Method and system for determining surrounding rock loosening zone based on neural network and borehole peeping

Abstract

The present invention discloses a method and system for determining the loosening zone of surrounding rock based on neural network and borehole peeping, by obtaining the characteristic image of the rock layer in the borehole at the distribution position of the loosening zone of surrounding rock in the tunnel surrounding rock photographed by the borehole peeping instrument; cutting the characteristic image of the rock layer in the borehole into sub-images along the depth direction of the borehole to obtain all sub-images; inputting all sub-images into a trained fracture image rating model, scoring all sub-images by the fracture image rating model, and obtaining the scores of all sub-images; performing weighted moving average on the scores of all sub-images to obtain the scores of the degree of surrounding rock fragmentation; comparing the scores of the degree of surrounding rock fragmentation with the threshold value for determining the loosening zone to determine the depth range of the loosening zone. The present invention gives full play to the high efficiency and accuracy of computers in image processing; based on neural networks, the efficiency and accuracy of identifying the loosening zone of surrounding rock based on drill peeping are improved.

Description

Technical Field

[0001] The invention relates to the field of mining engineering and computer vision technology, and in particular discloses a method and system for determining a surrounding rock loosening zone based on a neural network and borehole peeping. Background Art

[0002] In the field of tunnel surrounding rock control, the stability of the surrounding rock is the key to ensuring the safety of underground structures such as tunnels and chambers. The surrounding rock loosening zone is a fracture zone formed in a certain depth area around the surrounding area due to the deformation and destruction of the surrounding rock caused by secondary stress after the excavation of the underground space. Accurately identifying and quantifying the range of the loosening zone is crucial for evaluating the stability of the surrounding rock and formulating reasonable support measures. The use of borehole peep can directly observe the anchor drilling or geological drilling, providing a basis for judging the depth distribution of the loosening zone of the rock formation around the tunnel. When the existing technology determines the depth of the surrounding rock loosening zone, the images taken by the borehole peep instrument often need to be manually judged and processed, and the cracks and joints in the image are observed and judged, which requires the judge to spend a lot of energy and time, and requires the judge to have a solid reserve of relevant knowledge. At the same time, the determination of the range of the loosening zone is somewhat subjective, and different technicians may have inconsistent judgments on the range of the loosening zone, which will affect subsequent engineering applications.

[0003] Therefore, the above-mentioned defects existing in the prior art in determining the depth of the loose zone of surrounding rock are technical problems that need to be solved urgently. Summary of the invention

[0004] The present invention provides a method and system for determining the loosening zone of surrounding rock based on neural network and borehole peeping, aiming to solve at least one defect existing in the above-mentioned prior art.

[0005] One aspect of the present invention relates to a method for determining a surrounding rock loosening zone based on a neural network and borehole peeping, comprising the following steps:

[0006] Obtaining a characteristic image of rock formations in the borehole at the distribution position of the surrounding rock loosening zone in the tunnel surrounding rock taken by a borehole peep instrument;

[0007] The characteristic image of the rock formation in the borehole is cut into sub-images along the depth direction of the borehole to obtain all the sub-images;

[0008] Input all sub-images into the trained crack image rating model, and use the crack image rating model to score all sub-images to obtain the scores of all sub-images;

[0009] The weighted moving average of all sub-image scores is performed to obtain the surrounding rock fragmentation degree score; the surrounding rock fragmentation degree score is compared with the loose zone determination threshold to determine the loose zone depth range.

[0010] Furthermore, the characteristic image of the rock formation in the borehole is cut into sub-images along the depth direction of the borehole, and the steps of obtaining all the sub-images include:

[0011] Inputting the rock formation feature image in the borehole into the image cutting and preservation model;

[0012] The image cutting and saving model reads the size of the characteristic image of the rock formation in the borehole, crops and cuts the characteristic image of the rock formation in the borehole along the depth direction of the borehole, cuts out a sub-image at a set length, and names all the cut sub-images in numerical order and saves them to the file path of the same cut image.

[0013] Furthermore, all sub-images are input into the trained crack image rating model, and the crack image rating model scores all sub-images. The steps of obtaining the scores of all sub-images include:

[0014] Input the file path storing the cut image into the crack image rating model;

[0015] The crack image rating model reads all sub-images in the file path storing the cut image in numerical order and scores all the sub-images.

[0016] Furthermore, a weighted moving average is performed on all sub-image scores to obtain a surrounding rock fragmentation degree score; the surrounding rock fragmentation degree score is compared with a loose zone determination threshold, and the steps of determining the loose zone depth range include:

[0017] The fracture image rating model arranges all sub-image scores from shallow to deep in the borehole and writes them into a data list;

[0018] The score data written in the data list is weighted moving averaged to obtain the surrounding rock fragmentation degree score, and the surrounding rock fragmentation degree score obtained is compared with the loose zone judgment threshold to give the specific range of the loose zone.

[0019] Furthermore, a weighted moving average is performed on the score data written into the data list to obtain a surrounding rock fragmentation degree score, and the surrounding rock fragmentation degree score obtained is compared with the loosening zone determination threshold. In the step of giving a specific range of the loosening zone, a weighted moving average is performed using a weight function, and the weight function is:

[0020] in, is the weight function, is the first weight parameter, corresponding to the first crack level; is the second weight parameter, corresponding to the second crack level; is the third weight parameter, corresponding to the third crack level; is the fourth weight parameter, corresponding to the fourth crack level; is the fifth weight parameter, corresponding to the fifth crack level; Refers to the score obtained by the i-th cut image;

[0021] The weighted moving average calculation formula is:

[0022] in, Refers to the weighted moving average of the data from item t to item t+3. Refers to the score obtained by the i-th cut image; is the weight function.

[0023] Another aspect of the present invention relates to a system for determining a surrounding rock loosening zone based on a neural network and borehole peeping, comprising:

[0024] The first acquisition module is used to acquire the rock layer characteristic image in the borehole at the distribution position of the surrounding rock loosening zone in the tunnel surrounding rock taken by the borehole peep instrument;

[0025] The second acquisition module is used to cut and crop the rock formation feature image in the borehole into sub-images along the depth direction of the borehole to obtain all the sub-images;

[0026] The third acquisition module is used to input all sub-images into the trained crack image rating model, and the crack image rating model scores all sub-images to obtain scores of all sub-images;

[0027] The judgment module is used to perform weighted moving average on all sub-image scores to obtain the surrounding rock fragmentation degree score; compare the surrounding rock fragmentation degree score with the loose zone judgment threshold to determine the loose zone depth range.

[0028] Furthermore, the second acquisition module includes:

[0029] A first input unit is used to input the characteristic image of the rock formation in the borehole into the image cutting and saving model;

[0030] The cutting unit is used to read the size of the rock formation characteristic image in the borehole by the image cutting and saving model, crop and cut the rock formation characteristic image in the borehole along the depth direction of the borehole, cut out a sub-image at a set length, and name all the cut sub-images in numerical order and save them to the file path of the same cut image.

[0031] Furthermore, the third acquisition module includes:

[0032] A second input unit is used to input the file path storing the cut image into the crack image rating model;

[0033] The evaluation unit is used to read all sub-images in the file path storing the cut images in numerical order by the crack image rating model, and score all the sub-images.

[0034] Furthermore, the determination module includes:

[0035] A writing unit, used for writing all sub-image scores in order from shallow to deep parts of the borehole into a data list by the fracture image rating model;

[0036] The determination unit is used to perform weighted moving average on the score data written in the data list to obtain a surrounding rock fragmentation degree score, compare the obtained surrounding rock fragmentation degree score with the loosening zone determination threshold, and provide a specific range of the loosening zone.

[0037] Furthermore, in the determination unit, a weighted moving average is performed using a weight function, and the weight function is:

[0038] in, is the weight function, is the first weight parameter, corresponding to the first crack level; is the second weight parameter, corresponding to the second crack level; is the third weight parameter, corresponding to the third crack level; is the fourth weight parameter, corresponding to the fourth crack level; is the fifth weight parameter, corresponding to the fifth crack level; Refers to the score obtained by the i-th cut image;

[0039] The weighted moving average calculation formula is:

[0040] in, Refers to the weighted moving average of the data from item t to item t+3. Refers to the score obtained by the i-th cut image; is the weight function.

[0041] The beneficial effects achieved by the present invention are:

[0042] The present invention provides a method and system for determining the loosening zone of surrounding rock based on neural network and borehole peeping, by obtaining the characteristic image of the rock layer in the borehole at the distribution position of the loosening zone of surrounding rock in the tunnel surrounding rock photographed by the borehole peeping instrument; cutting the characteristic image of the rock layer in the borehole into sub-images along the depth direction of the borehole to obtain all sub-images; inputting all sub-images into the trained fracture image rating model, scoring all sub-images by the fracture image rating model, and obtaining the scores of all sub-images; performing weighted moving average on the scores of all sub-images to obtain the scores of surrounding rock fragmentation degree; comparing the scores of surrounding rock fragmentation degree with the loosening zone determination threshold to determine the depth range of the loosening zone. The method and system for determining the loosening zone of surrounding rock based on neural network and borehole peeping provided by the present invention combine the two fields of mining engineering and computer vision, and give play to the efficiency and accuracy of computers in image processing; the VGG19 network structure is adjusted in the process of model training, and the model depth is improved; the model after training will more comprehensively identify the characteristics of fracture performance, and improve the efficiency and accuracy of identifying the loosening zone of surrounding rock based on drill peeping. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 The present invention is a flowchart of an embodiment of a method for determining a surrounding rock loosening zone based on a neural network and borehole peeping. DETAILED DESCRIPTION

[0044] In order to better understand the above technical solution, the above technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0045] like Figure 1 As shown, the first embodiment of the present invention proposes a method for determining the surrounding rock loosening zone based on a neural network and borehole peeping, comprising the following steps:

[0046] Step S100, obtaining a characteristic image of rock formations in a borehole at a location where a loose zone of surrounding rock is determined in the surrounding rock of a tunnel photographed by a borehole peep instrument.

[0047] In the tunnel, holes are drilled at locations where the distribution of loose zones of surrounding rocks needs to be determined, and the characteristics of the rock formations in the holes are photographed with a borehole peep instrument, and images of the characteristics of the rock formations in the holes are exported.

[0048] Step S200: Cut and cut the characteristic image of the rock formation in the borehole into sub-images along the depth direction of the borehole to obtain all the sub-images.

[0049] The characteristic image of the rock formation in the borehole is input into the image cutting and saving model, and a file path for storing the cut image is returned, and the number of sub-images is recorded.

[0050] Step S300: input all sub-images into the trained fissure image rating model, and use the fissure image rating model to score all sub-images to obtain scores for all sub-images.

[0051] The file path storing the cut image is input into the crack image rating model, and the crack image rating model will score all sub-images under the path and save the scores.

[0052] The training set used in the present embodiment to train the crack image rating model must contain images of drill holes that have completed crack type identification or loose zone depth identification. The images are cut into sub-images of specified sizes, and each sub-image must complete crack rating based on actual conditions, and then sorted and classified into several (e.g., 1-5) types and levels for the establishment of the training set.

[0053] The training of the crack image rating model involved in this embodiment is based on a convolutional neural network, such as the VGG19 network. The VGG19 network has 16 convolutional layers and 3 fully connected layers. By adjusting the network structure and increasing the number of convolutional layers, the depth of the model can be improved, and more nonlinear transformations can be introduced, so that the model can learn more complex feature representations, thereby extracting deeper feature content of the crack image, which can improve the accuracy of the model in crack recognition, and at the same time better cope with the scene of evaluating individual blurred and unsmooth images.

[0054] The fissure image rating model trained in this embodiment needs to be verified using a validation set to check whether the trained model has excellent accuracy.

[0055] Step S400: Perform weighted moving average on all sub-graph scores to obtain a surrounding rock fragmentation degree score; compare the surrounding rock fragmentation degree score with a loose zone determination threshold to determine a loose zone depth range.

[0056] The score is imported into the loose zone depth determination module to finally determine the loose zone depth range. Among them, the "loose zone determination threshold" refers to a parameter set according to the actual situation in combination with the pre-determined crack grade score, which is used to determine the loose zone depth.

[0057] Furthermore, the method for determining the depth of the surrounding rock loosening zone based on a neural network provided in this embodiment, step S200 includes:

[0058] Step S210: input the characteristic image of the rock formation in the borehole into the image cutting and saving model.

[0059] Import the image path taken by the drilling peep instrument into the image cutting and saving model.

[0060] Step S220, the image cutting and saving model reads the size of the characteristic image of the rock formation in the borehole, cuts the characteristic image of the rock formation in the borehole along the depth direction of the borehole, cuts out a sub-image at a set length, and names all the cut sub-images in numerical order and saves them to the file path of the same cut image.

[0061] The image cutting and saving model will read the size of the rock formation feature image in the borehole, cut and crop the rock formation feature image in the borehole along the direction of the borehole depth, cut out a sub-image every 10cm, start with the Arabic numeral "1", name them in sequence and save them to the same file path, until the rock formation feature image in the borehole is cut to a length of less than 10cm, that is, the segment of the image is regarded as the last sub-image. After cutting, all sub-images will be stored in the folder of all cut images (sub-images), and the image cutting and saving model will return to the folder path.

[0062] Preferably, the method for determining the depth of the surrounding rock loosening zone based on a neural network provided in this embodiment, step S300 includes:

[0063] Step S310: input the file path storing the cut image into the crack image rating model.

[0064] Step S320: The crack image rating model reads all sub-images in the file path storing the cut image in numerical order and scores all sub-images.

[0065] The fracture image rating model will automatically read the sub-image named "1" from the file path, and use the trained fracture image rating model to score all the sub-images currently input to the fracture image rating model and save the scores. The fracture image rating model will then return to the file path, read the next sub-image in the order of the image naming in the folder, and continue to score and save until all the sub-images in the file path are traversed.

[0066] Furthermore, the method for determining the depth of the surrounding rock loosening zone based on a neural network provided in this embodiment, step S400 includes:

[0067] Step S410: The fracture image rating model arranges all sub-image scores in order from shallow to deep in the borehole and writes them into a data list.

[0068] Write all sub-image scores obtained by the crack image rating model into the data list in order, starting from the first data.

[0069] Step S420: Perform weighted moving average on the score data written into the data list to obtain a surrounding rock fragmentation degree score, compare the obtained surrounding rock fragmentation degree score with the loosening zone determination threshold, and provide a specific range of the loosening zone.

[0070] The weighted moving average of the score data written into the data list is performed with a window size of 5. The obtained surrounding rock fragmentation degree score is compared with the loose zone judgment threshold to give the specific range of the loose zone.

[0071] In order to make the surrounding rock loose zone depth determination method applicable to the loose zone analysis of various geological conditions, after the fracture image rating model scores each sub-image, a weight function is used for weighted moving average. The weight parameter used can be adjusted according to the actual situation. For each "pre-set fracture level", there is a corresponding weight parameter. , if 5 fracture levels are set in advance, the weight parameter is expressed as , , , and In this embodiment, the weight function is:

[0072] (1)

[0073] In formula (1), is the weight function, is the first weight parameter, corresponding to the first crack level; is the second weight parameter, corresponding to the second crack level; is the third weight parameter, corresponding to the third crack level; is the fourth weight parameter, corresponding to the fourth crack level; is the fifth weight parameter, corresponding to the fifth crack level; Refers to the score of the i-th cut image. According to different types of surrounding rock tunnels (coal tunnels, rock tunnels) and the development law of surrounding rock cracks in this mining area, weights are set according to different levels. ~ It can be adjusted according to actual conditions.

[0074] The weighted moving average calculation formula is:

[0075] (2)

[0076] In formula (2), Refers to the weighted moving average of the data from item t to item t+3. Refers to the score obtained by the i-th cut image; is the weight function. Combined with the general law of the depth distribution of the loose zone, if there is a loose zone in the entire range of the borehole, it is often close to the tunnel.

[0077] Specifically, starting from the surrounding rock fragmentation score obtained by the first weighted moving average, the surrounding rock fragmentation score is compared with the loosening zone determination threshold. If the surrounding rock fragmentation score is greater than the loosening zone determination threshold, the next data is compared. When the surrounding rock fragmentation score obtained by the nth weighted moving average is less than the loosening zone determination threshold, all sub-graphs participating in the nth weighted moving average calculation at this time and all sub-graphs participating in the weighted moving average calculation before can be regarded as part of the loosening zone, that is, the depth corresponding to the n+3th sub-graph is: 0.1×(n+3) meters.

[0078] Please see Figure 1 On the other hand, the present invention relates to a system for determining a loose zone of surrounding rock based on a neural network and borehole peeping, comprising a first acquisition module, a second acquisition module, a third acquisition module and a determination module, wherein the first acquisition module is used to acquire a characteristic image of the rock strata in the borehole at the distribution position of the loose zone of the surrounding rock of the tunnel photographed by the borehole peeping instrument; the second acquisition module is used to cut the characteristic image of the rock strata in the borehole into sub-images along the depth direction of the borehole to acquire all sub-images; the third acquisition module is used to input all sub-images into a trained fracture image rating model, and the fracture image rating model scores all sub-images to acquire scores of all sub-images; the determination module is used to perform a weighted moving average on the scores of all sub-images to acquire a score of the degree of surrounding rock fragmentation; the score of the degree of surrounding rock fragmentation is compared with the determination threshold of the loose zone to determine the depth range of the loose zone.

[0079] A hole is drilled at a position in the tunnel where the distribution of the loosened zone of the surrounding rock needs to be determined, and the rock characteristics in the hole are photographed with a borehole peep instrument. The first acquisition module obtains the rock characteristics image in the hole at the position where the distribution of the loosened zone of the surrounding rock is determined in the tunnel.

[0080] The second acquisition module acquires the rock layer characteristic image in the borehole at the distribution position of the surrounding rock loosening zone in the tunnel surrounding rock taken by the borehole peep instrument.

[0081] The third acquisition module inputs the file path storing the cut image into the crack image rating model, and the crack image rating model scores all sub-images under the path and saves the scores.

[0082] The determination module imports the score into the loose zone depth determination module, and finally determines the loose zone depth range. Among them, the "loose zone determination threshold" refers to a parameter set according to the actual situation in combination with the pre-established crack grade score, which is used to determine the loose zone depth.

[0083] Further, see Figure 1In the surrounding rock loosening zone determination system based on neural network and borehole peeping provided in this embodiment, the second acquisition module includes a first input unit and a cutting unit, wherein the first input unit is used to input the characteristic image of the rock formation in the borehole into the image cutting and saving model. The cutting unit is used to read the size of the characteristic image of the rock formation in the borehole from the image cutting and saving model, cut the characteristic image of the rock formation in the borehole along the direction of the borehole depth, cut out a sub-image at a set length, and name all the cut sub-images in numerical order and save them to the file path of the same cut image.

[0084] The first input unit imports the image path captured by the borehole peephole into the image cutting and saving model.

[0085] The cutting unit uses the image cutting and saving model to read the size of the characteristic image of the rock formation in the borehole, cuts the characteristic image of the rock formation in the borehole along the direction of the borehole depth, cuts out a sub-image every 10 cm, and names them in sequence starting from the Arabic numeral "1" and saves them to the same file path until the characteristic image of the rock formation in the borehole is cut to a length of less than 10 cm, that is, the segment of the image is regarded as the last sub-image. After cutting, all sub-images will be stored in the folder of all cut images (sub-images), and the image cutting and saving model will return to the folder path.

[0086] Preferably, see Figure 1 In the system for determining the loosening zone of surrounding rock based on neural network and borehole peeping provided in this embodiment, the third acquisition module includes a second input unit and an evaluation unit, wherein the second input unit is used to input the file path storing the cut image into the fracture image rating model. The evaluation unit is used to read all sub-images in the file path storing the cut image in numerical order by the fracture image rating model, and score all the sub-images.

[0087] The evaluation unit will automatically read the sub-image named "1" from the file path using the crack image rating model, and use the trained crack image rating model to score all the sub-images currently input into the crack image rating model and save the scores. The crack image rating model will then return to the file path, read the next sub-image in the order of the image naming in the folder, and continue to score and save until all the sub-images under the file path are traversed.

[0088] Further, see Figure 1The system for determining the loosening zone of surrounding rock based on neural network and borehole peeping provided in this embodiment includes a writing unit and a judging unit, wherein the writing unit is used to write all sub-image scores from shallow to deep parts of the borehole in a data list in sequence according to the fracture image rating model. The judging unit is used to perform weighted moving average on the score data in the written data list to obtain a score of the degree of surrounding rock fragmentation, compare the obtained score of the degree of surrounding rock fragmentation with the judgment threshold of the loosening zone, and give a specific range of the loosening zone.

[0089] The writing unit writes all sub-image scores obtained by the crack image rating model into the data list in order, starting from the first data.

[0090] The determination unit performs weighted moving average on the score data written into the data list, with a window size of 5, compares the obtained surrounding rock fragmentation degree score with the loose zone determination threshold, and gives the specific range of the loose zone.

[0091] In the judgment unit, a weight function is used for weighted moving average, and the weight function is:

[0092] (3)

[0093] In formula (3), is the weight function, is the first weight parameter, corresponding to the first crack level; is the second weight parameter, corresponding to the second crack level; is the third weight parameter, corresponding to the third crack level; is the fourth weight parameter, corresponding to the fourth crack level; is the fifth weight parameter, corresponding to the fifth crack level; Refers to the score of the i-th cut image. According to different types of surrounding rock tunnels (coal tunnels, rock tunnels) and the development law of surrounding rock cracks in this mining area, weights are set according to different levels. ~ It can be adjusted according to actual conditions.

[0094] The weighted moving average calculation formula is:

[0095] (4)

[0096] In formula (4), Refers to the weighted moving average of the data from item t to item t+3. Refers to the score obtained by the i-th cut image; is the weight function.

[0097] Specifically, the determination unit starts with the result of the first weighted moving average and compares it with the loose circle determination threshold. If the value is greater than the loose circle determination threshold, the next data is compared. When the nth weighted moving average result is less than the loose circle determination threshold, all sub-graphs participating in the nth weighted moving average calculation at this time and all sub-graphs participating in the weighted moving average calculation before can be regarded as loose circle parts, that is, the depth corresponding to the n+3th sub-graph is: 0.1×(n+3) meters.

[0098] The method and system for determining the surrounding rock loosening zone based on neural network and borehole peeping provided in this embodiment have the following working principles:

[0099] 1. Network construction and model training

[0100] (1) Training set establishment

[0101] In order to train a high-quality fracture image rating model, it is necessary to divide the fracture levels in advance according to the actual situation, such as level 1 to level 5, create folders corresponding to each level with Arabic numerals, and save them in the training set folder path named train.

[0102] The training set data can use a large number of images that have completed crack classification. The image of each borehole is cut along the direction of the borehole, and a sub-image is saved every 10 cm. The sub-images are manually graded and classified into folders of corresponding levels.

[0103] In the training set, no less than 10,000 crack images of each level and type should be prepared to ensure the effectiveness of model training. The scale of the training set data directly affects the accuracy of model recognition. The scale of the training set data should be increased as much as possible to improve the model quality.

[0104] (2) Neural network construction and training of crack image rating model

[0105] The training of the crack image rating model is based on the VGG19 network. In order to deepen the depth of the crack image rating model and enable it to distinguish deeper features of the crack image, the network structure is modified and a convolutional pooling structure is added.

[0106] The first convolution-pooling structure has two convolutional layers and one maximum pooling layer.

[0107] Convolution layer: The number of convolution kernels (filters) is 64, the convolution kernel size (kernel_size) is 3×3, the edge padding (padding) is the same, and the activation function (activation) uses ReLU.

[0108] The maximum pooling layer (MaxPooling2D) has a pooling window (pool_size) size of 2×2 and a sliding step size (strides) of 2×2.

[0109] The second convolution-pooling structure has two convolutional layers and one maximum pooling layer.

[0110] Convolution layer: The number of convolution kernels (filters) is 128, the convolution kernel size (kernel_size) is 3×3, the edge padding (padding) is the same, and the activation function (activation) uses ReLU.

[0111] The maximum pooling layer (MaxPooling2D) has a pooling window (pool_size) size of 2×2 and a sliding step size (strides) of 2×2.

[0112] The third convolution-pooling structure has four convolutional layers and one maximum pooling layer.

[0113] Convolution layer: The number of convolution kernels (filters) is 256, the convolution kernel size (kernel_size) is 3×3, the edge padding (padding) is the same, and the activation function (activation) uses ReLU.

[0114] The maximum pooling layer (MaxPooling2D) has a pooling window (pool_size) size of 2×2 and a sliding step size (strides) of 2×2.

[0115] The fourth convolution-pooling structure has four convolutional layers and one maximum pooling layer.

[0116] Convolutional layer: The number of convolution kernels (filters) is 512, the convolution kernel size (kernel_size) is 3×3, the edge padding (padding) is the same, and the activation function (activation) uses ReLU.

[0117] The maximum pooling layer (MaxPooling2D) has a pooling window (pool_size) size of 2×2 and a sliding step size (strides) of 2×2.

[0118] The fifth convolution-pooling structure has four convolutional layers and one maximum pooling layer.

[0119] Convolutional layer: The number of convolution kernels (filters) is 512, the convolution kernel size (kernel_size) is 3×3, the edge padding (padding) is the same, and the activation function (activation) uses ReLU.

[0120] The maximum pooling layer (MaxPooling2D) has a pooling window (pool_size) size of 2×2 and a sliding step size (strides) of 2×2.

[0121] Add the sixth convolution-pooling structure, which has two convolutional layers and one maximum pooling layer.

[0122] Convolutional layer: The number of convolution kernels (filters) is 512, the convolution kernel size (kernel_size) is 3×3, the edge padding (padding) is the same, and the activation function (activation) uses ReLU.

[0123] The maximum pooling layer (MaxPooling2D) has a pooling window (pool_size) size of 2×2 and a sliding step size (strides) of 2×2.

[0124] The full link layer has three layers:

[0125] The first layer contains 4096 nodes and uses the ReLU activation function. This layer receives the feature map from the convolutional layer and converts it into a fully connected form for subsequent classification tasks.

[0126] The second layer contains 4096 nodes and uses the ReLU activation function. This layer further processes the features from the previous layer and extracts higher-level abstract information.

[0127] The third layer contains 5 nodes (corresponding to the 5 fissure levels divided in the embodiment), and this layer outputs the predicted probability of each category, which is usually followed by a softmax activation function for multi-classification processing to complete the multi-classification project of fissure image rating.

[0128] After the neural network is built, the training set path is configured to train the crack image rating model.

[0129] (3) Validation set establishment

[0130] The images taken by the borehole peephole should be stored in the verification set, and these images do not need to be cut.

[0131] 2. Image cutting and saving module construction

[0132] Use the open cv library to write functions and complete the construction of the image cutting and saving module.

[0133] Set the image reading path → read the complete drilling image in the path → identify and obtain the image size → set the sub-image size (the height is along the drilling depth direction, the sub-image height is limited to 10cm, and the width is not required) → calculate the number of sub-images → set the counter → enter the loop with the number of sub-images as the upper limit → cut the image → calculate the sub-image coordinates → extract the sub-image → save the sub-image → the counter determines whether to continue the loop → return to the sub-image storage path after the loop is completed.

[0134] 3. Construction of loose circle depth determination module

[0135] Use the numpy library to write functions and complete the construction of the loose circle depth determination module.

[0136] Write all sub-image scores obtained by the crack image rating model into the score array, set the weight parameter array size to 5, and define the size of each weight value in the array. Define a loop, in which the data in the score array will be weighted moving averaged (starting from the first data, the current data and the following 4 data are weighted averaged, and then the second data in the array is located after the end, until the remaining data is less than 5, and the loop ends), and compare the calculated data with the loose zone judgment threshold to determine the specific depth of the loose zone.

[0137] 4. Method for determining the surrounding rock loosening zone and system quality verification

[0138] After the training of the crack image rating model is completed, the image data of the verification set is called using the image cutting and saving module. After the verification set images are cut, named, and saved, the file path storing the cut images is imported into the trained crack image rating model, and each sub-image that passes through the model is scored. The scores are imported into the loose circle depth determination module to determine the loose circle depth and verify the quality of the model.

[0139] The method and system for determining the loosening zone of surrounding rock based on neural network and borehole peeping provided in this embodiment are obtained by obtaining the characteristic image of the rock layer in the borehole at the distribution position of the loosening zone of surrounding rock in the tunnel surrounding rock photographed by the borehole peeping instrument; the characteristic image of the rock layer in the borehole is cut into sub-images along the depth direction of the borehole to obtain all sub-images; all sub-images are input into the trained fracture image rating model, and the fracture image rating model scores all sub-images to obtain the scores of all sub-images; the weighted moving average of all sub-image scores is performed to obtain the score of the degree of surrounding rock fragmentation; the score of the degree of surrounding rock fragmentation is compared with the threshold for determining the loosening zone to determine the depth range of the loosening zone. The method and system for determining the loosening zone of surrounding rock based on neural network and borehole peeping provided in this embodiment combine the two fields of mining engineering and computer vision, and give full play to the efficiency and accuracy of computers in image processing; the VGG19 network structure is adjusted during the model training process to improve the model depth; the model after training will recognize the characteristics of fracture performance more comprehensively, and improve the efficiency of borehole fracture image recognition.

[0140] Although preferred embodiments of the present invention have been described, additional changes and modifications may be made to these embodiments by those skilled in the art once the basic inventive concepts are known. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention. Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A method for determining the loose zone of surrounding rock based on neural network and borehole peeping, characterized in that: The following steps are involved: Obtaining a characteristic image of rock formations in the borehole at the distribution position of the surrounding rock loosening zone in the tunnel surrounding rock taken by a borehole peep instrument; Cutting the characteristic image of the rock formation in the borehole into sub-images along the depth direction of the borehole to obtain all the sub-images; Inputting all the sub-images into a trained fissure image rating model, and using the fissure image rating model to score all the sub-images to obtain scores for all the sub-images; Performing weighted moving average on all sub-image scores to obtain a surrounding rock fragmentation degree score; Compare the surrounding rock fragmentation degree score with the loose zone determination threshold to determine the depth range of the loose zone; Performing weighted moving average on all sub-image scores to obtain a surrounding rock fragmentation degree score; The steps of comparing the surrounding rock fragmentation degree score with the loose zone determination threshold and determining the loose zone depth range include: The fracture image rating model arranges all sub-image scores in order from shallow to deep in the borehole and writes them into a data list; Performing weighted moving average on the score data written into the input data list to obtain a surrounding rock fragmentation degree score, comparing the obtained surrounding rock fragmentation degree score with a loosening zone determination threshold, and providing a specific range of the loosening zone; The score data written into the input data list is weighted moving averaged to obtain a surrounding rock fragmentation degree score, and the surrounding rock fragmentation degree score obtained is compared with the loosening zone determination threshold value to give a specific range of the loosening zone. In the step of performing weighted moving average using a weight function, the weight function is: in, is the weight function, is the first weight parameter, corresponding to the first crack level; is the second weight parameter, used to correspond to the second crack level; is the third weight parameter, corresponding to the third crack level; is the fourth weight parameter, corresponding to the fourth crack level; is the fifth weight parameter, corresponding to the fifth crack level; Refers to the score obtained by the i-th cut image; The weighted moving average calculation formula is: in, Refers to the weighted moving average of the data from item t to item t+3. Refers to the score obtained by the i-th cut image; is the weight function.

2. The method for determining the surrounding rock loosening zone based on neural network and borehole peeping as claimed in claim 1 is characterized in that: The rock formation characteristic image in the borehole is cut into sub-images along the depth direction of the borehole, and the steps of obtaining all sub-images include: Inputting the rock formation feature image in the borehole into the image cutting and preservation model; The image cutting and saving model reads the characteristic image of the rock strata in the borehole, crops and cuts the characteristic image of the rock strata in the borehole along the depth direction of the borehole, cuts out a sub-image at a set length, and names all the cut sub-images in numerical order and saves them to the file path of the same cut image.

3. The method for determining the surrounding rock loosening zone based on neural network and borehole peeping as claimed in claim 2 is characterized in that: Inputting all the sub-images into a trained fissure image rating model, and using the fissure image rating model to score all the sub-images, the step of obtaining the scores of all the sub-images includes: Input the file path storing the cut image into the crack image rating model; The crack image rating model reads all sub-images in the file path storing the cut images in numerical order, and scores all the sub-images.

4. A system for determining the loosening zone of surrounding rock based on neural network and borehole peeping, characterized in that: include: The first acquisition module is used to acquire the rock layer characteristic image in the borehole at the distribution position of the surrounding rock loosening zone in the tunnel surrounding rock taken by the borehole peep instrument; A second acquisition module is used to cut and crop the rock formation feature image in the borehole into sub-images along the depth direction of the borehole to acquire all the sub-images; A third acquisition module is used to input all the sub-images into a trained fissure image rating model, and the fissure image rating model scores all the sub-images to obtain scores of all the sub-images; A determination module, used for performing weighted moving average on all sub-image scores to obtain a surrounding rock fragmentation degree score; Compare the surrounding rock fragmentation degree score with the loose zone determination threshold to determine the depth range of the loose zone; The determination module comprises: A writing unit, configured to write all sub-image scores in a data list in order from shallow to deep parts of the borehole by the fracture image rating model; A determination unit, configured to perform a weighted moving average on the score data written into the input data list to obtain a surrounding rock fragmentation degree score, compare the obtained surrounding rock fragmentation degree score with a loosening zone determination threshold, and provide a specific range of the loosening zone; In the determination unit, a weighted moving average is performed using a weight function, and the weight function is: in, is the weight function, is the first weight parameter, corresponding to the first crack level; is the second weight parameter, used to correspond to the second crack level; is the third weight parameter, corresponding to the third crack level; is the fourth weight parameter, corresponding to the fourth crack level; is the fifth weight parameter, corresponding to the fifth crack level; Refers to the score obtained by the i-th cut image; The weighted moving average calculation formula is: in, Refers to the weighted moving average of the data from item t to item t+3. Refers to the score obtained by the i-th cut image; is the weight function.

5. The system for determining the surrounding rock loosening zone based on neural network and borehole peeping as claimed in claim 4, characterized in that: The second acquisition module includes: A first input unit, used for inputting the rock formation feature image in the borehole into the image cutting and preservation model; The cutting unit is used to read the size of the rock stratum characteristic image in the borehole by the image cutting and saving model, crop and cut the rock stratum characteristic image in the borehole along the depth direction of the borehole, cut out a sub-image at intervals of a set length, and name all the cut sub-images in numerical order and save them to the file path of the same cut image.

6. The system for determining the loosening zone of surrounding rock based on neural network and borehole peeping as claimed in claim 5, characterized in that: The third acquisition module includes: A second input unit is used to input the file path storing the cut image into the crack image rating model; The evaluation unit is used to read all the sub-images in the file path storing the cut images in numerical order by the crack image rating model, and score all the sub-images.

Images