

[0017] Step S24: Repeat S21-S23 to calculate the L2, L3, ..., L... n The activation values ​​of the layer are a2, a3, ..., a n At the same time, each layer is standardized;

[0018] Step S25: Calculate the output layer L n Error term;

[0019] Step S26: Calculate the global error;

[0020] Step S27: Calculate the gradient values ​​of the weight parameters and the gradient values ​​of the bias parameters of the final layer, and calculate the gradient values ​​of the parameters of the m-th layer according to the backpropagation approach.

[0021] Step S28: Repeat the above steps, observe the learning rate and control the speed of gradient updates to achieve the optimal parameter selection.

[0022] A further technical solution is that the calculation formula in step S21 is:

[0023]

[0024] In the formula: x i For the input of the i-th data, is the activation value of the i-th neuron in the m-th layer.

[0025] A further technical solution is that the calculation formula in step S22 is:

[0026]

[0027]

[0028] In the formula: ω is the weighted sum of the i-th neuron in the m-th layer, b is the bias parameter, and f(·) is the activation function.

[0029] A further technical solution is that the mean μ in step S23 β and variance The calculation formula is:

[0030]

[0031]

[0032] Where: μ β The mean; Let Variance be the variance.

[0033] A further technical solution is that the formula for standardizing the data in step S23 is:

[0034]

[0035] Where: μ β This is the mean.

[0036] A further technical solution is that the formula for restoring the data features in step S23 is:

[0037]

[0038] In the formula: y i To restore the data.

[0039] A further technical solution is that the calculation formula in step S25 is:

[0040]

[0041] In the formula: This is the activation value of the i-th neuron in the n-th layer.

[0042] A further technical solution is that the calculation formula in step S26 is:

[0043]

[0044] In the formula: C is the global error.

[0045] A further technical solution is that the calculation formula in step S27 includes:

[0046]

[0047] Gradient value calculation formula:

[0048]

[0049] The gradient value of the weight parameters can then be expressed as:

[0050]

[0051] Similarly, a biased gradient value can be expressed as:

[0052]

[0053] For the gradient values ​​of the parameters of the m-th layer of the model:

[0054]

[0055]

[0056] In the formula: C is the global error; ω is the weight parameter; and b is the bias parameter.

[0057] The present invention has the following advantages: it has the advantages of fast processing time, simple operation and high precision. Attached Figure Description

[0058] Figure 1 Here is a diagram of the GPR-inv Net network structure;

[0059] Figure 2 This is a model drawing without reinforced concrete cavities;

[0060] Figure 3 This is a model diagram containing a cavity with half of its steel reinforcement.

[0061] Figure 4 A model diagram of a cavity fully covered by reinforcing steel.

[0062] Figure 5 A model diagram of a cavity under full reinforcement coverage in a lining structure;

[0063] Figure 6 A model diagram of cracks under full reinforcement coverage in a lining structure;

[0064] Figure 7 This is a model diagram of a porous structure under full steel reinforcement coverage in a lining structure. Detailed Implementation

[0065] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. 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.

[0066] The present invention provides a method for noise reduction of reinforcing steel bars in tunnel lining, comprising the following steps:

[0067] Step S1: Design a defect model for the tunnel lining structure under the steel mesh covering the tunnel lining to make it realistic. Obtain the forward simulation diagram through finite-difference time-domain forward modeling (FDTD) and design the real model diagram corresponding to the defect.

[0068] Step S2: Place the forward simulation graph as input and the real model graph as output into the GPR-inv Net network structure (e.g., Figure 1 The optimal GPR-inv Net network structure is obtained by training and parameter iteration in the (as shown) diagram.

[0069] Step S21: First, put the data x i Input into the first layer;

[0070]

[0071] In the formula: x i For the input of the i-th data, is the activation value of the i-th neuron in the m-th layer;

[0072] Step S22: The output layer is finally reached through the calculations of the intermediate layer;

[0073]

[0074]

[0075] In the formula: ω is the weighted sum of the i-th neurons in the m-th layer, b is the bias parameter, and f(·) is the activation function.

[0076] Step S23: Calculate data x i mean μ β and variance Then, for the data x i Standardize the data; finally, set the learnable parameters γ and β to restore the data characteristics.

[0077]

[0078]

[0079]

[0080]

[0081] Step S24: Repeat S21-S23 to calculate the L2, L3, ..., L... n The activation values ​​of the layer are a2, a3, ..., a n At the same time, each layer is standardized;

[0082]

[0083] Step S25: Calculate the output layer L n Error term;

[0084]

[0085] Step S26: Calculate the global error;

[0086]

[0087] Step S27: Calculate the gradient values ​​of the weight parameters and the gradient values ​​of the bias parameters of the final layer, and calculate the gradient values ​​of the parameters of the m-th layer according to the backpropagation approach.

[0088]

[0089] Gradient value calculation formula:

[0090]

[0091] The gradient value of the weight parameters can then be expressed as:

[0092]

[0093] Similarly, a biased gradient value can be expressed as:

[0094]

[0095] For the gradient values ​​of the parameters of the m-th layer of the model:

[0096]

[0097]

[0098] In the formula: C is the global error; ω is the weight parameter; and b is the bias parameter.

[0099] Step S28: Repeat the above steps, observe the learning rate and control the speed of gradient updates to achieve the best parameter selection;

[0100] Step S3: Output the denoised prediction result image based on the optimal GPR-inv Net network structure after training.

[0101]

[0102] In the formula: x i For the input of the i-th data, is the activation value of the i-th neuron in the m-th layer.

[0103] Example

[0104] First, the following settings were configured: Figure 2 The circular cavity model without reinforcing steel shown in figure a was used as a control experiment, and the results were obtained. Figure 2 The forward simulation diagram shown in b is used as a basis for further analysis, including the addition of half and complete steel meshes. Figure 3 and Figure 4 As shown. Comparison Figure 3 c and Figure 3 d、 Figure 4 c and Figure 4 Both studies found that this method can effectively remove the noise impact of reinforcing bars on the cavity.

[0105] Secondly, a model of void defects in tunnel lining under interference from a complete steel mesh was established, such as... Figure 5 As shown, analysis Figure 5 c and Figure 5 d found that it had good effects on both the removal of noise from the steel mesh and the inversion of the lining interface and void model.

[0106] Finally, models of other defects in tunnel lining under interference from complete steel mesh were set up, such as... Figure 6 , Figure 7 As shown, where Figure 6 This represents cracks or defects in tunnel lining. Figure 7 Representing loose material defects in tunnel lining, from Figure 6 d and Figure 7 As can be seen from d, this method is very suitable for removing noise from steel mesh, inverting lining interfaces, and modeling defects.

[0107] The above description is not intended to limit the present invention in any way. Although the present invention has been disclosed through the above embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some changes or modifications to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the present invention shall still fall within the scope of the present invention.

Claims

1. A method of denoising steel reinforcement in a tunnel lining, characterized in that, Includes the following steps: Step S1: Design a defect model for the tunnel lining structure under the steel mesh covering the tunnel lining to make it realistic. Obtain the forward model simulation diagram through finite difference time-domain forward modeling, and design the real model diagram corresponding to the defect. Step S2: Use the forward simulation graph as input and the real model graph as output in the GPR-inv Net network structure for training and parameter iteration to obtain the optimal GPR-inv Net network structure; Step S21, first input data into the first layer; In the formula: is the input of the first i layer, is the activation value of the first m neuron in the first layer, i is the activation value of the first Step S22: The output layer is finally reached through the calculations of the intermediate layer; Step S23: Calculate data mean and variance Then, on the data Standardize the parameters; finally, set the learnable parameters. and To restore data characteristics; The formula for restoring the data features is: In the formula: To restore the data; Step S24: Repeat S21-S23 to calculate the... layer activation value At the same time, each layer is standardized; Step S25: Calculate the output layer Error term; Step S26: Calculate the global error; Step S27: Calculate the gradient values ​​of the weight parameters and bias parameters of the final layer, and apply the backpropagation algorithm to the... m Calculate the gradient values ​​of the layer parameters; Step S28: Repeat the above steps, observe the learning rate and control the speed of gradient updates to achieve the best parameter selection; Step S3: Output the denoised prediction result image based on the optimal GPR-inv Net network structure after training.

2. The noise reduction method for reinforcing steel in tunnel lining according to claim 1, characterized in that, The calculation formula in step S22 is: In the formula: The weighted sum of the i-th neuron in the m-th layer. For weight parameters, b For bias parameters, This is the activation function.

3. The noise reduction method for reinforcing steel in tunnel lining according to claim 1, characterized in that, The mean in step S23 and variance The calculation formula is: In the formula: The mean; Let Variance be the variance.

4. The noise reduction method for reinforcing steel in tunnel lining according to claim 3, characterized in that, The formula for standardizing the data in step S23 is as follows: In the formula: This is the mean.

5. A method for noise reduction of reinforcing steel in tunnel lining according to claim 1, characterized in that, The calculation formula in step S25 is: In the formula: This is the activation value of the i-th neuron in the n-th layer.

6. A method for noise reduction of reinforcing steel in tunnel lining according to claim 1, characterized in that, The calculation formula in step S26 is: In the formula: C is the global error.

7. A method for noise reduction of reinforcing steel in tunnel lining according to claim 1, characterized in that, The calculation formula in step S27 includes: Gradient value calculation formula: The gradient value of the weight parameters can then be expressed as: Similarly, a biased gradient value can be expressed as: For the gradient values ​​of the parameters of the m-th layer of the model: In the formula: C represents the global error; For weight parameters, b This is the bias parameter.

Images