CN115731390A

CN115731390A - Method and equipment for identifying rock mass structural plane of limestone tunnel

Info

Publication number: CN115731390A
Application number: CN202211509993.6A
Authority: CN
Inventors: 宋冠先; 梁铭; 朱孟龙; 彭浩; 黄能豪; 韩玉; 解威威; 马必聪; 陈孝强; 赵婷婷
Original assignee: Guangxi Road and Bridge Engineering Group Co Ltd
Current assignee: Guangxi Road and Bridge Engineering Group Co Ltd
Priority date: 2022-11-29
Filing date: 2022-11-29
Publication date: 2023-03-03

Abstract

The invention relates to the technical field of rock mass structural plane identification, in particular to a method and equipment for identifying a rock mass structural plane of a limestone tunnel. The method comprises the following steps: s1, acquiring a limestone tunnel face image; s2, inputting the limestone tunnel face image into a pre-trained limestone structural face recognition model based on U-Net to obtain a limestone tunnel rock structural face image; s3, performing trace morphological processing on the image of the structural plane of the limestone tunnel rock body to obtain a skeleton line of the structural plane of the limestone tunnel rock body; s4, carrying out linear treatment on the skeleton line of the limestone tunnel rock mass structural plane to obtain a skeleton line parameter model of the limestone tunnel rock mass structural plane; and S5, carrying out pixel level statistics on the skeleton line parameter model, and calculating the length and the apparent dip angle of each skeleton line. The method of the invention finally forms a set of efficient and accurate structural surface parameter identification and statistics technology for the limestone area.

Description

Method and equipment for identifying rock mass structural plane of limestone tunnel

Technical Field

The invention relates to the technical field of rock mass structural plane identification, in particular to a method and equipment for identifying a rock mass structural plane of a limestone tunnel.

Background

The structural plane is a geological interface with a certain structure extension direction and length and relatively small thickness in a rock body, and is a main control factor of the stability degree of the tunnel rock body, the structural plane parameters are important indexes for evaluating the quality of the rock body, and the method for effectively obtaining the parameters is a problem which is very concerned by personnel in the rock engineering industry.

The rock mass geological condition of the limestone area is obviously different from that of the sandstone area, the layering of the sandstone rock mass is good, the structural plane network is obvious, and the identification is easy. The limestone is generally in a blocky and thick-layer shape, has better homogeneity, has smaller opening degree of a structural plane in a rock body with less karst development, has unobvious extending direction, and is difficult to identify. The limestone area is easily affected by karst, partial corrosion cracks are large in opening degree, structural surface network irregularity is strong, and structural surfaces are obviously affected by fillers, so that structural surface networks in the limestone area are complex and changeable, and the workload of manually identifying and measuring structural surface parameters is large.

In a limestone tunnel region, the traditional acquisition of rock mass structural plane parameters is generally carried out by adopting a manual on-site contact measurement method, which is represented by a line measurement method and a statistical window method. Under the background of modern tunnel rapid construction, structural surface parameter measurement is difficult to be carried out by a manual method, so that a novel non-contact measurement mode is created, and the current novel tunnel surrounding rock structural surface parameter measurement method comprises a three-dimensional laser scanning technology, a digital photogrammetry technology, a prism-free total station technology and the like. The three-dimensional laser scanning technology utilizes a laser ranging principle to obtain reflectivity signals and three-dimensional coordinates of each point of a measured object, and establishes a three-dimensional space model of a measured structural surface through data processing, although the technology has high accuracy, the requirement on the measured section condition is high, the limestone structural surface recognition rate for some sections which are smooth and flat and small in opening degree is low, equipment is expensive and inconvenient to carry, and the later-stage data processing is complex; the digital photogrammetry technology acquires three-dimensional information of measuring points based on a multi-view imaging principle, utilizes image processing software to carry out three-dimensional reconstruction, and extracts required information from images, is feasible for identifying rock mass structural planes theoretically at present, but is limited by the fact that image processing theoretical research is not easy to match with actual engineering application, and particularly has unsatisfactory effects on the image processing technology of a complex structural plane network of limestone, particularly the technologies of identification, splitting and connection, small branch rejection, network trace linearization and the like of the structural plane network in the actual engineering application, thereby bringing great challenges to the structural plane identification of a limestone area.

Benefiting from the rapid development of the artificial intelligence theory, the three-dimensional laser scanning technology and the digital photogrammetry technology are gradually applied to the actual tunnel engineering. The Unet convolutional neural network has great advantages for rock mass fracture identification (Zhang Ye and the like, 2021; liehanco and the like, 2021), and the Unet convolutional neural network can effectively identify and extract the features in the image by performing combined operation of downsampling and upsampling on a sample picture. Compared with the traditional edge detection technology based on gradient change, the network can greatly improve the robustness of image identification, reduce the risk of overfitting and increase the identification efficiency. He Peng (2020) proposes to adopt linear bunching extraction and magnetic tracking extraction, and researches and develops a multi-parameter characterization method based on linear detection, intelligent scissors and a morphological edge detection algorithm respectively. Because the structure surface structure identified by the neural network is not a straight line composed of single elements, fitting and skeletonization processing are required to be carried out on the joint, and the current commonly used methods include a region growing method (Zhaowei et al, 2022), a GMM-EM algorithm, a random sampling consensus algorithm (Zhang yew et al, 2021), a dark space curve structure enhancement (DRCSE) algorithm (Yudi Tang, 2021), a Zhang-Suen algorithm (Lynda, 2008) and the like. On the basis of joint line skeletonization, the cold powerful (2021) provides an automatic grouping algorithm for the rock fracture boundary line of the palm face according to the fracture identification result, and further perfects the structural face identification statistical theory based on image identification.

The prior art does not mention a rock mass structural plane identification method in different lithological regions, and although the manual on-site contact measurement method can be used for measuring structural plane parameters under various lithological properties, the applicability is obviously poor under the complex structural plane network background of the limestone region. Similarly, no network identification scheme specially aiming at the complex structural surface in the limestone region is available in the method for acquiring the characteristic parameters of the rock structural surface based on image identification. Because the structural surface recognition statistics based on the image recognition technology is a multi-step comprehensive process, each step is realized by different algorithms, including structural surface trace extraction based on a neural network, a multi-segment line fitting algorithm for trace connection, a trace splitting and connecting algorithm, a structural surface grouping algorithm, an algorithm for calculating length and distance and the like. A reasonable and reliable limestone structural plane identification method needs to be provided from an algorithm level.

The technical problems are mainly as follows:

1. the traditional acquisition of rock mass structural plane parameters is generally carried out by adopting a manual on-site contact measurement method, and the method has the unfavorable characteristics of high labor intensity, low working efficiency, high safety risk and the like;

2. the geological condition of the rock mass in the limestone area is obviously different from that in the sandstone area, the layering of the sandstone rock mass is good, the network of the structural surface is obvious and easy to identify, the limestone structural surface is easily affected by the development degree of karst, the network of the structural surface is complex and various, and the trace identification difficulty of the structural surface is high;

3. the image recognition technology based on threshold segmentation and edge detection has poor reliability and low accuracy for structural surface recognition in a tunnel complex environment;

4. aiming at a complex limestone structural surface network, if a structural surface with large opening degree needs to be subjected to skeletonization, irrelevant branch skeleton lines and noise points are easily generated after the structural surface trace is subjected to skeletonization, and a mature solution suitable for removing the branch skeleton lines and the noise points of the structural surface is not provided;

5. the structural plane trace obtained by recognition is generally an irregular curve, and parameters obtained after the structural plane trace is fitted into a straight line segment can be used for subsequent conversion of rock mass quality indexes, calculation of jointed rock mass and numerical simulation.

Disclosure of Invention

Based on the background and the technical problems, a set of U-Net deep learning models suitable for identifying the limestone structural plane are constructed on one hand, and the identified models are subjected to trace morphological processing and linearization processing on the other hand, so that a limestone tunnel rock structural plane identification method and equipment are provided, the limestone tunnel rock structural plane is identified, and the length and the apparent dip angle of the corresponding skeleton line are obtained.

In order to achieve the above object, the present invention provides the following technical solutions:

a method for identifying a structural plane of a limestone tunnel rock body comprises the following steps:

s1, acquiring a limestone tunnel face image;

s2, inputting the limestone tunnel face image into a pre-trained limestone structure face recognition model based on U-Net to obtain a limestone tunnel rock structure face image, wherein the limestone structure face recognition model based on U-Net comprises a left encoding part, a right decoding part, a lower convolutional layer and an active layer, and the right decoding part is a decoding part of a U-Net network; the left encoded portion is a VGG16 network;

s3, performing trace morphological processing on the limestone tunnel rock structural plane image to obtain a skeleton line of the limestone tunnel rock structural plane, wherein the trace morphological processing comprises structural plane trace skeletonization processing based on a skimming algorithm, limestone structural plane skeleton line splitting and limestone structural plane skeleton line connection;

s4, carrying out linearization treatment on the skeleton line of the limestone tunnel rock mass structural plane to obtain a skeleton line parameter model of the limestone tunnel rock mass structural plane, wherein the linearization treatment adopts a random sampling consensus algorithm;

and S5, carrying out pixel level statistics on the skeleton line parameter model, and calculating the length and the apparent dip angle of each skeleton line.

As a preferred scheme, the coding part on the left side of the U-Net based limestone structural plane recognition model is used for down-sampling to extract multi-scale structural plane features, and comprises a plurality of sub-modules, wherein each sub-module comprises 23 × 3 convolutional layers, a piecewise linear function active layer and 12 × 2 maximum pooling layer with the step length of 2.

Preferably, the training process of the limestone structural plane recognition model based on the U-Net comprises the following steps:

s21, setting model parameters of Epoch =200, batch_size =2, lr =0.0001;

s22, using an Adam optimizer to optimize network parameters, and simultaneously using binary cross entropy as a loss function to judge whether model training is converged;

and S23, after the training is converged, obtaining a trained limestone structural plane recognition model based on U-Net.

Preferably, the skeleton processing of the structural surface trace based on the skimage algorithm comprises the following steps:

a31, carrying out binarization processing on the limestone structural plane image of the limestone tunnel to obtain a structural plane trace binary image, and corroding the structural plane trace binary image to obtain a corroded structural plane trace;

a32, performing open operation on the corroded structural surface trace, wherein a deleted pixel part is a part of a skeleton during open operation processing, and adding the deleted pixel part into a limestone structural surface trace skeleton diagram;

and A33, repeating the steps A31-A32 to obtain a limestone structural plane trace skeleton diagram corresponding to the limestone structural plane image of the limestone tunnel.

Preferably, the method for splitting the skeleton line of the limestone structural plane comprises the following steps:

b31, converting the limestone structure surface trace skeleton map into a two-dimensional matrix point set;

b32, defining an eight-neighborhood matrix of the target pixel point P (x, y);

b33, setting the skeleton line pixel as 1 and the background as 0, traversing the two-dimensional matrix point set by adopting the eight-neighborhood matrix, and judging that a certain pixel point is a branch point when the situation conforming to the two-dimensional matrix point set appears in the eight neighborhoods of the pixel point;

and B34, modifying the pixel value of the branch point to be 0, and deleting the branch point to obtain a trace of the disconnected structure surface.

Preferably, the connection of the skeleton line of the limestone structure surface comprises the following steps:

c31, acquiring a skeleton diagram of the trace of the limestone structural plane after the skeleton line is split, and positioning and numbering the trace of the disconnected structural plane;

c32, performing inclination angle similarity analysis on the traces of the adjacent disconnection structure surfaces to judge whether the traces of the adjacent disconnection structure surfaces belong to the same structure surface, and if so, performing the step C33;

and C33, connecting the traces of the adjacent broken structure surfaces by a morphological operation method of expansion-corrosion.

Further, step S4 specifically includes the following steps:

s41, acquiring a data point set corresponding to a skeleton line of the limestone structural plane of the limestone tunnel, selecting 2 data points from the data point set, and establishing a sample data subset;

s42, fitting a straight line mathematical model according to the sample data subset;

s43, testing the remaining points in the data point set by using the linear mathematical model, judging the data point as an inner point if the tested data point is within an error allowable range, otherwise, judging the data point as an outer point, and obtaining the linear mathematical model and the corresponding number of the inner points;

s44, comparing the number of the inner points of the current linear mathematical model with the number of the inner points of the previously acquired linear mathematical model, and acquiring the maximum number of the inner points between the current linear mathematical model and the previously acquired linear mathematical model and the model parameters of the linear mathematical model corresponding to the maximum number of the inner points;

and S45, repeating the steps S41-S44 until the iteration is finished or the current straight line mathematical model meets the condition that the number of the inner points is more than a set threshold value, and obtaining a final skeleton line parameter model of the limestone structural plane of the limestone tunnel.

Further, in step S5, the length of the skeleton line is calculated by the following formula:

the formula of the tilt angle of the skeleton line is θ = arctan | a _i |，

Wherein the parameter model of the skeleton line of the rock mass structural plane of the limestone tunnel is y = a _i x+b _i ，a _i The gradient coefficient of the skeleton line parameter model is shown as (x, y) are two-dimensional coordinates of points on the skeleton line, max (x) is the maximum value of horizontal coordinates on the skeleton line, min (x) is the minimum value of horizontal coordinates on the skeleton line, max (y) is the maximum value of vertical coordinates on the skeleton line, and min (y) is the minimum value of vertical coordinates on the skeleton line.

Further, the step S1 also comprises a pretreatment process of the limestone tunnel face image, and the quantity and quality of the pretreated limestone tunnel face image are increased;

the pretreatment process comprises the following steps: establishing a palm surface profile database, and removing non-target information on pictures in the palm surface profile database by adopting a palm surface non-target information removal model trained on the basis of an open source U-Net network;

the increase in the number includes: randomly rotating the same face image and turning the same face image up, down, left and right to obtain limestone structural plane images with different angles; obtaining limestone structural plane images with the same size but different scales through image cutting; obtaining limestone structural plane images with different brightness and different contrast ratios by adjusting the image brightness and the image contrast ratio;

the mass improvement comprises: processing the limestone structural plane image with small opening degree by adopting a high-pass filtering method with convolution kernel of 3 multiplied by 3; and processing the limestone structural plane image with large opening and filled with fillers by adopting a Gaussian blur method with the Gaussian kernel size of 3 multiplied by 3.

Based on the same conception, the equipment for identifying the structural plane of the rock mass of the limestone tunnel is further provided, and comprises at least one processor and a memory which is in communication connection with the at least one processor; the memory stores instructions executable by the at least one processor to enable the at least one processor to perform a method of identifying a structural plane of a limestone tunnel rock mass according to any one of the preceding claims.

Compared with the prior art, the invention has the beneficial effects that:

the limestone tunnel rock mass structural plane identification method and device are used for solving the problem that complex limestone structural plane parameters are difficult to extract, and finally a set of efficient and accurate structural plane parameter identification statistical technology for a limestone area is formed.

Drawings

Fig. 1 is a flow chart of a method for identifying a structural plane of a limestone tunnel rock mass in embodiment 1;

fig. 2 is a picture of data acquired by a smartphone in image acquisition in embodiment 1;

FIG. 3 is a picture of data collected by a digital camera in the image collection in example 1;

fig. 4 is an original image of a working face of a limestone tunnel before image data of the working face is preprocessed in embodiment 1;

fig. 5 is an image from which non-target information is removed in the image data preprocessing of the face of the limestone tunnel in embodiment 1;

FIG. 6 is an original image of the structure surface in example 1 when labeled;

FIG. 7 is a label image when labeling the structure surface in example 1;

FIG. 8 is a framework diagram of the improved U-Net model in example 1;

fig. 9 is a picture of a tunnel face acquired by the smartphone in a tunnel field in embodiment 2;

FIG. 10 is a silhouette trimming image in the structural plane recognition effect map in example 2;

fig. 11 is a structural plane trace recognition effect diagram in the structural plane recognition effect diagram in example 2;

FIG. 12 is a graph showing the effect of morphological processing in a structural plane recognition effect graph in example 2;

fig. 13 is a RANSAC effect map in the structure plane recognition effect map in embodiment 2.

Detailed Description

The present invention will be described in further detail with reference to test examples and specific embodiments. It should be understood that the scope of the above-described subject matter is not limited to the following examples, and any techniques implemented based on the disclosure of the present invention are within the scope of the present invention.

Example 1

A method for identifying a structural plane of a limestone tunnel rock body is shown in a flow chart of figure 1 and specifically comprises the following steps:

1. image acquisition and image data preprocessing of face of limestone tunnel

(1) Limestone tunnel face image acquisition

The image recognition technology has certain requirements on image quality, different limestone structural plane images are included for improving the image acquisition efficiency and quality, a picture of data acquired by a smart phone during image acquisition is shown in figure 2, a picture of data acquired by a digital camera during image acquisition is shown in figure 3, and the following image acquisition requirements need to be met no matter how the image is acquired by adopting the method:

(1) selecting image acquisition equipment with more than 2000 ten thousand pixels, such as a smart phone, a digital camera and the like, and ensuring that the definition of an image meets the requirement of model training;

(2) considering that the environmental conditions are poor when the images are collected on a tunnel construction site and the man-made interference factors are more, the tunnel face image collection time is determined after the tunnel is tapped and danger is eliminated and before lining support is developed, the sufficient light of the tunnel face and the shielding of interference-free objects are ensured in the shooting process, the whole tunnel face is covered in the shooting range, and steps, side walls and lining are prevented from being brought into the shooting range as much as possible.

(3) The acquired images comprise limestone structural plane images under different karst development conditions, such as limestone images with smaller structural plane openness under the condition of no karst development, and images with large structural plane openness, no filler, filler and the like under the condition of karst development.

(2) Limestone tunnel face image data preprocessing method

The image preprocessing aims at eliminating non-target information (image cutting), standardizing the labeling of a structural surface, acquiring multi-angle and multi-scale limestone structural surface data, highlighting limestone structural surface characteristics in an image or reducing the influence of erosion crack fillers, irrelevant noise points and the like on the limestone structural surface, and finally realizing the improvement of the quantity and the quality of a data set.

(1) Image cropping

As the collected tunnel face image inevitably comprises tunnel steps and linings, the extraction of model features can be influenced by non-target information, and the information is removed firstly. According to the method, manual cutting is combined with an image segmentation technology (semantic segmentation) to establish a face profile database, non-target information on pictures in the database is removed, original faces of face images of a limestone tunnel before preprocessing are shown in figure 4, and images of face images of the limestone tunnel after non-target information is removed in preprocessing are shown in figure 5. The number of pictures in the database should be guaranteed to be more than 500 respectively. Model training is carried out by adopting an open-source U-Net network to obtain a palm surface non-target information removing model, so that the palm surface non-target information is removed.

(2) Limestone structural surface data annotation

The trace marking of the limestone structure surface can be combined by offline marking and online cloud marking. When the sample data volume is small, manually marking each structural surface trace on the palm surface image by using a labelme-based local marking tool. When the sample size of the palm surface image is large and the workload is large, the easy data online annotation tool can be adopted to realize the online annotation of the cloud of a plurality of people so as to improve the annotation efficiency. For the limestone structural surface which is not developed by karst, the extending direction of the structural surface with small opening degree is focused, and part of the structural surface is easily filled by cement, so that omission and error marking are avoided. For a limestone structure surface developing in a karst, particularly when a corrosion crack is filled with fillers, the structure surface is marked along a boundary line of the corrosion crack as much as possible. As a specific example, an original image when the structure surface is labeled is shown in fig. 6, and a label image when the structure surface is labeled is shown in fig. 7.

(3) Data set quantity and quality improvement

The structural surface recognition model training needs to be based on a certain number of limestone structural surface images, and under the condition of less image data, limestone structural surfaces with different angles and different scale ranges are difficult to cover. Therefore, the scheme obtains more comprehensive structural plane image data through the method of generating the network through image transformation and countermeasure. The same face image can obtain limestone structural planes with different angles through random rotation (the rotation angle is automatically generated by a code), up-down, left-right turning and the like; by means of image zooming (the image is zoomed to be 0.6, 0.8, 1.2, 1 and 4 times of the original image), and then by means of image cutting (the zoomed image is cut to be 224 multiplied by 224), the limestone structure surface images which are consistent in size and different in size are obtained, and the problem that the opening degree of part of the limestone structure surface is not easy to identify is solved; gray rock tunnel face images with different brightness and different contrast ratios are obtained by adjusting the image brightness and the image contrast ratio (the image brightness and the contrast ratio are respectively adjusted to be 0.7, 0.9, 1.1 and 1.3 times of the original image in the scheme), and gray rock tunnel face images with different brightness and contrast ratios acquired under different light backgrounds of the tunnel are simulated; in addition, the scheme introduces an open source generation confrontation network model, and generates a certain number of limestone working face images on the basis of an original data set.

Aiming at the problems that the limestone structural plane with small opening degree is not easy to identify, the limestone structural plane with large opening degree is easy to be influenced by fillers, and the like, an image enhancement technology is introduced to improve the quality of the data set. According to the scheme, the high-pass filtering method with the convolution kernel of 3 x 3 is adopted to process the structural plane image with small opening, and the Gaussian blur method with the Gaussian kernel of 3 x 3 is adopted to effectively process the limestone structural plane image with large opening and filled with fillers.

Therefore, by combining the image transformation and the image enhancement technology, the quantity and the quality of the data set can be further improved, and therefore the applicability and the generalization capability of the model in the limestone region are improved.

2. Construction and training of limestone structural plane recognition model based on U-Net

The original U-Net network structure was originally proposed by Ronneberger et al (2015), and the neural network was widely applied to the field of image segmentation, especially in the field of medical image recognition, and the application is becoming mature. The improved U-Net network is constructed aiming at the characteristics of the limestone structural surface, the left coding part of the improved model feature extraction network takes a VGG16 network as a framework, and the right decoding part of the improved model feature extraction network consists of an original U-Net network decoding part. The improved network model is composed of a left encoding part, a right decoding part and two convolution + activation layers at the lower side (as shown in FIG. 8).

(1) Left side coding part

The framework is composed of 4 repetitive structures: namely 23 × 3 convolutional layers + immediately following a piecewise linear function active layer (ReLU), 1 maximum pooling layer of 2 × 2 (max) with step size (stride) of 2; after each down-sampling, the number of the characteristic channels is doubled, the change process of the number of the channels is [3,64 ] - [6,128 ] - [12,256 ] - [24,512 ], and multi-scale structural surface characteristics are extracted.

(2) Right side decoding part

Like the encoding layer, the decoding layer also consists of 4 repeating structures: before each repetitive structure, deconvolution is used, the number of characteristic channels is halved after each deconvolution, and the size of a characteristic graph is doubled; after deconvolution, splicing the deconvolution result with the feature map of the corresponding step of the coding part (white/gray block), namely fusing the structural surface features under multiple scales; performing 3 × 3 convolution for 2 times on the spliced feature map; the convolution kernel of the last layer is a convolution kernel of 1 multiplied by 1, and the feature map of 64 channels is converted into a structural plane identification result.

(3) Model training

Model training involves data set introduction, model parameter adjustment, network parameter optimization, and training loss value calculation. Transmitting the preprocessed limestone structural plane image and the corresponding label image into a constructed improved U-Net model, setting model parameters of Epoch =200, batch \ size =2, lr =0.0001, performing network parameter optimization by using an Adam optimizer, and judging whether model training is converged by using Binary cross entropy (Binary cross entropy) as a loss function. And storing the weight model after the training convergence to finally obtain the optimal limestone structural plane prediction model.

3. Structural surface trace morphological processing

(1) Skeleton of structural surface trace based on sketch algorithm

After the predicted image is identified by the U-Net model, the limestone structural surface trace is marked and is segmented from the predicted image, the marked structural surface trace is a 'surface' formed by multiple pixels instead of a 'line' formed by connecting multiple single-phase pixel points, and if the structural surface trace parameters are further extracted, the identification result needs to be subjected to skeletonization processing to obtain the structural surface trace connected by the single-phase pixel points.

This patent adopts the skeleton of sketch image refinement algorithm to carry out structural plane trace skeletonization, and skeleton extraction of sketch belongs to the morphological processing category, realizes based on image expansion and corruption, and its principle is: let A be the target image to be processed, B be the structural element, S (A) represents the skeleton of A, S _k (A) For the kth skeleton subset of a, the skeleton expression of a is found as follows.

Based on the principle, the method comprises the following concrete implementation steps:

(1) corroding the binary image of the structural surface trace obtained by U-Net recognition, wherein the corroded structural surface trace becomes narrower and thinner;

(2) opening operation is carried out on the corroded image, the pixel part deleted in the opening operation processing is a part of the skeleton, and the pixel part is added into the skeleton image;

(3) and repeating the processes until the image is completely corroded, and finally obtaining a limestone structure surface trace skeleton diagram.

(2) Detachment and connection of skeleton line of limestone structural surface

The trace characteristics of the skeletonized limestone structure surface are obvious, namely the network structure is complex and obvious distribution rules are difficult to distinguish; a plurality of branches with different lengths exist in some structural surface traces, and some short branches are generated due to the influence of image noise points and structural surface traces unevenness in the skeletonization process; some structural surface traces have more local breakpoints, and the structural surface traces in the actual image are communicated into one trace. Therefore, the skeleton line of the limestone structural plane needs to be detached and connected.

The specific implementation method comprises the following steps:

1) Dismantling of skeleton line of limestone structural surface

(1) And (5) converting image data. Converting the skeletonized trace image into a corresponding two-dimensional matrix point set M through a related python image processing toolkit;

(2) defining the coordinate set of eight neighborhood pixel points of the target pixel point P (x, y) as P ₈ [(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)]

(3) Identifying the branch point of the trajectory of the limestone structural plane. Firstly, defining a branch point identification matrix, and deducing an eight-neighborhood matrix Q of pixel points belonging to a branch point by adopting an exhaustion method:

[[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],[1,0,1,0,1,1,0,1],[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]······]

and (3) setting the skeleton line pixel as 1 and the background as 0, traversing the skeleton line two-dimensional matrix point set M by adopting an eight-neighborhood matrix P8, and judging that a certain pixel point 8 is a branch point when the neighborhood of the pixel point meets the matrix Q. And the pixel point value is modified to be 0 so as to achieve the purpose of deleting branch points, and each skeleton line is separated from the structural plane network.

2) Limestone structural surface skeleton line connection

(1) Positioning and marking of structural surface traces

Firstly, positioning and numbering certain broken structure surface traces, wherein the traces are positioned and numbered by adopting cv2.DrawContours of opencv;

(2) similarity determination

To two adjacent structure surface traces L ₁ 、L ₂ And performing dip angle similarity analysis to judge whether the structural surface traces belong to the same structural surface. Respectively calculating the traces L of two adjacent structure surfaces ₁ 、L ₂ Inclination angle theta ₁ And theta ₂ When | θ ₁ -θ ₂ |<At 15 deg.C, judge L ₁ 、L ₂ Belonging to the same structural plane.

(3) Trace connections

By morphological manipulation of swelling-corrosion on L ₁ 、L ₂ Performing a treatment of the expansion-corrosion convolution kernel [ n × n]According to the size of L ₁ 、L ₂ Depending on the distance between the ends, n should be largeAt the end point distance. According to the scheme, the structural surface trace is connected through expansion-corrosion, and a structural surface network which is closer to the actual structural surface network is obtained after once skeletonization treatment.

(3) And deleting the branch point.

Aiming at the problems that the skeletonized image of the limestone structural surface has complex change and different lengths of skeleton trace graphs, more branch lines are generated particularly for the structural surface skeleton with large openness, and the branch lines belong to a false structural surface, namely noise point information. Therefore, the patent introduces an adaptive threshold segmentation method, firstly, the average length L of all skeleton lines is counted, the segmentation threshold is set to be 0.05L, and when the length L of a trace line is equal to the average length L of all skeleton lines _n <And when the number L is less than the preset threshold, judging the skeleton line as a noise point and deleting the skeleton line. The division threshold coefficient of 0.05 is obtained by counting a large number of true and false (noise point) structural plane data of the skeleton line of the limestone structural plane.

4. Structural plane skeleton line linearization processing based on random sampling consistency algorithm

The structural plane skeleton line after morphological processing is generally an irregular curve, parametric identification is difficult to perform, and linearization processing needs to be performed on the structural plane skeleton line. The method comprises the following implementation steps:

(1) each structural surface skeleton line corresponds to one data point set, one data set S is sequentially selected, the minimum number of samples required for building the model is selected, and because two-dimensional straight line segments need to be fitted, the number of samples required for modeling is 2, and the number of 2 sample data sets is recorded as S1;

(2) using the selected data set S1, a mathematical model M1 is calculated: y = ax + b;

(3) testing the rest points in the data set by using the calculated model M1, judging the data point as an inner point (inlier) if the tested data point is in an error allowable range, otherwise judging the data point as an outer point (outlier), and recording the data set formed by all the inner points as a consistency set S11, wherein S11 is called as a consistency set S1;

(4) comparing the number of "interior points" of the current model M1 and the best previously derived model Mi, recording the model parameters (a) at the maximum number of "interior points _i ,b _i ) And the number of "interior points" n;

(5) repeating the steps (1) and (4) until the iteration is finished or the current model meets the condition that the number of the inner points is more than the set threshold value, and finally obtaining the parameterized model y = a of the skeleton line of the structural plane _i x+b _i Wherein each structural plane skeleton line corresponds to a parameterized model M _i 。

5. Structural surface parameter statistics

On the basis of the work, the python graphic processing tool carries out pixel-level statistics on the geometric parameters of the skeleton lines of the structural surface, including the statistics of the total number of the traces, the length and the apparent dip angle of each skeleton line are calculated, and joint grouping and joint density calculation are carried out according to the statistics. The number N of the tracings of the limestone structure surface can be counted by a trace endpoint identification method, and the length and the apparent dip angle of each trace can be obtained by a parameterized model:

model: y = a _i x+b _i Apparent dip angle: θ = arctan | a _i |

Length:

example 2

Based on the technical scheme, a certain limestone tunnel in Guangxi is selected for technical verification, and a better technical effect is obtained. Fig. 9 is a photograph of a tunnel face collected on site by a smart phone, the contour of the tunnel face is cut as shown in fig. 10, the photograph is input into a structural face trace recognition model, the structural face trace recognition effect in the recognition effect graph is shown in fig. 11, the morphological processing effect graph after morphological processing and parameterization is shown in fig. 12, and the RANSAC effect graph after the linearization processing of the skeleton line of the limestone structural face is shown in fig. 13. The number of traces, length, and apparent dip of the resulting structure are shown in table 1.

TABLE 1 statistics of number, length, apparent dip of traces of structural surface

The above description is only for the purpose of illustrating the preferred embodiments of the present invention and is not to be construed as limiting the invention, and any modifications, equivalents and improvements made within the spirit and principle of the present invention are intended to be included within the scope of the present invention.

Claims

1. A method for identifying a rock mass structural plane of a limestone tunnel is characterized by comprising the following steps:

s1, acquiring a limestone tunnel face image;

s2, inputting the limestone tunnel face image into a pre-trained limestone structure face recognition model based on U-Net to obtain a limestone tunnel rock structure face image, wherein the limestone structure face recognition model based on U-Net comprises a left encoding part, a right decoding part, a lower convolution layer and an activation layer, and the right decoding part is a decoding part of a U-Net network; the left encoded portion is a VGG16 network;

s3, performing trace morphological processing on the limestone tunnel rock structural plane image to obtain a skeleton line of the limestone tunnel rock structural plane, wherein the trace morphological processing comprises structural plane trace skeletonization processing based on a skeleton algorithm, limestone structural plane skeleton line splitting and limestone structural plane skeleton line connection;

2. The method for identifying the rock mass structural plane of the limestone tunnel according to claim 1, wherein a coding part on the left side of the U-Net-based limestone structural plane identification model is used for down-sampling to extract multi-scale structural plane features, and the method comprises a plurality of sub-modules, wherein the sub-modules comprise 23 x 3 convolutional layers, a piecewise linear function active layer and 12 x 2 maximum pooling layer with the step length of 2.

3. The method for identifying the rock mass structural plane of the limestone tunnel as claimed in claim 2, wherein the training process of the U-Net based limestone structural plane identification model comprises the following steps:

s21, setting model parameters of Epoch =200, batch _size =2, lr =0.0001;

s22, optimizing network parameters by using an Adam optimizer, and meanwhile, using binary cross entropy as a loss function to judge whether model training is converged;

and S23, obtaining the trained limestone structural surface recognition model based on the U-Net after the training is converged.

4. The method for identifying the structural plane of the limestone tunnel rock body as claimed in claim 1, wherein the structural plane trace skeletonization processing based on the skimage algorithm comprises the following steps:

a31, performing binarization processing on the limestone tunnel rock mass structural plane image to obtain a structural plane trace binary image, and corroding the structural plane trace binary image to obtain a corroded structural plane trace;

5. The method for identifying the rock mass structural plane of the limestone tunnel according to claim 4, wherein the splitting of the skeleton line of the limestone structural plane comprises the following steps:

b31, converting the trace skeleton diagram of the limestone structural plane into a two-dimensional matrix point set;

b32, defining an eight-neighborhood matrix of the target pixel point P (x, y);

6. The method for identifying the limestone structural plane of the limestone tunnel according to claim 5, wherein the connection of the skeleton line of the limestone structural plane comprises the following steps:

c31, acquiring a skeleton drawing of the limestone structural plane trace after the skeleton line is split, and positioning and numbering the trace of the broken structural plane in the skeleton drawing;

7. The method for identifying the structural plane of the limestone tunnel rock body according to claim 1, wherein the step S4 specifically comprises the following steps:

s44, comparing the number of the inner points of the current linear mathematical model with the number of the inner points of the previously obtained linear mathematical model, and obtaining the maximum number of the inner points between the current linear mathematical model and the previously obtained linear mathematical model and the model parameters of the linear mathematical model corresponding to the maximum number of the inner points;

and S45, repeating the steps S41-S44 until iteration is finished or the current straight line mathematical model meets the condition that the number of the inner points is greater than a set threshold value, and obtaining a final skeleton line parameter model of the rock mass structural plane of the limestone tunnel.

8. The method for identifying the structural plane of the limestone tunnel rock body according to claim 1, wherein in the step S5, the length calculation formula of the skeleton line is as follows:

，

the apparent dip angle of the skeleton line is calculated by the formulaɵ=arctan| a _i |，

Wherein the parameter model of the skeleton line of the rock mass structural plane of the limestone tunnel is y = a _i x+ b _i ，a _i The gradient coefficient of the skeleton line parameter model is shown, wherein (x, y) are two-dimensional coordinates of points on the skeleton line, max (x) is the maximum value of the horizontal coordinate of the skeleton line, min (x) is the minimum value of the horizontal coordinate of the skeleton line, max (y) is the maximum value of the vertical coordinate of the skeleton line, and min (y) is the minimum value of the vertical coordinate of the skeleton line.

9. The method for identifying the rock mass structural plane of the limestone tunnel according to any one of claims 1 to 8, wherein the step S1 further comprises a pretreatment process of the limestone tunnel face image, and the quantity and quality of the pretreated limestone tunnel face image are increased;

the increase in the number includes: obtaining limestone structural plane images with different angles by randomly rotating the same face image and turning the same face image up, down, left and right; obtaining limestone structural plane images with the same size but different scales through image cutting; obtaining limestone structural plane images with different brightness and different contrast ratios by adjusting the image brightness and the image contrast ratio;

the mass improvement comprises: processing the limestone structural plane image with small opening degree by adopting a high-pass filtering method with convolution kernel of 3 multiplied by 3; and processing the limestone structural plane image which is large in opening and filled with filling materials by adopting a Gaussian blur method with the Gaussian kernel size of 3 multiplied by 3.

10. The equipment for identifying the rock mass structural plane of the limestone tunnel is characterized by comprising at least one processor and a memory which is in communication connection with the at least one processor; the memory stores instructions executable by the at least one processor to enable the at least one processor to perform a method of identifying a limestone tunnel rock mass structural plane according to any one of claims 1 to 9.