A tunnel surrounding rock grade automatic evaluation method, system and device

Abstract

This invention discloses an automated method, system, and device for assessing the surrounding rock grade of tunnels. The method includes the following steps: Step 1: Capturing images of the tunnel face using a drone and transmitting the images to a computer terminal; Step 2: Performing multi-dimensional preprocessing on the tunnel face to obtain effective image areas; Step 3: Calculating the fracture density image quantity D. f Crack direction dispersion H θ Surface roughness E r Step 4: Crack density image quantity D f Crack direction dispersion H θ Surface roughness E r Step 5: Perform normalization processing and calculate the surrounding rock structure complexity index RSCI; Step 6: Based on the surrounding rock structure complexity index RSCI, calculate the continuous surrounding rock quality score Q through a monotonic mapping function; Step 7: Map the continuous surrounding rock quality score Q to the surrounding rock grade according to a preset threshold interval and output the surrounding rock grade evaluation result.

Description

Technical Field

[0001] This invention relates to the fields of tunnel engineering and intelligent information processing technology, and in particular to an automated method, system and equipment for assessing the grade of surrounding rock in tunnels. Background Technology

[0002] During tunnel construction, the surrounding rock grade is a crucial parameter reflecting the engineering geological conditions of the surrounding rock. Its determination directly impacts the selection of construction methods, support structure design, and construction safety control. Determining the surrounding rock grade typically requires comprehensive consideration of multiple factors, including rock mass morphology, integrity, weathering degree, joint and fissure development, and overall stability of the tunnel face. This process is characterized by its multi-dimensional information and complexity. Factors such as rock mass morphology, the degree of joint and fissure development and their spatial distribution, and the surface undulations and fracture state of the tunnel face can be directly or indirectly reflected in tunnel face images and can be identified and quantified through image analysis. However, in practical engineering, indicators such as weathering degree and stability often require a comprehensive judgment based on visual characteristics of the tunnel face and relevant experience. Therefore, determining the surrounding rock grade is highly specialized and experience-dependent.

[0003] The existing technology still has the following technical problems that need to be solved: (1) The determination of surrounding rock grade relies heavily on human experience, and the evaluation results lack subjectivity and consistency: In current engineering practice, the determination of surrounding rock grade is still mainly based on manual on-site observation. Geological technicians conduct empirical analysis of the rock mass structure, fracture development, and surface fragmentation characteristics based on the exposure of the working face, and perform qualitative or semi-quantitative classification in conjunction with relevant specifications. This method has certain practicality in the early stages of engineering and under complex conditions, but it is highly dependent on the experience level and subjective judgment of personnel. The consistency and repeatability of evaluation results are difficult to guarantee between different construction sections, different times, or different personnel making the determination, which makes it difficult to meet the needs of stable and objective assessment of surrounding rock grade.

[0004] (2) Existing analysis methods based on tunnel face images lack systematic quantification and interpretable representation of surrounding rock structural characteristics: With the application of image processing and deep learning technologies in this field, existing studies have attempted to use tunnel face images to assist in the analysis or grade prediction of surrounding rock conditions. However, existing methods mostly focus on direct classification or regression prediction of surrounding rock grades, often mapping image features to the final grade results end-to-end, lacking explicit extraction and quantification of surrounding rock structural information such as the degree of fracture development, structural orientation distribution, and surface roughness characteristics. Although this "black box" prediction method can achieve certain accuracy on some datasets, its prediction process is difficult to interpret, and engineers find it difficult to understand the correspondence between image features and the engineering properties of surrounding rock, limiting its application value in actual construction decisions.

[0005] (3) Existing deep learning models are not adaptable to complex construction environments and heterogeneous surrounding rock structures: Tunnel construction environments are characterized by complex lighting conditions, narrow spaces, severe dust interference, and variable background structures, which place high demands on the quality and stability of tunnel face images. Although deep learning algorithms have been applied in this field, existing models are mostly based on conventional convolutional neural network structures, which have limited capabilities in multi-scale structural feature modeling and adapting to the heterogeneity of complex surrounding rock structures. At the same time, the model training methods are often relatively simple, and when the number of samples is limited or the geological conditions vary greatly, overfitting or insufficient generalization ability is likely to occur, affecting the reliability of the prediction results.

[0006] (4) The lack of standardization in image acquisition and preprocessing methods affects the accuracy of subsequent feature extraction: In the existing technology, the acquisition of tunnel face images mostly relies on fixed installation of camera equipment or manual shooting, which has problems such as limited shooting angle, unstable lighting conditions, and large differences in image quality. At the same time, the acquired images are usually only subjected to basic processing such as simple cropping, mean denoising, or brightness adjustment. There is a lack of standardized and refined preprocessing procedures for the characteristics of tunnel construction environment, which makes it difficult to effectively eliminate interference from invalid areas, uneven lighting, and noise pollution, resulting in limited accuracy of subsequent surrounding rock structure feature extraction.

[0007] In view of this, this invention is hereby proposed. Summary of the Invention

[0008] The purpose of this invention is to address the shortcomings of existing technologies by proposing an automated method, system, and device for evaluating the surrounding rock grade of tunnels.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: An automated method for assessing the surrounding rock grade of tunnels includes the following steps: Step 1: Take images of the working face using a drone and transmit the images to a computer terminal; Step 2: Perform multi-dimensional preprocessing on the face of the tunnel to obtain effective image regions, including image cropping based on MaskR-CNN semantic segmentation, non-local mean denoising, CLAHE adaptive illumination correction, Gamma brightness equalization, Laplacian sharpening enhancement, bilinear interpolation scale normalization, and effective region purification based on edge detection. Step 3: Perform pixel-level structure recognition on the preprocessed image using a crack segmentation network to obtain a crack probability distribution map, and calculate the crack density image quantity D. f By combining the set of fractures obtained from the fracture probability distribution map, a direction-aware feature learning network is used to model the direction-aware features of the preprocessed image, thereby obtaining the fracture direction distribution probability and calculating the fracture direction dispersion H. θThe surface undulation features in the preprocessed image are modeled using a texture response feature learning network to obtain a texture response feature map characterizing the roughness of the surrounding rock surface, and the surface roughness E is calculated. r ; Step 4: Crack density image quantity D f Crack direction dispersion H θ Surface roughness E r After normalization, the surrounding rock structure complexity index RSCI is calculated using the following formula: ; In the formula, , , These represent the normalized crack density image quantity, crack direction dispersion, and surface roughness, respectively, with α, β, and γ as weights. Step 5: Based on the surrounding rock structure complexity index RSCI, calculate the continuous surrounding rock quality score Q using a monotonic mapping function, as shown in the following formula: ; In the formula, k is a constant greater than 0, used to control the steepness of the scoring curve, and b is the bias parameter; Step 6: Map the continuous surrounding rock quality score Q to the surrounding rock grade according to the preset threshold range, and output the surrounding rock grade assessment result.

[0010] Furthermore, step 2 includes the following steps: Step 2.1 Image cropping based on Mask R-CNN semantic segmentation: Using a pre-trained Mask R-CNN model with a confidence threshold of 0.75, the non-rock areas in the image are identified and segmented, and the output is an initial valid image containing only the exposed rock area of ​​the working face. Step 2.2, Non-local means denoising: The non-local means denoising algorithm is implemented using the OpenCV library. The similarity window size is set to 7×7, the search window size is set to 21×21, and the denoising intensity parameter is set to h=10. The cropped image is then denoised. Step 2.3, CLAHE Adaptive Illumination Correction: Using the CLAHE function of OpenCV, the contrast gain is limited to 2.0 and the block size is 8×8. Histogram equalization is performed on the denoised image to solve the problem of local over-brightness / under-brightness caused by uneven lighting in the tunnel, balance the brightness distribution of the image, and make the surrounding rock structure features in the shadow area clearly visible. Step 2.4, Gamma Brightness Equalization: Adaptively adjust the Gamma coefficient (range 0.8-1.2) based on the average gray value of the image. When the gray value is below 128, take 0.8-0.9; when it is between 128 and 192, take 1.0; and when it is above 192, take 1.1-1.2. Optimize the overall brightness of the image through power-law transformation, enhance the detail expression in low-light areas, and avoid feature loss caused by extreme brightness values. Step 2.5, Laplacian sharpening enhancement: A 3×3 Laplacian operator (coefficient matrix [0,-1,0;-1,5,-1;0,-1,0]) is used to perform convolution operations on the illumination-corrected image to enhance edge features, improve detail recognition, and provide support for the extraction of crack development status features. Step 2.6, Bilinear Interpolation Scale Normalization: The bilinear interpolation algorithm is used to uniformly scale the sharpened image to 224×224 pixels. The bilinear interpolation algorithm is used to maintain the feature without distortion, so as to unify the input data format of the model in Step 3, eliminate the image scale difference caused by different shooting distances and angles, and ensure the consistency of feature extraction. Step 2.7: Purification of effective image region based on edge detection: The Canny edge detection algorithm (threshold range 50-150) is used to identify the edge contour of the surrounding rock area and remove a small number of non-surrounding rock edge pixels remaining after scale normalization to further purify the effective image region and ensure that the image input to the model in Step 3 contains only the core feature information of the surrounding rock.

[0011] Furthermore, in step 3, the crack density image quantity D is calculated through the following steps. f : Step A.1: Perform pixel-level structure recognition on the preprocessed image using a crack segmentation network to obtain a crack probability distribution map. The crack probability distribution map is a matrix with the same size as the preprocessed image, and each element of the matrix is ​​the probability value of the crack. Step A.2: Threshold the crack probability distribution map to obtain a binary crack mask, separating cracked and non-cracked regions; Step A.3: Perform morphological denoising and connected component filtering to obtain several segments of the fracture centerline; Step A.4: Obtain the projected length l of the centerline of each fracture segment, and calculate the fracture density image quantity D using the following formula. f : ; In the formula, A is the effective area of ​​the working face, and N... f For the number of cracks, l i Let be the projected length of the center line of the i-th crack.

[0012] Furthermore, in step 3, the crack direction dispersion H is calculated through the following steps.θ : Step B.1: Determine the number K of directional intervals and the range of each directional interval; Step B.2: Based on the set of fracture centerlines obtained from the fracture probability distribution map, the preprocessed image is modeled using a direction-aware feature learning network to obtain the local orientation angles of each fracture centerline; Step B.3: Count the number of samples in each directional interval, divide by the total number of samples, and obtain the probability distribution p of each directional interval. k ; Step B.4: Calculate the crack direction dispersion H using the following formula. θ : ; In the formula, It is a very small constant.

[0013] Furthermore, in step 3, the surface roughness E is calculated through the following steps. r : Step C.1: Model the surface undulation features in the preprocessed image using a texture response feature learning network to obtain a texture response feature map that characterizes the roughness of the surrounding rock surface; Step C.2: Calculate the surface roughness E using the following formula. r : ; In the formula, M and N are the pixel sizes of the texture response feature map. Let I(x,y) be the Laplacian operator, and let I(x,y) be the gray value of a pixel in the texture response feature map.

[0014] Furthermore, in step 6, the surrounding rock grade assessment results include Grade I, Grade II, Grade III, Grade IV, and Grade V; when Q ≥ 0.8, it is determined to be Grade I surrounding rock; when 0.6 ≤ Q < 0.8, it is determined to be Grade II surrounding rock; when 0.4 ≤ Q < 0.6, it is determined to be Grade III surrounding rock; when 0.2 ≤ Q < 0.4, it is determined to be Grade IV surrounding rock; and when Q < 0.2, it is determined to be Grade V surrounding rock.

[0015] Furthermore, the loss is calculated using the following steps to train the model constructed in steps 3-6: Step D.1: Calculate the fracture segmentation loss L using the following formula. seg : ; In the formula, Ω is the set of pixels in the preprocessed image, P(x,y) is the probability that the model predicts that pixel (x,y) belongs to the crack, and G(x,y) is the true crack label value of the corresponding pixel. Step D.2: Calculate the regression loss L of the surrounding rock structure image using the following formula. reg : ; In the formula, N is the number of training samples. , , The crack density image quantity, crack orientation dispersion, and surface roughness of the i-th sample predicted by the model are given. , , This corresponds to the actual labeled value; Step D.3: Calculate the consistency loss L of the surrounding rock quality score using the following formula. con : ; In the formula, The model predicts the surrounding rock quality score for the i-th sample. This corresponds to the actual labeled value; Step D.4: Calculate the surrounding rock grade classification loss L using the following formula. cls : ; In the formula, C represents the number of surrounding rock grade categories. When the model correctly predicts the surrounding rock grade... Set to 1 when the model's predicted surrounding rock grade is incorrect. Take 0, Predict the confidence level of the i-th sample belonging to class c for the model; Step D.5: Calculate the total loss L using the following formula: ; In the formula, λ1, λ2, λ3, and λ4 are weight coefficients used to balance the influence of different tasks during the training process. These weight coefficients can be set according to the importance of the tasks or through cross-validation.

[0016] To achieve the above objectives, the present invention also employs the following technical solution: An automated assessment system for the surrounding rock grade of a tunnel includes: a drone and a computer terminal, wherein the drone is used to execute step 1 of any one of the methods provided by the present invention, and the computer terminal is used to execute steps 2 to 6 of any one of the methods provided by the present invention. The drone is equipped with an autonomous control unit, an ambient light sensor, an automatic supplemental lighting unit, a high-definition camera unit, a wireless communication module, and a drone power supply unit; the computer terminal is equipped with a local processing unit, a storage unit, a human-computer interaction interface, and a terminal power supply unit.

[0017] To achieve the above objectives, the present invention also employs the following technical solution: An automated assessment device for the grade of surrounding rock in tunnels includes: an image preprocessing module: performing multi-dimensional preprocessing on the tunnel face to obtain effective image regions, including image cropping based on MaskR-CNN semantic segmentation, non-local mean denoising, CLAHE adaptive illumination correction, Gamma brightness equalization, Laplacian sharpening enhancement, bilinear interpolation scale normalization, and effective region purification based on edge detection. The surrounding rock structure feature learning module performs pixel-level structure recognition on the preprocessed image using a fracture segmentation network to obtain a fracture probability distribution map and calculate the fracture density image quantity D. f By combining the set of fractures obtained from the fracture probability distribution map, a direction-aware feature learning network is used to model the direction-aware features of the preprocessed image, thereby obtaining the fracture direction distribution probability and calculating the fracture direction dispersion H. θ The surface undulation features in the preprocessed image are modeled using a texture response feature learning network to obtain a texture response feature map characterizing the roughness of the surrounding rock surface, and the surface roughness E is calculated. r ; Module for calculating the complexity of surrounding rock structure: Fracture density image quantity D f Crack direction dispersion H θ Surface roughness E r After normalization, the surrounding rock structure complexity index RSCI is calculated. The surrounding rock quality scoring module calculates the continuous surrounding rock quality score Q using a monotonic mapping function based on the surrounding rock structure complexity index RSCI. Rock grade mapping module: Maps the continuous rock quality score Q to the rock grade according to a preset threshold range and outputs the rock grade assessment result.

[0018] Compared with the prior art, the beneficial effects of this invention are as follows: 1. By introducing the surrounding rock structure complexity index and the continuous surrounding rock quality scoring mechanism, interpretable and continuous assessment of the surrounding rock grade can be achieved from the tunnel face image, reducing reliance on human experience and improving the objectivity, consistency and efficiency of the surrounding rock grade evaluation, thus providing a reliable basis for tunnel construction safety and support decisions.

[0019] 2. To address the issues of high reliance on human experience and insufficient subjectivity and consistency in surrounding rock grade determination, this embodiment introduces unmanned aerial vehicles (UAVs) to automatically acquire tunnel face images. Combined with a deep learning model, the structural characteristics of the surrounding rock are automatically analyzed and quantified, reducing the involvement of manual on-site judgment. By objectively calculating surrounding rock image structural indicators such as fracture density, fracture direction dispersion, and surface roughness, the surrounding rock grade evaluation process is standardized and automated, effectively reducing the impact of human experience differences on the evaluation results and improving the objectivity and consistency of surrounding rock grade determination.

[0020] 3. To address the shortcomings of existing image-based or deep learning-based methods, such as the lack of systematic quantification of surrounding rock structure features and insufficient interpretability of prediction results, this embodiment explicitly extracts surrounding rock structure features with clear engineering significance from the face image, including fracture density, fracture direction dispersion, and surface roughness energy. It then constructs a surrounding rock structure complexity index and a continuous surrounding rock quality score, achieving an interpretable mapping relationship from image features to surrounding rock grade. Compared to existing end-to-end black-box prediction methods, the surrounding rock grade assessment process in this embodiment has clear physical meaning and engineering interpretability, facilitating understanding and application by construction personnel.

[0021] 4. To address the issue of insufficient adaptability of existing deep learning models to complex surrounding rock structures and construction environments, this embodiment adopts a standardized, multi-dimensional image preprocessing process, and combines a shared feature extraction backbone network with a multi-branch structural feature learning architecture to perform multi-scale modeling of surrounding rock fracture structures and surface texture features. At the same time, the model is trained through a multi-task joint learning approach, taking into account tasks such as fracture segmentation, structural quantity regression, quality scoring, and grade classification, thereby improving the model's adaptability to the heterogeneity of complex surrounding rock structures and changes in the construction environment, and enhancing the stability and reliability of the prediction results.

[0022] 5. To address the limitations of existing methods for acquiring tunnel face images, unstable image quality, and non-standard preprocessing, this embodiment utilizes a drone equipped with automatic supplemental lighting, ambient light perception, path planning, and obstacle avoidance capabilities to automatically capture images of the tunnel face. The acquired images undergo standardized preprocessing operations such as semantic segmentation and cropping, noise reduction, illumination correction, sharpening enhancement, and scale normalization. This effectively overcomes problems such as insufficient lighting, dust interference, and the influence of invalid areas in the tunnel construction environment, providing high-quality and stable image input for subsequent extraction of surrounding rock structure features and grade assessment. Attached Figure Description

[0023] Figure 1 This is a flowchart of an automated method for assessing the surrounding rock grade of tunnels. Detailed Implementation

[0024] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0025] Example 1: An automated method for assessing the surrounding rock grade of tunnels, such as... Figure 1 As shown, it includes the following steps: Step 1: Take images of the working face using a drone and transmit the images to a computer terminal.

[0026] In this embodiment, the drone used is equipped with an autonomous control unit, an ambient light sensor, an automatic illumination unit, a high-definition camera unit (resolution ≥ 4K), a wireless communication module (supporting 5G / WiFi), and a drone power supply unit. The autonomous control unit incorporates a path planning algorithm and an obstacle avoidance sensor, enabling automatic hovering, path planning, and obstacle avoidance. The ambient light sensor detects the light intensity inside the tunnel in real time and feeds it back to the automatic illumination unit. The automatic illumination unit adaptively adjusts the brightness (1000-10000 lux) and illumination angle (0-90°) according to the light intensity. The high-definition camera unit captures images of the tunnel face. The wireless communication module enables remote transmission of images and data. After capture, the image is connected to a computer terminal via the wireless communication unit (e.g., 5G / WiFi) to transmit the tunnel face image reflecting the overall structural state of the tunnel face to the computer terminal in real time. It should be noted that the core of step 1 is capturing images of the tunnel face, and the autonomous control unit, ambient light sensor, automatic illumination unit, high-definition camera unit (resolution ≥ 4K), wireless communication module (supporting 5G / WiFi), drone power supply unit, and control algorithm involved can be implemented using existing technologies.

[0027] Specifically, at the end of each construction cycle (after the surrounding rock exposure is complete), the drone receives a start command from the computer terminal and autonomously flies to a pre-set shooting area at the tunnel face (3-5m away). It uses an ambient light sensor to detect the light intensity inside the tunnel (detection range 0-2000 lux). When the light intensity is <500 lux, the LED supplementary lighting unit is automatically activated, adjusting the brightness to 3000-8000 lux and the illumination angle to 30-60° to ensure no shadows on the surrounding rock surface and clear visibility of cracks. During shooting, fixed exposure parameters are used (shutter speed 1 / 500s, ISO 100-400), and 3-5 images are continuously captured. These images are transmitted to the computer terminal via a 5G module, where the terminal automatically selects the clearest image as the sample for subsequent analysis. The acquired tunnel face images are stored in JPG format with a resolution of 4096×2160 pixels. The image names include information such as the construction date, tunnel mileage, and shooting time, facilitating sample data management and traceability.

[0028] Step 2: Perform multi-dimensional preprocessing on the face of the tunnel to obtain effective image regions, including image cropping based on MaskR-CNN semantic segmentation, non-local mean denoising, CLAHE adaptive illumination correction, Gamma brightness equalization, Laplacian sharpening enhancement, bilinear interpolation scale normalization, and effective region purification based on edge detection.

[0029] In this embodiment, the effective image area of ​​the face for analysis is obtained through step 2.

[0030] In this embodiment, step 2 includes the following steps: Step 2.1 Image cropping based on Mask R-CNN semantic segmentation: Using a pre-trained Mask R-CNN model with a confidence threshold of 0.75, non-rock areas (such as construction equipment, personnel, tunnel wall edges, etc.) in the image are identified and segmented, and the output is an initial valid image containing only the exposed rock area of ​​the tunnel face; that is, invalid interference information is removed, the rock analysis object is focused, and the computational redundancy of subsequent feature extraction is reduced.

[0031] Step 2.2, Non-local means denoising: The non-local means denoising algorithm is implemented using the OpenCV library. The similarity window size is set to 7×7, the search window size to 21×21, and the denoising intensity parameter h=10. The cropped image is denoised to eliminate image noise caused by construction dust and imaging sensor noise while preserving details such as rock fissures and textures, thereby improving image purity.

[0032] Step 2.3, CLAHE Adaptive Illumination Correction: Using the CLAHE function of OpenCV, the contrast gain is limited to 2.0 and the block size is 8×8. Histogram equalization is performed on the denoised image to solve the problem of local over-brightness / under-brightness caused by uneven lighting in the tunnel, balance the brightness distribution of the image, and make the surrounding rock structure features in the shadow area clearly visible.

[0033] Step 2.4, Gamma Brightness Equalization: Adaptively adjust the Gamma coefficient (range 0.8-1.2) based on the average gray value of the image. When the gray value is below 128, take 0.8-0.9; when it is between 128 and 192, take 1.0; and when it is above 192, take 1.1-1.2. Optimize the overall brightness of the image through power-law transformation, enhance the detail expression in low-light areas, and avoid feature loss caused by extreme brightness values.

[0034] Step 2.5, Laplacian sharpening enhancement: A 3×3 Laplacian operator (coefficient matrix [0,-1,0;-1,5,-1;0,-1,0]) is used to perform convolution operations on the illumination-corrected image to enhance edge features such as the boundaries of surrounding rock fractures and the texture of structural surfaces, improve detail recognition, and provide support for the extraction of fracture development status features.

[0035] Step 2.6, Bilinear Interpolation Scale Normalization: The bilinear interpolation algorithm is used to uniformly scale the sharpened image to 224×224 pixels. The bilinear interpolation algorithm is used to maintain the feature without distortion, so as to unify the input data format of the model in Step 3, eliminate the image scale difference caused by different shooting distances and angles, and ensure the consistency of feature extraction.

[0036] Step 2.7: Purification of effective image region based on edge detection: The Canny edge detection algorithm (threshold range 50-150) is used to identify the edge contour of the surrounding rock area and remove a small number of non-surrounding rock edge pixels remaining after scale normalization to further purify the effective image region and ensure that the image input to the model in Step 3 contains only the core feature information of the surrounding rock.

[0037] In this embodiment, steps 2.1 to 2.7 address issues such as interference from invalid areas in tunnel images, noise pollution, uneven illumination, blurred details, and inconsistent scales, thereby enhancing the identifiability of surrounding rock structures and fracture features in the images and improving the accuracy of surrounding rock grade prediction. The UAV equipment is equipped with automatic hovering, path planning, and obstacle avoidance functions, and can autonomously complete the task of shooting at the tunnel face during breaks in tunnel construction cycles.

[0038] Step 3: Perform pixel-level structure recognition on the preprocessed image using a crack segmentation network to obtain a crack probability distribution map, and calculate the crack density image quantity D. f By combining the set of fractures obtained from the fracture probability distribution map, a direction-aware feature learning network is used to model the direction-aware features of the preprocessed image, thereby obtaining the fracture direction distribution probability and calculating the fracture direction dispersion H. θ The surface undulation features in the preprocessed image are modeled using a texture response feature learning network to obtain a texture response feature map characterizing the roughness of the surrounding rock surface, and the surface roughness E is calculated. r .

[0039] In this embodiment, the gap segmentation network includes, but is not limited to, U-Net, DeepLabv3+, MaskR-CNN, and Transformer segmentation networks; the orientation-aware feature learning network includes, but is not limited to, orientation convolutional networks, relative position encoding attention mechanism networks, and window self-attention based Transformer networks; and the texture response feature learning network includes, but is not limited to, multi-scale convolutional networks, feature pyramid networks, and Transformer-based texture encoding networks.

[0040] In this embodiment, in step 3, the crack density image quantity D is calculated through the following steps. f : Step A.1: Perform pixel-level structure recognition on the preprocessed image using a crack segmentation network to obtain a crack probability distribution map. The crack probability distribution map is a matrix with the same size as the preprocessed image, and each element of the matrix is ​​the probability value of the crack.

[0041] Step A.2: Threshold the crack probability distribution map to obtain a crack binary mask, separating cracked and non-cracked regions.

[0042] Step A.3: Perform morphological denoising and connected component filtering to obtain several segments of the fracture centerline.

[0043] Step A.4: Obtain the projected length l of the centerline of each fracture segment, and calculate the fracture density image quantity D using the following formula. f : ; In the formula, A is the effective area of ​​the working face, and N... f For the number of cracks, l i Let be the projected length of the center line of the i-th crack.

[0044] In this embodiment, the crack probability distribution map characterizes the spatial distribution of cracks within the effective area of ​​the tunnel face. Through crack centerline and geometric statistical calculations, the total projected length of the cracks within a unit area of ​​the tunnel face is obtained, resulting in the crack density image quantity D. f .

[0045] In this embodiment, in step 3, the crack direction dispersion H is calculated through the following steps. θ : Step B.1: Determine the number of directional intervals K, and determine the range of each directional interval.

[0046] In practice, the local orientation angle is calculated for the center line by pixel segments (or equidistant sampling segments). The orientation angle range can be 0-180°, which is common in engineering. The number of orientation intervals K is specified. For example, K=6 means that 0-180° is divided into 6 intervals: 0-30°, 30-60°, ..., 150-180°.

[0047] Step B.2: Based on the set of fracture centerlines obtained from the fracture probability distribution map, the preprocessed image is modeled using a direction-aware feature learning network to obtain the local orientation angles of each fracture centerline.

[0048] Step B.3: Count the number of samples in each directional interval, divide by the total number of samples, and obtain the probability distribution p of each directional interval. k .

[0049] Step B.4: Calculate the crack direction dispersion H using the following formula. θ : ; In the formula, It is a very small constant, such as 10 -6 .

[0050] In this embodiment, based on the spatial continuity of the centerlines of each fracture, the distribution characteristics of the fractures in different directions are modeled using direction-aware features. The fracture direction dispersion H is then obtained through direction statistics and information entropy calculation. θ .

[0051] In this embodiment, in step 3, the surface roughness E is calculated through the following steps. r : Step C.1: Model the surface undulation features in the preprocessed image using a texture response feature learning network to obtain a texture response feature map that characterizes the roughness of the surrounding rock surface.

[0052] Step C.2: Calculate the surface roughness E using the following formula. r : ; In the formula, M and N are the pixel sizes of the texture response feature map. Let I(x,y) be the Laplacian operator, and let I(x,y) be the gray value of a pixel in the texture response feature map.

[0053] In this embodiment, the surface undulation features of the tunnel face image are modeled, and a texture response feature map characterizing the roughness of the surrounding rock surface is output. The surface roughness E is then calculated based on the texture response feature map. r .

[0054] Step 4: Crack density image quantity D f Crack direction dispersion H θ Surface roughness E r After normalization, the surrounding rock structure complexity index RSCI is calculated using the following formula: ; In the formula, , , α, β, and γ are the normalized crack density image quantity, crack direction dispersion, and surface roughness, respectively, with α, β, and γ as weights.

[0055] Step 5: Based on the surrounding rock structure complexity index RSCI, calculate the continuous surrounding rock quality score Q using a monotonic mapping function, as shown in the following formula: ; In the formula, k is a constant greater than 0, used to control the steepness of the scoring curve, and b is the bias parameter.

[0056] In this embodiment, the continuous surrounding rock quality score Q is used to comprehensively characterize the degree of fracture development, structural orientation disorder, and surface fracturing characteristics of the surrounding rock. It reflects the continuous changing trend of the surrounding rock condition. In practical applications, combined with a preset threshold range, the continuous score is mapped to a traditional surrounding rock grade to ensure compatibility with existing engineering specifications and support dynamic decision-making during construction.

[0057] Step 6: Map the continuous surrounding rock quality score Q to the surrounding rock grade according to the preset threshold range, and output the surrounding rock grade assessment result.

[0058] In this embodiment, the continuous surrounding rock quality score Q is mapped to the corresponding surrounding rock grade according to a preset threshold range, and the surrounding rock grade assessment result is output. During actual tunnel construction, the UAV periodically collects new images of the tunnel face and executes the above process, which can realize continuous and automated updating of the surrounding rock grade, providing real-time decision support for construction safety control and support parameter adjustment.

[0059] In this embodiment, in step 6, the surrounding rock grade assessment results include Grade I, Grade II, Grade III, Grade IV, and Grade V; when Q ≥ 0.8, it is determined to be Grade I surrounding rock; when 0.6 ≤ Q < 0.8, it is determined to be Grade II surrounding rock; when 0.4 ≤ Q < 0.6, it is determined to be Grade III surrounding rock; when 0.2 ≤ Q < 0.4, it is determined to be Grade IV surrounding rock; and when Q < 0.2, it is determined to be Grade V surrounding rock.

[0060] In an optional embodiment, a multi-task joint learning approach is employed for training. By simultaneously optimizing the fracture segmentation task, the surrounding rock structure image regression task, the surrounding rock quality score prediction task, and the surrounding rock grade classification task, the model can balance the prediction accuracy and consistency of each sub-task while sharing feature representations, thereby improving the stability and reliability of the surrounding rock grade assessment results. Specifically, the loss is calculated through the following steps and used to train the model constructed in steps 3-6: Step D.1: Calculate the fracture segmentation loss L using the following formula. seg : ; In the formula, Ω is the set of pixels in the preprocessed image, P(x,y) is the probability that the model predicts that pixel (x,y) belongs to the crack, and G(x,y) is the true crack label value (0 or 1) of the corresponding pixel.

[0061] Crack segmentation loss L seg This is used to constrain the consistency between the fracture segmentation results predicted by the model and the actual fracture annotations.

[0062] Step D.2: Calculate the regression loss L of the surrounding rock structure image using the following formula. reg : ; In the formula, N is the number of training samples. , , The crack density image quantity, crack orientation dispersion, and surface roughness of the i-th sample predicted by the model are given. , , This corresponds to the actual labeled value.

[0063] Regression loss L of surrounding rock structure image reg The error between the predicted fracture density, fracture direction dispersion, and surface roughness energy and their corresponding true values ​​is used to constrain the model.

[0064] Step D.3: Calculate the consistency loss L of the surrounding rock quality score using the following formula. con : ; In the formula, The model predicts the surrounding rock quality score for the i-th sample. This corresponds to the actual labeled value.

[0065] Loss of consistency in surrounding rock quality scoring L con The consistency between the surrounding rock quality score used to constrain the model prediction and the reference score calculated based on the surrounding rock structure complexity index.

[0066] Step D.4: Calculate the surrounding rock grade classification loss L using the following formula. cls : ; In the formula, C represents the number of surrounding rock grade categories. When the model correctly predicts the surrounding rock grade... Set to 1 when the model's predicted surrounding rock grade is incorrect. Take 0, The model predicts the confidence level of the i-th sample belonging to class c. The confidence level is determined based on the position of the continuous surrounding rock quality score within the corresponding grade interval, characterizing the reliability of the surrounding rock grade assessment results. Specifically, a function can be set such that when mapping the surrounding rock grade according to the preset threshold interval in step 6, a higher confidence level is obtained when Q is in the middle of the preset threshold interval, and a lower confidence level is obtained when it is at the edge. Different functions can be selected to accomplish the above function in practice; the form of the function is not limited here.

[0067] Rock mass classification loss L cls This is used to constrain the consistency between the predicted rock mass class and the actual rock mass class.

[0068] Step D.5: Calculate the total loss L using the following formula: ; In the formula, λ1, λ2, λ3, and λ4 are weight coefficients used to balance the influence of different tasks during the training process. These weight coefficients can be set according to the importance of the tasks or through cross-validation.

[0069] This embodiment of an automated assessment method for tunnel surrounding rock grade introduces a surrounding rock structure complexity index and a continuous surrounding rock quality scoring mechanism to achieve interpretable and continuous assessment from tunnel face images to surrounding rock grade. This reduces reliance on manual experience, improves the objectivity, consistency, and efficiency of surrounding rock grade evaluation, and provides a reliable basis for tunnel construction safety and support decisions.

[0070] To address the issues of high reliance on human experience and insufficient subjectivity and consistency in surrounding rock grade determination, this embodiment introduces unmanned aerial vehicles (UAVs) for automated acquisition of tunnel face images. Combined with a deep learning model, the structural characteristics of the surrounding rock are automatically analyzed and quantified, reducing the involvement of manual on-site judgment. By objectively calculating surrounding rock image structural indicators such as fracture density, fracture direction dispersion, and surface roughness, the surrounding rock grade evaluation process is standardized and automated, effectively reducing the impact of human experience differences on the evaluation results and improving the objectivity and consistency of surrounding rock grade determination.

[0071] To address the shortcomings of existing image-based or deep learning-based methods, such as the lack of systematic quantification of surrounding rock structure features and insufficient interpretability of prediction results, this embodiment explicitly extracts surrounding rock structure features with clear engineering significance from tunnel face images, including fracture density, fracture direction dispersion, and surface roughness energy. It then constructs a surrounding rock structure complexity index and a continuous surrounding rock quality score, achieving an interpretable mapping relationship from image features to surrounding rock grade. Compared to existing end-to-end black-box prediction methods, the surrounding rock grade assessment process in this embodiment has clear physical meaning and engineering interpretability, facilitating understanding and application by construction personnel.

[0072] To address the issue of insufficient adaptability of existing deep learning models to complex surrounding rock structures and construction environments, this embodiment adopts a standardized, multi-dimensional image preprocessing process and combines a shared feature extraction backbone network with a multi-branch structural feature learning architecture to perform multi-scale modeling of surrounding rock fracture structures and surface texture features. Simultaneously, the model is trained through a multi-task joint learning approach, taking into account tasks such as fracture segmentation, structural quantity regression, quality scoring, and grade classification, thereby improving the model's adaptability to the heterogeneity of complex surrounding rock structures and changes in the construction environment, and enhancing the stability and reliability of the prediction results.

[0073] To address the limitations of existing methods for acquiring tunnel face images, the instability of image quality, and the lack of standardized preprocessing, this embodiment utilizes a drone equipped with automatic supplemental lighting, ambient light perception, path planning, and obstacle avoidance capabilities to automatically capture images of the tunnel face. The acquired images undergo standardized preprocessing operations such as semantic segmentation and cropping, noise reduction, illumination correction, sharpening enhancement, and scale normalization. This effectively overcomes problems such as insufficient lighting, dust interference, and the influence of invalid areas in the tunnel construction environment, providing high-quality and stable image input for subsequent extraction of surrounding rock structural features and grade assessment.

[0074] Example 2: An automated assessment system for the surrounding rock grade of a tunnel, comprising: a drone and a computer terminal, wherein the drone is used to execute step 1 of the method in Example 1, and the computer terminal is used to execute steps 2 to 6 of the method in Example 1.

[0075] The drone is equipped with an autonomous control unit, an ambient light sensor, an automatic fill light unit, a high-definition camera unit (resolution ≥ 4K), a wireless communication module (supporting 5G / WiFi), and a drone power supply unit.

[0076] The autonomous control unit incorporates a path planning algorithm and obstacle avoidance sensors, enabling automatic hovering, path planning, and obstacle avoidance. The ambient light sensor detects the light intensity inside the tunnel in real time and feeds it back to the automatic supplemental lighting unit. The automatic supplemental lighting unit adaptively adjusts the brightness (1000-10000 lux) and illumination angle (0-90°) according to the light intensity. The high-definition camera unit captures images of the tunnel face. The wireless communication module enables remote transmission of images and data.

[0077] The computer terminal is equipped with a local processing unit, a storage unit, a human-computer interaction interface, and a terminal power supply unit.

[0078] The local processing unit adopts a CPU with ≥16 cores and a GPU computing unit, meeting the hardware configuration requirements for real-time inference of deep learning models; the storage unit is used to store the original images and pre-processed images; the human-computer interaction interface includes a touch screen and physical operation buttons, used to receive user operation commands such as start / stop shooting, model parameter adjustment, and data query, and to display the surrounding rock grade prediction results, confidence level, and construction decision suggestions in real time.

[0079] The drone and the computer terminal establish a two-way communication connection through the wireless communication module to realize real-time image transmission, control command issuance and status feedback. The whole equipment can autonomously complete the entire process of face image acquisition, analysis and prediction in the tunnel construction scenario without the need for manual on-site participation in the determination of the surrounding rock grade.

[0080] Example 3: An automated assessment device for tunnel surrounding rock grade, comprising: Image preprocessing module: Performs multi-dimensional preprocessing on the working face to obtain effective image regions, including image cropping based on Mask R-CNN semantic segmentation, non-local mean denoising, CLAHE adaptive illumination correction, Gamma brightness equalization, Laplacian sharpening enhancement, bilinear interpolation scale normalization, and effective region purification based on edge detection. The surrounding rock structure feature learning module performs pixel-level structure recognition on the preprocessed image using a fracture segmentation network to obtain a fracture probability distribution map and calculate the fracture density image quantity D. f By combining the set of fractures obtained from the fracture probability distribution map, a direction-aware feature learning network is used to model the direction-aware features of the preprocessed image, thereby obtaining the fracture direction distribution probability and calculating the fracture direction dispersion H. θ The surface undulation features in the preprocessed image are modeled using a texture response feature learning network to obtain a texture response feature map characterizing the roughness of the surrounding rock surface, and the surface roughness E is calculated. r .

[0081] Module for calculating the complexity of surrounding rock structure: Fracture density image quantity D f Crack direction dispersion H θ Surface roughness E r After normalization, the surrounding rock structure complexity index RSCI is calculated.

[0082] Surrounding rock quality scoring module: Based on the surrounding rock structure complexity index RSCI, the continuous surrounding rock quality score Q is calculated through a monotonic mapping function.

[0083] Rock grade mapping module: Maps the continuous rock quality score Q to the rock grade according to a preset threshold range and outputs the rock grade assessment result.

[0084] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An automated method for assessing the grade of surrounding rock in tunnels, characterized in that, Includes the following steps: Step 1: Take images of the working face using a drone and transmit the images to a computer terminal; Step 2: Perform multi-dimensional preprocessing on the face of the tunnel to obtain effective image regions, including image cropping based on MaskR-CNN semantic segmentation, non-local mean denoising, CLAHE adaptive illumination correction, Gamma brightness equalization, Laplacian sharpening enhancement, bilinear interpolation scale normalization, and effective region purification based on edge detection. Step 3: Perform pixel-level structure recognition on the preprocessed image using a crack segmentation network to obtain a crack probability distribution map, and calculate the crack density image quantity D. f By combining the set of fractures obtained from the fracture probability distribution map, a direction-aware feature learning network is used to model the direction-aware features of the preprocessed image, thereby obtaining the fracture direction distribution probability and calculating the fracture direction dispersion H. θ The surface undulation features in the preprocessed image are modeled using a texture response feature learning network to obtain a texture response feature map characterizing the roughness of the surrounding rock surface, and the surface roughness E is calculated. r ; Step 4: Crack density image quantity D f Crack direction dispersion H θ Surface roughness E r After normalization, the surrounding rock structure complexity index RSCI is calculated using the following formula: ； In the formula, , , These represent the normalized crack density image quantity, crack direction dispersion, and surface roughness, respectively, with α, β, and γ as weights. Step 5: Based on the surrounding rock structure complexity index RSCI, calculate the continuous surrounding rock quality score Q using a monotonic mapping function, as shown in the following formula: ； In the formula, k is a constant greater than 0, used to control the steepness of the scoring curve, and b is the bias parameter; Step 6: Map the continuous surrounding rock quality score Q to the surrounding rock grade according to the preset threshold range, and output the surrounding rock grade assessment result.

2. The automated assessment method for tunnel surrounding rock grade according to claim 1, characterized in that, Step 2 includes the following steps: Step 2.1 Image cropping based on Mask R-CNN semantic segmentation: Using a pre-trained Mask R-CNN model with a confidence threshold of 0.75, the non-rock areas in the image are identified and segmented, and the output is an initial valid image containing only the exposed rock area of ​​the working face. Step 2.2, Non-local means denoising: The non-local means denoising algorithm is implemented using the OpenCV library. The similarity window size is set to 7×7, the search window size is set to 21×21, and the denoising intensity parameter is set to h=10. The cropped image is then denoised. Step 2.3, CLAHE Adaptive Illumination Correction: Using the CLAHE function of OpenCV, the contrast gain is limited to 2.0 and the block size is 8×8. Histogram equalization is performed on the denoised image to solve the problem of local over-brightness / under-brightness caused by uneven lighting in the tunnel, balance the brightness distribution of the image, and make the surrounding rock structure features in the shadow area clearly visible. Step 2.4, Gamma Brightness Equalization: Adaptively adjust the Gamma coefficient (range 0.8-1.2) based on the average gray value of the image. When the gray value is below 128, take 0.8-0.9; when it is between 128 and 192, take 1.0; and when it is above 192, take 1.1-1.

2. Optimize the overall brightness of the image through power-law transformation, enhance the detail expression in low-light areas, and avoid feature loss caused by extreme brightness values. Step 2.5, Laplacian sharpening enhancement: A 3×3 Laplacian operator (coefficient matrix [0,-1,0;-1,5,-1;0,-1,0]) is used to perform convolution operations on the illumination-corrected image to enhance edge features, improve detail recognition, and provide support for the extraction of crack development status features. Step 2.6, Bilinear Interpolation Scale Normalization: The bilinear interpolation algorithm is used to uniformly scale the sharpened image to 224×224 pixels. The bilinear interpolation algorithm is used to maintain the feature without distortion, so as to unify the input data format of the model in Step 3, eliminate the image scale difference caused by different shooting distances and angles, and ensure the consistency of feature extraction. Step 2.7: Purification of effective image region based on edge detection: The Canny edge detection algorithm (threshold range 50-150) is used to identify the edge contour of the surrounding rock area and remove a small number of non-surrounding rock edge pixels remaining after scale normalization to further purify the effective image region and ensure that the image input to the model in Step 3 contains only the core feature information of the surrounding rock.

3. The automated assessment method for tunnel surrounding rock grade according to claim 1, characterized in that, In step 3, the crack density image quantity D is calculated through the following steps. f : Step A.1: Perform pixel-level structure recognition on the preprocessed image using a crack segmentation network to obtain a crack probability distribution map. The crack probability distribution map is a matrix with the same size as the preprocessed image, and each element of the matrix is ​​the probability value of the crack. Step A.2: Threshold the crack probability distribution map to obtain a binary crack mask, separating cracked and non-cracked regions; Step A.3: Perform morphological denoising and connected component filtering to obtain several segments of the fracture centerline; Step A.4: Obtain the projected length l of the centerline of each fracture segment, and calculate the fracture density image quantity D using the following formula. f : ； In the formula, A is the effective area of ​​the working face, and N... f For the number of cracks, l i Let be the projected length of the center line of the i-th crack.

4. The automated assessment method for tunnel surrounding rock grade according to claim 3, characterized in that, In step 3, the crack direction dispersion H is calculated through the following steps. θ : Step B.1: Determine the number K of directional intervals and the range of each directional interval; Step B.2: Based on the set of fracture centerlines obtained from the fracture probability distribution map, the preprocessed image is modeled using a direction-aware feature learning network to obtain the local orientation angles of each fracture centerline; Step B.3: Count the number of samples in each directional interval, divide by the total number of samples, and obtain the probability distribution p of each directional interval. k ; Step B.4: Calculate the crack direction dispersion H using the following formula. θ : ； In the formula, It is a very small constant.

5. The automated assessment method for tunnel surrounding rock grade according to claim 1, characterized in that, In step 3, the surface roughness E is calculated through the following steps. r : Step C.1: Model the surface undulation features in the preprocessed image using a texture response feature learning network to obtain a texture response feature map that characterizes the roughness of the surrounding rock surface; Step C.2: Calculate the surface roughness E using the following formula. r : ； In the formula, M and N are the pixel sizes of the texture response feature map. Let I(x,y) be the Laplacian operator, and let I(x,y) be the gray value of a pixel in the texture response feature map.

6. The automated assessment method for tunnel surrounding rock grade according to claim 1, characterized in that, In step 6, the surrounding rock grade assessment results include Grade I, Grade II, Grade III, Grade IV, and Grade V; when Q ≥ 0.8, it is judged as Grade I surrounding rock; when 0.6 ≤ Q < 0.8, it is judged as Grade II surrounding rock. When 0.4≤Q<0.6, it is classified as Class III surrounding rock; when 0.2≤Q<0.4, it is classified as Class IV surrounding rock; when Q<0.2, it is classified as Class V surrounding rock.

7. The automated assessment method for tunnel surrounding rock grade according to claim 1, characterized in that, The loss is calculated using the following steps and used to train the model built in steps 3-6: Step D.1: Calculate the fracture segmentation loss L using the following formula. seg : ； In the formula, Ω is the set of pixels in the preprocessed image, P(x,y) is the probability that the model predicts that pixel (x,y) belongs to the crack, and G(x,y) is the true crack label value of the corresponding pixel. Step D.2: Calculate the regression loss L of the surrounding rock structure image using the following formula. reg : ； In the formula, N is the number of training samples. , , The crack density image quantity, crack orientation dispersion, and surface roughness of the i-th sample predicted by the model are given. , , This corresponds to the actual labeled value; Step D.3: Calculate the consistency loss L of the surrounding rock quality score using the following formula. con : ； in the formula The model predicts the surrounding rock quality score for the i-th sample. This corresponds to the actual labeled value; Step D.4: Calculate the surrounding rock grade classification loss L using the following formula. cls : ； In the formula, C represents the number of surrounding rock grade categories. When the model correctly predicts the surrounding rock grade... Set to 1 when the model's predicted surrounding rock grade is incorrect. Take 0, Predict the confidence level of the i-th sample belonging to class c for the model; Step D.5: Calculate the total loss L using the following formula: ； In the formula, λ1, λ2, λ3, and λ4 are weight coefficients used to balance the influence of different tasks during the training process. These weight coefficients can be set according to the importance of the tasks or through cross-validation.

8. An automated assessment system for the surrounding rock grade of tunnels, characterized in that, include: The drone and the computer terminal, wherein the drone is used to perform step 1 of the method according to any one of claims 1 to 7, and the computer terminal is used to perform steps 2 to 6 of the method according to any one of claims 1 to 7; The drone is equipped with an autonomous control unit, an ambient light sensor, an automatic supplemental lighting unit, a high-definition camera unit, a wireless communication module, and a drone power supply unit; the computer terminal is equipped with a local processing unit, a storage unit, a human-computer interaction interface, and a terminal power supply unit.

9. An automated assessment device for the surrounding rock grade of tunnels, characterized in that, include: Image preprocessing module: Performs multi-dimensional preprocessing on the face of the tunnel to obtain effective image regions, including image cropping based on MaskR-CNN semantic segmentation, non-local mean denoising, CLAHE adaptive illumination correction, Gamma brightness equalization, Laplacian sharpening enhancement, bilinear interpolation scale normalization, and effective region purification based on edge detection. The surrounding rock structure feature learning module performs pixel-level structure recognition on the preprocessed image using a fracture segmentation network to obtain a fracture probability distribution map and calculate the fracture density image quantity D. f By combining the set of fractures obtained from the fracture probability distribution map, a direction-aware feature learning network is used to model the direction-aware features of the preprocessed image, thereby obtaining the fracture direction distribution probability and calculating the fracture direction dispersion H. θ The surface undulation features in the preprocessed image are modeled using a texture response feature learning network to obtain a texture response feature map characterizing the roughness of the surrounding rock surface, and the surface roughness E is calculated. r ; Module for calculating the complexity of surrounding rock structure: Fracture density image quantity D f Crack direction dispersion H θ Surface roughness E r After normalization, the surrounding rock structure complexity index RSCI is calculated. The surrounding rock quality scoring module calculates the continuous surrounding rock quality score Q using a monotonic mapping function based on the surrounding rock structure complexity index RSCI. Rock grade mapping module: Maps the continuous rock quality score Q to the rock grade according to a preset threshold range and outputs the rock grade assessment result.

Images