A method and equipment for identifying rock mass structure surfaces in limestone tunnels

Abstract

This invention relates to the field of rock mass structure surface recognition technology, and in particular to a method and device for recognizing rock mass structure surfaces in limestone tunnels. The method includes the following steps: S1, acquiring an image of the limestone tunnel face; S2, inputting the limestone tunnel face image into a pre-trained U-Net-based limestone structure surface recognition model to obtain an image of the limestone tunnel rock mass structure surface; S3, performing trace morphological processing on the limestone tunnel rock mass structure surface image to obtain the skeleton lines of the limestone tunnel rock mass structure surface; S4, straightening the skeleton lines of the limestone tunnel rock mass structure surface to obtain a skeleton line parameter model of the limestone tunnel rock mass structure surface; S5, performing pixel-level statistics on the skeleton line parameter model to calculate the length and apparent dip angle of each skeleton line. Using the method of this invention, a set of efficient and accurate structural surface parameter recognition statistical techniques for limestone areas is ultimately formed.

Description

Technical Field

[0001] This invention relates to the field of rock mass structure surface identification technology, and in particular to a method and equipment for identifying rock mass structure surfaces in limestone tunnels. Background Technology

[0002] Structural planes are geological interfaces within rock masses that have a certain structural extension direction and length, and are relatively thin. They are the main controlling factors for the stability of tunnel rock masses, and structural plane parameters are important indicators for evaluating rock mass quality. How to effectively obtain these parameters has always been a major concern for people in the rock mass engineering industry.

[0003] The geological conditions of limestone areas differ significantly from those of sandstone areas. Sandstone masses are well-stratified with a clear network of structural planes, making them easy to identify. Limestone masses, on the other hand, are generally massive, thick-layered, and relatively homogeneous. In masses with less karst development, the opening of structural planes is smaller, their extension direction is less obvious, and they are difficult to identify. Limestone areas are susceptible to karst influence, with some dissolution fissures exhibiting large openings, resulting in a highly irregular network of structural planes. The structural planes are also significantly affected by infill materials, leading to a complex and variable network of structural planes in limestone areas, making manual identification and measurement of structural plane parameters a significant undertaking.

[0004] In limestone tunnel areas, traditional methods for obtaining rock mass structural parameters typically involve manual, on-site contact measurement, such as the survey line method and the statistical window method. These methods are characterized by high labor intensity, low efficiency, and high safety risks. In the context of modern rapid tunnel construction, it is difficult to measure structural parameters manually. Therefore, new non-contact measurement methods have emerged, including 3D laser scanning technology, digital photogrammetry, and prism-free total station technology. Three-dimensional laser scanning technology uses the principle of laser ranging to obtain the reflectivity signal and three-dimensional coordinates of each point on the measured object. After data processing, a three-dimensional spatial model of the measured structural surface is established. Although this technology has high accuracy, it has high requirements for the conditions of the measured cross-section. It has a low recognition rate for some limestone structural surfaces with flat and smooth cross-sections and small opening. Moreover, the equipment is expensive, inconvenient to carry, and the subsequent data processing is complex. Digital photogrammetry technology is based on the principle of multi-view imaging to obtain the three-dimensional information of the measurement points and uses image processing software to perform three-dimensional reconstruction and extract the required information from the image. At present, this technology is theoretically feasible for rock mass structural surface identification. However, it is limited by the difficulty in matching the theoretical research of image processing with actual engineering applications. In particular, the image processing technology for complex structural surface networks of limestone, especially the identification, splitting and connection of structural surface networks, small branch removal, and straightening of network traces, has not achieved ideal results in practical engineering applications, thus bringing great challenges to the identification of structural surfaces in limestone areas.

[0005] Thanks to the rapid development of artificial intelligence theory, 3D laser scanning technology and digital photogrammetry technology are gradually being applied to practical tunnel engineering. The Unet convolutional neural network has significant advantages in identifying rock fractures (Zhang et al., 2021; Li et al., 2021). The Unet network effectively identifies and extracts features from images by combining downsampling and upsampling operations on sample images. Compared with traditional gradient-based edge detection techniques, this network can greatly improve the robustness of image recognition, reduce the risk of overfitting, and increase recognition efficiency. He Peng (2020) proposed a multi-parameter characterization method using linear clustering extraction and magnetic tracking extraction, and based on line detection, smart scissors, and morphological edge detection algorithms, respectively. Since the structural surfaces identified by neural networks are not straight lines composed of single-phase elements, joints need to be fitted and skeletonized. Commonly used methods include region growing (Zhang et al., 2022), GMM-EM algorithm, random sampling consensus algorithm (Zhang et al., 2021), dark area curve structure enhancement (DRCSE) algorithm (Yudi Tang, 2021), and Zhang-Suen algorithm (Lynda, 2008). Based on the skeletonization of joint lines, Leng Biao (2021) proposed an automatic grouping algorithm for rock mass fracture boundaries at the working face based on fracture identification results, further improving the statistical theory of structural surface identification based on image recognition.

[0006] The aforementioned existing technical solutions do not mention methods for identifying rock mass structural surfaces in different lithological regions. Although manual field contact measurement methods can be used to measure structural surface parameters under various lithologies, their applicability is significantly poor in the context of complex structural surface networks in limestone regions. Similarly, no specific scheme for identifying complex structural surface networks in limestone regions has been found among image recognition-based methods for acquiring rock mass structural surface feature parameters. Since structural surface identification and statistics based on image recognition technology is a multi-step integrated process, each step is implemented by different algorithms, including neural network-based structural surface trace extraction, multi-segment line fitting algorithms for trace connection, trace splitting and connection algorithms, structural surface grouping algorithms, and algorithms for calculating length and spacing, a reasonable and reliable method for identifying limestone structural surfaces must be proposed at the algorithmic level.

[0007] The main technical issues are as follows:

[0008] 1. Traditional methods for obtaining rock mass structural parameters generally employ manual on-site contact measurement, which is characterized by high labor intensity, low work efficiency, and high safety risks.

[0009] 2. The geological conditions of limestone areas are significantly different from those of sandstone areas. Sandstone rock masses have good stratification and obvious structural surface networks, making them easy to identify. In contrast, the structural surfaces of limestone are easily affected by the degree of karst development, resulting in complex and diverse structural surface networks and making it difficult to identify structural surface traces.

[0010] 3. Image recognition technology based on threshold segmentation and edge detection has poor reliability and low accuracy in recognizing structural surfaces in complex tunnel environments;

[0011] 4. For complex limestone structural surface networks, such as structural surfaces with large opening, they need to be skeletonized. However, after the structural surface traces are skeletonized, they are prone to generating incoherent branch skeleton lines and noise. There is no mature solution for removing branch skeleton lines and noise in structural surfaces.

[0012] 5. The structural surface traces obtained from identification are generally irregular curves. They need to be fitted into straight line segments before the parameters can be used for subsequent rock mass quality index conversion, jointed rock mass calculation and numerical simulation. At present, there is still a problem with the inaccuracy of straight line fitting of structural surface traces. Conventional fitting methods are prone to missing some traces, resulting in the phenomenon of missing structural surface traces after fitting. Summary of the Invention

[0013] Based on the above background and technical issues, on the one hand, a U-Net deep learning model suitable for limestone structural surface identification is constructed, and on the other hand, trace morphology processing and straightening processing are performed on the identified model. A method and device for identifying limestone tunnel rock mass structural surfaces are proposed, which identifies the limestone tunnel rock mass structural surfaces and obtains the length and apparent dip angle of the corresponding skeleton lines.

[0014] To achieve the above-mentioned objectives, the present invention provides the following technical solution:

[0015] A method for identifying rock mass structure surfaces in limestone tunnels includes the following steps:

[0016] S1, Obtain the image of the face of the limestone tunnel;

[0017] S2, the image of the limestone tunnel face is input into a pre-trained U-Net-based limestone structure surface recognition model to obtain an image of the limestone tunnel rock mass structure surface. The U-Net-based limestone structure surface recognition model consists of an encoding part on the left, a decoding part on the right, and convolutional and activation layers on the bottom. The decoding part on the right is the decoding part of the U-Net network; the encoding part on the left is a VGG16 network.

[0018] S3, perform trace morphological processing on the image of the rock mass structure surface of the limestone tunnel to obtain the skeleton line of the rock mass structure surface of the limestone tunnel. The trace morphological processing includes skeletonization of the trace of the structure surface based on the skimage algorithm, splitting of the skeleton line of the limestone structure surface, and connecting of the skeleton line of the limestone structure surface.

[0019] S4, the skeleton lines of the limestone tunnel rock mass structure surface are straightened to obtain the skeleton line parameter model of the limestone tunnel rock mass structure surface. The straightening process adopts the random sampling consensus algorithm.

[0020] S5, perform pixel-level statistics on the skeleton line parameter model to calculate the length and viewing angle of each skeleton line.

[0021] As a preferred embodiment, the encoding part on the left side of the U-Net-based limestone structure surface recognition model is used for downsampling to extract multi-scale structure surface features, including several sub-modules. Each sub-module includes two 3×3 convolutional layers, a piecewise linear function activation layer, and a 2×2 max pooling layer with a stride of 2.

[0022] As a preferred embodiment, the training process of the U-Net-based limestone structure surface recognition model includes the following steps:

[0023] S21, set the model parameters to Epoch=200, batch_size=2, lr=0.0001;

[0024] S22 uses the Adam optimizer to optimize network parameters and uses binary cross-entropy as the loss function to determine whether the model training has converged.

[0025] S23. After training convergence, a well-trained U-Net-based limestone structure surface recognition model is obtained.

[0026] As a preferred embodiment, the skeletonization process based on the skimage algorithm for structural surface traces includes the following steps:

[0027] A31, the image of the rock mass structure surface of the limestone tunnel is binarized to obtain a binary image of the structure surface trace, and the binary image of the structure surface trace is eroded to obtain the eroded structure surface trace.

[0028] A32, perform an opening operation on the eroded structural surface traces. The deleted pixel portion during the opening operation is part of the skeleton. Add the deleted pixel portion to the limestone structural surface trace skeleton diagram.

[0029] A33. Repeat steps A31 to A32 to obtain the skeleton diagram of the limestone structural surface traces corresponding to the image of the limestone tunnel rock mass structural surface.

[0030] As a preferred embodiment, the splitting of the limestone structural plane skeleton line includes the following steps:

[0031] B31, convert the limestone structural surface trace skeleton diagram into a two-dimensional matrix point set;

[0032] B32 defines the eight-neighbor matrix of the target pixel P(x,y);

[0033] B33, set the skeleton line pixel to 1 and the background to 0, use the eight-neighbor matrix to traverse the two-dimensional matrix point set, and when a pixel has a situation that matches the two-dimensional matrix point set in its eight-neighbor matrix, determine that the pixel is a branch point.

[0034] B34, modify the pixel value of the branch point to 0, delete the branch point, and obtain the trace of the broken structure surface.

[0035] As a preferred embodiment, the connection of the limestone structural surface skeleton line includes the following steps:

[0036] C31, obtain the skeleton diagram of limestone structural surface traces after the skeleton lines are split, and locate and number the traces of the broken structural surfaces.

[0037] C32, perform inclination similarity analysis on the traces of adjacent broken structural surfaces to determine whether the traces of adjacent broken structural surfaces belong to the same structural surface. If so, proceed to step C33.

[0038] C33 connects the traces of the adjacent broken structural surfaces using an expansion-erosion morphological manipulation method.

[0039] Furthermore, step S4 specifically includes the following steps:

[0040] S41, obtain the data point set corresponding to the skeleton line of the rock mass structure surface of the limestone tunnel, and select 2 data points from the data point set to establish a sample data subset;

[0041] S42, Fit a linear mathematical model based on the subset of sample data;

[0042] S43, use the linear mathematical model to test the remaining points in the data point set. If the tested data point is within the allowable error range, then the data point is judged as an interior point; otherwise, it is judged as an exterior point. The linear mathematical model and the corresponding number of interior points are obtained.

[0043] S44, compare the number of interior points of the current linear mathematical model with the number of interior points of the previously obtained linear mathematical model, and obtain the maximum number of interior points between the two and the model parameters of the linear mathematical model corresponding to the maximum number of interior points.

[0044] S45. Repeat steps S41 to S44 until the iteration ends or the current linear mathematical model has satisfied the condition that the number of interior points is greater than the set threshold, and obtain the final skeleton line parameter model of the limestone tunnel rock mass structure surface.

[0045] Furthermore, in step S5, the formula for calculating the length of the skeleton line is:

[0046]

[0047] The formula for calculating the apparent tilt angle of the skeleton line is θ = arctan|a i |,

[0048] Among them, the skeleton line parameter model of the rock mass structure plane of the limestone tunnel is y = a i x+b i a i is the slope coefficient of the skeleton line parameter model, (x,y) is the two-dimensional coordinate of a point on the skeleton line, max(x) is the maximum value of the abscissa on the skeleton line, min(x) is the minimum value of the abscissa on the skeleton line, max(y) is the maximum value of the ordinate on the skeleton line, and min(y) is the minimum value of the ordinate on the skeleton line.

[0049] Furthermore, step S1 also includes a preprocessing process for the limestone tunnel face image, as well as increasing the quantity and improving the quality of the preprocessed limestone tunnel face image.

[0050] The preprocessing process includes: establishing a face contour database and using a face non-target information removal model trained on the open-source U-Net network to remove non-target information from the images in the face contour database;

[0051] The increase in quantity includes: obtaining limestone structural surface images at different angles by randomly rotating, flipping up, down, left, and right from the same face image; obtaining limestone structural surface images of the same size but different scales by cropping the image; and obtaining limestone structural surface images with different brightness and contrast by adjusting the image brightness and image contrast.

[0052] The quality improvement includes: using a high-pass filtering method with a 3×3 convolution kernel to process images of limestone structural surfaces with small openings; and using a Gaussian blurring method with a 3×3 Gaussian kernel to process images of limestone structural surfaces with large openings and filling material.

[0053] Based on the same concept, a limestone tunnel rock mass structure surface identification device is also proposed, including at least one processor and a memory communicatively connected to the at least one processor; the memory stores instructions executable by the at least one processor, which are executed by the at least one processor to enable the at least one processor to perform any of the above-described limestone tunnel rock mass structure surface identification methods.

[0054] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0055] The present invention provides a method and equipment for identifying the rock mass structure surface of a limestone tunnel, which is used to solve the problem of difficulty in extracting parameters of complex limestone structure surfaces, and ultimately forms a set of efficient and accurate structural surface parameter identification and statistical technology for limestone areas. Attached Figure Description

[0056] Figure 1 This is a flowchart of a method for identifying the rock mass structure surface of a limestone tunnel in Example 1;

[0057] Figure 2 The image is a picture captured using a smartphone during image acquisition in Example 1;

[0058] Figure 3 The images are those captured using a digital camera during image acquisition in Example 1;

[0059] Figure 4 This is the original image of the tunnel face before preprocessing in Example 1;

[0060] Figure 5 This is the image after removing non-target information during the preprocessing of the limestone tunnel face image data in Example 1;

[0061] Figure 6 This is the original image from when the structural surfaces were annotated in Example 1;

[0062] Figure 7 This is the label image used for structural surface annotation in Example 1;

[0063] Figure 8 This is a diagram of the improved U-Net model framework in Example 1;

[0064] Figure 9 These are photos of the tunnel face taken by a smartphone at the tunnel site in Example 2;

[0065] Figure 10 This is the outline cropping image in the structural surface recognition effect diagram of Example 2;

[0066] Figure 11 This is a diagram showing the recognition effect of structural surface traces in Example 2.

[0067] Figure 12 This is a morphological processing effect diagram of the structural surface recognition effect diagram in Example 2;

[0068] Figure 13 This is the RANSAC result diagram in the structural surface recognition effect diagram of Example 2. Detailed Implementation

[0069] The present invention will be further described in detail below with reference to experimental examples and specific embodiments. However, this should not be construed as limiting the scope of the above-mentioned subject matter of the present invention to the following embodiments; all technologies implemented based on the content of the present invention fall within the scope of the present invention.

[0070] Example 1

[0071] A method for identifying rock mass structure surfaces in limestone tunnels, the flowchart of which is shown below. Figure 1 As shown, the specific steps include:

[0072] 1. Image acquisition and image data preprocessing at the face of limestone tunnels

[0073] (1) Image acquisition of the face of limestone tunnel

[0074] Image recognition technology has certain requirements for image quality. To improve image acquisition efficiency and quality, images of different limestone structural surfaces are included. Images are acquired using smartphones, such as... Figure 2 As shown, the images acquired using a digital camera during image acquisition are as follows: Figure 3 As shown, regardless of the method used to acquire images, the following image acquisition requirements must be met:

[0075] ① Select image acquisition devices with a resolution of 20 megapixels or higher, such as smartphones or digital cameras, to ensure that the image clarity meets the model training requirements;

[0076] ② Considering the poor environmental conditions and numerous human interference factors when collecting images at the tunnel construction site, the time for collecting images of the tunnel face is set from the time the tunnel muck removal and hazard mitigation are completed until the lining support is carried out. During the shooting process, it should be ensured that the tunnel face is well-lit and free from obstructions. The shooting range should cover the entire tunnel face, while minimizing the inclusion of too many steps, sidewalls, and linings in the shooting range.

[0077] ③ The collected images should include images of limestone structural surfaces under different karst development conditions, such as images of limestone with small structural surface opening under karst non-development conditions, and images of structural surfaces with large opening and no filling material or with filling material under karst development conditions.

[0078] (2) Preprocessing of image data from the face of limestone tunnels

[0079] The purpose of image preprocessing is to remove non-target information (image cropping), standardize the annotation of structural surfaces, obtain multi-angle and multi-scale limestone structural surface data, highlight the characteristics of limestone structural surfaces in the image, or reduce the influence of dissolution fissure filling materials, irrelevant noise, etc. on limestone structural surfaces, and ultimately improve the quantity and quality of the dataset.

[0080] ①Image cropping

[0081] Since the collected tunnel face images inevitably contain tunnel steps and linings, such non-target information can affect the model's feature extraction. Therefore, this information must be removed first. This patent uses manual cropping combined with image segmentation technology (semantic segmentation) to establish a tunnel face contour database and removes non-target information from the images in the database. The original tunnel face image before preprocessing of the limestone tunnel face image data is shown below. Figure 4 As shown, the image after removing non-target information during the preprocessing of limestone tunnel face image data is as follows. Figure 5 As shown. The database should contain at least 500 images each. An open-source U-Net network is used for model training to obtain a non-target information removal model for the tunnel face, thereby achieving the removal of non-target information from the tunnel face.

[0082] ② Labeling of limestone structural surfaces

[0083] The annotation of limestone structural surfaces can be achieved through a combination of offline and online cloud-based annotation. When the sample data volume is small, manual annotation of each structural surface trace on the face image can be performed using the Labelme local annotation tool. When the sample size of the face image is large and the workload is significant, the Easydata online annotation tool can be used to enable multi-person online annotation to improve annotation efficiency. For limestone structural surfaces with underdeveloped karst, special attention should be paid to the extension direction of structural surfaces with small openings, as some structural surfaces are easily filled with cementing material, to avoid omissions and annotation errors. For limestone structural surfaces with developed karst, especially when the dissolution fissures are filled with material, the structural surfaces should be annotated along the boundary lines of the dissolution fissures as much as possible. As a specific example, the original image for structural surface annotation is shown below. Figure 6 As shown, the label image during structural surface annotation is as follows: Figure 7 As shown.

[0084] (3) Improvement in the quantity and quality of datasets

[0085] Training a structural surface recognition model requires a certain number of limestone structural surface images. However, with limited image data, it's difficult to cover limestone structural surfaces at different angles and scales. Therefore, this scheme uses image transformation and generative adversarial networks to obtain more comprehensive structural surface image data. The same tunnel face image can be randomly rotated (rotation angle automatically generated by code), flipped vertically and horizontally, to obtain limestone structural surfaces at different angles. By scaling the image (this scheme scales the image to 0.6, 0.8, 1.2, 1, and 4 times the original image) and then cropping the image (cropping the scaled image to 224×224 size), images of limestone structural surfaces of the same size but different scales are obtained, compensating for the problem that the opening of some limestone structural surfaces is not easy to identify. By adjusting the image brightness and contrast (this scheme adjusts the image brightness and contrast to 0.7, 0.9, 1.1, and 1.3 times the original image, respectively), limestone tunnel face images with different brightness and contrast are obtained, simulating limestone tunnel face images with different brightness and contrast acquired under different lighting backgrounds in the tunnel. In addition, this scheme introduces an open-source generative adversarial network model to generate a certain number of limestone tunnel face images based on the original dataset.

[0086] To address the challenges of identifying limestone structural surfaces with small openings, and the susceptibility of larger openings to infilling material, image enhancement techniques are introduced to improve dataset quality. This scheme employs a high-pass filter with a 3×3 convolution kernel to process images of structural surfaces with small openings, while a Gaussian blurring method with a 3×3 Gaussian kernel effectively handles images of limestone structural surfaces with large openings and infilling material.

[0087] Therefore, by combining the above-mentioned image transformation and image enhancement techniques, the quantity and quality of the dataset can be further improved, thereby enhancing the applicability and generalization ability of the model in limestone areas.

[0088] 2. Construction and Training of a U-Net-based Limestone Structure Surface Recognition Model

[0089] The original U-Net network structure was first proposed by Ronneberger et al. (2015). This neural network has since been widely applied in image segmentation, particularly in medical image recognition. This patent, tailored to the characteristics of limestone structural surfaces, constructs an improved U-Net network. The improved model's feature extraction network uses a VGG16 network as its backbone on the left side of the encoding portion, while the right side decoding portion consists of the original U-Net network's decoding portion. The improved network model as a whole consists of a left-side encoding portion, a right-side decoding portion, and two convolutional + activation layers at the bottom (e.g., ...). Figure 8 (As shown).

[0090] (1) Left-side coding section

[0091] The architecture consists of four repeating structures: two 3×3 convolutional layers followed by a piecewise linear activation function (ReLU) layer and a 2×2 max pooling layer with a stride of 2. After each downsampling, the number of feature channels doubles, with the number of channels changing from [3,64,64] to [6,128,128] to [12,256,256] to [24,512,512], and multi-scale structural surface features are extracted.

[0092] (2) Decoding section on the right

[0093] Similar to the encoding layer, the decoding layer also consists of four repeating structures: each repeating structure is preceded by a deconvolution, after which the number of feature channels is halved and the feature map size is doubled; after deconvolution, the deconvolution result is concatenated with the feature map of the corresponding step in the encoding part (white / gray blocks), that is, the structural surface features under multiple scales are fused; the concatenated feature map is then subjected to two 3×3 convolutions; the last layer has a 1×1 convolution kernel, which transforms the 64-channel feature map into the result of structural surface recognition.

[0094] (3) Model Training

[0095] Model training involves inputting the dataset, tuning model parameters, optimizing network parameters, and calculating the training loss. Preprocessed images of limestone structural surfaces and their corresponding label images are input into a pre-built, improved U-Net model. Model parameters are set to Epoch = 200, batch_size = 2, and lr = 0.0001. The Adam optimizer is used to optimize network parameters, and binary cross-entropy is used as the loss function to determine if the model training has converged. The converged weighted model is then saved, resulting in the optimal limestone structural surface prediction model.

[0096] 3. Morphological processing of structural surface traces

[0097] (1) Skeletonization of structural surface traces based on skimage algorithm

[0098] After the predicted image is identified by the U-Net model, the limestone structural surface traces will be marked and segmented from the predicted image. The marked structural surface traces are "surfaces" composed of multiple pixels, rather than "lines" connected by multiple single-phase pixels. If the structural surface trace parameters are to be further extracted, the recognition results need to be skeletonized to obtain structural surface traces connected by single-phase pixels.

[0099] This patent employs the skimage image thinning algorithm for skeletonization of structural surface traces. Skeleton extraction in skimage falls under the category of morphological processing, based on image dilation and erosion. Its principle is as follows: Let A be the target image to be processed, B be the structuring element, and S(A) represent the skeleton of A. k (A) is the k-th skeleton subset of A. The skeleton expression of A is given below.

[0100]

[0101]

[0102] Based on the above principles, the specific implementation steps are as follows:

[0103] ① The binary image of the structural surface traces obtained by U-Net recognition is eroded, and the eroded structural surface traces become narrower and thinner;

[0104] ② Perform an opening operation on the eroded image. The pixels deleted during the opening operation are part of the skeleton. Add them to the skeleton image.

[0105] ③ Repeat the above process until the image is completely eroded, and finally obtain the skeleton map of the limestone structural surface trace.

[0106] (2) Decomposition and connection of limestone structural surface skeleton lines

[0107] The characteristics of the limestone structural surface traces after skeletonization are quite obvious: the network structure is complex, making it difficult to distinguish obvious distribution patterns; some structural surface traces have many branches of varying lengths, with some short branches caused by image noise and unevenness of the structural surface traces during the skeletonization process; some structural surface traces show many local discontinuities, while in the actual image these structural surface traces should be connected into a single trace. Therefore, it is necessary to split and connect the limestone structural surface skeleton lines.

[0108] The specific implementation method is as follows:

[0109] 1) Deconstruction of limestone structural plane skeleton lines

[0110] ① Image data conversion. The skeletonized trace image is converted into a corresponding two-dimensional matrix point set M using relevant Python image processing toolkits;

[0111] ② Define the set of coordinates of the eight neighboring pixels of the target pixel P(x, y) as P8[(x-1, y), (x-1, y+1), (x, y+1), (x+1, y+1), (x+1, y), (x+1, y-1), (x, y-1), (x-1, y-1)]

[0112] ③ Identification of branch points on limestone structural surface traces. First, a branch point identification matrix is ​​defined. This patent uses an exhaustive method to derive the eight-neighbor matrix Q of pixels belonging to branch points:

[0113] [[0,1,0,1,0,0,1,0],[0,0,1,0,1,0,0,1],[1,0,0,1,0,1,0,0],[0,1,0,0,1,0,1,0],[0,0,1,0,0,1,0,1],[1,0,0,1,0,0,1,0],[0,1,0,0,1,0,0,1],[1,0,1,0,0,1,0,0],[0,1,0,0,0,1,0,1],[0,1,0,1,0,0,0,1], [0,1,0,1,0,1,0,0],[0,0,0,1,0,1,0,1],[1,0,1,0,0,0,1,0],[1,0,1,0,1,0,0,0],[0,0,1,0,1,0,1,0],[1,0,0,0,1,0,1,0],[1,0,0,1,1,1,0,0],[0,0,1,0,0,1,1,1],[1,1,0,0,1,0,0,1],[0,1,1,1,0,0,1,0],[1 [0,1,1,0,0,1,0],[1,0,1,0,0,1,1,0],[1,0,1,1,0,1,1,0],[0,1,1,0,1,0,1,1],[1,1,0,1,1,0,1,0],[1,1,0,0,1,0,1,0],[0,1,1,0,1,0,1,0],[0,0,1,0,1,0,1,1],[1,0,0,1,1,0,1,0],[1,0,1,0,1,0,1,1],[1,0 [1,0,1,1,0,0],[1,0,1,0,1,0,0,1],[0,1,0,0,1,0,1,1],[0,1,1,0,1,0,0,1],[1,1,0,1,0,0,1,0],[0,1,0,1,1,0,1,0],[0,0,1,0,1,1,0,1],[1,0,1,0,0,1,0,1],[1,0,0,1,0,1,1,0],[1,0,1,1,0,1,0,0]······]

[0114] Assuming the skeleton line pixel is 1 and the background is 0, the 8-neighbor matrix P8 is used to traverse the 2D matrix point set M of the skeleton line. When a pixel has a value in its 8-neighborhood that matches matrix Q, that pixel is considered a branch point. The value of that pixel is then changed to 0 to delete the branch point, thereby separating each skeleton line from the structural surface network.

[0115] 2) Connection of limestone structural surface skeleton lines

[0116] ① Location and Marking of Structural Plane Traces

[0117] To determine whether several discontinuous structural plane traces belong to the same trace in reality, it is necessary to locate and number them first. In this solution, cv2.drawContours in opencv is used to achieve trace location and numbering;

[0118] ② Similarity Judgment

[0119] Perform dip angle similarity analysis on two adjacent structural plane traces L1 and L2 to determine whether these structural plane traces belong to the same structural plane. Calculate the dip angles θ1 and θ2 of the two adjacent structural plane traces L1 and L2 respectively. When |θ1 - θ2| < 15°, it is judged that L1 and L2 belong to the same structural plane.

[0120] ③ Trace Connection

[0121] Process L1 and L2 through the morphological operation method of dilation - erosion. The size of the convolution kernel [n×n] for dilation - erosion depends on the distance between the endpoints of L1 and L2, and n should be greater than the endpoint distance. In this solution, the connection of structural plane traces is achieved through dilation - erosion, and after a skeletonization process, a more realistic structural plane network is obtained.

[0122] (3) Removal of Branch Points

[0123] For the complex changes in the skeletonized image of the limestone structural plane, the lengths of the skeleton trace lines are uneven. Especially for the structural plane skeletons with large opening degrees, there are many branch lines generated, and these branch lines belong to false structural planes, that is, noise information. Therefore, this patent introduces an adaptive threshold segmentation method. First, calculate the average length L of all skeleton lines, set the segmentation threshold to 0.05L. When the trace length L n < L, determine that this skeleton line is noise and delete it. The segmentation threshold coefficient 0.05 is obtained by statistically analyzing a large amount of true and false (noise) structural plane data of the limestone structural plane skeleton lines.

[0124] 4. Straightening Processing of Structural Plane Skeleton Lines Based on the Random Sample Consensus Algorithm

[0125] After morphological processing, the structural plane skeleton lines are generally irregular curves, which are difficult to be parameterized and recognized. It is necessary to straighten them. This patent bypasses the conventional least - squares straight - line fitting method and proposes to apply the Random Sample Consensus (RANSAC) algorithm to the straightening processing of limestone structural plane skeleton lines. The advantage of the RANSAC algorithm is that it can estimate the parameters of a mathematical model iteratively from a set of observation data containing a large number of "outlier (noise)" points. The implementation steps are as follows:

[0126] ① Each structural plane skeleton line corresponds to a data point set. Select a dataset S in sequence and choose the minimum number of samples required to build the model. Since it is necessary to fit a two-dimensional straight line segment, the number of samples required for modeling is 2. Let S1 be the dataset with 2 samples.

[0127] ② Using the selected dataset S1, a mathematical model M1 is calculated: y = ax + b;

[0128] ③ Use the calculated model M1 to test the remaining points in the dataset. If the tested data point is within the allowable error range, then the data point is judged as an inlier; otherwise, it is judged as an outlier. The dataset consisting of all inliers is called S11, and S11 is called the consistency set of S1.

[0129] ④ Compare the number of "interior points" of the current model M1 with the previously best model Mi, and record the model parameters (a) when the maximum number of "interior points" is reached. i ,b i ) and the number of "interior points" n;

[0130] ⑤ Repeat steps ①-④ until the iteration ends or the current model satisfies "the number of interior points is greater than the set threshold", finally obtaining the parametric model y = a of the structural surface skeleton line. i x+b i Each structural plane skeleton line corresponds to a parametric model M. i .

[0131] 5. Structural Surface Parameter Statistics

[0132] Based on the above work, this patent's Python graphics processing tool performs pixel-level statistical analysis of the geometric parameters of the structural surface skeleton lines obtained above, including counting the total number of traces, calculating the length and apparent dip angle of each skeleton line, and accordingly performing joint grouping and joint density calculation. The number N of limestone structural surface traces can be counted through a trace endpoint identification method, and the length and apparent dip angle of each trace can be obtained through a parametric model.

[0133] Model: y=a i x+b i Apparent tilt angle: θ = arctan|a i |

[0134] length:

[0135] Example 2

[0136] Based on the above technical solution, a limestone tunnel in Guangxi was selected for technical verification, and good technical results were achieved. Figure 9 These are photos of the tunnel face taken at the tunnel site using a smartphone. The outline of the tunnel face is cropped as follows: Figure 10 As shown, when a photograph is input into a structural surface trace recognition model, the resulting structural surface trace recognition effect is shown in the image below. Figure 11 As shown, the morphological processing effect after morphological processing and parameterization is as follows: Figure 12 As shown, the RANSAC effect diagram after straightening the skeleton lines of the limestone structural surface is as follows: Figure 13 As shown in Table 1, the number, length, and apparent tilt angle of the obtained structural surface traces are as follows.

[0137] Table 1. Statistics on the number, length, and apparent tilt angle of structural surface traces

[0138]

[0139] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for identifying rock mass structure surfaces in limestone tunnels, characterized in that, Includes the following steps: S1, Obtain the image of the face of the limestone tunnel; S2, input the image of the face of the limestone tunnel into the pre-trained limestone structure surface recognition model based on U-net to obtain the image of the rock mass structure surface of the limestone tunnel. The limestone structure surface recognition model based on U-net consists of an encoding part on the left, a decoding part on the right, and a convolutional layer and activation layer on the bottom. The decoding part on the right is the decoding part of the U-net network. The encoding portion on the left is a VGG16 network; S3, perform trace morphological processing on the image of the rock mass structure surface of the limestone tunnel to obtain the skeleton line of the rock mass structure surface of the limestone tunnel. The trace morphological processing includes skeletonization of the trace of the structure surface based on the skimage algorithm, splitting of the skeleton line of the limestone structure surface, and connecting of the skeleton line of the limestone structure surface. S4, the skeleton lines of the limestone tunnel rock mass structure surface are straightened to obtain the skeleton line parameter model of the limestone tunnel rock mass structure surface. The straightening process adopts the random sampling consensus algorithm. S5, perform pixel-level statistics on the skeleton line parameter model to calculate the length and viewing angle of each skeleton line; The encoding part on the left side of the U-net-based limestone structure surface recognition model is used for downsampling to extract multi-scale structure surface features. It includes several sub-modules, each of which includes two 3×3 convolutional layers, a piecewise linear function activation layer, and a 2×2 max pooling layer with a stride of 2. The decoding layer consists of four repeating structures: each repeating structure is preceded by a deconvolution for upsampling, after which the number of feature channels is halved and the feature map size is doubled; after deconvolution, the deconvolution result is concatenated with the feature map of the corresponding step in the encoding part to fuse structural surface features at multiple scales; the concatenated feature map is then subjected to two 3×3 convolutions; the last layer has a 1×1 convolution kernel, which transforms the 64-channel feature map into the result of structural surface recognition; The skeletonization process based on the skimage algorithm for structural surface traces includes the following steps: A31, the image of the rock mass structure surface of the limestone tunnel is binarized to obtain a binary image of the structure surface trace, and the binary image of the structure surface trace is eroded to obtain the eroded structure surface trace. A32, perform an opening operation on the eroded structural surface traces. The deleted pixel portion during the opening operation is part of the skeleton. Add the deleted pixel portion to the limestone structural surface trace skeleton diagram. A33, repeat steps A31 to A32 to obtain the limestone structure surface trace skeleton diagram corresponding to the limestone tunnel rock mass structure surface image; Step S4 specifically includes the following steps: S41, obtain the data point set corresponding to the skeleton line of the rock mass structure surface of the limestone tunnel, and select 2 data points from the data point set to establish a sample data subset; S42, Fit a linear mathematical model based on the subset of sample data; S43, use the linear mathematical model to test the remaining points in the data point set. If the tested data point is within the allowable error range, then the data point is judged as an interior point; otherwise, it is judged as an exterior point. The linear mathematical model and the corresponding number of interior points are obtained. S44, compare the number of interior points of the current linear mathematical model with the number of interior points of the previously obtained linear mathematical model, and obtain the maximum number of interior points between the two and the model parameters of the linear mathematical model corresponding to the maximum number of interior points. S45. Repeat steps S41 to S44 until the iteration ends or the current linear mathematical model has satisfied the condition that the number of interior points is greater than the set threshold, and obtain the final skeleton line parameter model of the limestone tunnel rock mass structure surface.

2. The method for identifying rock mass structure surfaces in limestone tunnels as described in claim 1, characterized in that, The training process of the U-net-based limestone structure surface recognition model includes the following steps: S21, set the model parameters to Epoch=200, batch_size=2, lr=0.0001; S22 uses the Adam optimizer to optimize network parameters and uses binary cross-entropy as the loss function to determine whether the model training has converged. S23. After training convergence, a well-trained limestone structure surface recognition model based on U-net is obtained.

3. The method for identifying rock mass structure surfaces in limestone tunnels as described in claim 1, characterized in that, The disassembly of the limestone structural plane skeleton line includes the following steps: B31, convert the limestone structural surface trace skeleton diagram into a two-dimensional matrix point set; B32 defines the eight-neighbor matrix of the target pixel P(x,y); B33, set the skeleton line pixel to 1 and the background to 0, use the eight-neighbor matrix to traverse the two-dimensional matrix point set, when a pixel has a situation in its eight-neighbor matrix that matches the two-dimensional matrix point set, determine that the pixel is a branch point; B34, modify the pixel value of the branch point to 0, delete the branch point, and obtain the trace of the broken structure surface.

4. The method for identifying rock mass structure surfaces in limestone tunnels as described in claim 3, characterized in that, The connection of the limestone structural surface skeleton line includes the following steps: C31, obtain the skeleton diagram of limestone structural surface traces after the skeleton lines are split, and locate and number the traces of the broken structural surfaces. C32, perform inclination similarity analysis on the traces of adjacent broken structural surfaces to determine whether the traces of adjacent broken structural surfaces belong to the same structural surface. If so, proceed to step C33. C33, through expansion The morphological manipulation method of corrosion connects the traces of the adjacent broken structural surfaces.

5. The method for identifying the rock mass structure surface of a limestone tunnel as described in claim 1, characterized in that, In step S5, the formula for calculating the length of the skeleton line is: ， The formula for calculating the apparent tilt angle of the skeleton line is ɵ=arctan|a i |, Among them, the skeleton line parameter model of the limestone tunnel rock mass structure surface is y= a i x+ b i a i is the slope coefficient of the skeleton line parameter model, (x, y) are the two-dimensional coordinates of the points on the skeleton line, max(x) is the maximum value of the abscissa on the skeleton line, min(x) is the minimum value of the abscissa on the skeleton line, max(y) is the maximum value of the ordinate on the skeleton line, and min(y) is the minimum value of the ordinate on the skeleton line.

6. As claimed in claim 1 5. A method for identifying the rock mass structure surface of a limestone tunnel as described in any one of the above, characterized in that, Step S1 also includes a preprocessing process for the limestone tunnel face image, as well as increasing the quantity and improving the quality of the preprocessed limestone tunnel face image. The preprocessing process includes: establishing a face contour database and using a face non-target information removal model trained on the open-source U-net network to remove non-target information from the images in the face contour database; The increase in quantity includes: obtaining limestone structural surface images at different angles by randomly rotating, flipping up, down, left, and right from the same face image; obtaining limestone structural surface images of the same size but different scales by cropping the image; and obtaining limestone structural surface images with different brightness and contrast by adjusting the image brightness and image contrast. The quality improvement includes: using a high-pass filtering method with a 3×3 convolution kernel to process images of limestone structural surfaces with small openings; and using a Gaussian blurring method with a 3×3 Gaussian kernel to process images of limestone structural surfaces with large openings and filling material.

7. A device for identifying the rock mass structure surface of a limestone tunnel, characterized in that, The method includes at least one processor and a memory communicatively connected to the at least one processor; the memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enable the at least one processor to perform a method for identifying the rock mass structure surface of a limestone tunnel according to any one of claims 1 to 6.

Images