An intelligent identification method of surrounding rock cracks based on borehole peek

Through a fully convolutional neural network based on the attention mechanism, the problem of difficulty in identifying complex features in ore drilling fracture images is solved, and accurate and rapid identification of fractures in the surrounding rock of coal mine tunnels and underground engineering guidance is achieved.

CN117392548BActive Publication Date: 2024-11-26XIAN UNIV OF SCI & TECH +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202311596497.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-11-27
Publication Date
2024-11-26
Estimated Expiration
2043-11-27

AI Technical Summary

Technical Problem

Traditional crack detection algorithms perform poorly in ore drilling crack images, and cannot effectively identify complex structural features and noise interference, resulting in poor detection results.

Method used

A fully convolutional neural network based on the attention mechanism is used to obtain crack images through drilling and peek, preprocess and annotate, train the model to identify complex crack features, and enhance model performance using Resnet18 and CBAM attention mechanisms.

Benefits of technology

It realizes accurate and rapid identification of complex cracks in surrounding rocks of coal mine tunnels, can identify underground cracks and guide underground engineering practices, improving the accuracy and efficiency of detection.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117392548B_ABST
    Figure CN117392548B_ABST
Patent Text Reader

Abstract

The present invention discloses a method for intelligent identification of surrounding rock fissures based on borehole peeping, the steps of which include: obtaining a fissure image, which includes a sample image and a test image; constructing a fissure image recognition model by means of an improved fully convolutional neural network, and training the model with the sample image to obtain a well-trained fissure image recognition model; then, inputting the test image into the well-trained fissure image recognition model to obtain a fissure recognition result. The present invention is based on an improved fully convolutional neural network to accurately and quickly identify fissures in the surrounding rock of an actual tunnel.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the technical field of intelligent identification of surrounding rock fissures in coal mine tunnel boreholes, and in particular to an intelligent identification method of surrounding rock fissures based on borehole peeping. Background Art

[0002] Nowadays, computer vision technology is widely used in the field of crack detection, and mainly uses edge detection and threshold segmentation, which can accurately identify crack images with common and obvious features and show good results. However, the characteristics of mine borehole crack images are complex and changeable, such as complex and changeable background textures, numerous noise types, disordered crack distribution, abnormal and irregular surface shapes, and complex contact methods. These make traditional crack detection algorithms perform poorly in the detection of ore borehole crack images and cannot effectively detect cracks.

[0003] Therefore, a fully convolutional neural network based on the attention mechanism was proposed and used in the intelligent recognition method of cracks in the surrounding rock of coal mine tunnels. Summary of the invention

[0004] The purpose of the present invention is to provide an intelligent recognition method for drilling peek cracks in surrounding rock of coal mine tunnels based on a fully convolutional neural network with an attention mechanism, so as to solve the complex structural features of the drilling peek crack images that cannot be observed by the naked eye.

[0005] This innovation proposes a method for intelligently identifying surrounding rock cracks by drilling and peering, which mainly includes the following steps:

[0006] Acquire a crack image, wherein the crack image includes a sample image and a test image;

[0007] We built a fracture image recognition model based on an optimized fully convolutional neural network and trained it with sample images, ultimately obtaining a trained fracture image segmentation model.

[0008] The test image is imported into the trained crack image recognition model to obtain the crack recognition results.

[0009] We can also optionally preprocess these sample images before training the above crack recognition model. The preprocessing steps include:

[0010] Performing gama transformation on the sample image based on Python+OpenCV to obtain a first image;

[0011] Performing sharpening processing on the preprocessed sample image based on a sharpening operator to obtain a second image;

[0012] An erosion operation is performed on the sharpened image to obtain a preprocessed sample image.

[0013] Optionally, the preprocessed sample image is manually annotated based on labelme to obtain a label map.

[0014] Optionally, the process of training the crack image recognition model based on the sample image includes:

[0015] S1, initialize the crack image recognition model;

[0016] S2. Training the preprocessed sample image based on the initialized crack image recognition model to obtain a feature map;

[0017] S3, performing a deconvolution operation on the feature map to obtain a prediction map;

[0018] S4, calculating the error between the prediction image and the label image to obtain a pixel matching error;

[0019] S5, using a back propagation algorithm to iterate the pixel matching error so that the error is continuously reduced;

[0020] S6. Execute S2-S5 in a loop to obtain optimal model parameter values, and form a crack image recognition model based on the optimal model parameter values.

[0021] Optionally, the mentioned crack image recognition model is an optimized fully convolutional neural network.

[0022] Among them, the improved fully convolutional neural network uses Resnet18 as the front-end basic network.

[0023] Optionally, the improved fully convolutional neural network includes a CBAM attention mechanism.

[0024] Optionally, the process of processing based on the improved fully convolutional neural network includes:

[0025] Perform several convolution operations on the preprocessed sample image to obtain a low-resolution heat map;

[0026] Performing upsampling and up-pooling operations on the low-resolution heat map to obtain a high-resolution image;

[0027] The resolution size of the high-resolution image is the same as that of the original image.

[0028] The present invention has the following technical effects:

[0029] Based on the source of the surrounding rock fracture data set and fracture characteristics in deep learning, the present invention identifies the surrounding rock fractures from aspects such as fracture scenes and fracture background interferences.

[0030] The present invention constructs a crack image recognition model to accurately and quickly identify cracks in the surrounding rock of an actual tunnel, and guides underground engineering practice by identifying underground cracks. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Next, the diagrams required for the implementation examples will be basically interpreted to more clearly present the technical process of the present invention or the prior art. Obviously, the following diagrams are only for some implementation examples of the present invention. For general technical industry personnel, other diagrams can be imagined according to the existing diagrams without the need for innovative efforts.

[0032] Figure 1 It is a schematic flow chart of a method for intelligently identifying surrounding rock fractures based on borehole peeping in an embodiment of the present invention;

[0033] Figure 2 This is a structural diagram of an FCN model with an attention mechanism added and a schematic diagram of convolutional layer parameter settings in an embodiment of the present invention; DETAILED DESCRIPTION

[0034] Next, the technical planning in the examples will be explained in detail and thoroughly with the help of illustrations of the examples of this invention. Obviously, these examples are partial examples of this invention, but not all examples. Based on the examples of this invention, all other examples obtained by ordinary technicians in this field without innovative work are all within the protection of this invention.

[0035] Embodiment 1

[0036] The present invention takes the surrounding rock drilling peeping image data of coal mine tunnel as the research object, establishes the fracture division standard according to the quantitative information in the image, and comprehensively uses artificial intelligence deep learning technology, computer vision technology and other methods to propose an intelligent identification method of surrounding rock fractures based on drilling peeping. The steps are as follows:

[0037] like Figure 1 As shown, (1) based on the JL-IDOI (A) intelligent borehole television imager, the crack image of the tunnel surrounding rock is obtained. The crack image includes sample images and test images. The image resolution size is 288px×256px, which ensures the uniformity of image size and can collect images in weak light and high dust environment.

[0038] The study used an intelligent borehole television imager to batch acquire images of cracks on the tunnel roof and both sides.

[0039] (2) Research on crack image preprocessing and postprocessing

[0040] By using existing image preprocessing and postprocessing methods through Python+OpenCV, the sample image is subjected to gama transformation based on Python+OpenCV to obtain a first image; the image is subjected to sharpening processing using a sharpening operator to obtain a second image; the image is subjected to an erosion operation to obtain a preprocessed sample image. A labeling tool (labelme) is used for manual annotation to obtain a label map.

[0041] (3) Crack image recognition method based on fully convolutional neural network with attention mechanism

[0042] A model capable of identifying crack images is established and trained with sample images to obtain a model capable of identifying crack images. The specific implementation process includes the following:

[0043] according to Figure 2 In the demonstration, Resnet18 demonstrated the ability of CNN to handle complex models and achieved training results in a reasonable time using GPU, which promoted the progress of supervised deep learning.

[0044] The fully convolutional network (FCN) that uses Resnet18 as the basic front-end network will perform multi-level convolution processing, thereby reducing the clarity of the image. In order to restore the original image size to perform classification predictions for each pixel, we need to perform a reverse convolution on the heatmap, which is an upsampling process. This step involves an up-pooling process, followed by a deconvolution to improve the clarity of the image and return it to its original size. The difference between the deconvolved image and the label map is then calculated, and the total error in the network is minimized through the back-propagation mechanism.

[0045] As one of the dual methods in hybrid attention, CBAM can generate attention feature map information in both channel and spatial dimensions when processing feature maps. Then, these two types of feature map information are multiplied with the original input feature map to perform adaptive feature adjustment and finally generate the final feature map. Since the CBAM module is small, it can be inserted into any mainstream network to optimize its performance.

[0046] Generally, we use CNN to classify images, however, FCN has the function of performing pixel-level image classification, thus solving the problem of semantic segmentation of images. The main difference between FCN and typical CNN is that traditional CNN adds a fully connected layer after the convolution layer to generate a fixed-length feature vector for classification, while FCN can process input images of any size. It upsamples the feature map of the last convolution layer through the deconvolution layer and restores it to the size of the input image, which enables each pixel to generate a prediction and store the spatial information of the input image. Finally, each pixel is classified on the upsampled feature map. Fully convolutional neural networks mainly rely on the following three technologies:

[0047] (1) Implementing convolution operations: It can adapt to inputs of any size and produce low-resolution cut images;

[0048] Deconvolution operation: enlarge the low-resolution image and obtain the original image. Figure 1 Crop of the same resolution.

[0049] Skip structure (also known as skip layer or Skip Layer): This combines upsampling and upper-level convolutional pooling data to complete the missing detailed data to repair and restore the image.

[0050] There is no need to require the same size of training and test images. Fully convolutional neural networks (FCNs) can handle input images of various sizes. This approach avoids the problem of redundant storage and convolution calculation caused by using pixel blocks, thus improving efficiency.

[0051] Therefore, the intelligent identification of cracks in the surrounding rock of the tunnel based on drilling peeping has a relatively complex structure and needs to be divided in a complex way, so it can only be classified as a semantic segmentation problem. The fully convolutional network is specifically used for semantic segmentation of images, which can determine the location of the cracks. The study improves and optimizes the fully convolutional neural network (FCN). A ReLU layer and a maximum pooling layer are set after each convolution layer. The improved FCN network model adds the CBAM attention mechanism to enhance the model's perception of the global information of the surrounding rock cracks and the details of the crack edges. CBAM enhances the performance of FCN at a relatively low cost by giving different weights to each feature channel and generation space, thereby enhancing useful features and suppressing the continued transmission of useless features. The expression ability of FCN is improved by strengthening the ability to allocate the generation space and channel weights. The CBAM component is mainly formed by connecting two parts, CAM and SAM. CAM, namely the channel attention module, has the main function of giving different attention feature weights to different channels. The spatial attention module SAM generates the spatial attention feature weights, and finally concatenates them with the input feature map to obtain the final attention mechanism features and returns them to the improved FCN for training. Specifically, it focuses on the important features of the image and suppresses unnecessary regional influences. Use pytorch to build an FCN model with an added attention mechanism for training. Set it to use GPU, and set the batch according to the video memory. This method sets it to 16, the training round is 2000, and the initial learning rate is set to 0.0001.

[0052] After 2000 iterations, the pixel matching error tends to converge, and the optimal model parameter value is obtained and saved.

[0053] The process of training the crack image recognition model based on sample images includes:

[0054] S1, initialize the crack image recognition model;

[0055] The feature map obtained by S2 is the product of training the initial crack image recognition model on the given sample image.

[0056] S3, performing a deconvolution operation on the feature map to obtain a prediction map;

[0057] S4, calculating the error between the prediction image and the label image to obtain a pixel matching error;

[0058] S5, using a back propagation algorithm to iterate the pixel matching error so that the error is continuously reduced;

[0059] S6. Execute S2-S5 in a loop to obtain optimal model parameter values, and form a crack image recognition model based on the optimal model parameter values.

[0060] The present invention selects Pytorch as the learning framework and implements crack image segmentation based on deep learning and computer vision through programming.

[0061] The present invention focuses on the identification of cracks based on the research on the development map of the drilling peek image of the surrounding rock of the coal mine tunnel.

[0062] Based on the source of the surrounding rock fracture data set and fracture characteristics in deep learning, the present invention identifies the surrounding rock fractures from aspects such as fracture scenes and fracture background interferences.

[0063] The present invention constructs a crack image recognition model to accurately and quickly identify cracks in the surrounding rock of an actual tunnel, and guides underground engineering practice by identifying underground cracks.

[0064] The above content has summarized and explained the basic concept, core features and advantages of this invention. Technical experts in this field should understand that this invention is not limited by the examples listed above, and the examples and explanations listed above are only used to illustrate the concept of this invention. Without departing from the core and limitations of this invention, this invention may also undergo numerous adjustments and updates, which fall within the scope of protection of this invention to be protected. The scope of protection of this invention is defined by the attached claims and their equivalents.

Claims

1. An intelligent identification method of surrounding rock cracks based on borehole peek, It is characterized in that include: Acquire a crack image, wherein the crack image includes a sample image and a test image; An improved fully convolutional neural network is used to create a crack image recognition model, and the model is trained with sample images to obtain a trained crack image recognition model. Input the test image into the pre-trained crack image recognition model to obtain the crack recognition result; Based on the drilling peek, it also includes preprocessing the sample image before training the crack image recognition model. This preprocessing process includes preprocessing the sample image; Performing gama transformation on the sample image based on Python+OpenCV to obtain a first image; Performing sharpening processing on the first image based on a sharpening operator to obtain a second image; Performing an erosion operation on the second image to obtain a preprocessed sample image; Manually annotate the preprocessed sample image based on labelme to obtain a label map; The process of training the crack image recognition model based on the sample image includes: S1, initialize the crack image recognition model; S2. Extracting features from the preprocessed sample image based on the initialized crack image recognition model to obtain a feature map; S3, performing a deconvolution operation on the feature map to obtain a prediction map; S4, calculating the error between the prediction image and the label image to obtain a pixel matching error; S5, using a back propagation algorithm to iterate the pixel matching error so that the error is continuously reduced; S6, looping through S2-S5 to obtain optimal model parameter values, and forming a crack image recognition model based on the optimal model parameter values; The crack image recognition model is an improved fully convolutional neural network; The improved fully convolutional neural network uses Resnet18 as the front-end basic network; The improved fully convolutional neural network includes a CBAM attention mechanism; The process of processing the preprocessed sample image based on the improved fully convolutional neural network includes: Perform several convolution operations on the preprocessed sample image to obtain a low-resolution heat map; Performing upsampling and up-pooling operations on the low-resolution heat map to obtain a high-resolution image; The resolution of the high-resolution image is the same as that of the original image.