Tunnel surrounding rock dynamic grading method based on geological radar and transfer learning

Abstract

The invention belongs to the field of tunnel surrounding rock investigation, and relates to a tunnel face image and tunnel face front rock mass information are obtained by taking surrounding rock in front of a tunnel face in a tunnel construction process as a research object and adopting a mobile photographing device and a geological radar advanced geological forecast technology. And a deep learning method is adopted to perform feature extraction and recognition on the tunnel face image and the geological radar map, so that the tunnel surrounding rock dynamic grading method based on the geological radar and transfer learning is established, and rapid and accurate surrounding rock dynamic grading on the tunnel face and the surrounding rock in front in the tunnel construction excavation process is realized.

Description

Technical Field

[0001] The present invention belongs to the field of tunnel surrounding rock survey, and in particular relates to a tunnel surrounding rock dynamic classification method based on geological radar and transfer learning. Background Art

[0002] Tunnel surrounding rock quality classification is a method to reflect the geological conditions of the surrounding rock during geological survey. The surrounding rock is tested by instruments and equipment to obtain several main evaluation indicators. Rock masses with different geological conditions but similar stability are classified into the same grade, and unlimited rock masses are divided into limited grades with different stability. Accurate surrounding rock quality classification plays a vital role in guiding tunnel excavation methods, support structure design, safety and stability of the construction process, and cost control. If the surrounding rock classification results are inconsistent with the actual situation, it may lead to the tunnel support structure design not meeting the actual requirements of the project, posing a safety risk in construction; or the excavation plan and support design are too conservative, increasing construction costs. However, in the process of tunnel geological survey, it is difficult to accurately grasp the overall situation of the tunnel surrounding rock geological environment due to the limitations of the survey technology level and complex geological conditions. In addition, the poor geology of tunnels such as faults and fracture zones, karst areas, and water-rich fracture zones has a certain degree of concealment, making it difficult to accurately reveal the poor geological conditions of deep and long mountain tunnels. According to statistics, in the process of engineering geological survey of mountain tunnels in my country, about 35% of the surrounding rock classification results are inconsistent with the actual situation. Therefore, during the tunnel excavation process, accurately exploring the changes in rock properties in front of the heading face and efficiently and reasonably assessing the surrounding rock grade are of great practical significance for guiding the further construction plan of the tunnel and improving the support design.

[0003] At present, there are two main methods for surrounding rock classification during the tunnel construction stage: one is to evaluate the surrounding rock level by observing the tunnel face after tunnel excavation; the other is to use tunnel advanced geological prediction technologies such as seismic reflection wave method and geological radar method to measure and evaluate the grade of the surrounding rock in front of the tunnel face. However, no matter which method is used to classify the surrounding rock, due to the different work experience and understanding level of each geological surveyor, when they use personal subjective factors to classify the same surrounding rock grade, even if the same set of evaluation standards are used, different conclusions may be drawn. This uncertainty is particularly significant under the influence of factors such as increasingly complex geological environment and instrument and equipment operation errors. In addition, the traditional tunnel surrounding rock classification method requires a lot of manpower and material resources, which is not conducive to saving construction time and economic costs.

[0004] In recent years, with the improvement of computer computing performance and the enhancement of data training capabilities, deep learning in the field of artificial intelligence machine learning has begun to receive widespread attention. Due to its good generalization ability and robustness, deep learning has been widely used in fields such as computer vision. Deep learning is a technology that uses computer neural networks for analysis and learning. It achieves the effect of data classification by inputting data into convolutional neural networks for automatic training and feature extraction. It does not require manual construction of image features, but outputs results in an end-to-end manner. Chen Peishuai et al. used deep learning convolutional neural networks to locate and predict water-rich fracture zone information in tunnel advance geological forecasts, providing a basis for engineering. Liu Houxiang et al. used the deep learning AlexNet model to identify the rock features of the tunnel face, and used color models and digital image processing methods to determine the rock lithology and the development of rock joints and fissures, thereby establishing a surrounding rock intelligent classification system.

[0005] At the same time, China has published the following related patents:

[0006] 1. A method and device for intelligent grading of tunnel face surrounding rock based on deep learning, application (patent) number: CN202111481523.9. The invention first labels and preprocesses the tunnel face surrounding rock photo samples to establish a trained face image database; then, a surrounding rock grading prediction model based on deep learning is built, and the model is trained using the face image database, thereby constructing a prediction model with more accurate recognition effect. In addition, the present invention uses image cutting technology to segment the face photos to be detected and independently predict each sub-image. Finally, a comprehensive evaluation and prediction is performed based on the prediction results of each sub-image, thereby achieving efficient and accurate discrimination of the face surrounding rock.

[0007] 2. A cloud computing-based highway tunnel surrounding rock classification method, application (patent) number: CN201911133442.2. Aiming at the characteristics of highway tunnel surrounding rock mass properties and occurrence conditions are very complex, the number of parameter tests is limited, and the classification is difficult to implement, the present invention develops a cloud-based classification method. The method of obtaining the parameters of the degree of development of the surrounding rock structure surface by digital photography and deep learning technology can greatly reduce the time of obtaining surrounding rock classification indicators during the construction stage. At the same time, the rock mass strength is determined by rebound test, and the integrity of the rock mass is determined by cloud computing technology, which improves the reliability of surrounding rock classification indicators during the construction stage. The tunnel face pictures can be saved in the cloud, which is convenient for reference, communication, verification and collaboration of similar projects.

[0008] Although the above method is an improvement over the previous method, the efficiency and accuracy of dynamic classification of the tunnel face and the surrounding rock in front during tunnel construction and excavation need to be further improved. The existing patent only uses photos of the tunnel face, which makes it difficult to achieve three-dimensional prediction of geological radar.

[0009] Therefore, the present invention provides a method for dynamic classification of tunnel surrounding rock based on geological radar and transfer learning, which combines the tunnel face photos with the geological radar images 20m in front of the tunnel face for prediction and forecasting, and can realize the three-dimensional prediction of geological radar. Summary of the invention

[0010] In view of this, the purpose of the present invention is to provide a method for dynamic classification of tunnel surrounding rock based on geological radar and transfer learning. The present invention aims to solve the problem that the existing method only uses the tunnel face photos, does not consider the geological radar images within 20m in front of the tunnel face, and does not achieve true prediction.

[0011] To achieve the above object, the present invention provides a tunnel surrounding rock dynamic classification method based on geological radar and transfer learning, comprising the following steps:

[0012] S1. Image acquisition: Use a mobile camera to take pictures of the tunnel face to obtain photos of the tunnel face;

[0013] S2. Extract the surface characteristics of the tunnel surrounding rock;

[0014] S2.1 performs data enhancement on the captured tunnel face image;

[0015] S2.2 Create a characterization data set: classify and annotate tunnel face photos according to rock lithology, weathering degree, water content, rock mass structure type, and fracture development degree;

[0016] S2.3 improves the MobileNet network model and builds the MobileNetV3 convolutional neural network model;

[0017] S2.4 trains a MobileNetV3 convolutional neural network model, and then uses the MobileNetV3 convolutional neural network model to detect the rock lithology, weathering degree, water content, rock structure type and fracture development degree of the tunnel face;

[0018] S3. Identify and quantify weak interlayers and cracks in the tunnel face;

[0019] S3.1 Create a weak interlayer dataset: perform data expansion and screening on the captured tunnel face images, and then use the deep learning semantic segmentation and annotation tool Labelme to annotate the weak interlayer zones in the images, that is, outline the locations of the weak interlayers in the images;

[0020] S3.2 Build semantic segmentation network model: Unet network model;

[0021] S3.3 trains the Unet network model, and then uses the Unet network model to identify and extract the weak interlayer of the tunnel face;

[0022] S3.4 Identification and quantification of rock cracks at the tunnel face;

[0023] S3.5 Calculate the rock crushing coefficient and evaluate the rock integrity of the tunnel face;

[0024] S4. Based on the rock information of the tunnel face obtained in steps S1 to S3, extract and convert the surrounding rock classification index to obtain the surrounding rock classification result of the current tunnel face;

[0025] S5. Geological radar advanced geological forecast;

[0026] S5.1 During tunnel excavation, advance geological prediction is carried out on the tunnel face to obtain radar images of the geology in front of the face;

[0027] S5.2 Compare the interpretation results according to the tunnel excavation situation, annotate the confirmed results in the geological radar image, and then train the deep learning target detection network model of the geological radar;

[0028] S6. Extract the surrounding rock features in front of the tunnel face based on deep learning, and compare and verify them with the actual situation to continuously improve the interpretation accuracy.

[0029] Furthermore, the data enhancement methods include geometric transformation, clarity transformation, color gamut conversion and noise addition;

[0030] Geometric transformation: Process the image through one or more operations such as rotation, flipping, translation, stretching, and mirroring, thereby simulating a mobile camera device taking photos of the tunnel face from different angles;

[0031] Clarity transformation: blur the image through median filtering / mean filtering / Gaussian filtering;

[0032] Color gamut conversion: By adjusting the brightness of the tunnel face image, the different light and dark levels of the tunnel face can be simulated;

[0033] Add Noise: Add salt and pepper noise or Gaussian noise to simulate dust in the air.

[0034] Furthermore, in step S2.3, the improvement of the MobileNetV3 convolutional neural network model includes:

[0035] (1) Introducing the reverse residual structure;

[0036] (2) Add lightweight SE module;

[0037] The SE module consists of two parts: compression and excitation. The compression part performs global average pooling on the output feature matrix of the deep convolution, and the excitation part outputs the compressed feature matrix after passing it through two fully connected layers.

[0038] (3) Improve the activation function;

[0039] The original activation function of the MobileNet network model is improved, and the H-swish activation function is proposed. The calculation expression is as follows:

[0040] ReLU6(x)=min(max(x,0),6)

[0041] H-swish(x) = x ReLU6(x+3) / 6

[0042] Where x is the input of the nonlinear activation function in the i-th channel, and the ReLU6 function is the ReLU function with a maximum output limit of 6;

[0043] (4) Optimize network structure.

[0044] Furthermore, the step S3.4, identifying and quantifying rock cracks at the tunnel face, includes the following sub-steps:

[0045] I. Use image segmentation to identify and extract cracks in the rock image of the tunnel face;

[0046] II. Based on the distance between the camera and the rock on the tunnel face, or the presence of a reference object with known geometric parameters on the tunnel face, the actual size of the fracture is calculated according to the pixel size of the target area in the image and the corresponding ratio;

[0047] III. Calculate the crack length, crack width, crack area and crack ratio;

[0048] Crack length: Use the Canny operator to extract the edge of the image after rock crack threshold segmentation, and perform skeleton processing on the rock crack edge detection image, and make an approximate calculation based on the length of the two edges of the crack divided by 2;

[0049] Crack area: based on the number of black pixels in the crack binary image, obtained by proportional conversion;

[0050] Crack width: calculated by approximation, the crack area divided by the crack length;

[0051] Crack ratio: expressed as the numerical value of surface crack ratio;

[0052] The face crack ratio is expressed by the ratio of the total area of ​​all cracks on the rock surface of the tunnel face to the corresponding rock area.

[0053] Furthermore, the image segmentation method is an inter-class maximum distance method.

[0054] Furthermore, in step S3.4, before image segmentation, the grayscale value of the tunnel face rock image is adjusted to within 30% of the original image.

[0055] Furthermore, in step S3.5, the calculation expression of the rock crushing coefficient is as follows:

[0056]

[0057] Where: A is the total area of ​​all cracks in the rock image; w and h are the width and height of the image; b is the average width of the rock cracks; K p is the crushing coefficient of rock.

[0058] Furthermore, in step S3.5, the evaluation relationship between the rock crushing coefficient and the integrity of the rock at the tunnel face is as follows:

[0059] K p ∈0.01~0.04, indicating that the rock on the face is relatively intact; K p ∈0.04~0.08, indicating that the rock on the face is relatively broken; K p >0.08, indicating that the rock at the tunnel face is broken / extremely broken.

[0060] Further, the specific steps of step S4 are as follows:

[0061] S4.1 Based on the rock lithology and weathering degree information obtained in step S2, the range of uniaxial saturated compressive strength of various types of rocks at different weathering degrees is obtained according to the specification, and then the median value of the range is used as the rock strength index of the tunnel face;

[0062] S4.2 Based on the rock fracture development degree and rock mass structure type obtained in step S2, and the fine fractures in the rock of the tunnel face quantitatively identified in step S3, the rock integrity index is determined in proportion according to the rock crushing coefficient and the corresponding classification interval;

[0063] S4.3 based on the water content obtained in step S2, obtain a water content adjustment coefficient;

[0064] S4.4 Use a geological compass to measure the inclination of the rock structure surface at the tunnel face and its angle with the tunnel axis to obtain the structural surface attitude adjustment coefficient;

[0065] The S4.5 rock strength index, rock integrity index, water content adjustment coefficient and structural surface attitude adjustment coefficient are calculated using the BQ surrounding rock classification calculation formula to obtain the surrounding rock quality index, and then the surrounding rock classification result of the current tunnel face is obtained based on the surrounding rock classification standard.

[0066] Further, the specific steps of step S5 are as follows:

[0067] S6.1 Geological radar image acquisition and interpretation;

[0068] S6.2 Prepare geological radar signal dataset;

[0069] S6.3 Build the MobileNet-SSD target detection model;

[0070] Based on the SSD model, the MobileNet lightweight convolutional neural network is introduced as the backbone feature extraction network of the SSD model to build the MobileNet-SSD target detection model.

[0071] S6.4 trains the MobileNet-SSD target detection model, and then uses the MobileNet-SSD target detection model to extract the surrounding rock features in front of the tunnel face.

[0072] The beneficial effects of the present invention are:

[0073] This paper adopts the deep learning method, combines the geological conditions of the tunnel face at the construction site with the results of geological radar advanced geological prediction, and proposes a method for dynamic classification of surrounding rock during tunnel construction. It has important research value and engineering significance, which is mainly reflected in the following points:

[0074] (1) In theory, by combining advanced geological prediction technology with artificial intelligence technology, the surrounding rock in front of the tunnel face during tunnel construction is graded, and multidisciplinary cross-integration exploratory and cutting-edge research is carried out, which is conducive to promoting the development of advanced geological prediction technology towards intelligence, and at the same time, it also broadens the application scope of artificial intelligence in the field of tunnel engineering.

[0075] (2) In terms of engineering significance, the deep learning method is used to extract tunnel surrounding rock classification indicators, making the tunnel geological survey and surrounding rock classification process more convenient, accurate and intelligent, reducing the influence of subjective factors in the tunnel geological survey and surrounding rock classification process, improving work efficiency and reducing economic costs.

[0076] The present invention takes the surrounding rock in front of the tunnel face during tunnel construction as the research object, obtains the tunnel face image and the rock mass information in front of the tunnel face by adopting mobile photographing equipment and geological radar advanced geological prediction technology, and uses deep learning method to extract and identify the features of the tunnel face image and geological radar atlas, thereby establishing a surrounding rock classification method based on geological radar advanced geological prediction technology and transfer learning, and realizing fast and accurate dynamic classification of the tunnel face and the surrounding rock in front during tunnel construction and excavation.

[0077] The dynamic here means that every 20m, a geological radar advance geological prediction is required to obtain image data and update the network. The surrounding rock in front refers to the geological conditions of the 20m unexcavated rock mass in front of the tunnel face that can be predicted by the geological radar advance geological prediction technology. A three-dimensional prediction of the geological conditions is achieved, including the tunnel face and the 20m in front. The bottom area is the size of the tunnel face, and the height is 20m in front of the tunnel face.

[0078] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] Figure 1 This is a flow chart of a method for dynamic classification of tunnel surrounding rock based on geological radar and transfer learning according to the present invention;

[0080] Figure 2 It is the module structure of the MobileNetV3 network model;

[0081] Figure 3 This is the training process of the MobileNetV3 network model;

[0082] Figure 4 It is the Unet network model structure;

[0083] Figure 5 It is the Unet network model training process;

[0084] Figure 6 Weak interlayer at the tunnel face used for Unet network model testing;

[0085] Figure 7 Identification results for the Unet network model;

[0086] Figure 8 This is a comparison chart of rock fracture images after contrast adjustment and threshold segmentation;

[0087] Fig. 9This is the MobileNet-SSD model training process;

[0088] Fig.10 This is the surrounding rock image sample of the YK18+048 face of Shitang Tunnel;

[0089] Fig.11 This is the geological radar image of Shitang Tunnel YK18+048~YK18+018. DETAILED DESCRIPTION

[0090] In order to make the technical solutions, advantages and purposes of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the described embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work belong to the protection scope of this application.

[0091] like Figure 1 As shown, the present invention provides a tunnel surrounding rock dynamic classification method based on geological radar and transfer learning, comprising the following steps:

[0092] S1. Image acquisition: Use a mobile camera to take pictures of the tunnel face to obtain photos of the tunnel face;

[0093] The photos of the tunnel faces used in this embodiment are taken from actual on-site projects, including the Jiangjunshan Tunnel in Liannan, Guangdong, the Kuiba Road Tunnel in Shenzhen, Guangdong, the Lingsha Road Tunnel in Wenzhou, Zhejiang, a tunnel in Taojiang, Hunan, a tunnel in Xincheng, Guangxi, and a tunnel in Rucheng, Hunan. After each section of the tunnel is excavated, a smartphone is used to photograph the tunnel face. After the tunnel is blasted and excavated and the gravel is discharged, the on-site construction personnel will conduct a safety inspection of the tunnel face and perform ventilation and dust reduction treatment at the same time. The best time to shoot is after the tunnel is ventilated and before the initial support concrete is sprayed on the tunnel surrounding rock. At this time, the dust content in the tunnel environment is low, and the surrounding rock of the face is still in its original state;

[0094] S2. Extract the surface characteristics of the tunnel surrounding rock;

[0095] S2.1 performs data enhancement on the captured tunnel face image;

[0096] The deep learning convolutional neural network model requires a large number of samples during the sample learning training process to improve the learning effect of the model, increase the robustness of the model and prevent overfitting. However, in actual engineering shooting, the number of images obtained is often limited. If these small amounts of image data are directly used for training, the model may have low accuracy and insufficient generalization ability.

[0097] Therefore, data enhancement of tunnel face images before model training is an effective strategy that helps solve the above problems. Data enhancement is to increase the diversity and richness of the data set through specific algorithms and techniques, so that the model can learn and generalize better. In the data enhancement of tunnel face images, it can be divided into geometric transformation, clarity transformation, color gamut conversion and noise addition. Geometric transformation mainly includes operations such as rotation, flipping, translation, stretching and mirroring. By processing images through geometric transformation, the diversity of image data can be increased and the situation of smartphones shooting the face from different angles can be simulated. Clarity transformation includes median filtering, mean filtering and Gaussian filtering, etc., which mainly blur the image to improve the robustness of the model. Color gamut transformation is to simulate the different brightness and darkness of the tunnel face in actual engineering by adjusting the brightness of the face image. Adding noise is divided into salt and pepper noise and Gaussian noise. Since there may be a lot of dust in the air during tunnel excavation, adding noise processing to the face photos appropriately can make the model closer to the real tunnel excavation environment and improve the accuracy and generalization ability of the model. Each type of tunnel face image dataset is screened, and the tunnel face images with good quality are retained and data enhancement is performed.

[0098] S2.2 Create a characterization data set: classify and annotate tunnel face photos according to rock lithology, weathering degree, water content, rock mass structure type, and fracture development degree;

[0099] According to factors such as rock lithology, weathering degree, water content, rock structure type and degree of crack development of the tunnel face photos, each photo is classified and annotated. The collected tunnel face images include different rock types such as dolomitic limestone, slate, brecciated glassy tuff, flint nodule-bearing limestone, potassium feldspar granite, tuff, sandstone and limestone; according to the rock structure type, it can be divided into overall block structure, layered structure, fragmented structure and loose structure. According to the degree of weathering of the rock mass, it can be divided into strong differentiation, medium weathering and slightly weathering. For the degree of development of joints and fissures, it is divided into very developed fissures (extremely broken), relatively developed fissures (relatively broken), developed fissures (broken) and less developed fissures (relatively complete). The water content of the face is divided into dry, infiltrated and flowing water. When annotating each face photo, the results of the geological sketch and geological survey report shall prevail, and relevant specifications and expert guidance shall be followed to ensure the accuracy of the annotation of each photo.

[0100] In addition, if the number of images in each category in the data set differs too much, it will have an adverse effect on the training effect of the convolutional neural network model. Therefore, in the process of data enhancement, the number of images in each category must be controlled to ensure the balance of model training. In order to meet the requirements of the convolutional neural network model for the input image size, the images in each category of the data set will be compressed to improve the learning and training efficiency of the model for samples. For each category of tunnel face image data set, 70% of the images will be randomly selected as training sets, 20% of the images will be used for verification during the training process, and the remaining images will be used as test sets to test the recognition effect of the model.

[0101] S2.3 improves the MobileNet network model and builds the MobileNetV3 convolutional neural network model;

[0102] The present invention uses the MobileNetV3 convolutional neural network model to carry out feature extraction and intelligent recognition research on the tunnel face. The network model is partially improved on the basis of the MobileNet network model, making the model more lightweight and enhancing the computing power, while improving the recognition performance. The MobileNetV3 network model is mainly improved in the following aspects:

[0103] (1) Propose a reverse residual structure. The MobileNetV3 network model borrows the residual idea from the ResNet convolutional neural network model and makes improvements on it. During the convolution process, when the distance the filter moves each time is 1 and the input and output feature matrices have the same size, the reverse residual structure is used. This structure first lifts the input feature matrix to a high-dimensional space through a dimensional convolution with a convolution kernel size of 1×1, and then continues to use a depth-wise separable convolution operation. Through the reverse residual structure, the MobileNetV3 network model enhances the expressive power of low-dimensional features.

[0104] (2) Adding a lightweight SE module. After deep convolution, the MobileNetV3 network model adds a SE module, which consists of two parts: compression (Squeeze) and excitation (Excitation). The compression part performs global average pooling on the output feature matrix of deep convolution, and the excitation part outputs the compressed feature matrix through two fully connected layers. The SE module focuses on each channel of the output feature and assigns different weights according to the importance of different channels.

[0105] (3) Improved activation function. In the SE module, the MobileNetV3 network model improves the original activation function and proposes the H-swish activation function. This function helps to simplify the calculation process and improve the computing performance of the network model. Its calculation formula is as follows:

[0106] ReLU6(x)=min(max(x,0),6)

[0107] H-swish(x) = x ReLU6(x+3) / 6

[0108] Where x is the input of the nonlinear activation function in the i-th channel, and the ReLU6 function is the ReLU function with a maximum output limit of 6;

[0109] (4) Optimize the network structure. The MobileNetV3 network model analyzes and calculates the operation time of each layer through neural network architecture search, and simplifies and optimizes redundant and time-consuming layers. The module structure of the MobileNetV3 network model is as follows: Figure 2 shown.

[0110] S2.4 trains a MobileNetV3 convolutional neural network model, and then uses the MobileNetV3 convolutional neural network model to detect the rock lithology, weathering degree, water content, rock structure type and fracture development degree of the tunnel face;

[0111] When using the MobileNetV3 network model to train tunnel face images, the training parameters were adjusted several times to achieve the best training effect. The final training parameters are as follows: the batch size is 32, the number of training iterations is 20, the parameter γ of the loss function is 2, the value of α is 1, and the learning rate of the optimizer is 0.005. In addition, in order to prevent overfitting, 20% of the neurons in the network model are randomly turned off. The accuracy and loss value changes of the MobileNetV3 network model in the training process of the tunnel face rock lithology, rock structure type, weathering degree, fracture development degree and water content data set are shown in Figure 3 .

[0112] Depend on Figure 3It can be seen that the MobileNetV3 network model has been successfully trained, and the training accuracy and loss of different characteristic data sets of the tunnel face have stabilized. Specifically, the training accuracy of the MobileNetV3 network model for rock lithology, weathering degree and water content data sets is as high as 100%, and the verification accuracy during the training process is 99.6%, 99.4% and 98.2%, respectively. In addition, although the training accuracy of the rock structure type and fracture development degree data sets did not reach 100%, it also reached 99.8% and 99.9%, respectively, and the verification set accuracy was 95.7% and 96.1%, respectively. These data show that the MobileNetV3 network model has a good recognition effect on the above data sets, and has a good recognition ability for the rock lithology, water content, weathering degree, fracture development degree and rock structure type of the tunnel face. The trained MobileNetV3 network model was tested on the rock lithology, rock structure type, fracture development degree, weathering degree and water content data sets of the tunnel face. In order to comprehensively evaluate the overall recognition results of the MobileNetV3 network model for various test sets, the macro-average calculation of precision, recall rate and F1 score of different indicators of each test set was performed during the test. The test results of the MobileNetV3 network model are shown in Table 1.

[0113] Table 1 MobileNetV3 network model test results

[0114]

[0115]

[0116] S3. Identify and quantify weak interlayers and cracks in the tunnel face;

[0117] S3.1 Create a weak interlayer dataset;

[0118] During the cracking process of geological structures, if significant displacement occurs on both sides of the fracture surface, a fault zone geological structure is formed. Under relative movement, weathering and dissolution, the fault may form a weak interlayer. During the tunnel construction process, when encountering such adverse geological phenomena, it is necessary to adjust the tunnel excavation and support design according to its location, scale and nature. Therefore, the present invention uses a deep learning semantic segmentation method to identify and extract the weak interlayer belt in the tunnel face for the weak interlayer belt existing in the supporting project. A smart phone is used to shoot the weak interlayer that appears during the tunnel excavation process, and the captured images are expanded and screened for data, and finally 183 image data samples are obtained, and the data samples are randomly divided into training samples and test samples at a ratio of 9:1. The deep learning semantic segmentation annotation tool Labelme is used to annotate the weak interlayer belt in the sample, that is, to outline the position of the weak interlayer in the picture.

[0119] S3.2 build semantic segmentation network model;

[0120] In the study of identifying and extracting soft interlayers in tunnel faces, the present invention adopts the Unet network model. The network model consists of an encoder and a decoder, and is distributed in a symmetrical manner. A jump connection is used between the encoder and the decoder, so that the network model has an overall U-shaped structure. This structure can effectively integrate shallow spatial information with deep semantic features, and has a good effect on the semantic segmentation of small data sets.

[0121] In the encoder structure, the Unet network model extracts features from the input image and performs downsampling operations through two convolutional layers with a convolution kernel size of 3×3 and an activation function of ReLU, and a maximum pooling layer with a pooling kernel size of 2×2. Each downsampling operation doubles the number of channels of the input feature map. This process is performed a total of 4 times, and finally the high-resolution feature map is converted into a low-resolution feature map. In the decoder structure, an upsampling operation is performed through a deconvolution with a convolution kernel size of 2×2, two convolutional layers with a convolution kernel size of 3×3, and feature concatenation. After each upsampling operation, the number of channels of the feature map is halved. After 4 upsamplings and corresponding downsamplings, the features are fused, so that the low-resolution feature map is restored to a high resolution, and the image segmentation result is obtained. Finally, a convolution layer with a convolution kernel size of 1×1 is used to reduce the dimension and output the final target result. See the Unet network model structure. Figure 4 .

[0122] S3.3 trains the Unet network model, and then uses the Unet network model to identify and extract the weak interlayer of the tunnel face;

[0123] The Unet network model was used to train the soft interlayer dataset of the tunnel face. During the training process, some training parameters are as follows: the number of iterations is 100, the batch size is 4, the initial learning rate is 0.00001, and the learning rate is reduced by 20% in each iteration to enable the network model to better extract features and improve recognition accuracy. Momentum is a hyperparameter used to update the network model parameters, which is set to 0.9 here. When the network weights are updated during model training, the RMSProp optimizer is used to optimize the loss. The optimizer calculates the cumulative gradient through an exponentially weighted sliding average algorithm, which solves the problem of difficulty in training convergence due to the continuous attenuation of the learning rate being too low during the training process. The training and verification loss value results of the Unet network model after each round of training are retained, and the change in loss value during the training process can be seen in Figure 5 .

[0124] After 100 rounds of iterative calculations, the loss value has converged. The trained Unet network model is used to test the data on the test set. In addition to precision and recall, the average intersection over union (MIoU) and mean pixel accuracy (MPA) indicators are added to evaluate the test results of the network model. The average intersection over union refers to the ratio of the intersection and union of the true value and the predicted value of the two images. The average pixel accuracy is the ratio of the number of accurately identified pixels in each category among all categories. The calculation formulas for the average intersection over union and the average pixel accuracy are as follows:

[0125]

[0126]

[0127] Where: k is the number of categories; p ij It is represented by the number of the jth class predicted as the ith class. The test results of the Unet network model on the test set are shown in Table 2.

[0128] Table 2 Unet network model test results

[0129]

[0130] The test results show that the Unet network model has an accuracy of 0.95, a recall rate of 0.83, an IoU ratio of 0.80, and a pixel accuracy of 0.83 for soft interlayers. Overall, the average IoU ratio and average pixel accuracy of the Unet network model are 0.88 and 0.91, and the average accuracy and average recall rate are 0.96 and 0.91. In order to intuitively display and evaluate the recognition results of the Unet network model, some image samples are randomly selected from the soft interlayer test set for testing. The Unet network model is used to identify the following: Figure 6 The weak interlayer is identified, and the identification results are shown in Figure 7 .

[0131] from Figure 7 It can be seen that the Unet network model can identify and extract the weak interlayer in the tunnel face. However, due to the lack of training data and the complexity of the shape of the weak interlayer, the average intersection ratio and average pixel accuracy of the Unet network model are only about 0.9, which means that the model is not accurate enough in extracting the contour of the weak interlayer. In addition, the clay filled in the interlayer is disturbed and overflows during excavation, which will interfere with the recognition results of the model and cause misjudgment in some areas. In general, the Unet network model can identify the shape of the weak interlayer to a certain extent, meeting the requirements for rapid identification and quantitative analysis of the weak interlayer in engineering.

[0132] S3.4 Identification and quantification of rock cracks at the tunnel face;

[0133] When the rock cracks under the action of external force but no obvious displacement occurs, it is called a fissure. Since the light in the fissure part of the face rock is darker than that in other parts, the pixel grayscale value here is the lowest in the whole picture. Therefore, the traditional image segmentation method can be used to identify and extract the fissures in the face rock image. Threshold segmentation is a segmentation method based on image region, which can be divided into single threshold segmentation and multi-threshold segmentation according to the number of types of regions of interest in the image. The basic principle of threshold segmentation is to divide a single or multiple grayscale thresholds according to the grayscale characteristics of the region of interest in the image, and compare the grayscale values ​​of all pixels in the image with the divided grayscale thresholds, and represent the part below the grayscale threshold in black, and the rest in white, so as to realize the extraction of the region of interest. At present, there are many commonly used threshold segmentation methods in image processing, such as iterative threshold segmentation method, histogram bimodal method, maximum distance between classes method, etc. After processing the image with multiple threshold segmentation methods and comparing the results, the present invention finally selects the maximum distance between classes method to perform threshold segmentation processing on the face rock image to extract the fissure characteristics therein. However, rocks in their natural state may be weathered, or the mineral components in the rocks may undergo chemical reactions such as oxidation and silicification, which may cause the rock to appear black in some parts, which will have a negative impact on the effect of threshold segmentation. Therefore, before threshold segmentation, it is necessary to adjust the contrast of the image and control the grayscale value range of the adjusted image to within 30% of the original image, thereby weakening the interference of some black substances in the rock. The results of contrast adjustment and threshold segmentation of rock crack images are shown in Figure 2. Figure 8 .

[0134] The integrity of the rock at the tunnel face is an important indicator for evaluating the grade of the surrounding rock. When calculating the geometric characteristics of rock cracks, the following method is adopted: First, when photographing the rock at the tunnel face to obtain samples, it is necessary to ensure that the camera lens is parallel to the rock at the face. Then, by determining the distance between the camera and the rock at the face, or using a reference object with known geometric parameters in the face, the actual size is converted according to the pixel size of the target area in the image according to the corresponding ratio. When analyzing the geometric characteristics of rock cracks at the face, the main calculation parameters are: crack length, crack width, crack area and crack ratio. In order to obtain these parameters, the number of corresponding black pixels in the crack binary image is counted to obtain the geometric information of the cracks in the image.

[0135] When calculating the crack length in the image, the Canny operator is used to extract the edge of the image after the rock crack threshold segmentation, and the rock crack edge detection image is skeletonized, and the approximate calculation is performed based on the length of the two edges of the crack divided by 2. The area of ​​the crack is calculated as the number of black pixels in the crack binary image, and the actual area of ​​the crack is obtained by proportional conversion. The crack width is the opening degree of the crack structure surface. Since the crack opening degree is small in reality, when calculating the crack width, the crack area is divided by the length through approximate calculation. The crack ratio is divided into line crack ratio, surface crack ratio and body crack ratio, which is one of the important indicators for evaluating the development degree of rock cracks. Among them, the surface crack ratio is expressed by the ratio of the total area of ​​all cracks on the rock surface of the tunnel face to the corresponding rock area. In the rock photos of the tunnel face, the calculation index of the surface crack ratio is easy to obtain and has better applicability, while the crack depth of the body crack ratio is difficult to determine, so the crack ratio of the rock is expressed by the numerical value of the surface crack ratio. In addition, the size of the surface crack ratio is calculated by the ratio of the number of crack pixels in the image to the pixel size of the image. The crack length, crack width, crack area and crack ratio indicators are used to quantify the rock cracks extracted above, and the calculation results are shown in Table 3.

[0136] Table 3 Quantitative calculation results of rock fractures

[0137]

[0138] S3.5 Calculate the rock crushing coefficient and evaluate the rock integrity of the tunnel face;

[0139] The rock crack rate index is calculated by extracting and quantitatively analyzing the cracks in the face of the tunnel. This index can reflect the crack development of the face of the tunnel to a certain extent, which is helpful to evaluate the rock structure type and the integrity of the tunnel face. However, due to the limitations of the surface crack rate itself, it can only reflect the extension of the cracks on the rock surface, and cannot reflect its extension in the direction perpendicular to the tunnel face. Therefore, the present invention adjusts the surface crack rate according to relevant literature, and adds the ratio of the average width of the rock cracks to the diagonal length of the rock image to the result obtained by the surface crack rate to represent the crushing of the face of the tunnel rock. The calculation formula is as follows:

[0140]

[0141] Where: A is the total area of ​​all cracks in the rock image; w and h are the width and height of the image; b is the average width of the rock cracks; K p is the crushing coefficient of rock. At the same time, by investigating the actual situation of tunnel engineering and combining the research of relevant scholars, an evaluation scheme for the integrity of the tunnel face corresponding to the crushing coefficient is summarized, and the evaluation scheme is shown in Table 4.

[0142] Table 4 Evaluation of rock integrity at the tunnel face

[0143]

[0144] According to the above evaluation scheme, the development of rock fractures at the tunnel face in the project was evaluated to prove the feasibility of the scheme. The identification and evaluation results are shown in Table 5.

[0145] Table 5 Evaluation of rock integrity at the tunnel face

[0146]

[0147] S4. Based on the rock information of the tunnel face obtained in steps S1 to S3, extract and convert the surrounding rock classification index to obtain the surrounding rock classification result of the current tunnel face;

[0148] The engineering rock mass classification standard (BQ method) is a common rock mass classification method in my country's highway tunnel projects. Its main principle is to calculate the initial value of the basic rock mass quality index BQ through the rock strength index and the rock mass integrity index; and calculate the corresponding adjustment coefficient based on the actual water content of the surrounding rock of the face, the occurrence of weak structural surfaces and the ground stress state. The initial value of BQ is corrected by these coefficients, and finally the corrected value of the basic rock mass quality index is obtained; different surrounding rock grades have a corresponding correction value range, and the surrounding rock grade parameters of the tunnel are obtained by consulting the surrounding rock grades corresponding to the different correction value ranges.

[0149] The present invention adopts a deep learning image recognition method to identify and extract surrounding rock classification indicators in the tunnel face. First, the tunnel face is photographed and acquired through a smart phone, and then the rock lithology and weathering degree information in the image are obtained by a deep learning image recognition model. According to relevant specifications, the range of uniaxial saturated compressive strength values ​​of various types of rocks at different weathering degrees is obtained, and the median of the range is used as the rock strength index of the face. At the same time, the deep learning image recognition model is continued to be used to extract the degree of rock crack development and rock structure type information in the tunnel face image. For small cracks in the face rock, digital image processing technology is used to identify and quantify them, and their integrity index is determined according to the proportional relationship based on the rock crushing coefficient and the corresponding classification interval. If the face has unfavorable geological conditions such as weak interlayers, a deep learning semantic segmentation model is used to extract and quantify them, and the rock integrity index is calculated.

[0150] When correcting the surrounding rock quality indicators, the deep learning method is also used to identify the surrounding rock water content characteristics such as dryness, seepage and flow in the tunnel face, so as to obtain the water content adjustment coefficient. However, for the acquisition of the structural surface attitude adjustment coefficient, since the face image contains less information on various indicators, it is difficult to directly extract it only through the deep learning method. It is necessary to use tools such as geological compasses on site to measure the inclination of the structural surface and its angle with the tunnel axis. Given that the relevant indicators of the structural surface attitude adjustment coefficient are relatively easy to obtain and the acquisition speed is fast, it is sufficient to simply measure the inclination of the structural surface and its angle with the tunnel axis while taking pictures of the face at the tunnel site.

[0151] By substituting the above-mentioned rock strength index, rock mass integrity index, water content adjustment coefficient and structural surface attitude adjustment coefficient into the BQ surrounding rock classification calculation formula for simple calculation, the surrounding rock quality index can be obtained. Then, according to the relevant surrounding rock classification standards, the final classification result of the surrounding rock of the current tunnel face can be obtained.

[0152] S5. Geological radar advanced geological forecast;

[0153] S5.1 During tunnel excavation, advance geological prediction is carried out on the tunnel face to obtain radar images of the geology in front of the face;

[0154] S5.2 Compare the interpretation results according to the tunnel excavation situation, annotate the confirmed results in the geological radar image, and then train the deep learning target detection network model of the geological radar;

[0155] S6. Extract the surrounding rock features in front of the tunnel face based on deep learning, and compare and verify with the actual situation to continuously improve the interpretation accuracy;

[0156] S6.1 Geological radar image acquisition and interpretation;

[0157] The geological radar data used in this embodiment were collected by the team when conducting advanced geological forecasting in tunnel projects, including the Jiangjunshan Tunnel in Liannan, Guangdong, the Kuiba Road Tunnel in Shenzhen, Guangdong, and the tunnels along the Lingsha Highway in Wenzhou, Zhejiang. After collecting the signal data, the geological radar images of each survey line need to be processed to better reflect the engineering geological conditions in front of the tunnel face, so as to facilitate the interpretation of the geological radar images.

[0158] The interpretation of the geological radar image is to make a reasonable judgment on the geological information in front of the tunnel face based on the waveform, amplitude, intensity and other characteristics of the signal in the image, combined with the geological conditions of the tunnel face and the vault and side walls behind it. When the dielectric constant of the geological body in front of the tunnel face changes, the corresponding reflection waveform change characteristics can be found in the geological radar image, and the reflection waves with the same characteristics form the phase axis for identification. After each judgment on the information in the geological radar image, it is compared and verified with the actual situation after excavation to continuously improve the accuracy of the interpretation.

[0159] S6.2 Prepare geological radar signal dataset;

[0160] During the tunnel excavation process, in order to make an advance geological prediction of the tunnel face, 1 to 2 survey lines were arranged according to the actual size of the tunnel face, and a total of more than 100 geological radar images were obtained. After processing the geological radar images, they were analyzed and interpreted in combination with the on-site conditions of the tunnel project. Then, the interpretation results were compared and verified according to the actual conditions after tunnel excavation. Finally, the confirmed results were annotated in the geological radar image for the subsequent training of the deep learning target detection network model.

[0161] For data sets that need to be trained with target detection models, the LabelImg annotation tool is usually used to annotate the target area in the data set image. The annotation method of this tool is to draw a rectangular bounding box, and this type of box is consistent with the final output form of the target detection network model. The LabelImg annotation tool has different annotation formats and storage methods according to different target detection network models. The present invention adopts the annotation format of the VOC data set for the type of target detection network model, which is mainly divided into image data folders, label data folders, and data set partition folders. In the label data folder, the data is stored in xml format. In order to further improve the recognition performance of the deep learning target detection network, the geological radar image data is expanded. However, due to the particularity of the geological radar image, its expansion method is limited, and it is only processed by simple geometric transformation methods such as random cropping and scaling. After data expansion and screening of the geological radar image, a total of 169 image data were obtained, and training, verification and test set division were performed on it at the same time.

[0162] S6.3 Build the MobileNet-SSD target detection model;

[0163] The deep learning target detection model is a fully supervised representation learning model. It uses a convolutional neural network to train the labeled target categories and location information in the input image, automatically extracts feature information of different dimensions from low to high layers in the image, and finally outputs the location of the target area in the image and the corresponding category and its confidence. SSD is a single-stage convolutional neural network target detection model that divides the input image into grids of different sizes to detect target objects of different sizes. In the low-level case, the image size is large and the divided grid is small; in the high-level case, the image size is small after multiple convolutions and the divided grid is large. When the center of the object is located in a certain area of ​​the image, it is determined by the grid in this area and the corresponding category and confidence are output.

[0164] The backbone feature extraction network of the SSD model is the VGG network. Since the network has a large number of parameters and computational complexity, the overall size of the SSD model is relatively large. In tunnel engineering, equipment is often in an offline environment. Large-volume models pose a great challenge to equipment hardware calculations, reducing the portability and detection speed of the model. In view of the above problems, the present invention improves and optimizes the SSD model, introduces the MobileNet lightweight convolutional neural network as the backbone feature extraction network of the SSD model, and constructs a lightweight deep learning target detection network MobileNet-SSD to identify poor geological bodies in geological radar images.

[0165] S6.4 trains the MobileNet-SSD target detection model, and then uses the MobileNet-SSD target detection model to extract the surrounding rock features in front of the tunnel face.

[0166] The MobileNet-SSD model was used to train the geological radar image dataset. During the training process, the transfer learning method was used to call the pre-trained weights of different models to obtain better feature extraction effects. In the freezing stage of the backbone network, the initial number of iterations of different model training was 100 rounds, and only the top-level classifier weights of the model were trained. Since the training parameters of the model were small at this time, the batch size of the training was set to 16. After 100 rounds of iterative training, the model was fine-tuned, that is, the backbone of the model was unfrozen and trained with the top-level classifier for 200 rounds, so that the model could more fully extract image features. In this process, due to the large number of model parameters, the batch size of the training was adjusted to 8. The Adam optimization algorithm was used during the model training process, and the initial learning rate was set to 0.0001. At the beginning of the training, the pre-trained weights of the model were randomly initialized, resulting in unstable changes in the accuracy and loss value of the training. Therefore, the cosine decay learning rate adjustment method was adopted. This adjustment method makes the size of the learning rate change like a cosine function, that is, it decays slowly first, then accelerates the decay, and finally decays slowly as the number of iterations increases. Through the cosine decay method, the model uses a lower learning rate in the initial stage of training, and then increases to the initial learning rate when the model training reaches stability, which has the effect of warming up the training. In addition, during model training, the validation set is verified once every 10 rounds of iterative operations to check the training status of the model and guide the optimization direction of the model. At the same time, the training weight file of the model is saved to find the best model weight after the training is completed. After a total of 300 rounds of iterative training, the loss of the MobileNet-SSD model has stabilized. The training and validation losses of the MobileNet-SSD model training process can be seen in Fig. 9 .

[0167] After the training, in order to evaluate the recognition performance of different target detection models on geological radar images, the test set was used to test different models, and the precision, recall, F1 score and average precision (AP) were used to evaluate the test results. Precision and recall are a set of mutually opposing and mutually restrictive indicators. Usually, when the precision of the model is high, its recall will be low; and when the recall is high, the precision will be low. In order to more comprehensively measure the recognition performance of the model, the relationship curve between the two is drawn with the recall rate under different confidence thresholds as the horizontal axis and the corresponding precision rate as the vertical axis, and the area under the curve is taken as the average precision. The calculation formula for average precision is as follows:

[0168]

[0169] Where: p(r) is the curve of the change of precision with respect to recall. The test results of the MobileNet-SSD model on the geological radar image test set are shown in Table 6.

[0170] Table 6 MobileNet-SSD model test results

[0171]

[0172] Use geological radar to detect the surrounding rock conditions in front of the tunnel face. Based on the current surrounding rock grade of the tunnel face and combined with the interpretation results of the geological radar image by the deep learning target detection model, the surrounding rock grade information within a certain distance in front of the tunnel face can be determined.

[0173] Example 1

[0174] The tunnel surrounding rock classification method studied in the present invention was tested and verified in the following projects, namely, the tunnels along Lingsha Highway in Wenzhou City, the Kuiba Road Tunnel in Shenzhen City, and the Jiangjunshan Tunnel on Qinglian Highway in Liannan County. The present invention first uses the trained MobileNetV3 network model to intelligently identify the rock lithology, weathering degree, rock structure type, fracture development degree and water content of the tunnel face, and uses digital image processing methods to quantitatively identify rock fractures, and converts the acquired feature information into classification indicators according to relevant specifications, so as to calculate the surrounding rock grade of different tunnel faces; then, the geological radar advanced geological prediction of the tunnel face is carried out, and the MobileNet-SSD target detection model is used to intelligently interpret the obtained geological radar images, so as to infer the surrounding rock grade in front of the tunnel face. Taking the Shitang Tunnel of Linglong Highway with pile number mileage YK18+048~YK18+018 as an example, the present invention classifies the surrounding rock of this section. The rock image of the tunnel face is as follows Fig.10 As shown, the geological radar images of Shitang Tunnel YK18+048~YK18+018 are as follows Fig.11 The specific process and results of the classification are shown in Tables 7 and 8.

[0175] Table 7 Intelligent identification information of surrounding rock at YK18+048 face of Shitang Tunnel

[0176]

[0177] Table 8 BQ surrounding rock classification of YK18+048 face of Shitang Tunnel

[0178]

[0179] According to the intelligent interpretation of the geological radar detection results, it was found that the rock cracks within 13m in front of the tunnel face were developed and the rock was still broken, so the surrounding rock grade in this area was rated as Grade IV; while within the range of 13 to 20m in front of the face, the surrounding rock integrity was good and the cracks were relatively undeveloped, so the surrounding rock grade in this area was rated as Grade III. According to the above surrounding rock classification method and steps, the different pile number intervals in the supporting project were classified, and the classification results are shown in Table 9.

[0180] Table 9 Classification results of surrounding rock of different tunnels

[0181]

[0182] Results Analysis

[0183] The surrounding rock classification method studied by the present invention was used to classify some sections of the supporting project. It can be seen from Table 9 that the accuracy of this method is 85.7%. During the test, a few sections had surrounding rock classification errors and insufficient classification position accuracy. After analyzing the tunnel sections with misjudgment, the main reasons are as follows:

[0184] (1) When judging the strength of the rock at the tunnel face, the strength information of common rocks in the project is obtained by consulting relevant specifications in advance. Due to the different formation ages and development environments of rocks, the same type of rock can also have large lithological differences. The different mineral contents in the rocks will lead to large differences in the strength of the rocks.

[0185] (2) Due to the relatively poor tunnel construction environment, it is inevitable that the image will be unclear when shooting the tunnel face image, which affects the network model's judgment of its integrity. In addition, for the complex geological environment in front of the tunnel face, only the approximate location of the poor geological area can be marked, so there may be an error of about 1 m from the actual situation in the application.

[0186] The present invention was applied to the surrounding rock classification test of the tunnels along Wenzhou Lingsha Highway, Shenzhen Kuiba Road Tunnel and Liannan Jiangjunshan Tunnel, and the obtained classification prediction results were compared with the actual surrounding rock grade situation, and the accuracy of the method was obtained to be 85.7%. In general, the predicted results of the surrounding rock classification method of the present invention are relatively close to the actual results, with certain accuracy and reliability, and can assist technicians in the rapid and accurate identification of the surrounding rock of the tunnel face in practical applications.

[0187] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the protection scope of the present invention.

Claims

1. A tunnel surrounding rock dynamic classification method based on geological radar and transfer learning, characterized in that: The following steps are involved: S1. Image acquisition: Use a mobile camera to take pictures of the tunnel face to obtain photos of the tunnel face; S2. Extract the surface characteristics of the tunnel surrounding rock; S2.1 performs data enhancement on the captured tunnel face image; S2.2 Create a characterization data set: classify and annotate tunnel face photos according to rock lithology, weathering degree, water content, rock mass structure type, and fracture development degree; S2.3 improves the MobileNet network model and builds the MobileNetV3 convolutional neural network model; S2.4 trains a MobileNetV3 convolutional neural network model, and then uses the MobileNetV3 convolutional neural network model to detect the rock properties, weathering degree, water content, rock structure type, and fracture development degree of the tunnel face; S3. Identify and quantify weak interlayers and cracks in the tunnel face; S3.1 Create a weak interlayer dataset: perform data expansion and screening on the captured tunnel face images, and then use the deep learning semantic segmentation and annotation tool Labelme to annotate the weak interlayer zones in the images, that is, outline the locations of the weak interlayers in the images; S3.2 Build semantic segmentation network model: Unet network model; S3.3 trains the Unet network model, and then uses the Unet network model to identify and extract the weak interlayer of the tunnel face; S3.4 Identification and quantification of rock cracks at the tunnel face; S3.5 Calculate the rock crushing coefficient and evaluate the rock integrity of the tunnel face; S4. Based on the rock information of the tunnel face obtained in steps S1 to S3, extract and convert the surrounding rock classification index to obtain the surrounding rock classification result of the current tunnel face; S5. Geological radar advanced geological forecast; S5.1 During tunnel excavation, advance geological prediction is carried out on the tunnel face to obtain radar images of the geology in front of the face; S5.2 Compare the interpretation results according to the tunnel excavation situation, annotate the confirmed results in the geological radar image, and then train the deep learning target detection network model of the geological radar; S6. Extract the surrounding rock features in front of the tunnel face based on deep learning, and compare and verify them with the actual situation to continuously improve the interpretation accuracy.

2. According to claim 1, a tunnel surrounding rock dynamic classification method based on geological radar and transfer learning is characterized by: The data enhancement methods include geometric transformation, clarity transformation, color gamut conversion and noise addition; Geometric transformation: Process the image through one or more operations such as rotation, flipping, translation, stretching, and mirroring, thereby simulating a mobile camera device taking photos of the tunnel face from different angles; Clarity transformation: blur the image through median filtering / mean filtering / Gaussian filtering; Color gamut conversion: By adjusting the brightness of the tunnel face image, the different light and dark levels of the tunnel face can be simulated; Add Noise: Add salt and pepper noise or Gaussian noise to simulate dust in the air.

3. According to claim 1, a tunnel surrounding rock dynamic classification method based on geological radar and transfer learning is characterized by: In step S2.3, the improvements of the MobileNetV3 convolutional neural network model include: (1) Introducing the reverse residual structure; (2) Add lightweight SE module; The SE module consists of two parts: compression and excitation. The compression part performs global average pooling on the output feature matrix of the deep convolution, and the excitation part outputs the compressed feature matrix after passing it through two fully connected layers. (3) Improve the activation function; The original activation function of the MobileNet network model is improved, and the H-swish activation function is proposed. The calculation expression is as follows: ReLU6(x)=min(max(x,0),6) H-swish(x) = x ReLU6(x+3) / 6 Where x is the input of the nonlinear activation function in the i-th channel, and the ReLU6 function is the ReLU function with a maximum output limit of 6; (4) Optimize network structure.

4. The method for dynamic classification of tunnel surrounding rock based on geological radar and transfer learning according to claim 1 is characterized in that: The step S3.4, identifying and quantifying rock cracks at the tunnel face, comprises the following sub-steps: I. Use image segmentation method to identify and extract cracks in the rock image of the tunnel face; II. Based on the distance between the camera and the rock on the tunnel face, or the presence of a reference object with known geometric parameters on the tunnel face, the actual size of the fracture is calculated according to the pixel size of the target area in the image and the corresponding ratio; III. Calculate the crack length, crack width, crack area and crack ratio; Crack length: Use the Canny operator to extract the edge of the image after rock crack threshold segmentation, and perform skeleton processing on the rock crack edge detection image, and make an approximate calculation based on the length of the two edges of the crack divided by 2; Crack area: based on the number of black pixels in the crack binary image, obtained by proportional conversion; Crack width: calculated by approximation, the crack area divided by the crack length; Crack ratio: expressed as the numerical value of surface crack ratio; The face crack ratio is expressed by the ratio of the total area of ​​all cracks on the rock surface of the tunnel face to the corresponding rock area.

5. The method for dynamic classification of tunnel surrounding rock based on geological radar and transfer learning according to claim 4 is characterized in that: The image segmentation method is the maximum distance between classes method.

6. The method for dynamic classification of tunnel surrounding rock based on geological radar and transfer learning according to claim 4 is characterized in that: In step S3.4, before image segmentation, the grayscale value of the tunnel face rock image is adjusted to within 30% of the original image.

7. The method for dynamic classification of tunnel surrounding rock based on geological radar and transfer learning according to claim 1 is characterized by: In step S3.5, the calculation expression of the rock crushing coefficient is as follows: Where: A is the total area of ​​all cracks in the rock image; w and h are the width and height of the image; b is the average width of the rock cracks; K p is the crushing coefficient of rock.

8. The method for dynamic classification of tunnel surrounding rock based on geological radar and transfer learning according to claim 1 is characterized by: In step S3.5, the evaluation relationship between the rock crushing coefficient and the integrity of the rock at the tunnel face is as follows: K p ∈0.01~0.04, indicating that the rock on the face is relatively intact; K p ∈0.04~0.08, indicating that the rock on the face is relatively broken; K p >0.08, indicating that the rock at the tunnel face is broken / extremely broken.

9. The method for dynamic classification of tunnel surrounding rock based on geological radar and transfer learning according to claim 1, characterized in that: The specific steps of step S4 are as follows: S4.1 Based on the rock lithology and weathering degree information obtained in step S2, the range of uniaxial saturated compressive strength of various types of rocks at different weathering degrees is obtained according to the specification, and then the median value of the range is used as the rock strength index of the tunnel face; S4.2 Based on the rock fracture development degree and rock mass structure type obtained in step S2, and the fine fractures in the rock of the tunnel face quantitatively identified in step S3, the rock integrity index is determined in proportion according to the rock crushing coefficient and the corresponding classification interval; S4.3 based on the water content obtained in step S2, obtain a water content adjustment coefficient; S4.4 Use a geological compass to measure the inclination of the rock structure surface at the tunnel face and its angle with the tunnel axis to obtain the structural surface attitude adjustment coefficient; The S4.5 rock strength index, rock integrity index, water content adjustment coefficient and structural surface attitude adjustment coefficient are calculated using the BQ surrounding rock classification calculation formula to obtain the surrounding rock quality index, and then the surrounding rock classification result of the current tunnel face is obtained based on the surrounding rock classification standard.

10. The method for dynamic classification of tunnel surrounding rock based on geological radar and transfer learning according to claim 1, characterized in that: The specific steps of step S6 are as follows: S6.1 Geological radar image acquisition and interpretation; S6.2 Prepare geological radar signal dataset; S6.3 Build the MobileNet-SSD target detection model; Based on the SSD model, the MobileNet lightweight convolutional neural network is introduced as the backbone feature extraction network of the SSD model to build the MobileNet-SSD target detection model. S6.4 trains the MobileNet-SSD target detection model, and then uses the MobileNet-SSD target detection model to extract the surrounding rock features in front of the tunnel face.

Images