Surrounding rock grade identification method based on deep learning and geological radar detection

Abstract

The invention relates to a surrounding rock grade identification method based on deep learning and geological radar detection. The method comprises the following steps: acquiring geological radar data to be identified; the geological radar data to be recognized are input into a surrounding rock grade classification model, a surrounding rock grade prediction result is output, and the surrounding rock grade classification model is obtained by training a pre-trained ResNet-50 model in a transfer learning mode according to a data set. The method can rapidly and accurately predict the grade of the surrounding rock in front of the tunnel excavation face through the geological radar detection result, and improves the tunnel construction efficiency.

Description

Technical Field

[0001] This invention relates to the field of tunnel engineering technology, and in particular to a method for identifying the grade of surrounding rock based on deep learning and ground-penetrating radar detection. Background Technology

[0002] Tunnel construction often faces geological hazards such as water and mud inrush, rock collapse, and large plastic deformation. The mechanisms of these hazards are significantly related to the instability of the surrounding rock structure. To address this, by developing a surrounding rock condition identification technology, the geological parameters of the tunnel face area can be dynamically collected and analyzed in real time, enabling the prediction and prevention of geological hazards and ensuring the smooth progress of tunnel construction.

[0003] However, traditional methods for determining the surrounding rock grade mainly rely on manual on-site observation and experience, which suffers from low efficiency, strong subjectivity, and difficulty in adapting to complex geological environments. While existing technologies have incorporated ground-penetrating radar (GPR) detection, data analysis still relies on manual interpretation, which is time-consuming and easily affected by operator skill. In recent years, although deep learning-based image classification technology has opened new paths for automated identification, its practical application in GPR data analysis still faces multiple technical bottlenecks: GPR images exhibit strong noise interference due to the complex reflective characteristics of geological bodies; limited professional annotation resources lead to insufficient sample diversity in model training, resulting in decreased generalization performance; and existing algorithm architectures have not yet established a precise mapping relationship between the features of the detected data and the actual surrounding rock grade parameters.

[0004] Therefore, proposing a method for identifying the surrounding rock grade based on deep learning and ground-penetrating radar detection, and constructing an end-to-end surrounding rock grade identification model to further improve the objectivity and accuracy of ground-penetrating radar image discrimination is an urgent problem to be solved. Summary of the Invention

[0005] The purpose of this invention is to provide a method for identifying the surrounding rock grade based on deep learning and ground-penetrating radar detection, applicable to the assessment of surrounding rock stability and early warning of geological hazards during tunnel construction.

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

[0007] A method for identifying the grade of surrounding rock based on deep learning and ground-penetrating radar detection includes:

[0008] Acquire the ground-penetrating radar data to be identified;

[0009] The geological radar data to be identified is input into the surrounding rock grade classification model, and the surrounding rock grade prediction result is output. The surrounding rock grade classification model is obtained by training a pre-trained ResNet-50 model based on the dataset through transfer learning.

[0010] Optionally, obtaining the dataset includes:

[0011] Survey lines were laid out in the tunnel test section to collect ground-penetrating radar detection data;

[0012] The ground-penetrating radar data is preprocessed to construct a dataset that includes seismic radar images and corresponding surrounding rock grade labels.

[0013] Optionally, the layout of survey lines in the tunnel test section includes: laying survey lines along the tunnel face and sidewalls, and determining the spacing of the survey lines according to the frequency of the ground-penetrating radar antenna.

[0014] Optionally, preprocessing the ground-penetrating radar detection data includes: denoising the ground-penetrating radar detection data through wavelet transform and histogram equalization.

[0015] Optionally, the seismic radar image and the corresponding surrounding rock grade label are obtained at the same spatial location.

[0016] Optionally, training the pre-trained ResNet-50 model using transfer learning based on the dataset includes: freezing the first 20 convolutional weights of the pre-trained ResNet-50 model and adding Dropout layers, and training the pre-trained ResNet-50 model based on the dataset, wherein the pre-trained ResNet-50 model is obtained through ImageNet pre-training.

[0017] Optionally, training the pre-trained ResNet-50 model using transfer learning based on the dataset also includes training with a focus loss function;

[0018] The focus loss function is:

[0019] FL(p t )=-α t (1-p t ) γ log(p t );

[0020] Among them, FL(p t ) is the focus loss function, α t For the category weight coefficients, γ = 2, p t This represents the model's predicted probability for the correct category.

[0021] Optionally, after outputting the surrounding rock grade prediction result, the following steps are included: comparing the surrounding rock grade prediction result with the expert manual judgment; if the consistency is lower than a preset value, then optimizing the training parameters of the surrounding rock grade classification model, and re-labeling the data to iteratively train the surrounding rock grade classification model.

[0022] The beneficial effects of this invention are as follows: Compared with the prior art, this invention discloses a method for identifying surrounding rock grades based on deep learning and ground-penetrating radar (GPR). Through data acquisition, transfer learning, and dynamic feedback mechanisms, it achieves automated and high-precision grading of surrounding rock grades based on GPR; through dataset preprocessing and focus loss function design, the model still achieves high recognition accuracy under complex geological conditions; using the ResNet-50 classification model, it predicts surrounding rock grades in real time, significantly outperforming manual methods; wavelet transform and histogram equalization techniques ensure the model maintains high robustness in noisy environments; transfer learning strategies reduce the amount of labeled data required, and data augmentation techniques further reduce the dependence on sample size; through equipment consistency constraints and dynamic iteration mechanisms, the model can adapt to the surrounding rock grading requirements of different tunnel engineering projects. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart of a method for identifying the grade of surrounding rock based on deep learning and ground-penetrating radar, according to an embodiment of the present invention.

[0025] Figure 2 This refers to the surrounding rock condition exposed at the tunnel excavation face in an embodiment of the present invention.

[0026] Figure 3 The images show the ground-penetrating radar detection results for Class IV and Class V surrounding rocks in this embodiment of the invention, wherein (a) is the ground-penetrating radar detection result for Class IV surrounding rocks and (b) is the ground-penetrating radar detection result for Class V surrounding rocks.

[0027] Figure 4 The diagram shows a comparison between the actual detection results and the identification results of the method in this embodiment of the invention. In this diagram, (a) is a training / validation set loss curve and (b) is an accuracy curve. Detailed Implementation

[0028] 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.

[0029] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0030] This embodiment provides a method for identifying the surrounding rock grade based on deep learning and ground-penetrating radar detection, including:

[0031] Acquire the ground-penetrating radar data to be identified;

[0032] The geological radar data to be identified is input into the surrounding rock grade classification model, and the surrounding rock grade prediction result is output. The surrounding rock grade classification model is obtained by training a pre-trained ResNet-50 model based on the dataset through transfer learning.

[0033] Furthermore, obtaining the dataset includes:

[0034] Survey lines were laid out in the tunnel test section to collect ground-penetrating radar detection data;

[0035] The ground-penetrating radar data is preprocessed to construct a dataset that includes seismic radar images and corresponding surrounding rock grade labels.

[0036] Furthermore, the layout of survey lines in the tunnel test section includes: laying survey lines along the tunnel face and sidewalls, and determining the spacing of the survey lines according to the frequency of the ground-penetrating radar antenna.

[0037] Furthermore, preprocessing of ground-penetrating radar (GPR) data includes denoising the data through wavelet transform and histogram equalization.

[0038] Specifically, in this embodiment, wavelet transform is used to remove high-frequency noise. The Daubechies wavelet basis function is selected, and the decomposition level is 3 to filter out high-frequency noise while retaining the features of the stratigraphic interface and fractures. Histogram equalization is used to enhance image contrast and highlight the details of the surrounding rock structure.

[0039] Furthermore, seismic radar images and corresponding surrounding rock grade labels are obtained in the same spatial location.

[0040] Specifically, in this embodiment, after excavation to expose the surrounding rock, the actual surrounding rock grade (I-VI) is determined according to the current "Highway Tunnel Design Specification", "Highway Engineering Geological Survey Specification" and the performance of the ground-penetrating radar, and then bound to the radar data of the corresponding survey line segment.

[0041] Furthermore, training the pre-trained ResNet-50 model using transfer learning based on the dataset includes: freezing the first 20 convolutional weights of the pre-trained ResNet-50 model and adding Dropout layers; training the pre-trained ResNet-50 model based on the dataset, wherein the pre-trained ResNet-50 model was obtained through ImageNet pre-training.

[0042] Specifically, this embodiment preserves the ability to extract low-level features by freezing the weights of the first 20 convolutional layers; and adds a Dropout layer to prevent the model from overfitting during training.

[0043] Furthermore, training the pre-trained ResNet-50 model using transfer learning based on the dataset also includes training with a focus loss function;

[0044] The focus loss function is:

[0045] FL(p t )=-α t (1-p t ) γ log(p t );

[0046] Among them, FL(p t ) is the focus loss function, α t For the category weight coefficients, γ = 2, p t This represents the model's predicted probability for the correct category.

[0047] Furthermore, after outputting the surrounding rock grade prediction results, the following steps are taken: the surrounding rock grade prediction results are compared with the expert manual judgment. If the consistency is lower than the preset value, the training parameters of the surrounding rock grade classification model are optimized, and the data is re-labeled and iteratively trained to train the surrounding rock grade classification model.

[0048] The selected ground-penetrating radar model should remain unchanged in the above steps, including the antenna model and sampling frequency, and ensure that the detection depth and resolution meet the requirements for tunnel surrounding rock feature identification.

[0049] The following description, in conjunction with the accompanying drawings, further illustrates this embodiment:

[0050] This embodiment proposes a method for identifying the surrounding rock grade based on deep learning and ground-penetrating radar detection, such as... Figure 1 As shown, the method includes the following steps:

[0051] S1. Select a tunnel test section, lay out survey lines, and collect ground-penetrating radar detection data.

[0052] Survey lines are laid out along the tunnel face and sidewalls. The spacing between the survey lines is determined based on the frequency of the ground-penetrating radar antenna. A uniform model of ground-penetrating radar is used, and the antenna model and sampling frequency must be fixed. During data acquisition, the mileage of the survey lines at the tunnel face must be recorded to ensure that the data matches the subsequent location of the exposed surrounding rock.

[0053] S2. Preprocess the collected ground-penetrating radar images; construct a dataset containing ground-penetrating radar images and corresponding surrounding rock grade labels by using noise reduction methods such as wavelet transform and histogram equalization.

[0054] The Daubechies wavelet basis function was used to perform a three-level decomposition on the original radar image, preserving low-frequency components and filtering out high-frequency noise. The denoised image was then subjected to contrast-limited adaptive histogram equalization (CLAHE) with a window size of 8×8 and a contrast limit of 0.01 to enhance the contrast of weakly reflected signals and highlight the surrounding rock fissures and bedding structure. The preprocessed ground-penetrating radar image was then bound to the exposed surrounding rock grade labels. These labels were independently determined by three experts (grades I-VI) according to the "Highway Tunnel Design Specifications," and the majority result was used as the final label.

[0055] S3. A rock mass classification model is constructed based on the ResNet-50 network. This model is trained using a dataset to obtain a prediction model. The network input goes through 5 stages to obtain the output. Stage 0 can be regarded as the preprocessing of the input. The last 4 stages (Stage 1 to Stage 5) are all composed of Bottlenecks. Stage 1 contains 3 Bottlenecks, and the remaining 3 stages contain 4, 6, and 3 Bottlenecks (BTNK) respectively.

[0056] For model training, first set up the initial configuration file for the ResNet-50 model, including training parameters, loss function, etc. Use the focal loss function, with the parameter set to α. t =0.25, γ=2, focusing on hard-to-classify samples. A Dropout layer (rate=0.5) is added before the fully connected layer to prevent overfitting.

[0057] S4. Input the geological radar data to be identified into the prediction model for reasoning, output the surrounding rock grade prediction results, and combine them with manual judgment to carry out anomaly warning and construction plan adjustment.

[0058] The radar image to be identified is input into the prediction model, which outputs the probability distribution of each level. The level corresponding to the maximum value is taken as the prediction result.

[0059] If the consistency between the prediction results and expert judgment is less than 90%, then optimization of the model training parameters should be considered, such as reducing the number of frozen layers and lowering the model training learning rate.

[0060] This embodiment relies on the ground-penetrating radar detection results of the exposed surrounding rock and Class IV and V surrounding rock collected from on-site engineering, such as... Figure 2 , Figure 3 As shown in (a)-(b), these results are compared with the identification results of the method in this embodiment, as follows: Figure 4 As shown in (a)-(b), the actual detection results are consistent with the identification results of this invention, demonstrating the superiority of the method in this embodiment.

[0061] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for identifying the grade of surrounding rock based on deep learning and ground-penetrating radar detection, characterized in that, include: Acquire the ground-penetrating radar data to be identified; The geological radar data to be identified is input into the surrounding rock grade classification model, and the surrounding rock grade prediction result is output. The surrounding rock grade classification model is obtained by training a pre-trained ResNet-50 model based on the dataset through transfer learning.

2. The method for identifying the surrounding rock grade based on deep learning and ground-penetrating radar detection according to claim 1, characterized in that, Obtaining the dataset includes: Survey lines were laid out in the tunnel test section to collect ground-penetrating radar detection data; The ground-penetrating radar data is preprocessed to construct a dataset that includes seismic radar images and corresponding surrounding rock grade labels.

3. The method for identifying the grade of surrounding rock based on deep learning and ground-penetrating radar detection according to claim 2, characterized in that, The layout of survey lines in the tunnel test section includes: laying survey lines along the tunnel face and sidewalls, and determining the spacing of the survey lines according to the frequency of the ground-penetrating radar antenna.

4. The method for identifying the surrounding rock grade based on deep learning and ground-penetrating radar detection according to claim 2, characterized in that, The preprocessing of the ground-penetrating radar detection data includes: denoising the ground-penetrating radar detection data through wavelet transform and histogram equalization.

5. The method for identifying the grade of surrounding rock based on deep learning and ground-penetrating radar detection according to claim 2, characterized in that, The seismic radar images and corresponding surrounding rock grade labels are obtained in the same spatial location.

6. The method for identifying surrounding rock grade based on deep learning and ground-penetrating radar detection according to claim 1, characterized in that, Training a pre-trained ResNet-50 model using transfer learning based on a dataset includes: freezing the first 20 convolutional weights of the pre-trained ResNet-50 model and adding a Dropout layer; training the pre-trained ResNet-50 model based on the dataset; wherein the pre-trained ResNet-50 model was obtained through ImageNet pre-training.

7. The method for identifying the grade of surrounding rock based on deep learning and ground-penetrating radar detection according to claim 1, characterized in that, Training a pre-trained ResNet-50 model using transfer learning based on a dataset also includes training with a focus loss function. The focus loss function is: FL(p t )=-a t (1-p t ) γ log(p t ); Among them, FL(p t ) is the focus loss function, α t For the category weight coefficients, γ = 2, p t This represents the model's predicted probability for the correct category.

8. The method for identifying surrounding rock grade based on deep learning and ground-penetrating radar detection according to claim 1, characterized in that, After outputting the surrounding rock grade prediction results, the following steps are taken: the surrounding rock grade prediction results are compared with the expert manual judgment. If the consistency is lower than the preset value, the training parameters of the surrounding rock grade classification model are optimized, and the data is re-labeled and iteratively trained to train the surrounding rock grade classification model.

Images