A deep learning-based intelligent rapid regional classification method for tunnel rock mass

Abstract

The present application relates to a kind of based on deep learning's tunnel face rock mass intelligent fast regional grading method, comprising: dividing tunnel face area;Drill rig positioning, drill rig is fixed in preset position;Drilling parameter is obtained by drilling process;Through establishing the relationship model of drilling parameter and rock uniaxial compressive strength, rock uniaxial compressive strength is solved using drilling parameter;Rock uniaxial compressive strength in step (four) is substituted into BQ method model using improved BQ method, and BQ method correction model is obtained by Pearson linear correlation coefficient calculation formula;BQ method correction model is combined with depth learning to identify tunnel face fissure information, and then the intelligent fast regional grading of rock mass is carried out.The present application solves the problems of low accuracy of artificial data collection, low identification efficiency, inaccurate surrounding rock grade, low construction efficiency and other problems of existing grading method.

Description

Technical Field

[0001] This invention belongs to the field of tunnel engineering, specifically relating to a deep learning-based intelligent and rapid regional classification method for rock mass at the tunnel face during tunnel blasting. Background Technology

[0002] In recent years, my country has constructed a large number of important underground projects, including transportation tunnels, underground mining, underground hydropower plants, and underground energy storage caverns. Underground projects such as tunnels are located in complex geological environments. Current design theories and construction technologies are still incomplete, and the construction process involves significant uncertainties, making it a complex and high-risk systems engineering project. Conventional surrounding rock classification methods are slow and influenced by many qualitative factors. There is a lack of intelligent and rapid classification methods for tunnel surrounding rock during construction. Furthermore, the collection and analysis of quantitative indicators are difficult. As a result, surrounding rock classification during construction is mainly determined by human observation at the tunnel and underground engineering site, relying on subjective judgment and lacking scientific rigor. This has led to accidents caused by incorrect surrounding rock classification, and the difficulty in collecting and analyzing quantitative indicators has resulted in loss of life and property.

[0003] The continuous improvement in the speed, quantification, and scientific nature of surrounding rock classification through modern science and technology provides a basis and guidance for the dynamic design and construction of tunnel engineering, and also promotes the development of tunnel surrounding rock classification technology towards speed and intelligence. Generally speaking, deep learning can quickly and accurately judge the state, properties, and other elements of rock, greatly improving the efficiency of engineering projects while ensuring the safety of construction sites. Using convolutional neural networks to analyze point cloud images can quickly obtain feature information of the tunnel face. Compared with traditional methods that rely on manual methods, using computers to obtain tunnel face information is a leading approach, with advantages such as shorter image acquisition time and convenient information processing. This reduces the working time of construction personnel in the tunnel and significantly lowers safety hazards, which is highly beneficial to on-site construction personnel. Furthermore, the working parameters obtained by the drilling rig during drilling can effectively reflect geological changes. The pressure and vibration generated during drilling also vary with different geological conditions, which is of great reference value for predicting the surrounding rock grade. During the formation of rock masses, geological interfaces constantly change and develop, forming their own distribution patterns and scale. Real-time identification of the tunnel face state plays a crucial role in engineering construction.

[0004] The distribution of fractures at tunnel faces varies greatly, and their physical and mechanical properties often differ. Closed fractures, micro-tensioned fractures, open fractures, and wide-tensioned fractures are common types of fractures at tunnel faces. Since the drilling tool is in direct contact with the rock mass during drilling, the tool's response information reflects the rock mass's mechanical characteristics. Research on extracting fracture image information from tunnel faces and drilling parameters during the drilling process using a rock drilling rig is crucial. By utilizing the correlation between tunnel face fractures and the surrounding rock grade, more accurate determination of the surrounding rock grade can be achieved. This can also aid in disaster early warning and improve work efficiency, which is of significant importance.

[0005] In summary, existing classification methods suffer from problems such as low accuracy of manual data collection, low identification efficiency, inaccurate prediction of surrounding rock grade, and low construction efficiency. Summary of the Invention

[0006] This invention proposes a deep learning-based intelligent and rapid regional classification method for rock mass at tunnel face, aiming to solve the problems of low accuracy of manual data collection, low identification efficiency, inaccurate prediction of surrounding rock grade, and low construction efficiency in existing classification methods.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A deep learning-based intelligent and rapid regional classification method for tunnel rock mass includes the following steps:

[0009] Step (1) Divide the tunnel face area;

[0010] Step (2) Positioning the rock drilling rig: Fix the rock drilling rig in the preset position;

[0011] Step (3) Obtain drilling parameters through the drilling process;

[0012] Step (iv) By establishing a model relating drilling parameters to the uniaxial compressive strength of rock, the uniaxial compressive strength of rock is determined using the drilling parameters.

[0013] Step (5) The improved BQ method is used to substitute the uniaxial compressive strength of the rock in step (4) into the BQ method model, and the modified BQ method model is obtained by calculating the Pearson linear correlation coefficient.

[0014] Step (VI) combines the BQ method correction model from step (V) with deep learning to identify crack information at the tunnel face, thereby enabling intelligent and rapid regional classification of the rock mass.

[0015] Furthermore, the model relating drilling parameters to the uniaxial compressive strength of rock is as follows:

[0016]

[0017] Where: σ c Ω is the uniaxial compressive strength of the rock, MPa; W is the axial pressure on the drill bit, N; Ω is the rotational speed, r / min; V is the drilling rate, m / min; D is the drill bit diameter, m; f is the energy transfer rate; μ is the drill bit slip friction coefficient.

[0018] Furthermore, the specific steps of the improved BQ method are as follows:

[0019] Step (1) Collect images of cracks at the tunnel face, generate a training set, and begin network training until the convolutional neural network is mature;

[0020] Step (2) convert the cracks at the tunnel face to be tested into a point cloud image;

[0021] Step (3) Point cloud image preprocessing;

[0022] Step (4): Based on the mature convolutional neural network trained in step (1), feature recognition is performed on the point cloud image;

[0023] Step (5) The residual network module based on the convolutional neural network performs secondary processing on the point cloud image, performs deconvolution on the identified tunnel face crack point cloud image, transforms the tunnel face crack point cloud image into a clearer and more feature-clear point cloud image, and then identifies the clear point cloud image. With each iteration, the accuracy of identifying the tunnel face crack point cloud image is further improved.

[0024] Step (6) Use the drilling parameters and the uniaxial compressive strength of rock relationship model to obtain the uniaxial compressive strength of rock and apply it to the BQ method model;

[0025] Step (7) uses the identification results obtained in step (5) to output the information on the types of cracks at the tunnel face, and uses the Pearson linear correlation coefficient calculation formula to obtain the influence coefficient of cracks at the tunnel face, and applies it to the BQ method correction model to correct the BQ method model.

[0026] Furthermore, the residual network module steps of the convolutional neural network in step (5) are as follows:

[0027] a. Selecting locations of cracks at the tunnel face and outputting feature layers through feature recognition;

[0028] b. Input the output feature layer into the deconvolution module;

[0029] c. After the deconvolution operation, the output feature layer is added to the initial feature layer to obtain the final output feature layer. This process is repeated multiple times until the output result meets the actual requirements.

[0030] Furthermore, the BQ method model is as follows:

[0031] BQ = 90 + 3σ c +250K v

[0032] K v =(V pm / V pv ) 2

[0033] Where: σ c V represents the uniaxial compressive strength of the rock obtained using a drill bit. pm V represents the elastic longitudinal wave velocity of the rock mass, in km / s. pr Let be the elastic longitudinal wave velocity of the rock, in km / s.

[0034] Furthermore, the formula for calculating the Pearson linear correlation coefficient is as follows:

[0035]

[0036] In the formula: n is the sample size, X i Y i These are the observation values ​​at point i corresponding to variables X and Y. It is the sample mean of X. is the sample mean of Y, and r is the correlation coefficient.

[0037] Furthermore, the modified BQ method model is as follows:

[0038] [BQ] 1 =BQ-100(K1+K2+K3+r1+r2+r3+r4)

[0039] In the formula: K1 is the groundwater influence correction coefficient; K2 is the main weak structural plane attitude influence correction coefficient; K3 is the initial stress state influence correction coefficient; r1 is the closed crack influence coefficient; r2 is the micro-tension crack influence coefficient; r3 is the open crack influence coefficient; and r4 is the wide-tension crack influence coefficient.

[0040] The beneficial effects of this invention are:

[0041] This invention is based on the recognition of point cloud image information of tunnel face. It uses equipment to capture images of the tunnel face to obtain point cloud images, and uses a computer to generate tunnel face point cloud data. The acquired point cloud images are processed using algorithms such as image enhancement, grayscale conversion, sharpening, denoising, RANSAC algorithm, and Hough transform. Then, a convolutional neural network is used to recognize the tunnel face point cloud images. By identifying the type of cracks on the tunnel face surface, such as closed cracks, micro-open cracks, open cracks, and wide-open cracks, and by optimizing the recognition effect using algorithms, the crack recognition rate and effect are greatly improved. By dividing the tunnel face area, the strength parameters obtained during the drilling process of the rock drilling rig are used to study the response during the drilling process and extract information, such as axial pressure on the drill bit, drill bit rotation speed, drilling rate, drill bit diameter, energy transfer rate, and drill bit slip friction coefficient. Combined with the tunnel face fracture information and the drilling parameters of the rock drilling rig, the rock strength is obtained using the drilling parameters. By combining the correlation between tunnel face fractures and surrounding rock grade, rapid regional classification of the tunnel surrounding rock grade can be achieved.

[0042] The improved residual network module can more accurately and realistically identify tunnel face cracks, greatly improving identification efficiency. By comprehensively analyzing tunnel face crack information and drilling parameter information, the surrounding rock grade can be intelligently, efficiently, and accurately predicted. It can effectively overcome the impact of low accuracy in manual data collection, greatly reducing the time spent on manual data collection and improving construction efficiency. The information and data obtained can provide important references and guidance for subsequent projects, which is of great benefit to tunnel construction. Attached Figure Description

[0043] Figure 1 This is a schematic diagram of the process of this invention.

[0044] Figure 2 This is a schematic diagram of the image recognition process based on convolutional neural networks.

[0045] Figure 3 This is a schematic diagram of the point cloud image preprocessing process.

[0046] Figure 4 This is the improved residual network module diagram.

[0047] Figure 5 This is a simplified diagram of the tunnel face area division. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of this invention clearer, the following detailed description of this patent is provided in conjunction with the accompanying drawings.

[0049] This invention utilizes a convolutional neural network (CNN) to identify processed point cloud images. The softmax function is used to calculate the data generated by the CNN, and the recognition result of each rock image is evaluated to determine the probability that the detected tunnel face cracks within the range (0,1) match the actual situation. This allows for accurate category classification of the tunnel face cracks. Furthermore, the invention employs a YOLOv3 algorithm incorporating an improved residual network module. Based on a well-trained neural network, the tunnel face image is processed through a series of different convolutional layers, linear rectified layers, pooling layers, fully connected layers, and residual networks for feature recognition, resulting in more accurate and faster image recognition of tunnel face cracks.

[0050] The Pearson linear correlation coefficient formula was used to calculate the influence coefficients of different fractures and surrounding rock classification indices, and the influence coefficients of different fractures were determined. Specifically, these include the influence coefficients r1 for closed fractures, r2 for micro-open fractures, r3 for open fractures, and r4 for wide-open fractures. Based on the different types and numbers of fractures in the region, different fracture influence coefficients were substituted into the improved BQ method correction formula.

[0051] Analysis of relevant information on the geological conditions of the tunnel surrounding rock reveals that rock integrity (fractures), forces, drilling efficiency, and geological information are the main reference data, while drill bit type and wear degree have indirect correlations. An improved BQ method replaces the traditional manual method for obtaining the uniaxial compressive strength of the rock, rapidly obtaining it using a functional relationship between drilling parameters during the drilling process on a rock drilling rig and the uniaxial compressive strength of the rock. Regarding the modification of the BQ method, this invention primarily considers the factor of fractures at the tunnel face, incorporating a fracture influence coefficient into the BQ method modification, thereby improving the accuracy of identifying the surrounding rock zone grade.

[0052] By combining data generated from point cloud images identified by convolutional neural networks, the final results are summarized and analyzed to obtain more accurate results. Furthermore, the data analysis allows for adjustments to the locations of blast holes at the tunnel face. By integrating information on fractures at the tunnel face and drilling parameters from the drilling rig, the surrounding rock grade is determined. Based on the acquired parameters and fracture information, subsequent blasting design can be guided, ensuring the blasting effect meets expectations and achieves safety and efficiency. Identifying the surrounding rock grade at the tunnel face allows for the prediction of the amount of explosives used in blasting. By predicting the development of fractures at the tunnel face, appropriate blast hole locations can be selected for blasting, ensuring the safe and rational use of the blasted energy, which greatly benefits construction.

[0053] like Figure 1 As shown, this invention is a deep learning-based intelligent and rapid regional classification method for rock mass at the tunnel face in tunnel blasting, comprising the following steps:

[0054] Step (1) Divide the tunnel face into five regions, S1 to S5. Select several drilling points within each region as the preset positions for fixing the rock drilling rig. Set the hole drilled by the rock drilling rig as the center and specify the radius that can basically cover the divided region.

[0055] Step (II) Positioning the rock drilling rig: Fix the rock drilling rig in the preset position and use the rock drilling rig to drill the tunnel face. This can be divided into impact, propulsion, rotation and flushing processes.

[0056] Step (3) Obtain drilling parameters through the drilling process; select a suitable drill bit according to the geological conditions of the tunnel face. In this invention, a toothed drill bit is selected, and parameters such as axial pressure on the drill bit, rotational speed of the drill bit, drilling rate, drill bit diameter, energy transfer rate, and drill bit slip friction coefficient are obtained through the drilling process.

[0057] Step (iv) By establishing a model relating drilling parameters to the uniaxial compressive strength of rock, the uniaxial compressive strength of rock is calculated using the drilling parameters. By establishing a functional relationship between drilling parameters and the uniaxial compressive strength of rock, specific data on the uniaxial compressive strength of rock are obtained.

[0058] The model relating drilling parameters to the uniaxial compressive strength of rock is as follows:

[0059]

[0060] Where: σ c Ω is the uniaxial compressive strength of the rock, MPa; W is the axial pressure on the drill bit, N; Ω is the rotational speed, r / min; V is the drilling rate, m / min; D is the drill bit diameter, m; f is the energy transfer rate; μ is the drill bit slip friction coefficient.

[0061] Step (V) employs the improved BQ method, substituting the uniaxial compressive strength of the rock from Step (IV) into the BQ method model. A modified BQ method model is obtained using the Pearson linear correlation coefficient calculation formula. The Pearson linear correlation coefficient formula is then used to calculate the grading indices for different fractures and surrounding rock types, determining the influence coefficients for different fractures, specifically including the influence coefficient r1 for closed fractures, r2 for micro-open fractures, r3 for open fractures, and r4 for wide-open fractures. Depending on the type and number of fractures within the region, different fracture influence coefficients are substituted into the improved BQ method correction formula.

[0062] The improved BQ method has the following specific steps:

[0063] Step (1) Collect images of cracks at the tunnel face, generate a training set, and begin network training until the convolutional neural network is mature;

[0064] Images of tunnel face cracks, including closed cracks, micro-cracks, open cracks, and wide-cracks, were collected and their sizes were standardized to control the number of cracks appearing in each image. A dataset was created by accumulating a large amount of tunnel face crack image data. This dataset was then processed, and a convolutional neural network was trained until it reached a high level of accuracy. The training set was then used for crack recognition.

[0065] like Figure 2 As shown, the network structure is first constructed, specifying the number and size of convolutional layers, pooling layers, and fully connected layers, primarily based on the YOLOv3 algorithm. To increase accuracy, an auxiliary process is added to the residual network, and deconvolution is incorporated to further refine the recognition results. The convolutional layers mainly enhance image features, dynamically extracting them and reducing noise. If the output image size is n×n, with edges p-pixel padding, a convolutional kernel size of f×f, and a stride of s, then the feature map size is... When performing a convolution operation, the number of weight parameters that need to be trained under weight sharing is f. 2 indivual.

[0066] Pooling layers primarily reduce the dimensionality of extracted features, making the feature map smaller and simplifying computation. Using pooling layers not only improves speed but also avoids overfitting. Pooling operations impart scale invariance and a degree of rotation invariance to the features. Commonly used pooling methods include max pooling, random pooling, and average pooling.

[0067] The fully connected layer primarily compresses and flattens multidimensional image data. First, the resulting feature maps are concatenated to obtain one-dimensional features. Then, these features are weighted and summed, and finally, an activation function is applied.

[0068] Convolution operation refers to the calculation of a filter by a certain stride, which is the sum of the products of the numbers in the filter and the corresponding numbers in the input image. Its definition is as follows:

[0069] s(t)=(x×w)(t) (10)

[0070] Let t be the integer time point. The convolution operation formula for discrete time points is defined as follows:

[0071]

[0072] The first parameter x is usually the input, the second parameter w is the kernel function, and the output is called the feature map.

[0073] Convolutional neural networks use Softmax as their output function. When applied to classification, the output data lies between (0,1), which can be viewed as the probability of a particular class corresponding to the true situation. The output function of Softmax can be expressed as:

[0074]

[0075] In the formula: y k a represents the output signal of the k-th neuron; k represents the input signal of the k-th neuron; i represents the input signal of the ith neuron; n represents the total number of input neurons. Each calculation uses the ReLU activation function. When the input to the ReLU function is greater than 0, the value is directly input; when the input to the ReLU function is less than 0, the output is 0.

[0076] The backbone feature extraction network is Darknet53, which has two important characteristics:

[0077] (1) A key feature of Darknet53 is its use of residual networks. Residual networks are easy to optimize and can improve accuracy by increasing their depth. The residual convolution in Darknet53 first performs a 3×3 convolution with a stride of 2, which compresses the width and height of the input feature layer, resulting in a feature layer. Then, a 1×1 convolution and a 3×3 convolution are performed on this feature layer, and the result is added to the feature layer, thus forming the residual structure.

[0078] (2) Each convolutional part of Darknet53 uses a unique DarknetConv2D structure. Regularization is performed during each convolution, and after the convolution is completed, BatchNormalization and LeakyReLU are performed. Ordinary ReLU sets all negative values ​​to zero, while LeakyReLU assigns a non-zero slope to all negative values.

[0079] In mathematical terms, we can express it as:

[0080]

[0081] Local receptive field, weight sharing, and max pooling are three key characteristics of Convolutional Neural Networks (CNNs). These three characteristics combine to give CNNs geometric distortion invariance in image recognition. Local receptive field refers to the connection of neurons between each layer through a local receptive field; weight sharing means that each convolutional kernel in a convolutional layer repeatedly operates within the receptive field to process the image; and max pooling performs dimensionality reduction on the convolutional features.

[0082] The training method for CNNs is as follows, involving two steps:

[0083] Step 1) Forward propagation of working signal

[0084] Take a sample (X) p Y p Where X p For input, Y p For the ideal output, then X p Input network; calculate actual output O p .

[0085] In this stage, the signal is transformed from the input layer to the output layer, and the process is repeated during operation after training. This stage involves the network performing computational operations, as shown in the following formula:

[0086] O p =F n (...F2(F1(X p W (1) W (2) ...W (n) (14)

[0087] Step 2) Backpropagation of error signal

[0088] Calculate O p and Y p The error is reduced by backpropagating the error and modifying the weight matrix layer by layer until the error no longer decreases. The total network error E of the CNN and the error E of the p-th sample are also considered. p The definition is as follows:

[0089] E = ΣE p (15)

[0090]

[0091] Step (2) convert the cracks at the tunnel face to be tested into a point cloud image;

[0092] Step (3) Point cloud image preprocessing;

[0093] The equipment acquires point cloud images of cracks at the tunnel face, including operations such as grayscale conversion, sharpening, noise reduction, image enhancement, RANSAC algorithm, and Hough transform.

[0094] Specifically as follows: Figure 3 The image shown is a data acquisition diagram based on 3D point clouds. First, the tunnel face is scanned using equipment to acquire 3D point cloud images. Then, the point cloud data is preprocessed, including the following steps:

[0095] First, a grayscale conversion is performed. The resulting image is usually a color image, but due to interference from color and sunlight, it's prone to generating unnecessary information. Therefore, many image processing algorithms convert the image to grayscale before designing and improving the algorithm. This can be calculated using the formula:

[0096] Gray=0.30×R+0.59×G+0.11×B (17)

[0097] During crack image acquisition, various factors can cause crack blurring, affecting crack recognition and subsequently impacting the accuracy of output parameters. Therefore, image sharpening is commonly used. Image sharpening is a mathematical calculation method to compensate for image contours, aiming to enhance image edges and obtain a clearer image. This process is implemented using a convolution algorithm. First, an odd-dimensional matrix is ​​set as a template, and each pixel is multiplied by its corresponding element in the template. The sum of the resulting data is then assigned to the center point of the region. The calculation formula is as follows:

[0098]

[0099]

[0100] The following are commonly used image sharpening templates, with the 3×3 template being the most common.

[0101]

[0102] Because median filtering is effective at reducing noise while maintaining good image clarity, it is chosen for image denoising.

[0103] The basic principle of median filtering is to replace the pixel value of each point in a digital image with the median value of its neighborhood points, using a window with a neighborhood size of 3×3 or larger. This generates pixel values ​​that more closely approximate reality. The calculation formula is as follows:

[0104] Y(x,y)=Med{y(s,t)|(s,t)∈Sxy} (19)

[0105] Histogram equalization is used to enhance image quality and improve visual effects. Histogram equalization essentially expands the grayscale range of an image, transforming concentrated grayscale levels into a more dispersed state through calculations, thus enhancing the image. Considering various practical conditions, histogram equalization is a suitable method for processing tunnel face images. The grayscale statistical histogram function is as follows:

[0106]

[0107] In the formula: p s (s k ) represents the probability of the k-th gray level appearing in the original image, where n is the probability of the k-th gray level appearing. k It is a grayscale value of s k The number of pixels, where n is the total number of pixels in the image, 0 represents black, and L-1 represents white.

[0108] Fractures are categorized into banded fractures and irregular fractures for identification. Due to the diverse nature of fractures on the tunnel face, only the more significant fractures can be analyzed. These fractures are often located in accident-prone areas and are easily detected. These fractures typically appear in groups with uniform angles. Banded fractures are mostly composed of interlayers and fissures within the rock mass structure; they are yellowish-brown in color and grow laterally in groups. Irregular fractures are those that cannot be described using a uniform standard. While numerous, only a small number of these fractures actually require extraction; the focus is on extracting larger and deeper fractures.

[0109] The RANSAC algorithm is used to simplify and enhance the structure of the tunnel face, and the segmentation effect is determined based on a set threshold. The segmentation process based on the RANSAC algorithm is as follows:

[0110] (1) Given the original point cloud P, the minimum sampling set size is m. Randomly select a subset P' of P containing m points from p to initialize the model M.

[0111] (2) Traverse all points in the remainder set R = PP` and the error between them and the model M. Points that are less than a certain set threshold are combined with P` to form P*, which is considered to be the effective set of the model M.

[0112] If the number of points in P* reaches the specified threshold N, then M can be considered a correct model, and steps 1 and 2 can be repeated.

[0113] If the algorithm fails to find the correct model after completing the specified number of samplings (MAX), it will otherwise select the model with the largest effective set.

[0114] This invention employs Hough transform to extract features from images, primarily for identifying gaps at tunnel face. The basic idea of ​​Hough transform is to utilize the duality of points and lines, meaning that collinear points in image space correspond to intersecting lines in parameter space. All straight lines or curves intersecting at the same point in parameter space have corresponding collinear points in image space.

[0115] In the image space XY, all collinear points (x, y) can be described by the equation of a straight line:

[0116] y = mx + c (21)

[0117] Where m is the slope of the line and c is the intercept, the above formula can also be rewritten as:

[0118] c = -xm + y (22)

[0119] The above formula can be viewed as the equation of a straight line in the parameter space MC, where the slope of the line is x and the intercept is y. Comparing the above formulas, a point (x, y) in the image space corresponds to a straight line in the parameter space, and a straight line in the image space is determined by a point (m, c) in the parameter space.

[0120] This invention combines the results of convolutional neural networks and 3D point cloud data generation for analysis, and transitions 2D features to 3D features. This not only solves the problem of unstable recognition results caused by the rotation of 3D point clouds, but also further improves the accuracy of convolutional neural networks in recognizing 2D images, and demonstrates the robustness of convolution in processing point cloud data.

[0121] Step (4): Based on the mature convolutional neural network trained in step (1), feature recognition is performed on the point cloud image;

[0122] This invention combines the results of convolutional neural networks and 3D point cloud data generation for analysis, and transitions 2D features to 3D features. This not only solves the problem of unstable recognition results caused by the rotation of 3D point clouds, but also further improves the accuracy of convolutional neural networks in recognizing 2D images, and demonstrates the robustness of convolution in processing point cloud data.

[0123] Specifically, the method of combining convolutional neural networks to recognize point cloud images has the following advantages: When using convolutional neural networks to recognize point cloud image data, it has the characteristics of accuracy and efficiency. It can perform feature recognition of point cloud images from multiple directions and angles. Compared with using convolutional neural networks to recognize two-dimensional images, more accurate recognition results can be obtained by recognizing three-dimensional images.

[0124] The equipment scans the tunnel to acquire a large amount of point cloud data. By preprocessing the point cloud images, a simplified point cloud data image is obtained, which specifically reflects the characteristics of the scanned material. The potential direction of subsequent crack development in the tunnel face is predicted. The predicted cracks in the tunnel face are then displayed on the computer using an algorithm.

[0125] Step (5) Secondary processing of point cloud images: Deconvolution is performed on the identified point cloud images of tunnel face cracks to transform them into clearer and more distinctive point cloud images. The clear point cloud images are then identified. With each iteration, the accuracy of identifying point cloud images of tunnel face cracks is further improved.

[0126] like Figure 4As shown, based on the direct recognition of point cloud images using convolutional neural networks, to increase accuracy, the following auxiliary method is added to the residual network Residual. Furthermore, by incorporating deconvolution, the point cloud images are processed a second time. The recognized tunnel face crack point cloud images are deconvolved, transforming them into clearer, more feature-rich point cloud images. These clearer point cloud images are then used for recognition. With each iteration, the accuracy of identifying tunnel face crack point cloud images is further improved, resulting in more precise recognition results. The specific steps are as follows:

[0127] a. Selecting locations of cracks at the tunnel face and outputting feature layers through feature recognition;

[0128] The farthest point method is used to collect the crack features of the tunnel face. An initial point is randomly selected, and the point farthest from the initial point is added as the starting point until the maximum lateral and longitudinal distances of the cracks in the tunnel face are obtained through iteration.

[0129] Using the intersection of the lines with the maximum horizontal and vertical distances as the center, the tunnel face cracks are divided into four regions. One region is selected, and the point within the region with the intersection as the center point is selected, thus completing the point selection within one region.

[0130] The same approach is then applied to the other three areas, using the nearest point method to obtain points that meet the requirements. This completes the selection of points for a tunnel face crack. Feature recognition is then performed, using the intersection of the maximum horizontal and vertical distances as the center. For cracks that are thin and curved, the two ends of the crack are directly selected as the starting points, and the middle area of ​​the crack as the center point.

[0131] b. Input the output feature layer into the deconvolution module;

[0132] The output feature layer is input into the deconvolution module, using a 3×3 deconvolution layer to recover the feature map pixels from the input image. A 3×3 convolution layer with a stride of 2 compresses the width and height of the point cloud image, resulting in a feature layer. This feature layer is then passed through a ReLU function. When the input to the ReLU function is greater than 0, the input is used directly; when the input is less than 0, the output is 0.

[0133] Then, perform a 1×1 convolution and a 3×3 convolution on the above output feature layer, and then pass it through a ReLU function to output the result.

[0134] Compared to upsampling methods, deconvolution not only fills in neighboring pixels but also involves an additional parameter learning step, which increases resolution while obtaining more comprehensive feature information. Therefore, by applying deconvolution to the tunnel face cracks, a clearer image of the tunnel face cracks can be obtained.

[0135] c. After the deconvolution operation, the output feature layer is added to the initial feature layer to obtain the final output feature layer. This process is repeated multiple times until the output result meets the actual requirements.

[0136] By collecting a large amount of data on tunnel face fissures after blasting, including the number, type, and distribution of fissures, and after classifying the surrounding rock using the BQ method, the data on tunnel face fissures can be collected in real time and identified through the aforementioned deep learning.

[0137] Step (6) Use the drilling parameters and the uniaxial compressive strength of rock relationship model to obtain the uniaxial compressive strength of rock and apply it to the BQ method model;

[0138] like Figure 5 The diagram shown is a simplified map of the tunnel face area. It mainly describes how drilling parameters are obtained by using a rock drilling rig, the uniaxial compressive strength of the rock is obtained, and the correlation between tunnel face fissures and surrounding rock grade is used to correct the BQ method surrounding rock classification, thereby classifying the surrounding rock area.

[0139] A model relating drilling parameters to the uniaxial compressive strength of rock is used to obtain the uniaxial compressive strength of the rock, which is then applied to the BQ method model. Firstly, the calculation of the basic quality index BQ of the surrounding rock using the BQ method is based on the quantitative index of the uniaxial compressive strength R of the rock. c and rock mass integrity coefficient K v Calculate using the following formula:

[0140] BQ = 90 + 3R c +250K v (twenty three)

[0141] When applying R c >90K v At +30, it should be R c =90K v Substitute +30 into the calculation of BQ, if K v >0.04R c When +0.4, K should be used. v =0.04R c Substitute +0.4 into the calculation of BQ.

[0142] R c The calculation is performed using the formula relating drilling parameters to the uniaxial compressive strength of the rock, as follows:

[0143]

[0144] Where: σ cΩ is the uniaxial compressive strength of the rock, MPa; W is the axial pressure on the drill bit, N; Ω is the rotational speed, r / min; V is the drilling rate, m / min; D is the drill bit diameter, m; f is the energy transfer rate; μ is the drill bit slip friction coefficient.

[0145] The improved BQ method formula is as follows:

[0146] BQ = 90 + 3σ c +250K v (25)

[0147] Rock integrity coefficient K v The value is generally calculated using the elastic wave test value of the surrounding rock:

[0148] K v =(V pm / V pv ) 2 (26)

[0149] In the formula: V pm V represents the elastic longitudinal wave velocity of the rock mass, in km / s. pr Let be the elastic longitudinal wave velocity of the rock, in km / s.

[0150] Step (7) Based on the identification results, output the information on the types of cracks at the tunnel face, and obtain the influence coefficient of cracks at the tunnel face using the Pearson linear correlation coefficient calculation formula, and apply it to the BQ method correction model;

[0151] Considering the influence of groundwater, unfavorable structures, and in-situ stress on the surrounding rock condition, the BQ value is corrected according to the following formula:

[0152] [BQ] = BQ - 100(K1 + K2 + K3) (27)

[0153] In the formula: K1 is the correction factor for the influence of groundwater; K2 is the correction factor for the influence of the attitude of the main weak structural planes; K3 is the correction factor for the influence of the initial stress state.

[0154] The correlation coefficient between tunnel face fissures and surrounding rock classification indicators was calculated using the Pearson linear correlation coefficient formula, which is as follows:

[0155]

[0156] In the formula: n is the sample size, X i Y i These are the observation values ​​at point i corresponding to variables X and Y. It is the sample mean of X. This is the sample mean of Y. The calculated correlation coefficient r indicates whether the two variables are positively or negatively correlated. Its value is between (0,1). Generally, a value greater than 0.3 indicates a correlation, and the larger the value, the stronger the correlation.

[0157] The influence of closed fractures, micro-open fractures, open fractures, and wide-open fractures on the classification of surrounding rock is calculated using the above formulas. Thus, the influence coefficients r1 for closed fractures, r2 for micro-open fractures, r3 for open fractures, and r4 for wide-open fractures are obtained.

[0158] The influence coefficients are applied to the surrounding rock classification, and the tunnel face is roughly divided into five regions, S1 to S5. Multiple drilling points are selected as centers within each region, and a radius R is defined to basically cover the region. This radius can be adjusted according to actual conditions. The types of fractures within radius R are identified, specifically closed fractures, micro-tensioned fractures, open fractures, and wide-tensioned fractures. Based on the proposed influence coefficients r1, r2, r3, and r4, the surrounding rock classification within each region is corrected. The specific formula is as follows:

[0159] [BQ] 1 =BQ-100(K1+K2+K3+r1+r2+r3+r4) (29)

[0160] In the formula: K1 is the groundwater influence correction coefficient; K2 is the main weak structural plane attitude influence correction coefficient; K3 is the initial stress state influence correction coefficient; r1 is the closed crack influence coefficient; r2 is the micro-tension crack influence coefficient; r3 is the open crack influence coefficient; and r4 is the wide-tension crack influence coefficient.

[0161] Step (VI) combines the BQ method correction model from step (V) with deep learning to identify the crack information at the tunnel face. The BQ value obtained through the BQ method model is applied to the BQ method correction model to obtain the corrected BQ value, and then the rock mass is intelligently and rapidly classified into regions.

[0162] The BQ value obtained from the BQ method model is applied to the modified BQ method model to obtain the corrected BQ value, thus completing the surrounding rock classification. By dividing the tunnel face into regions and considering the influence of different fissures within each region on the surrounding rock grade, the surrounding rock grade is regionally classified. Using the modified BQ method formula, the rock mass at the tunnel face is then regionally classified, offering advantages such as high accuracy and ease of operation.

[0163] Based on the above-described preferred embodiments of the present invention, and through the foregoing description, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. A deep learning-based intelligent rapid regional classification method for tunnel rock mass, characterized by the following steps Comprise: Step (I) divides the tunnel face area; Step (II) position the drilling jumbo at the preset position; Step (III) obtain drilling parameters through the drilling process; Step (IV) obtain the uniaxial compressive strength of the rock by establishing a relationship model between the drilling parameters and the uniaxial compressive strength of the rock; Step (V) use the improved BQ method to substitute the uniaxial compressive strength of the rock in step (IV) into the BQ method model, and obtain the BQ method correction model through the Pearson linear correlation coefficient calculation formula; Step (VI) combine the BQ method correction model in step (V) with deep learning to identify the tunnel face fracture information, and then intelligently and quickly classify the rock mass; The relationship model between the drilling parameters and the uniaxial compressive strength of the rock is: ； wherein: σ c is the uniaxial compressive strength of the rock, MPa; W is the axial load on the bit, N; Ω is the rotational speed, r / min; V is the penetration rate, m / min; D is the bit diameter, m; f is the energy transfer rate; and μ is the bit slip friction coefficient. The specific steps of the improved BQ method are as follows: Step (1) collect the tunnel face fracture images, generate a training set, and start network training until the convolutional neural network is trained maturely; Step (2) convert the tunnel face fracture to be tested into a point cloud image; Step (3) point cloud image preprocessing; Step (4), based on the convolutional neural network trained maturely in step (1), identify the features of the point cloud image; Step (5) based on the residual network module of the convolutional neural network, process the point cloud image twice, deconvolve the identified tunnel face fracture point cloud image, transform the tunnel face fracture point cloud image into a clearer and more characteristic point cloud image, and then identify the clear point cloud image. Each iteration further improves the accuracy of tunnel face fracture point cloud image identification; Step (6) obtain the uniaxial compressive strength of the rock using the relationship model between the drilling parameters and the uniaxial compressive strength of the rock, and apply it to the BQ method model; Step (7) output the tunnel face fracture type information through the identification result obtained in step (5), and obtain the tunnel face fracture influence coefficient through the Pearson linear correlation coefficient calculation formula, and apply it to the BQ method correction model. The BQ method correction model is used to modify the BQ method model; The BQ method model is as follows: ； ； wherein: σ c is the uniaxial compressive strength of the rock obtained using the drill bit; V pm is the elastic longitudinal wave velocity of the rock mass rock, km / s; V pr is the elastic longitudinal wave velocity of the rock, km / s; The BQ method correction model is as follows: ； In the formula: K1 is the groundwater influence correction coefficient; K2 is the main weak structure plane occurrence influence correction coefficient; K3 is the initial stress state influence correction coefficient, r1 is the closed fracture influence coefficient, r2 is the micro-tension fracture influence coefficient, r3 is the tension fracture influence coefficient, and r4 is the wide tension fracture influence coefficient.

2. The deep learning-based intelligent rapid regional classification method for tunnel rock mass according to claim 1, characterized in that: The steps of the residual network module of the convolutional neural network in step (5) are as follows: a. Select points of the tunnel face fracture, and output feature layers through feature identification; b. input the output feature layer into the deconvolution module; c. After the deconvolution operation, add the output feature layer to the initial feature layer to obtain the final output feature layer. Through multiple loop iterations, the output result meets the actual requirements.

3. The deep learning-based intelligent rapid regional classification method for tunnel rock mass according to claim 1, characterized in that: The Pearson linear correlation coefficient calculation formula is as follows: ； where n is the number of samples, X i , Y i are the i-th observation of the variables X, Y, is the sample mean of X, is the sample mean of Y, and r is the correlation coefficient.

Images