G-UNet-based drilling surrounding rock fracture detection and classification method

Abstract

The invention discloses a G-UNet-based drilling surrounding rock fracture detection and classification method, which comprises the steps of constructing a drilling surrounding rock fracture data set and performing data enhancement, and constructing a G-UNet network model which comprises a feature encoder module, a multi-dimensional feature adaptive aggregation module, a feature decoder module and an ECA channel feature enhancement module, the multi-dimensional feature self-adaptive aggregation module is connected with the feature encoder module and the feature decoder module, and has better feature extraction capability on fractures which are different in size and length and are difficult to distinguish from the background on the hole wall of the drilling surrounding rock through a plurality of hole convolution with different expansion rates and a plurality of pooling operations with different pooling kernel sizes; and inputting the data set into the established G-UNet network for training to obtain a training optimal weight, and realizing accurate segmentation detection and classification of the drilling surrounding rock fracture by using the optimal weight. According to the method, pixel-level segmentation can be carried out on the crack, automatic giving and classification are carried out according to the trend of the crack, the segmentation detection precision of the crack is high, and the effect is good.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent detection of cracks in surrounding rocks of tunnel boreholes, and in particular to a detection and classification method of cracks in surrounding rocks of tunnel boreholes based on G-UNet. Background Art

[0002] During tunnel development and underground engineering construction, observing the development of cracks in the unexcavated rock mass ahead is of great significance for safe construction. The rock mass itself has a complex structural surface, and during the tunnel excavation process, it will cause the destruction of the stress balance of the surrounding rock, resulting in more complex cracks. Blind excavation will lead to tunnel geological disasters, posing a great threat to the safety of construction. Therefore, it is necessary to accurately identify the development of cracks in the unexcavated rock mass ahead in order to make corresponding support measures and excavation methods.

[0003] The main means of conventional geotechnical engineering survey is the borehole observation method, which involves drilling and coring, conducting indoor experimental observations, and recording drilling conditions, rock properties, etc. This traditional survey method is time-consuming and laborious. With the development of machine vision and artificial intelligence, pixel-level segmentation of images has achieved object detection and recognition. At present, the traditional methods for crack detection in machine vision mainly include threshold segmentation method, region growing method, etc. However, these traditional image segmentation algorithms are greatly affected by the environment, do not have good robustness, and cannot actively detect cracks. Semantic segmentation based on deep learning can achieve autonomous segmentation of objects by learning the features of objects to be identified. Its robustness is higher than that of traditional methods. Therefore, methods based on deep learning and machine vision have been widely used in tunnel engineering. However, due to the complex situations of the connectivity, different widths and lengths, and similarities with the background area of ​​the cracks in the borehole surrounding rock wall, the segmentation effect of the traditional network model is generally poor. For example, finer cracks may not be detected. Therefore, a method that can accurately identify cracks in the borehole surrounding rock is urgently needed to be studied by those skilled in the art. Summary of the invention

[0004] In order to solve the above-mentioned problems, the present invention provides a method for detecting and classifying borehole surrounding rock cracks based on G-UNet. The method extracts and fuses global feature information and local feature information through a deep learning semantic segmentation network model, actively adapts to changes in the scale of crack image information, and can accurately detect and segment cracks that are connected, have different widths and lengths, and are similar to background areas.

[0005] The present invention provides a method for detecting and classifying cracks in surrounding rocks of boreholes based on G-UNet, comprising the following steps:

[0006] S1: Construct a dataset of borehole surrounding rock fractures, and divide them into three categories: transverse fractures, longitudinal fractures, and background according to the development characteristics of the fractures;

[0007] S2: Data enhancement processing is performed on the borehole surrounding rock fracture dataset;

[0008] S3: Build the G-UNet network model, including the feature encoder module, the multi-dimensional feature adaptive aggregation module, the feature decoder module, and the ECA channel feature enhancement module;

[0009] S4: Input the enhanced processed borehole surrounding rock fracture data set into the G-UNet network model for training to obtain the training model weight, verify the training model weight on the test set and save the optimal training model weight;

[0010] S5: Use the optimal training model weights to achieve pixel-level segmentation detection and classification of borehole surrounding rock cracks.

[0011] Optionally, the cracks are divided into three categories according to their development characteristics: transverse cracks, longitudinal cracks, and background. The cracks whose angle with the central axis of the borehole is greater than 60 degrees are marked as transverse cracks; the cracks whose angle with the central axis of the borehole is less than or equal to 60 degrees are marked as longitudinal cracks.

[0012] Optionally, the data enhancement processing is performed on the borehole surrounding rock fracture data set, and the get_random_data function is called to randomly process the original image of the borehole surrounding rock fracture and the corresponding label image in the data set, including: scaling, cropping, color gamut transformation, and horizontal flipping. When the color gamut is transformed, only the original image is transformed, and the label image is not transformed.

[0013] Optionally, the G-UNet network model includes a feature encoder module, a multi-dimensional feature adaptive aggregation module, a feature decoder module, and an ECA channel feature enhancement module, and specifically includes the following steps:

[0014] a: The feature decoder module consists of the pooling layer and convolutional layer before the fully connected layer of the VGG16 network;

[0015] b: Multi-dimensional feature adaptive aggregation module connects the feature encoder module and the feature decoder module;

[0016] c: An ECA channel feature enhancement module is inserted before each upsampling operation in the feature decoder module, and an ECA channel feature enhancement module is inserted at the jump connection between the feature encoder module and the feature decoder module.

[0017] Optionally, the multi-dimensional feature adaptive aggregation module consists of three sub-modules: a hole feature capture sub-module DFC, a global feature enhancement sub-module GFE, and a multi-scale feature aggregation sub-module MSFA. The three sub-modules are combined as follows: the hole feature capture sub-module is connected in parallel with the global feature enhancement sub-module, the feature maps output from the two sub-modules are spliced ​​together through concat, and then a 1×1 convolution is performed to restore the number of channels, and finally the multi-scale feature aggregation sub-module is connected in series.

[0018] Optionally, the void feature capture submodule includes multiple void convolutions with different expansion rates, and the expansion rates are r=1, r=3, r=5, and r=7 respectively. The borehole surrounding rock fracture feature map input to the void feature capture submodule is respectively extracted with void convolutions with different expansion rates to extract feature information of different scales. The output feature map is spliced ​​together using concat to fuse feature information of different scales together.

[0019] Optionally, the global feature enhancement submodule extracts global feature information of the surrounding rock fracture feature map and learns difference information between the fracture and the rock mass pore wall. The module contains two branches, specifically including the following steps:

[0020] A: The first branch inputs the borehole surrounding rock fracture feature map, which is first subjected to global average pooling, then 1×1 convolution, and finally the feature map size is restored by upsampling;

[0021] B: The other branch takes the borehole surrounding rock fracture feature map as input and directly splices it with the output feature map of the first branch through residual connection.

[0022] Optionally, the multi-scale feature aggregation submodule is composed of multiple pooling operations and residual connections. The borehole surrounding rock crack feature map input to the module is subjected to pooling operations with different pooling kernels respectively. The pooling operations with different pooling kernels can adapt to the sizes of different cracks. The output feature maps after different pooling will be subjected to 1×1 convolution respectively, and the number of channels will be adjusted to 1. Then, the feature map size will be restored through upsampling. At the same time, the input borehole surrounding rock crack feature map will be directly aggregated with all feature maps output from the pooling operation through residual connections. The feature map finally output from the module is also the feature map output from the multi-dimensional feature adaptive aggregation module. Compared with the feature map input to the multi-dimensional feature adaptive aggregation module, the number of channels of the output feature map is increased by four, and the size remains unchanged.

[0023] Compared with the prior art, the present invention has the following beneficial effects:

[0024] (1) The present invention can effectively capture contextual information of different scales, pay attention to local details and global semantic features, learn the differences between cracks and backgrounds in images, and greatly improve the segmentation accuracy of the model in complex scenes.

[0025] (2) The present invention is superior to other mainstream segmentation network models in terms of average intersection-over-union ratio, average pixel accuracy and segmentation effect, such as UNet3+, crack segmentation network DeepCrack, SegNet, and DeepLab3+.

[0026] (3) The present invention can accurately detect, identify and classify cracks in the borehole walls of different surrounding rocks and has high robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is a system flow diagram of a method for detecting and classifying borehole surrounding rock cracks based on G-UNet of the present invention.

[0028] Figure 2 It is a G-UNet network model structure diagram of a G-UNet-based borehole surrounding rock crack detection and classification method of the present invention.

[0029] Figure 3 It is an original image of the borehole surrounding rock wall and a segmentation effect image of a borehole surrounding rock crack detection and classification method based on G-UNet of the present invention. DETAILED DESCRIPTION

[0030] The following combination Figure 1 , 2 3. Further explanation of the present invention.

[0031] Figure 1 The present invention is a flow chart of a method for detecting and classifying cracks in surrounding rocks of a borehole based on G-UNet, comprising the following steps:

[0032] S1: Construct a dataset of borehole surrounding rock fractures, and divide them into three categories: transverse fractures, longitudinal fractures, and background according to the development characteristics of the fractures;

[0033] S2: Data enhancement processing is performed on the borehole surrounding rock fracture dataset;

[0034] S3: Build the G-UNet network model, including the feature encoder module, the multi-dimensional feature adaptive aggregation module, the feature decoder module, and the ECA channel feature enhancement module;

[0035] S4: Input the enhanced processed borehole surrounding rock fracture data set into the G-UNet network model for training to obtain the training model weight, verify the training model weight on the test set and save the optimal training model weight;

[0036] S5: Use the optimal training model weights to achieve pixel-level segmentation detection and classification of borehole surrounding rock cracks.

[0037] Furthermore, the constructed dataset of borehole surrounding rock fractures is divided into three categories according to the development characteristics of the fractures: transverse fractures, longitudinal fractures, and background. First, obtain the image of the borehole surrounding rock fractures and use Labelme software to annotate them. Since the smaller the inclination angle between the fracture structure surface and the central axis, the more dangerous the surrounding rock construction is, the fractures are classified as follows: fractures with an angle greater than 60 degrees to the central axis of the borehole are automatically classified as transverse fractures; fractures with an angle less than or equal to 60 degrees to the central axis of the borehole are automatically classified as longitudinal fractures; thus, according to the fracture development characteristics, they are divided into three labels: background, transverse fractures, and longitudinal fractures;

[0038] After that, the original image and the labeled image after the original image are divided into training set, validation set and test set in the ratio of 8:1:1. The training set is used in the learning stage of the model. The model learns the characteristics of the data through the training set; the validation set is used to evaluate the performance of the model. The parameters and hyperparameters of the model are adjusted during the training process to improve the performance of the model; the test set is used to evaluate the final performance of the model. The test set is not trained. The divided training set, validation set and test set together constitute the borehole surrounding rock fracture dataset.

[0039] Furthermore, the data enhancement processing of the borehole surrounding rock fracture dataset is specifically performed by:

[0040] Call the get_random_data function to randomly process the original image of the borehole surrounding rock cracks and the corresponding label image in the dataset, input the original image and the corresponding label image in the borehole surrounding rock cracks dataset, scale the image according to the input size, and fill the excess part with a gray background when the size of the scaled image and label is inconsistent with the target; flip the image and label horizontally with a probability of 50% to increase the diversity of samples; convert the image from RGB to HSV, randomly adjust the hue, saturation and brightness to increase the change of the image, and do not change the label image; convert the segmentation label to one-hot encoding format to facilitate subsequent model training; return the enhanced image data, original label and one-hot encoded segmentation label.

[0041] Figure 2 It is a G-UNet network model structure diagram of a borehole surrounding rock crack detection and classification method based on G-UNet of the present invention, and the specific implementation method includes the following steps:

[0042] a: The feature decoder module consists of the pooling layer and convolutional layer before the fully connected layer of the VGG16 network;

[0043] b: Multi-dimensional feature adaptive aggregation module connects the feature encoder module and the feature decoder module;

[0044] c: An ECA channel feature enhancement module is inserted before each upsampling operation in the feature decoder module, and an ECA channel feature enhancement module is inserted at the jump connection between the feature encoder module and the feature decoder module.

[0045] Furthermore, the multi-dimensional feature adaptive aggregation module consists of three submodules: the hole feature capture submodule DFC, the global feature enhancement submodule GFE, and the multi-scale feature aggregation submodule MSFA. The three submodules are combined in the following way: the hole feature capture submodule is connected in parallel with the global feature enhancement submodule, and then the multi-scale feature aggregation submodule is connected in series. First, the feature maps output from the feature encoder module are respectively input into the hole feature capture submodule and the global feature enhancement submodule. The feature maps output after the two submodules are spliced ​​together by concat. At this time, the number of channels of the feature map is the sum of the number of channels of the feature maps output by the two submodules; then, the spliced ​​feature maps are subjected to a convolution operation with a convolution kernel size of 1×1 to restore the number of channels of the feature map; finally, the feature map after the number of channels is restored is input into the multi-scale feature aggregation submodule.

[0046] Furthermore, the hole feature capture submodule is composed of 4 branches containing hole convolutions with different expansion rates. The expansion rates of the hole convolutions are r=1, r=3, r=5, and r=7 respectively. The hole convolution increases the receptive field by introducing holes in the convolution kernel without increasing the amount of calculation, which enables the model to capture contextual information in a larger range, and the hole convolution can extract features from different scales. The combination of multiple hole convolutions with different expansion rates increases the model's ability to learn multi-scale information. After the convolution operation, each branch will be processed with batch normalization and ReLu activation function to improve the stability and convergence speed of model training, while avoiding the gradient vanishing problem. The feature maps output from the four branches are spliced ​​together by concat. At this time, the number of channels of the feature map output from the hole feature capture submodule is 512×4.

[0047] Furthermore, the global feature enhancement submodule extracts the global feature information of the surrounding rock fracture feature map and learns the difference information between the fracture and the rock mass pore wall. The module contains two branches, specifically including the following steps:

[0048] A: The feature map input by the first branch is first subjected to an adaptive global average pooling operation to change the feature map size to 1×1 to obtain global feature information. It is then passed through a convolution layer with a convolution kernel of 1×1 and finally upsampled to restore the feature map size.

[0049] B: The input feature map of another branch is directly fused with the output feature map of the first branch through a residual connection. At this time, the number of channels of the feature map output from the global feature enhancement submodule is 512×2.

[0050] Furthermore, the multi-scale feature aggregation submodule consists of five pooling operation branches, four of which contain pooling operations with different pooling kernels, which are 2×2, 3×3, 6×6, and 7×7. The pooling kernels of 2×2 and 3×3 are used to capture small-scale features, and the pooling kernels of 6×6 and 7×7 are used to capture large-scale features. After each pooling operation, a 1×1 convolution is performed to change the number of feature map channels to 1, and then the size of the feature map is restored by upsampling. The last branch fuses the input feature map with the pooled feature map through a residual connection, thereby realizing the aggregation of multi-scale features. This module can adapt to different sizes of cracks and effectively extract crack features of different sizes. The number of feature map channels output from the multi-scale feature aggregation submodule is 512+4.

[0051] Figure 3 The invention discloses a borehole surrounding rock crack detection and classification method based on G-UNet, and an original image of the borehole surrounding rock wall and a segmentation effect image. After the network model training is completed, the training weight is obtained, and the obtained weight is verified on the test set and the optimal model training weight is saved. The borehole surrounding rock crack image to be tested is input into the model using the optimal weight, and a segmentation effect image of the crack is output, and the crack is classified into transverse cracks and longitudinal cracks by pixel masks of different colors.

[0052] The present invention uses a G-UNet-based borehole surrounding rock crack detection and classification method to accurately detect, identify and classify surrounding rock cracks in front of tunnel excavation, allowing workers to judge the development of cracks and thus make corresponding support measures and construction methods.

[0053] 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 above examples, which are only used to illustrate the concept of this invention. Without departing from the core and limitations of this invention, this invention may also have many adjustments and updates, which fall within the scope of protection of this invention that needs to be protected.

Claims

1. A method for detecting and classifying borehole surrounding rock cracks based on G-UNet, characterized in that: The following steps are involved: S1: Construct a dataset of borehole surrounding rock fractures, and divide them into three categories: transverse fractures, longitudinal fractures, and background according to the development characteristics of the fractures; S2: Data enhancement processing is performed on the borehole surrounding rock fracture dataset; S3: Build the G-UNet network model, including the feature encoder module, the multi-dimensional feature adaptive aggregation module, the feature decoder module, and the ECA channel feature enhancement module; S4: Input the enhanced processed borehole surrounding rock fracture data set into the G-UNet network model for training to obtain the training model weight, verify the training model weight on the test set and save the optimal training model weight; S5: Use the optimal training model weights to achieve pixel-level segmentation detection and classification of borehole surrounding rock cracks.

2. A method for detecting and classifying borehole surrounding rock cracks based on G-UNet as claimed in claim 1, characterized in that: In step S1, the cracks are divided into three categories according to their development characteristics: transverse cracks, longitudinal cracks, and background cracks. The cracks whose angle with the central axis of the borehole is greater than 60 degrees are marked as transverse cracks; and the cracks whose angle with the central axis of the borehole is less than or equal to 60 degrees are marked as longitudinal cracks.

3. The method for detecting and classifying borehole surrounding rock cracks based on G-UNet as claimed in claim 1, characterized in that: In the step S2, data enhancement processing is performed on the borehole surrounding rock fracture data set, and the get_random_data function is called to randomly process the original image of the borehole surrounding rock fracture and the corresponding label image in the data set, including: scaling, cropping, color gamut transformation, and horizontal flipping. When the color gamut is transformed, only the original image is transformed, and the label image is not transformed.

4. The method for detecting and classifying borehole surrounding rock cracks based on G-UNet according to claim 1, characterized in that: The G-UNet network model in step S3 includes a feature encoder module, a multi-dimensional feature adaptive aggregation module, a feature decoder module, and an ECA channel feature enhancement module, and specifically includes the following steps: a: The feature decoder module consists of the pooling layer and convolutional layer before the fully connected layer of the VGG16 network; b: Multi-dimensional feature adaptive aggregation module connects the feature encoder module and the feature decoder module; c: An ECA channel feature enhancement module is inserted before each upsampling operation in the feature decoder module, and an ECA channel feature enhancement module is inserted at the jump connection between the feature encoder module and the feature decoder module.

5. The method for detecting and classifying borehole surrounding rock cracks based on G-UNet as claimed in claim 4, characterized in that: The multidimensional feature adaptive aggregation module in step b is composed of three submodules: a hole feature capture submodule DFC, a global feature enhancement submodule GFE, and a multi-scale feature aggregation submodule MSFA. The three submodules are combined in the following way: the hole feature capture submodule is connected in parallel with the global feature enhancement submodule, the feature maps output from the two submodules are spliced ​​together by concat, and then a 1×1 convolution is performed to restore the number of channels, and finally the multi-scale feature aggregation submodule is connected in series.

6. A method for detecting and classifying borehole surrounding rock cracks based on G-UNet as claimed in claim 5, characterized in that: The void feature capture submodule includes multiple void convolutions with different expansion rates, and the expansion rates are r=1, r=3, r=5, and r=7 respectively. The borehole surrounding rock fracture feature map input to the void feature capture submodule is respectively extracted with void convolutions with different expansion rates to extract feature information of different scales. The output feature map is spliced ​​together using concat to fuse feature information of different scales together.

7. The method for detecting and classifying borehole surrounding rock cracks based on G-UNet as claimed in claim 5, characterized in that: The global feature enhancement submodule extracts the global feature information of the surrounding rock fracture feature map and learns the difference information between the fracture and the rock mass pore wall. The module contains two branches, specifically including the following steps: A: The first branch inputs the borehole surrounding rock fracture feature map, which is first subjected to global average pooling, then 1×1 convolution, and finally the feature map size is restored by upsampling; B: The other branch takes the borehole surrounding rock fracture feature map as input and directly splices it with the output feature map of the first branch through residual connection.

8. The method for detecting and classifying borehole surrounding rock cracks based on G-UNet as claimed in claim 5, characterized in that: The multi-scale feature aggregation submodule is composed of multiple pooling operations and residual connections. The borehole surrounding rock crack feature map input to the module is subjected to pooling operations with different pooling kernels. The pooling operations with different pooling kernels can adapt to the sizes of different cracks. The output feature maps after different pooling will be subjected to 1×1 convolution respectively, and the number of channels will be adjusted to 1. Then, the size of the feature map will be restored by upsampling. At the same time, the input borehole surrounding rock crack feature map will be directly aggregated with all feature maps output from the pooling operation through residual connections. The feature map finally output from the module is also the feature map output from the multi-dimensional feature adaptive aggregation module. Compared with the feature map input to the multi-dimensional feature adaptive aggregation module, the number of channels of the output feature map is increased by four, and the size remains unchanged.

Images