A Method for Predicting Water Content in Tunnel Rock Mass Based on Multi-Scale Residual Denoising Network

Abstract

This invention discloses a method for predicting the water content of tunnel rock mass based on a multi-scale residual denoising network, belonging to the field of tunnel geological prediction technology. It includes the following steps: a dataset construction stage, constructing a transient electromagnetic method water content sample dataset; a sample dataset partitioning stage, dividing the dataset into a training set, a validation set, and a test set, and performing randomization during partitioning; a model construction stage, constructing a multi-scale sampling residual denoising network model based on a self-attention mechanism; a model training stage, using the training set for model training, employing an early stopping strategy and the Adam optimizer during training, and using the signal-to-noise ratio as an evaluation metric to optimize the model; and a water content prediction stage, using the trained model to predict water content, validating the model's imaging effect through tunnel data in a real-world scenario, and further optimizing the model based on the validation results. This invention can achieve high-precision inversion of tunnel rock mass water content under complex noise environments.

Description

Technical Field

[0001] This invention relates to the field of tunnel geological prediction technology, and in particular to a method for predicting the water content of tunnel rock mass based on a multi-scale residual denoising network. Background Technology

[0002] The water content of underground rock masses is a crucial indicator of their physical properties, directly impacting their mechanical behavior, stability, and engineering safety. For underground engineering, mineral resource development, and groundwater resource surveys, accurately determining the water content of underground rock masses is not only a key issue in scientific research but also a core task in engineering practice. Changes in rock mass water content affect not only the distribution, migration, and storage of groundwater but also the strength, toughness, and deformation characteristics of the rock mass. Therefore, accurately assessing the water content of underground rock masses is essential for the design, extraction, and monitoring of underground engineering projects.

[0003] Traditional methods for measuring rock mass moisture content primarily rely on sampling and laboratory analysis, such as drying methods and chemical analysis. While these methods can yield accurate results to a certain extent, they also have several limitations. First, traditional sampling methods can only obtain moisture information for localized areas, failing to achieve comprehensive monitoring of large-scale underground rock masses. Second, these methods typically require considerable time and are costly, making them unsuitable for the rapid and efficient measurement demands of modern engineering projects. Therefore, the application of traditional methods faces significant limitations, necessitating the development of new technologies to address this issue.

[0004] To address the limitations of traditional methods, many geophysical exploration techniques have emerged. Among them, the transient electromagnetic method (TEM), as a non-invasive, real-time underground detection technique, has significant application value. TEM measures the response of the underground medium to electromagnetic waves, allowing for real-time acquisition of the medium's electrical characteristics. The propagation characteristics of electromagnetic waves are closely related to the physical properties of the underground medium, such as conductivity and permittivity, thus reflecting the water distribution within the underground rock mass. Therefore, TEM has become an important tool in fields such as groundwater resource exploration and geotechnical engineering investigation.

[0005] However, traditional transient electromagnetic method (TEM) data analysis methods largely rely on simple physical models and empirical formulas. The accuracy of these methods is often affected by the complexity and variability of the underground medium, making it difficult to effectively extract moisture information from underground rock masses. Accurately identifying and extracting moisture characteristics from complex TEM data has always been a challenge in geological exploration technology. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for predicting the water content of tunnel rock mass based on a multi-scale residual denoising network.

[0007] The objective of this invention is achieved through the following technical solution: a method for predicting the water content of tunnel rock mass based on a multi-scale residual denoising network, comprising the following steps:

[0008] In the dataset construction phase, forward modeling was performed using various rock mass parameters, and data augmentation was performed by introducing various noises, setting variable parameters, and adding random factors, thereby constructing a transient electromagnetic method water content sample dataset.

[0009] In the sample dataset partitioning stage, the transient electromagnetic method moisture content sample dataset is divided into a training set, a validation set, and a test set, and randomization is performed during the partitioning process.

[0010] In the model building phase, a multi-scale sampling residual denoising network model based on the self-attention mechanism is constructed, which includes a self-attention module, a multi-scale sampling module, and a residual denoising autoencoder module.

[0011] During the model training phase, the training set is used to train the model. The early stopping strategy and Adam optimizer are used during the training process, and the signal-to-noise ratio is used as the evaluation metric to optimize the model.

[0012] In the moisture content prediction stage, the trained model is used to predict the moisture content of tunnel electromagnetic data under different noise conditions. The root mean square error is used to evaluate the prediction accuracy, and the model imaging effect is verified by tunnel data in real-world scenarios. Based on the verification results, the model is optimized again.

[0013] Preferably, the dataset construction phase further includes the following steps:

[0014] The electrical conductivity, water content, signal-to-noise ratio, waveform type, and geological structure parameters of the tunnel were used as variables in the forward modeling.

[0015] By using a weighted exponential attenuation model, the attenuation effect of signals caused by multiple reflections and scattering in a tunnel environment is simulated to generate attenuated noise. Random pulses are inserted into the original signal to simulate transient interference caused by equipment failure or external activities, resulting in burst noise. Gaussian noise is simulated using a normal distribution. An electromagnetic interference and equipment noise are randomly added to the data using a random fluctuation model to simulate environmental noise. Attenuated noise, burst noise, Gaussian noise, and environmental noise are introduced, and variable parameters are set to add random factors for data augmentation.

[0016] The SimPEG library was used for tunnel electromagnetic forward modeling. Parallel computing methods were used in conjunction with a hierarchical conductivity variation mechanism to generate a preset number of samples to obtain a transient electromagnetic water content sample dataset.

[0017] Preferably, the mechanism by which the layered conductivity varies over time is as follows:

[0018] The simulation depth is set to a preset depth, and the thickness of each layer is set to a preset thickness, and the layers are divided into layers. The conductivity of the first layer is randomly selected from the preset conductivity range, and the coefficient of variation of conductivity of adjacent layers is controlled within the preset coefficient range.

[0019] Preferably, the preset depth is 96 meters, the preset thickness is 4 meters, and the preset coefficient ranges from 0.3 to 0.5.

[0020] Preferably, the ratio of the training set, validation set, and test set is 75%, 15%, and 10%, respectively.

[0021] Preferably, the model building stage further includes the following steps:

[0022] Based on the Transformer architecture, the TensorFlow framework is used for self-attention computation. Hyperparameters are optimized through grid search to obtain an initial model with X attention heads and Y layers in depth.

[0023] A multi-scale sampling module is introduced into the initial model for feature enhancement. First, a self-attention mechanism is applied to the input features for calculation. The feature relationship weights in the results are sorted, and the top K features are selected for downsampling. Through three layers of downsampling, four different levels of feature representation are obtained. Then, during the upsampling operation, a pseudo-feature vector with all zeros is constructed. The feature information from the previous downsampling layer is filled into the pseudo-feature vector, and it is concatenated with the features obtained from downsampling at the same level. Finally, the concatenated features are input into the attention mechanism of this layer for calculation, resulting in the modified model.

[0024] A residual denoising autoencoder module is introduced into the modified model to understand complex features and eliminate the influence of noise. The original noise-free data and random noise data are input into the residual denoising autoencoder module to obtain the predicted noise-free original data. The cosine similarity between the original noise-free data and the predicted noise-free original data is calculated as the loss function. The loss function is used to train the model to recover the noise-free original data. The output of the residual denoising autoencoder module is used as the input of the multi-scale sampling module for feature extraction. Finally, a multi-scale sampling residual denoising network model based on the self-attention mechanism is obtained.

[0025] Preferably, X is 8, Y is 4, and K is 30% of the total number of features.

[0026] Preferably, the early stopping strategy is to terminate training when the validation set loss does not decrease for 15 consecutive rounds.

[0027] Preferably, the imaging effect of the model is verified by comparing the imaging effects of the Transformer method, the multi-scale sampling method, and the multi-scale sampling residual denoising network model based on the self-attention mechanism.

[0028] The beneficial effects of this invention are:

[0029] 1) This invention achieves high-precision inversion of tunnel rock mass water content under complex noise environment by constructing a multi-scale sampling residual denoising network model based on self-attention mechanism.

[0030] 2) Experiments show that the multi-scale sampling residual denoising network model based on the self-attention mechanism improves the overall signal-to-noise ratio (SNR) by 10.28 dB compared with the original denoising model, and the average RMSE of water content prediction is as low as 0.002657, which effectively solves the problems of insufficient feature extraction and low prediction accuracy of traditional methods in strong noise environment. Attached Figure Description

[0031] Figure 1 This is a flowchart of the method of the present invention;

[0032] Figure 2 This is a schematic diagram illustrating the partitioning and application of the sample set;

[0033] Figure 3 This is a schematic diagram of a multi-scale sampling residual denoising network model based on a self-attention mechanism;

[0034] Figure 4 This is a schematic diagram illustrating the denoising effect of the model on the attenuation noise test set;

[0035] Figure 5 This is a schematic diagram illustrating the denoising effect of the model on a burst noise test set.

[0036] Figure 6 This is a schematic diagram illustrating the denoising effect of the model on a Gaussian noise test set.

[0037] Figure 7 This is a schematic diagram illustrating the denoising effect of the model on an environmental noise test set.

[0038] Figure 8 This is a schematic diagram illustrating the model's performance on a noisy test set.

[0039] Figure 9 This is a schematic diagram showing the comparison of 90° imaging along the side of the tunnel face. Detailed Implementation

[0040] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. 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.

[0041] Deep learning algorithms possess powerful data fitting and pattern recognition capabilities, automatically extracting patterns from massive datasets and achieving significant results in areas such as image recognition, speech processing, and nonlinear problems. Deep learning technology has also demonstrated immense potential in the field of geological exploration. This invention combines deep learning with traditional geophysical exploration techniques to achieve accurate prediction and analysis of the water content of underground rock masses.

[0042] See Figures 1-9 This invention provides a technical solution: a method for predicting the water content of tunnel rock mass based on a multi-scale residual denoising network, comprising the following steps:

[0043] In the dataset construction phase, forward modeling was performed using various rock mass parameters, and data augmentation was performed by introducing various noises, setting variable parameters, and adding random factors, thereby constructing a transient electromagnetic method water content sample dataset.

[0044] In the sample dataset partitioning stage, the transient electromagnetic method moisture content sample dataset is divided into a training set, a validation set, and a test set, and randomization is performed during the partitioning process.

[0045] In the model building phase, a multi-scale sampling residual denoising network model based on the self-attention mechanism is constructed, which includes a self-attention module, a multi-scale sampling module, and a residual denoising autoencoder module.

[0046] During the model training phase, the training set is used to train the model. The early stopping strategy and Adam optimizer are used during the training process, and the signal-to-noise ratio is used as the evaluation metric to optimize the model.

[0047] In the moisture content prediction stage, the trained model is used to predict the moisture content of tunnel electromagnetic data under different noise conditions. The root mean square error is used to evaluate the prediction accuracy, and the model imaging effect is verified by tunnel data in real-world scenarios. Based on the verification results, the model is optimized again.

[0048] In this embodiment, forward modeling is first performed using the SimPEG library in conjunction with parameters such as underground rock conductivity, water content, waveform type, SNR, and tunnel geological structure. By introducing various types of noise, including attenuation noise, burst noise, Gaussian noise, and environmental noise, and setting variable parameters and random factors, over 50,000 diverse datasets closely resembling actual tunnel scenarios are generated. Subsequently, the datasets are divided into training, validation, and test sets in a 75%:15%:10% ratio and randomized. A network model integrating a self-attention mechanism, a multi-scale sampling module, and a residual denoising autoencoder (RDAE) is then constructed. An early stopping strategy is used to train the model, and the denoising performance is evaluated using the signal-to-noise ratio (SNR). Finally, the trained model is used to process tunnel electromagnetic data under different noise conditions and invert the water content, with the root mean square error (RMSE) measuring the prediction error.

[0049] Comparison of the original noise-free signal and the noise curves before and after denoising in the noise reduction test set. Figure 4 As shown in (a); a comparison of the relative errors of predicted values ​​before and after denoising in the noise reduction test set. Figure 4 As shown in (b). Comparison of the original noise-free signal and the noise curves before and after denoising in the burst noise test set. Figure 5 As shown in (a); a comparison of the relative errors of the predicted values ​​before and after denoising in the burst noise test set. Figure 5 As shown in (b), a comparison is made between the original noise-free signal and the noise curves before and after denoising in the Gaussian noise test set. Figure 6 As shown in (a); a comparison of the relative errors of the predicted values ​​before and after denoising in the Gaussian noise test set. Figure 6 As shown in (b). A comparison of the original noise-free signal and the noise curves before and after denoising in the environmental noise test set. Figure 7 As shown in (a); a comparison of the relative errors of predicted values ​​before and after denoising in the environmental noise test set. Figure 7 As shown in (b).

[0050] Table 1 shows the SNR comparison of the multi-scale sampling residual denoising network model based on self-attention mechanism before and after denoising on the test set:

[0051] Table 1

[0052]

[0053] Table 2 shows the RMSE comparison of the multi-scale sampling residual denoising network model based on self-attention mechanism in water content prediction:

[0054] Table 2

[0055]

[0056] In some embodiments, the dataset construction phase further includes the following steps:

[0057] The electrical conductivity, water content, signal-to-noise ratio, waveform type, and geological structure parameters of the tunnel were used as variables in the forward modeling.

[0058] By using a weighted exponential attenuation model, the attenuation effect of signals caused by multiple reflections and scattering in a tunnel environment is simulated to generate attenuated noise. Random pulses are inserted into the original signal to simulate transient interference caused by equipment failure or external activities, resulting in burst noise. Gaussian noise is simulated using a normal distribution. An electromagnetic interference and equipment noise are randomly added to the data using a random fluctuation model to simulate environmental noise. Attenuated noise, burst noise, Gaussian noise, and environmental noise are introduced, and variable parameters are set to add random factors for data augmentation.

[0059] The SimPEG library was used for tunnel electromagnetic forward modeling. Parallel computing methods were used in conjunction with a hierarchical conductivity variation mechanism to generate a preset number of samples to obtain a transient electromagnetic water content sample dataset.

[0060] In some embodiments, the mechanism for the variation of layered conductivity over time is specifically as follows:

[0061] The simulation depth is set to a preset depth, and the thickness of each layer is set to a preset thickness, and the layers are divided into layers. The conductivity of the first layer is randomly selected from the preset conductivity range, and the coefficient of variation of conductivity of adjacent layers is controlled within the preset coefficient range.

[0062] In some embodiments, the preset depth is 96 meters, the preset thickness is 4 meters, and the preset coefficient ranges from 0.3 to 0.5.

[0063] In some embodiments, the ratio of the training set, validation set, and test set is 75%, 15%, and 10%, respectively.

[0064] In this embodiment, the partitioning and application of the sample set are as follows: Figure 2 As shown, the training set is used for training and optimizing the network model, adjusting the fitting parameters to obtain the trained model; the validation set is used to adjust the hyperparameters of model complexity to help optimize the model's performance; the test set is used to measure the final model's performance and does not participate in the training process. The sample set was randomized during the partitioning process to avoid the network remembering the correlation between samples during training, thus ensuring the model's generalization ability.

[0065] In some embodiments, the model building phase further includes the following steps:

[0066] Based on the Transformer architecture, the TensorFlow framework is used for self-attention computation. Hyperparameters are optimized through grid search to obtain an initial model with X attention heads and Y layers in depth.

[0067] A multi-scale sampling module is introduced into the initial model for feature enhancement. First, a self-attention mechanism is applied to the input features for calculation. The feature relationship weights in the results are sorted, and the top K features are selected for downsampling. Through three layers of downsampling, four different levels of feature representation are obtained. Then, during the upsampling operation, a pseudo-feature vector with all zeros is constructed. The feature information from the previous downsampling layer is filled into the pseudo-feature vector, and it is concatenated with the features obtained from downsampling at the same level. Finally, the concatenated features are input into the attention mechanism of this layer for calculation, resulting in the modified model.

[0068] A residual denoising autoencoder module is introduced into the modified model to understand complex features and eliminate the influence of noise. The original noise-free data and random noise data are input into the residual denoising autoencoder module to obtain the predicted noise-free original data. The cosine similarity between the original noise-free data and the predicted noise-free original data is calculated as the loss function. The loss function is used to train the model to recover the noise-free original data. The output of the residual denoising autoencoder module is used as the input of the multi-scale sampling module for feature extraction. Finally, a multi-scale sampling residual denoising network model based on the self-attention mechanism is obtained.

[0069] In this embodiment, the constructed network model is as follows: Figure 3 As shown, the Residual Denoising Autoencoder (RDAE) module includes a Denoising Autoencoder (DAE) and introduces residual structures to help the model better understand complex features and eliminate the influence of noise.

[0070] In some embodiments, X is 8, Y is 4, and K is 30% of the total number of features.

[0071] In some embodiments, the early stopping strategy is to terminate training when the validation set loss does not decrease for 15 consecutive rounds.

[0072] In this embodiment, an early stopping strategy is employed during training to ensure stable convergence of the network. The early stopping condition is that training terminates if the validation set loss does not decrease for 15 consecutive rounds. The optimizer is Adam, and the learning rate is set to 0.005. The signal-to-noise ratio (SNR) is used as the evaluation metric for model performance, and four types of noise test samples are selected for testing. The SNR directly relates to the effective detection range and accuracy. At a low SNR, the signal is easily masked by noise, leading to a shorter detection distance and reduced reliability of the results; a high SNR, on the other hand, helps to accurately extract the electrical characteristics of the underground medium.

[0073] In some embodiments, the imaging performance of the model is verified by comparing the imaging performance of the Transformer method, the multi-scale sampling method, and the multi-scale sampling residual denoising network model based on the self-attention mechanism.

[0074] In this embodiment, a trained network model is used to predict the water content of tunnel electromagnetic data under different noise conditions. The predictive ability of the model in different noise environments is analyzed by comparing it with the original geological data. RMSE is used as an important indicator to measure the model's prediction error. Furthermore, the water content prediction in a real-world scenario is verified using advanced geological prediction data from the left tunnel entrance. Data from the 90° side line of the tunnel face is extracted for verification. Imaging verification and comparison are performed using the Transformer method, multi-scale sampling method, and residual denoising model, respectively. The comparison results are as follows: Figure 9 As shown.

[0075] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A method for predicting the water content of tunnel rock mass based on a multi-scale residual denoising network, characterized in that: Includes the following steps: In the dataset construction phase, forward modeling was performed using various rock mass parameters, and data augmentation was performed by introducing various noises, setting variable parameters, and adding random factors, thereby constructing a transient electromagnetic method water content sample dataset. In the sample dataset partitioning stage, the transient electromagnetic method moisture content sample dataset is divided into a training set, a validation set, and a test set, and randomization is performed during the partitioning process. In the model building phase, a multi-scale sampling residual denoising network model based on the self-attention mechanism is constructed, which includes a self-attention module, a multi-scale sampling module, and a residual denoising autoencoder module. During the model training phase, the training set is used to train the model. The early stopping strategy and Adam optimizer are used during the training process, and the signal-to-noise ratio is used as the evaluation metric to optimize the model. In the water content prediction stage, the trained model is used to predict the water content of tunnel electromagnetic data under different noise conditions. The root mean square error is used to evaluate the prediction accuracy. The model imaging effect is verified by tunnel data in real-world scenarios. The model is then optimized again based on the verification results. The model building phase also includes the following steps: Based on the Transformer architecture, the TensorFlow framework is used for self-attention computation. Hyperparameters are optimized through grid search to obtain an initial model with X attention heads and Y layers in depth. A multi-scale sampling module is introduced into the initial model for feature enhancement. First, a self-attention mechanism is applied to the input features for calculation. The feature relationship weights in the results are sorted, and the top K features are selected for downsampling. Through three layers of downsampling, four different levels of feature representation are obtained. Then, during the upsampling operation, a pseudo-feature vector with all zeros is constructed. The feature information from the previous downsampling layer is filled into the pseudo-feature vector, and it is concatenated with the features obtained from downsampling at the same level. Finally, the concatenated features are input into the attention mechanism of this layer for calculation, resulting in the modified model. A residual denoising autoencoder module is introduced into the modified model to understand complex features and eliminate the effects of noise; The original noise-free data and random noise data are input into the residual denoising autoencoder module to obtain the predicted noise-free original data. The cosine similarity between the original noise-free data and the predicted noise-free original data is calculated as the loss function. The loss function is used to train the model to recover the noise-free original data. The output of the residual denoising autoencoder module is used as the input of the multi-scale sampling module for feature extraction. Finally, a multi-scale sampling residual denoising network model based on the self-attention mechanism is obtained.

2. The method for predicting the water content of tunnel rock mass based on a multi-scale residual denoising network according to claim 1, characterized in that: The dataset construction phase also includes the following steps: The electrical conductivity, water content, signal-to-noise ratio, waveform type, and geological structure parameters of the tunnel were used as variables in the forward modeling. By using a weighted exponential attenuation model, the attenuation effect of signals caused by multiple reflections and scattering in a tunnel environment is simulated to generate attenuated noise. Random pulses are inserted into the original signal to simulate transient interference caused by equipment failure or external activities, resulting in burst noise. Gaussian noise is simulated using a normal distribution. An electromagnetic interference and equipment noise are randomly added to the data using a random fluctuation model to simulate environmental noise. Attenuated noise, burst noise, Gaussian noise, and environmental noise are introduced, and variable parameters are set to add random factors for data augmentation. The SimPEG library was used for tunnel electromagnetic forward modeling. Parallel computing methods were used in conjunction with a hierarchical conductivity variation mechanism to generate a preset number of samples to obtain a transient electromagnetic water content sample dataset.

3. The method for predicting the water content of tunnel rock mass based on a multi-scale residual denoising network according to claim 2, characterized in that: The mechanism by which the layered conductivity varies over time is specifically as follows: The simulation depth is set to a preset depth, and the thickness of each layer is set to a preset thickness, and the layers are divided into layers. The conductivity of the first layer is randomly selected from the preset conductivity range, and the coefficient of variation of conductivity of adjacent layers is controlled within the preset coefficient range.

4. The method for predicting the water content of tunnel rock mass based on a multi-scale residual denoising network according to claim 3, characterized in that: The preset depth is 96 meters, the preset thickness is 4 meters, and the preset coefficient ranges from 0.3 to 0.

5.

5. The method for predicting the water content of tunnel rock mass based on a multi-scale residual denoising network according to claim 1, characterized in that: The training set, validation set, and test set are divided into three proportions: 75%, 15%, and 10%, respectively.

6. The method for predicting the water content of tunnel rock mass based on a multi-scale residual denoising network according to claim 1, characterized in that: X is 8, Y is 4, and K is 30% of the total number of features.

7. The method for predicting the water content of tunnel rock mass based on a multi-scale residual denoising network according to claim 1, characterized in that: The early stopping strategy is to terminate training when the validation set loss does not decrease for 15 consecutive rounds.

8. The method for predicting the water content of tunnel rock mass based on a multi-scale residual denoising network according to claim 1, characterized in that: The imaging performance of the model was verified by comparing the imaging effects of the Transformer method, the multi-scale sampling method, and the multi-scale sampling residual denoising network model based on the self-attention mechanism.

Images