Rock image-based tmb surrounding rock state real-time perception method and system

Abstract

This invention discloses a real-time perception method and system for the surrounding rock condition of a TBM (Tunnel Boring Machine) based on rock debris images, belonging to the field of geological perception technology for TBM construction. The method includes acquiring images of rock debris on the conveyor belt during TBM tunneling, performing boundary recognition and contour extraction on the rock debris particles, and calculating multiple morphological parameters for individual particles. Based on the morphological parameters, the particles are clustered to classify their respective rock debris types, and the proportion of each type of particle in the current cycle is statistically analyzed. Simultaneously, the global texture roughness of the rock debris image is calculated. The particle quantity proportion and global texture roughness are used as input features and input to a pre-trained surrounding rock condition classifier, outputting the current surrounding rock condition category at the TBM face. This invention achieves real-time and automatic identification of the surrounding rock condition through joint analysis of rock debris morphology and texture features, providing a reliable basis for TBM tunneling parameter control.

Description

Technical Field

[0001] This invention relates to the field of geological sensing technology for tunnel boring machine (TBM) construction, specifically to a method and system for real-time sensing of the surrounding rock condition of TBMs based on rock debris images. Background Technology

[0002] During tunnel boring machine (TBM) construction, real-time and accurate perception of the surrounding rock condition at the tunnel face is a crucial prerequisite for ensuring tunneling safety and efficiency. The traditional method of relying on on-site geologists to manually observe and sketch the tunnel walls or face is highly subjective and lags behind the tunneling progress, failing to meet the timeliness requirements of automated and intelligent construction. Utilizing images of rock debris generated during tunneling to determine the surrounding rock condition is a promising non-contact real-time sensing method, as the rock debris originates directly from the cut and broken rock mass at the tunnel face and image acquisition is convenient.

[0003] Existing techniques for analyzing surrounding rock conditions based on rock debris images largely rely on describing the overall characteristics of the rock debris pile, such as extracting global texture parameters or statistically averaged particle size distribution from the entire image to infer the integrity of the rock mass. This approach, which relies solely on overall statistical values, has significant limitations when dealing with complex rock structures and rock debris exhibiting multimodal morphological distributions. When the tunnel face contains both intact rock blocks and weak interlayers, the rock debris produced by fracturing often encompasses multiple types, including blocky, flaky, and powdery forms. Simply using global texture averages or particle size distribution curves cannot effectively decouple heterogeneous particle groups generated by different fracturing mechanisms. This results in the overall characteristics being insensitive to changes in the proportion of different particle types, making it difficult to accurately reflect the gradual transition between the degree of joint development and fracturing in the surrounding rock.

[0004] Furthermore, current quantitative analysis of rock debris particle morphology often stops at particle size or a single shape factor, lacking a systematic approach for multi-dimensional morphological clustering and typological classification of particle groups. When a rock mass evolves from a jointed state to a fractured state, the relationship between massive particles and flaky / powdered particles in the rock debris is a core indicator characterizing the trend of surrounding rock deterioration. However, overall statistical methods cannot quantify this relationship, making it difficult to capture precursory information about the evolution of the surrounding rock state. For these reasons, effectively decoupling multi-type mixed rock debris particle groups at the image level and extracting multi-dimensional characterization information that accurately reflects the compositional relationships of different particle types and the overall degree of fracture and roughness has become a challenging problem to solve. Summary of the Invention

[0005] The purpose of this invention is to provide a real-time sensing method and system for the surrounding rock condition of a TBM based on rock slag images, in order to solve the problem that the overall statistical features in the prior art are insufficient in their ability to decouple multi-type mixed rock slag particle groups, resulting in limited accuracy in judging the surrounding rock condition.

[0006] To achieve the above objectives, the present invention provides the following technical solution: The present invention provides a real-time perception method for the surrounding rock state of a TBM based on rock debris images, comprising the following steps: acquiring rock debris images on the conveyor belt during TBM tunneling; identifying the boundaries of rock debris particles in the rock debris images, extracting the contours of individual rock debris particles, and calculating multiple morphological parameters of the individual rock debris particles based on the contours; clustering the individual rock debris particles according to the multiple morphological parameters to classify the rock debris types to which the rock debris particles belong; statistically analyzing the proportion of particles of each rock debris type within the current image acquisition period, and calculating the global texture roughness of the rock debris image; using the proportion of particle numbers and the global texture roughness as input features, inputting them into a pre-trained surrounding rock state classifier, and outputting the current state category of the surrounding rock at the TBM face. This method directly extracts two complementary features—particle morphology and global texture—from the rock debris images on the conveyor belt, eliminating the need for manual sampling and laboratory analysis, and achieving real-time automatic perception of different states such as intact surrounding rock, jointed surrounding rock, fractured surrounding rock, and extremely fractured surrounding rock.

[0007] As a preferred technical solution of the present invention, the process of identifying the boundaries of rock slag particles and extracting the contours of individual rock slag particles from the rock slag image specifically includes: converting the rock slag image into a grayscale image, and performing binarization processing on the grayscale image using an adaptive threshold segmentation algorithm to generate a binary image in which the foreground and background regions of the rock slag are separated; performing morphological opening operations on the binary image to eliminate noise and small connecting bridges within the foreground region of the rock slag, thereby separating mutually adhered rock slag particles; marking connected components in the binary image after the morphological opening operation to identify each independent connected component as a candidate rock slag particle region; and removing candidate rock slag particle regions with an area smaller than a preset area threshold, using the boundaries of the remaining candidate rock slag particle regions as the contours of the individual rock slag particles. Preferably, the preset area threshold is dynamically set according to the acquisition resolution of the rock slag image and the running speed of the conveyor belt, thereby adapting to changes in the size of rock slag particles under different working conditions, reducing interference from fine debris or noise, and improving the reliability of particle extraction.

[0008] Furthermore, the step of calculating multiple morphological parameters of the individual rock debris particle based on the contour includes: calculating the projected area, perimeter, and major and minor axis lengths of the equivalent ellipse of the individual rock debris particle based on the contour; calculating the roundness of the individual rock debris particle based on the projected area and perimeter; calculating the aspect ratio of the individual rock debris particle based on the major and minor axis lengths; and calculating the convexity of the individual rock debris particle based on the projected area and the convex hull area of ​​the contour. These morphological parameters characterize the rock debris particle from dimensions such as size, shape regularity, and angular features, providing a multi-dimensional basis for the accurate classification of subsequent rock debris types. The equivalent ellipse is an ellipse with the same second-order central moment as the individual rock debris particle. The major and minor axis lengths are obtained by eigenvalue decomposition of the covariance matrix of the contour, ensuring rotational invariance in the calculation of the major and minor axes, thus stably describing the extension direction and size of the particle.

[0009] As a preferred embodiment of the present invention, the specific method for clustering and classifying the rock debris particles according to the multiple morphological parameters is as follows: The projected area, roundness, aspect ratio, and convexity of each individual rock debris particle are combined to form a feature vector for that particle; multiple cluster centers are set, each corresponding to a rock debris type, including blocky rock debris, flaky rock debris, powdery rock debris, and elongated rock debris; the Euclidean distance from each feature vector to each cluster center is calculated, and the rock debris particles are assigned to the rock debris type corresponding to the cluster center with the smallest Euclidean distance; after each division, the mean of all feature vectors under each rock debris type is recalculated, and the mean is used as the new cluster center, and the division steps are repeated until the cluster centers no longer change. Preferably, when the sum of the displacements of the cluster centers in two consecutive iterations is less than a preset displacement threshold, it is determined that the cluster centers no longer change and the iteration is terminated. This clustering process does not require pre-labeling of a large number of samples and can automatically adapt to the overall changes in the morphological characteristics of rock debris under different working conditions, and complete the rapid classification of rock debris types online.

[0010] After determining the types of rock debris, the specific steps for calculating the proportion of particles of each rock debris type within the current image acquisition period include: calculating the total number of all rock debris particles identified within the current image acquisition period; separately calculating the number of rock debris particles belonging to massive rock debris, flaky rock debris, powdery rock debris, and elongated rock debris; and dividing the number of rock debris particles belonging to each type of rock debris by the total number to obtain the proportion of particles of each type of rock debris. This proportion reflects the macroscopic statistical characteristics of rock debris composition and is directly related to the integrity and failure mode of the surrounding rock.

[0011] In another preferred embodiment of the present invention, the step of calculating the global texture roughness of the rock slag image includes: dividing the rock slag image into multiple sub-windows of the same size; calculating the standard deviation of the pixel grayscale values ​​within each sub-window as the local texture variation of that sub-window; calculating the average of the local texture variation of all sub-windows, and using the average as the global texture roughness of the rock slag image. The global texture roughness quantifies the overall degree of grayscale and texture variation on the surface of the rock slag pile, and has a good ability to distinguish between the fine texture when the proportion of powdery rock slag is high and the rough texture when the proportion of blocky rock slag is high.

[0012] Preferably, before inputting the particle number ratio and the global texture roughness as input features into the pre-trained surrounding rock state classifier, the particle number ratio and the global texture roughness are normalized to map them to the same numerical range, thereby eliminating the influence of dimensional differences on the classifier's decision and improving classification stability.

[0013] Regarding the output of the surrounding rock state classifier, the present invention preferably adopts the following method: the surrounding rock state classifier calculates the probability value of the current surrounding rock belonging to intact surrounding rock, jointed surrounding rock, fractured surrounding rock, and extremely fractured surrounding rock based on the input features, and outputs the surrounding rock state category with the highest probability value as the current TBM face surrounding rock state category. Further, the surrounding rock state classifier adopts a decision tree model, which contains multiple internal nodes and leaf nodes, each internal node corresponding to a feature partitioning condition; starting from the root node of the decision tree model, the input features are judged according to the feature partitioning condition of the current node, selecting to enter the corresponding child node, until reaching the leaf node, where the leaf node outputs the probability value. This decision tree model has a simple structure, fast calculation speed, and can be directly deployed on the TBM field control equipment to meet real-time perception requirements.

[0014] To improve the quality of rock debris images, after acquiring images of rock debris on the conveyor belt during TBM tunneling, illumination compensation processing is performed on the rock debris images to correct brightness differences caused by uneven illumination. Then, the illumination-compensated rock debris images undergo color space conversion to extract the luminance component from the HSV color space, which is used as the basis for subsequent image processing. This preprocessing effectively suppresses the interference of complex lighting conditions within the tunnel on image segmentation and texture analysis, enhancing the robustness of the entire perception method to the on-site environment.

[0015] This invention also provides a real-time TBM surrounding rock condition sensing system based on rock debris images, including a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of the aforementioned real-time TBM surrounding rock condition sensing method based on rock debris images. This system can acquire rock debris images on the conveyor belt online, calculate rock debris particle morphology parameters, particle type proportions, and global texture roughness in real time, and quickly output the surrounding rock condition category at the tunnel face using a surrounding rock condition classifier, providing immediate decision-making basis for TBM tunneling parameter control and disaster early warning.

[0016] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0017] For each individually extracted rock debris particle, multiple morphological parameters, including projected area, roundness, aspect ratio, and convexity, are calculated. Based on these parameters, iterative clustering automatically classifies the particles into types with clear geological fracturing implications, such as blocky, flaky, powdery, and elongated shapes, effectively decoupling the originally mixed particle groups in the feature space. Furthermore, the proportion of each type of rock debris particle within the current image acquisition period is statistically analyzed, converting the single average particle size or overall shape factor of the entire image into a vector representing the proportion of each particle type. As the surrounding rock deteriorates from an intact state to a fractured state through joint development, the proportion of blocky particles gradually decreases, while the proportions of flaky and powdery particles increase synergistically. This dynamic relationship is directly quantified by the particle quantity proportion vector, thus overcoming the insensitivity of the overall statistical method to changes in the internal structure of multimodal particle groups and enabling the timely capture of subtle gradual changes in the surrounding rock condition.

[0018] Simultaneously, the proportion of different rock debris types obtained after decoupling and the global texture roughness of the rock debris image are used together as input features for the surrounding rock state classifier, realizing the fusion of particle type composition information and overall texture information at the decision level. The global texture roughness is obtained by dividing sub-windows and calculating the mean of the standard deviation of gray values, reflecting the macroscopic undulation degree of the rock debris surface formed by the stacking of broken blocks. During operation, the classifier considers both the increase and decrease of the proportion of each type of particle and the disorder of the stacked texture. When the proportion of flaky particles increases but the texture roughness remains at a high level, it can distinguish jointed surrounding rocks; while when powdery particles dominate and the texture roughness decreases significantly, it points to extremely broken surrounding rocks. This multi-dimensional feature combination method avoids the ambiguity of single-class features in critical states, making the boundaries between intact surrounding rocks, jointed surrounding rocks, broken surrounding rocks, and extremely broken surrounding rocks clearer, and improving the accuracy and robustness of the output state category. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0020] Figure 1 This is a flowchart of a real-time sensing method for the surrounding rock condition of a TBM based on rock debris images;

[0021] Figure 2 This is a flowchart of rock slag image processing and particle morphology parameter calculation;

[0022] Figure 3 This is a flowchart of a rock condition classification method based on a decision tree model;

[0023] Figure 4 This is a flowchart of the TBM surrounding rock condition real-time sensing method. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] See Figure 1 This invention provides a real-time perception method for the surrounding rock condition of a TBM based on rock debris images. The method first acquires images of rock debris on the conveyor belt during TBM tunneling. Boundary recognition of rock debris particles is performed on the acquired rock debris images to extract the contours of individual rock debris particles. Based on the extracted contours, multiple morphological parameters of each individual rock debris particle are calculated. The individual rock debris particles are clustered according to the calculated morphological parameters to classify their rock debris type. After classifying all particles, the proportion of each rock debris type within the current image acquisition period is statistically analyzed, and the global texture roughness of the rock debris image is calculated. The statistically analyzed particle proportion and the calculated global texture roughness are used as input features and fed into a pre-trained surrounding rock condition classifier, which outputs the current state category of the surrounding rock at the TBM face.

[0026] Example 1

[0027] In specific implementation, please refer to Figure 2The acquired rock debris image was converted into a grayscale image using a weighted average method. The red, green, and blue channel values ​​of each pixel in the rock debris image were assigned a first weighting coefficient, a second weighting coefficient, and a third weighting coefficient, respectively. The weighted sum was then used as the grayscale value of the corresponding pixel in the grayscale image. The first weighting coefficient was set to 0.299, the second weighting coefficient to 0.587, and the third weighting coefficient to 0.114.

[0028] After generating the grayscale image, an adaptive thresholding segmentation algorithm is used to binarize it. The adaptive thresholding segmentation algorithm uses the mean or Gaussian weighted mean of the grayscale values ​​in the neighborhood of each pixel in the grayscale image as a local threshold. When the grayscale value of a pixel is greater than the local threshold, the pixel is classified as a rock debris foreground region; when the grayscale value of a pixel is less than or equal to the local threshold, the pixel is classified as a background region, thus forming a binary image with separated rock debris foreground and background regions. The neighborhood size in the adaptive thresholding segmentation algorithm is set based on the average size of the rock debris particles within the field of view during image acquisition; the neighborhood size is an odd value greater than the average diameter of the rock debris particles.

[0029] A morphological opening operation is performed on the binary image, consisting of two cascaded operations: erosion and dilation. The erosion operation uses a circular structuring element, the radius of which is set based on the minimum size of interest in the rock debris image, and is set to 3 pixels. The erosion operation iterates through every pixel in the binary image; if all pixels within the area covered by the circular structuring element are rock debris foreground regions, the central pixel remains a rock debris foreground region; otherwise, it is set to the background region. The dilation operation uses the same circular structuring element, iterating through every pixel in the binary image after the erosion operation. If at least one pixel within the area covered by the circular structuring element is a rock debris foreground region, the central pixel is set to a rock debris foreground region. The morphological opening operation eliminates fine noise within the rock debris foreground region and breaks the tiny connecting bridges between rock debris particles, separating the originally adhered rock debris particles in the image.

[0030] Connectivity labeling is performed on the binary image after morphological opening operations, using either 4-neighborhood or 8-neighborhood connectivity. The labeling process scans pixels line by line in the binary image. When an unlabeled pixel belonging to the slag foreground region is scanned, a new connectivity label is assigned. Connectivity growth is then performed starting from this pixel, labeling all neighboring pixels belonging to the slag foreground region with the same connectivity label, until all connected slag foreground region pixels are labeled. After scanning, each independent connectivity region is identified as a candidate slag particle region.

[0031] After obtaining all candidate rock debris particle regions, those with an area smaller than a preset area threshold are discarded. The preset area threshold is dynamically set based on the acquisition resolution of the rock debris image and the operating speed of the conveyor belt. The dynamic setting method is as follows: multiply the conveyor belt speed by a time reference to obtain the distance the conveyor belt travels within that time reference; multiply this distance by the acquisition resolution to obtain the corresponding pixel travel distance in the image; and the preset area threshold is proportional to the square of this pixel travel distance. The boundary of the remaining candidate rock debris particle regions is used as the contour of a single rock debris particle, and the contour is represented by a sequence of boundary point coordinates.

[0032] After extracting the contour of a single rock debris particle, the projected area, perimeter, major axis length, and minor axis length of the equivalent ellipse of the single rock debris particle are calculated based on the contour of the single rock debris particle.

[0033] The projected area is obtained by counting the total number of pixels contained within the outline of a single rock debris particle and multiplying the total number of pixels by the physical area of ​​a single pixel. The physical area of ​​a single pixel is determined by the acquisition resolution of the rock debris image. The perimeter is obtained by sequentially traversing the sequence of boundary points of the outline of a single rock debris particle, calculating the Euclidean distance between adjacent boundary points, and summing the Euclidean distances between all adjacent boundary points; when the outline is a closed outline, the Euclidean distance between the first and last boundary points is also included in the summation value.

[0034] The equivalent ellipse is an ellipse with the same second-order central moment as the contour of a single rock debris particle. The major and minor axis lengths of the equivalent ellipse are obtained by eigenvalue decomposition of the covariance matrix of the contour of the single rock debris particle. The covariance matrix of the contour is constructed as follows: the coordinates of all pixels inside the contour of a single rock debris particle are considered as a set of points in two-dimensional space; the mean of the point set in the row and column directions is calculated, and the covariance matrix of the row and column coordinates is constructed. Eigenvalue decomposition is then performed on the covariance matrix to obtain the first eigenvalue. Second eigenvalue ,and The length of the major axis of the equivalent ellipse. Values The length of the minor axis of the equivalent ellipse Values .

[0035] After obtaining the projected area and perimeter, the roundness of individual rock fragments is calculated. The formula for calculating roundness is:

[0036]

[0037] in, Indicates roundness. Indicates the perimeter. Represents the projected area. This represents pi, which is 3.1415926.

[0038] The aspect ratio of a single rock fragment is calculated based on its major axis length and minor axis length, with the aspect ratio being equal to the length of the major axis. With minor axis length The ratio of .

[0039] The convexity of a single rock debris particle is calculated based on its projected area and the area of ​​the convex hull of its contour. The convex hull area is calculated as follows: The convex hull is obtained by taking the coordinate sequence of the boundary points of the single rock debris particle's contour. The convex hull is the smallest convex polygon containing all boundary points. The total number of pixels inside the convex polygon is calculated and multiplied by the physical area of ​​a single pixel to obtain the convex hull area. Convexity is the projected area. Area of ​​convex hull The ratio, i.e. .

[0040] Example 2

[0041] In practice, the projected area, roundness, aspect ratio, and convexity of each individual rock slag particle are combined into a four-dimensional feature vector. The first feature component of the four-dimensional feature vector is the projected area, the second is the roundness, the third is the aspect ratio, and the fourth is the convexity. The four-dimensional feature vectors of all individual rock slag particles constitute a feature vector set.

[0042] To perform clustering, four initial cluster centers are defined in the feature space, each corresponding to a type of rock debris: massive rock debris, flaky rock debris, powdery rock debris, and elongated rock debris. Each of the four initial cluster centers is a four-dimensional vector, and the component values ​​of each initial cluster center are determined by the mean of the feature vectors of manually labeled typical rock debris particle samples from historical rock debris image data. In the absence of historical data, four four-dimensional feature vectors are randomly selected from all four-dimensional feature vectors of the current rock debris image as the four initial cluster centers.

[0043] For each four-dimensional feature vector in the feature vector set, calculate the Euclidean distance between the four-dimensional feature vector and the four cluster center vectors. The formula for calculating the Euclidean distance is:

[0044]

[0045] in, Indicates the first The four-dimensional eigenvector of a single rock debris particle to the first... Euclidean distance between the cluster center vectors corresponding to each type of rock slag; Indicates the first The projected area of ​​a single rock debris particle; Indicates the first The projected area components of each cluster center vector; Indicates the first The roundness of individual rock debris particles; Indicates the first The circularity component of each cluster center vector; Indicates the first The aspect ratio of a single rock debris particle; Indicates the first The aspect ratio components of the cluster center vectors; Indicates the first The convexity of a single rock debris particle; Indicates the first The convexity components of the cluster center vectors.

[0046] After completing the Euclidean distance calculation, the first... Individual rock debris particles are divided into... The first to obtain the minimum value Among the rock slag types corresponding to the cluster center vectors.

[0047] After the four-dimensional feature vectors of all individual rock debris particles in the feature vector set have been partitioned once, for each type of rock debris, the four-dimensional feature vectors of all individual rock debris particles classified under that type are collected, and the arithmetic mean of these four-dimensional feature vectors on each feature component is calculated. The arithmetic mean of all calculated feature components is used to form a new four-dimensional cluster center vector, which replaces the original cluster center vector. The arithmetic mean is calculated as follows: for massive rock debris, the projected area component of the new cluster center vector is the average projected area of ​​all individual rock debris particles classified as massive rock debris; the roundness component of the new cluster center vector is the average roundness of all individual rock debris particles classified as massive rock debris; the aspect ratio component of the new cluster center vector is the average aspect ratio of all individual rock debris particles classified as massive rock debris; and the convexity component of the new cluster center vector is the average convexity of all individual rock debris particles classified as massive rock debris. The calculation method for the new cluster center vectors of flaky, powdery, and elongated rock debris types is the same as that for massive rock debris type.

[0048] After updating the cluster center vectors, the sum of the displacements of the four cluster center vectors before and after the update is calculated. The sum of displacements is calculated by calculating the Euclidean distance between the new and old cluster center vectors for each of the four rock debris types, and then summing the four Euclidean distances. The sum of displacements is compared with a preset displacement threshold, which is set to 0.001. The reason for setting the preset displacement threshold to 0.001 is that the typical numerical range of the projected area feature component is in the hundreds to thousands of pixels range, and the roundness, aspect ratio, and convexity feature components are all ratios, with numerical ranges typically between 0 and 1 or within a finite range greater than 1. When the sum of the displacements of the four cluster center vectors is less than 0.001, it indicates that the positional change of the cluster center vectors in the four-dimensional feature space is very small, and the impact of continued iteration on the type classification result of individual rock debris particles is negligible; the clustering result has reached a stable state.

[0049] When the total displacement is less than the preset displacement threshold of 0.001, the iteration process terminates, and the rock debris type of each individual rock debris particle determined in this partition is taken as the final rock debris type partitioning result. When the total displacement is greater than or equal to the preset displacement threshold of 0.001, based on the new four cluster center vectors, the steps of calculating the Euclidean distance for each four-dimensional feature vector and partitioning, as well as the subsequent steps of updating the cluster center vectors, are repeated until the total displacement satisfies the condition of being less than the preset displacement threshold of 0.001.

[0050] After classifying all individual rock fragments into their corresponding types within the current image acquisition period, the total number of individual rock fragments identified within that period is counted. The number of individual rock fragments ultimately classified as blocky, flaky, powdery, and elongated is counted separately. The percentage of blocky rock fragments is calculated by dividing the number of individual rock fragments by the total number of individual rock fragments; the percentage of flaky rock fragments is calculated by dividing the number of individual rock fragments by the total number of individual rock fragments; the percentage of powdery rock fragments is calculated by dividing the number of individual rock fragments by the total number of individual rock fragments; and the percentage of elongated rock fragments is calculated by dividing the number of individual rock fragments by the total number of individual rock fragments. The sum of the percentages of the four rock fragment types is 1.

[0051] Example 3

[0052] In practice, the rock debris image is divided into multiple sub-windows of the same size. Each sub-window is rectangular, with equal pixel dimensions in both the row and column directions. The row and column pixel dimensions of each sub-window are set to 64 pixels. This 64-pixel size is chosen because its proportion to the overall rock debris image ensures that the number of sub-windows is statistically representative of the texture space variations, while each sub-window contains enough pixels to guarantee the statistical stability of the grayscale standard deviation calculation.

[0053] The partitioning method is as follows: starting from the top left corner pixel of the rock debris image, rectangular regions of 64 pixels in the row direction and 64 pixels in the column direction are successively extracted as sub-windows, following the order of row-first and column-second. When the number of remaining pixels in the row or column direction of the rock debris image is less than 64 pixels, the edge regions with less than 64 pixels are discarded. The rock debris image is divided into M sub-windows, where M is equal to the product of the number of complete sub-windows in the row direction and the number of complete sub-windows in the column direction.

[0054] For each sub-window obtained from the partitioning, the standard deviation of the grayscale values ​​of all pixels within the sub-window is calculated, and this standard deviation is used as the local texture variation of the sub-window. The formula for calculating the local texture variation is as follows:

[0055]

[0056] in, Indicates the first The amount of local texture variation in each sub-window This is the index number of the child window. The value range is from 1 to Integers; This represents the total number of pixels contained within a child window. The value is 4096, which is obtained by multiplying the row-direction pixel size of the sub-window by the column-direction pixel size of 64; Indicates the index number of the pixel within the child window. The value range is an integer from 1 to 4096; Indicates the first The first sub-window The grayscale value of each pixel; Indicates the first The arithmetic mean of the grayscale values ​​of all pixels within a sub-window. By the first The grayscale values ​​of all pixels within each sub-window are summed and then divided by . get.

[0057] After calculating the local texture variations in all M sub-windows, the arithmetic mean of the M local texture variations is calculated, and this arithmetic mean is used as the global texture roughness of the rock debris image. The global texture roughness is represented by the symbol... express, The calculation method is as follows: arrive Accumulate, then divide the accumulated value by .

[0058] Example 4

[0059] In specific implementation, please refer to Figure 3 Before inputting the particle number proportion and global texture roughness as input features into the pre-trained surrounding rock state classifier, the particle number proportion and global texture roughness are normalized to map all feature values ​​to the same numerical range. The input features for normalization include five components: the particle number proportion of blocky rock debris, the particle number proportion of flaky rock debris, the particle number proportion of powdery rock debris, the particle number proportion of elongated rock debris, and global texture roughness. Linear normalization is used, performing a normalization operation on each feature component separately, mapping the original value of each feature component to the [0,1] interval. Specifically, for each feature component, the minimum and maximum values ​​are obtained. These values ​​are extracted from historical statistical records. If no historical statistical records exist, the global minimum and global maximum values ​​of the corresponding feature components for all rock debris images within the current image acquisition period are used. For an original value in a feature component, the normalized value is equal to the original value minus the minimum value, then divided by the difference between the maximum and minimum values. After normalization, the values ​​of each feature component of all input features are in the range [0,1].

[0060] After receiving normalized input features, the surrounding rock state classifier calculates the probability values ​​of whether the current surrounding rock state belongs to intact rock, jointed rock, fractured rock, or extremely fractured rock. The classifier employs a decision tree model, which contains multiple internal nodes and multiple leaf nodes. Each internal node corresponds to a feature splitting condition, set based on a comparison between a feature component of the input features and a threshold. The splitting condition takes the form: the feature component is less than or equal to the threshold. Leaf nodes do not contain splitting conditions; they output the probability values ​​of whether the surrounding rock state belongs to one of the four rock state categories.

[0061] The decision tree model is constructed through offline training. The training dataset consists of multiple historically collected samples. Each sample includes five input feature components and a manually labeled rock condition category label, which can be one of the following: intact rock, jointed rock, fractured rock, or extremely fractured rock. The decision tree model is trained using a classification and regression tree algorithm, with the Gini coefficient used as a measure of impurity for node partitioning. A binary decision tree is recursively constructed starting from the root node. At each internal node, the algorithm iterates through all possible thresholds for the five input feature components, calculates the Gini coefficient after partitioning the current node's sample set into left and right subsets based on the threshold, and selects the feature component that minimizes the Gini coefficient and the threshold as the feature partitioning condition. The formula for calculating the Gini coefficient is:

[0062]

[0063] in, Represents the sample set The Gini coefficient, This represents the set of samples contained in the current node. Represents the sample set The total number of samples in the sample, The index number representing the surrounding rock condition category, with values ​​ranging from 1 to 4. Corresponding to intact surrounding rock, Corresponding to the development of joints in the surrounding rock. Corresponding to fractured surrounding rock, Corresponding to extremely fractured surrounding rock. Represents the sample set The middle belongs to the first A subset of samples with similar surrounding rock conditions. Represents a subset of samples The number of samples in the sample.

[0064] The recursive construction process of the decision tree model terminates when preset stopping conditions are met. These conditions include: the number of samples in the current node is less than a preset minimum sample size parameter; all samples in the current node belong to the same surrounding rock state category; or the depth of the decision tree model reaches a preset maximum depth parameter. The preset minimum sample size parameter is set to 10, and the preset maximum depth parameter is set to 5. The reason for setting the preset minimum sample size parameter to 10 is that when the number of samples in a node is less than 10, the sample size is insufficient to reliably estimate the probability distribution of each category, and further division will lead to overfitting. The reason for setting the preset maximum depth parameter to 5 is that the input feature dimension is five-dimensional, and limiting the decision tree depth to 5 can avoid an overly complex model structure while ensuring that the model has sufficient nonlinear expressive power. After the initial generation of the decision tree is completed, a cost complexity pruning strategy is used to prune the decision tree. During the pruning process, an independent validation sample set is used to evaluate the error change before and after replacing the subtree with a leaf node, and the subtree with the best performance on the validation sample set is selected as the final decision tree model.

[0065] In the leaf nodes of the decision tree model, the probability value of the surrounding rock state belonging to each category is obtained by statistically analyzing the proportion of samples of each category in the training samples arriving at the leaf node. For any leaf node, let... This represents the set of training samples that have reached the leaf nodes. Represents the training sample set The number of samples, Represents the training sample set The middle belongs to the first The number of samples representing different rock conditions determines the leaf node's output rock condition belonging to the first category. The probability value of the class is .

[0066] During real-time perception, a trained decision tree model is used for inference. Starting from the root node of the decision tree model, the feature partitioning conditions corresponding to the root node are obtained, specifying specific input feature components and thresholds. The values ​​of the corresponding components in the normalized input features are compared with the thresholds. If the value is less than or equal to the threshold, the process moves to the left child node of the root node; if the value is greater than the threshold, the process moves to the right child node of the root node. At the new node, if the new node is an internal node, the above decision process is repeated; if the new node is a leaf node, the movement stops, and the leaf node outputs four probability values, corresponding to intact surrounding rock, jointed surrounding rock, fractured surrounding rock, and extremely fractured surrounding rock, respectively. The four probability values ​​are compared, and the surrounding rock state category with the highest probability value is output as the current TBM face surrounding rock state category.

[0067] Example 5

[0068] In specific implementation, please refer to Figure 4 After acquiring images of rock debris on the conveyor belt during TBM tunneling, illumination compensation processing was performed on the rock debris images. The illumination compensation processing adopted the single-scale Retinex algorithm based on Retinex theory. The single-scale Retinex algorithm treats the rock debris image as the product of the incident light component and the reflected light component. By estimating the incident light component in the rock debris image and separating it from the rock debris image, the reflected light component, which only reflects the reflection characteristics of the rock mass surface, is obtained to correct the brightness difference caused by uneven illumination.

[0069] The illumination compensation process is implemented as follows: the rock debris image is converted from the RGB color space to the logarithmic domain. The natural logarithm of the red, green, and blue channel values ​​of each pixel in the rock debris image is taken to obtain the logarithmic domain red, green, and blue channel values, respectively. A Gaussian wraparound function is then convolved with each channel value in the logarithmic domain to estimate the logarithmic domain incident light component. The expression for the Gaussian wraparound function is:

[0070]

[0071] in, This indicates that the Gaussian wraparound function has an Euclidean distance of 0.5 from the center pixel. The weight value at the position; This represents the Euclidean distance from a given pixel to the center pixel within the neighborhood. The unit is pixels. The value ranges from 0 to the maximum distance within the neighborhood radius; The scaling parameter represents the Gaussian wrap function. Set to 80. The basis for setting it to 80 is that when the scale parameter is on the order of 80 pixels, the neighborhood coverage of the Gaussian wrap function can effectively capture the low-frequency brightness changes caused by uneven illumination in the rock debris image, while preserving high-frequency detail information such as rock debris particle texture and edges, thus avoiding severe loss of detail due to an excessively small scale parameter or insufficient illumination compensation due to an excessively large scale parameter. The base of the natural logarithm is 2.71828. The original values ​​of each channel of the rock debris image are subtracted from the estimated incident light components in the logarithmic domain to obtain the logarithmic domain reflected light components. Then, the exponentiation of the logarithmic domain reflected light components is performed to obtain the illumination-compensated rock debris image.

[0072] Color space conversion was performed on the rock debris image after illumination compensation processing. The color space conversion transformed the rock debris image from the RGB color space to the HSV color space. The HSV color space consists of hue, saturation, and lightness components. The conversion from RGB to HSV involved extracting the red, green, and blue channel values ​​for each pixel in the rock debris image. The difference between the maximum and minimum values ​​of these three channels was used as the chromaticity range value. The hue component was calculated based on six cases: whether the chromaticity range value was zero, and the comparison between the maximum value and the red, green, and blue channel values. The saturation component was calculated as follows: when the maximum value was zero, the saturation component was set to zero; when the maximum value was greater than zero, the saturation component was equal to the chromaticity range value divided by the maximum value. The lightness component was calculated as the maximum value among the red, green, and blue channel values. The luminance component in the HSV color space is extracted and used as the basis for subsequent image processing steps. Specifically, the luminance component is subjected to all operations, including conversion to grayscale image, adaptive threshold segmentation, morphological opening operation, connected component labeling, candidate rock debris particle region screening, contour extraction, morphological parameter calculation, rock debris type clustering, particle number ratio statistics, and global texture roughness calculation.

[0073] A real-time sensing system for the surrounding rock condition of a TBM based on rock debris images includes a memory, a processor, and a computer program stored in the memory and executable on the processor. The memory is a high-speed random access memory or non-volatile memory, and the processor is a central processing unit or a graphics processing unit. When the processor executes the computer program, it performs the following steps: acquiring images of rock debris on the conveyor belt during TBM tunneling; performing illumination compensation processing on the rock debris images to correct brightness differences caused by uneven illumination; performing color space conversion on the illumination-compensated rock debris images to extract the luminance component from the HSV color space of the rock debris images, and using the luminance component as the basis for subsequent image processing; converting the luminance component into a grayscale image, and using an adaptive threshold segmentation algorithm to binarize the grayscale image, generating a binary image separating the foreground and background regions of the rock debris; and performing morphological opening operations on the binary image to eliminate internal features within the foreground region of the rock debris. The noise and tiny connecting bridges separate the sticky rock debris particles; connected component labeling is performed on the binary image after morphological opening operation to identify each independent connected component as a candidate rock debris particle region; candidate rock debris particle regions with an area smaller than a preset area threshold are removed, and the boundaries of the remaining candidate rock debris particle regions are used as the contours of individual rock debris particles; the projected area, perimeter, and major and minor axis lengths of the equivalent ellipse of an individual rock debris particle are calculated based on the contour; the roundness of an individual rock debris particle is calculated based on the projected area and perimeter, where roundness is the ratio of the square of the perimeter to four times pi of the projected area; the roundness of an individual rock debris particle is calculated based on the major and minor axis lengths. The aspect ratio is calculated as the ratio of the major axis length to the minor axis length. The convexity of a single rock debris particle is calculated based on its projected area and the convex hull area of ​​its contour; the convexity is the ratio of the projected area to the convex hull area. The projected area, roundness, aspect ratio, and convexity of each individual rock debris particle are combined into a four-dimensional feature vector. Four cluster centers are defined, each corresponding to a type of rock debris, including blocky rock debris, flaky rock debris, powdery rock debris, and elongated rock debris. The Euclidean distance from each four-dimensional feature vector to each cluster center is calculated, and the rock debris particles are assigned to the rock debris type corresponding to the cluster center with the smallest Euclidean distance. After each partitioning, the number of particles assigned to each rock debris type is recalculated. The mean of the four-dimensional feature vector is used as the new cluster center, and the division process is repeated until the cluster centers no longer change. The total number of all individual rock debris particles in the current image acquisition period is counted. The number of individual rock debris particles belonging to blocky rock debris, flaky rock debris, powdery rock debris, and elongated rock debris is counted separately. The number of individual rock debris particles belonging to each type of rock debris is divided by the total number to obtain the proportion of each type of rock debris particle. The rock debris image is divided into multiple sub-windows of the same size. The standard deviation of the pixel gray value in each sub-window is calculated as the local texture change. The average of the local texture changes of all sub-windows is calculated as the global texture roughness.The particle count ratio and global texture roughness are normalized and mapped to the same numerical range. These normalized particle count ratios and global texture roughness are then used as input features and fed into a pre-trained decision tree model-based rock condition classifier. The decision tree model-based rock condition classifier calculates the probability values ​​of the current rock condition as intact, jointed, fractured, or extremely fractured based on the input features. The rock condition category with the highest probability value is output as the current TBM face rock condition category.

[0074] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for real-time sensing of TBM surrounding rock condition based on rock debris images, characterized in that, Includes the following steps: Acquire images of rock debris on the conveyor belt during TBM tunneling; The boundary of the rock debris particles is identified in the rock debris image, the contour of a single rock debris particle is extracted, and multiple morphological parameters of the single rock debris particle are calculated based on the contour. The individual rock debris particles are clustered based on the multiple morphological parameters to classify the rock debris types to which the rock debris particles belong. The percentage of particles of each type of rock debris within the current image acquisition period is statistically analyzed, and the global texture roughness of the rock debris image is calculated. The particle quantity ratio and the global texture roughness are used as input features and fed into a pre-trained surrounding rock state classifier to output the current state category of the surrounding rock at the TBM face.

2. The real-time sensing method for TBM surrounding rock condition based on rock debris images according to claim 1, characterized in that, The steps of identifying the boundaries of rock debris particles and extracting the contours of individual rock debris particles from the rock debris image specifically include: The rock debris image is converted into a grayscale image, and an adaptive threshold segmentation algorithm is used to binarize the grayscale image to generate a binary image in which the foreground and background regions of the rock debris are separated. Morphological opening operations are performed on the binary image to eliminate noise and small connecting bridges within the foreground region of the rock debris, thereby separating the mutually adhering rock debris particles. Connected component labeling is performed on the binary image after morphological opening operation, and each independent connected component is identified as a candidate rock debris particle region; The candidate rock slag particle regions with an area smaller than a preset area threshold are removed, and the boundaries of the remaining candidate rock slag particle regions are used as the outline of the individual rock slag particle.

3. The real-time sensing method for TBM surrounding rock condition based on rock debris images according to claim 2, characterized in that, The preset area threshold is dynamically set based on the acquisition resolution of the rock slag image and the running speed of the conveyor belt.

4. The real-time sensing method for TBM surrounding rock condition based on rock debris images according to claim 2, characterized in that, The steps of calculating multiple morphological parameters of a single rock debris particle based on the profile specifically include: The projected area, perimeter, major axis length, and minor axis length of the equivalent ellipse of the individual rock debris particle are calculated based on the contour. The roundness of a single rock debris particle is calculated based on the projected area and the perimeter, where the roundness is the ratio of the square of the perimeter to four times pi of the projected area. The aspect ratio of a single rock slag particle is calculated based on the length of its major axis and the length of its minor axis, wherein the aspect ratio is the ratio of the length of its major axis to the length of its minor axis. The convexity of a single rock debris particle is calculated based on the projected area and the convex hull area of ​​the contour, where the convexity is the ratio of the projected area to the convex hull area.

5. The real-time sensing method for TBM surrounding rock condition based on rock debris images according to claim 4, characterized in that, The equivalent ellipse is an ellipse with the same second-order central moment as the individual rock slag particle. The lengths of the major axis and the minor axis are obtained by eigenvalue decomposition of the covariance matrix of the profile.

6. The real-time sensing method for TBM surrounding rock condition based on rock debris images according to claim 4, characterized in that, The step of clustering individual rock debris particles based on the multiple morphological parameters to classify the rock debris types to which the rock debris particles belong specifically includes: The projected area, roundness, aspect ratio, and convexity of each individual rock debris particle are combined to form the feature vector of that rock debris particle; Multiple cluster centers are set, and each cluster center corresponds to a type of rock debris, including blocky rock debris, flaky rock debris, powdery rock debris, and strip-shaped rock debris; Calculate the Euclidean distance from each feature vector to each cluster center, and classify the rock debris particles into the rock debris type corresponding to the cluster center with the smallest Euclidean distance; After each division, the mean of all feature vectors under each type of rock slag is recalculated, and the mean is used as the new cluster center. The division steps are repeated until the cluster center no longer changes.

7. The real-time sensing method for TBM surrounding rock condition based on rock debris images according to claim 6, characterized in that, In the clustering step, when the sum of the displacements of the cluster centers in two consecutive iterations is less than a preset displacement threshold, it is determined that the cluster centers no longer change and the iteration is terminated.

8. The real-time sensing method for TBM surrounding rock condition based on rock debris images according to claim 6, characterized in that, The step of statistically analyzing the proportion of particles of each type of rock debris within the current image acquisition period specifically includes: Count the total number of all the rock debris particles identified within the current image acquisition period; The number of rock debris particles belonging to blocky rock debris, flaky rock debris, powdery rock debris, and elongated rock debris were counted separately. The percentage of each type of rock debris particle is obtained by dividing the number of each type of rock debris particle by the total number.

9. The real-time sensing method for TBM surrounding rock condition based on rock debris images according to claim 1, characterized in that, The steps for calculating the global texture roughness of the rock slag image specifically include: The rock debris image is divided into multiple sub-windows of the same size; Calculate the standard deviation of pixel grayscale values ​​within each sub-window, and use it as the local texture variation of that sub-window; Calculate the average of the local texture variations of all the sub-windows, and use the average as the global texture roughness of the rock debris image.

10. A real-time sensing system for the surrounding rock condition of a TBM based on rock debris images, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the TBM surrounding rock state real-time perception method based on rock slag images as described in any one of claims 1 to 9.

Images