Coal-rock combination fracture three-dimensional reconstruction method and device based on deep learning, computer equipment and storage medium

Abstract

A coal-rock combination fracture three-dimensional reconstruction method based on deep learning comprises the following steps: acquiring a CT scanning image of a coal-rock combination, and carrying out standardization processing to obtain an analysis image; a pre-trained three-dimensional reconstruction model is used for processing the analysis image to generate fracture three-dimensional data, the three-dimensional reconstruction model comprises an encoder and a decoder, the encoder comprises five encoding convolution modules arranged in sequence, the decoder comprises an AC-ASPP module and four decoding block modules arranged in sequence, and the decoding block modules are used for decoding the fracture three-dimensional data. The method for processing the analysis image through the three-dimensional reconstruction model comprises the steps that five coding convolution modules of a decoder are used for conducting step-by-step feature extraction on the analysis image, and five feature maps arranged in sequence are generated; the AC-ASPP module is used for optimizing the fifth feature map and then up-sampling the fifth feature map to a fourth decoding block module; and performing feature recognition on the feature map by using four decoding block modules, and outputting fracture three-dimensional data. According to the method, identification and three-dimensional reconstruction can be more accurately carried out on the fractures of the coal-rock combination.

Description

Technical Field

[0001] The present invention relates to the field of three-dimensional reconstruction of cracks in coal-rock assemblies, and specifically to a method, device, computer equipment and storage medium for three-dimensional reconstruction of cracks in coal-rock assemblies based on deep learning. Background Art

[0002] During coal mining, as the depth of coal mining continues to increase, the ground stress increases significantly, resulting in a significant increase in the frequency and intensity of rock bursts. Rock burst is a mine disaster caused by the instantaneous release of a large amount of elastic energy of a coal-rock combination under the action of mine pressure. Studying the destruction of coal seams or rock formations alone is difficult to reflect the objective reality of coal-rock dynamic disasters induced by rock bursts. Therefore, conducting research on the deformation and destruction mechanism of coal-rock combinations is of great significance for correctly understanding the occurrence mechanism of rock bursts and the coal-rock dynamic disasters induced by rock bursts.

[0003] In order to determine the impact of coal-rock combination cracks on coal-rock dynamic disasters and clarify their mechanism of action during the occurrence process, CT scanning combined with Avizo software 3D reconstruction technology is traditionally used to obtain the structural characteristics of coal-rock combinations and analyze them. However, the light and dark background structure of rock layers and coal layers in coal-rock combinations makes it impossible for the traditional Avizo software's built-in threshold segmentation method to accurately define the threshold of cracks.

[0004] In recent years, thanks to the development of graphics processing technology, semantic segmentation networks have been widely used in the field of coal-rock fracture detection, achieving higher detection accuracy. The principle of the semantic segmentation model is to classify the image at the pixel level and output the category of each pixel, avoiding the problem of difficult threshold determination when dealing with coal-rock assemblies using threshold segmentation methods. However, existing semantic segmentation models are often used to segment images with simple target backgrounds, while coal-rock assemblies have complex fracture structures and the color of the fractures is similar to that of the coal matrix and difficult to distinguish. Therefore, the existing methods of three-dimensional reconstruction of coal-rock assemblies based on speech segmentation models generally have the problem of insufficient accuracy. Summary of the invention

[0005] In order to address the deficiencies in the prior art, the present invention provides a method, device, computer equipment and storage medium for three-dimensional reconstruction of cracks in a coal-rock combination based on deep learning, which can more accurately identify and three-dimensionally reconstruct the cracks in the coal-rock combination.

[0006] In order to achieve the above object, the specific scheme adopted by the present invention is:

[0007] A three-dimensional reconstruction method of coal-rock combination fractures based on deep learning, comprising the following steps:

[0008] Collect CT scan images of coal-rock assemblages and perform standardized processing to obtain analytical images;

[0009] The pre-trained 3D reconstruction model is used to process the analysis image to generate the 3D data of the crack. The 3D reconstruction model includes an encoder and a decoder, wherein the encoder includes five sequentially arranged encoding convolution modules, and the decoder includes an AC-ASPP module and four sequentially arranged decoding block modules. The method for the 3D reconstruction model to process the analysis image includes:

[0010] The five coded convolution modules of the decoder are used to perform step-by-step feature extraction on the analysis image to generate five sequentially arranged feature maps;

[0011] The fifth feature map is optimized using the AC-ASPP module and then upsampled to the fourth decoding block module;

[0012] Four decoding block modules are used to perform feature recognition on the feature map and output the three-dimensional data of the crack.

[0013] Preferably, the method of performing step-by-step feature extraction on the analysis image using the five encoding convolution modules of the decoder comprises:

[0014] The first feature extraction is performed on the analysis image using the two first convolution layers of the first coding convolution module to obtain the first feature map, and the first feature map is subjected to maximum average pooling and then input into the second coding convolution module;

[0015] The two second convolutional layers of the second encoding convolutional module are used to perform a second feature extraction on the input data to obtain a second feature map, and the second feature map is subjected to maximum average pooling and then input into the third encoding convolutional module;

[0016] The third convolution layers of the third encoding convolution module are used to perform a third feature extraction on the input data to obtain a third feature map, and the third feature map is subjected to maximum average pooling and then input into the fourth encoding convolution module;

[0017] The three fourth convolutional layers of the fourth encoding convolutional module are used to perform a fourth feature extraction on the input data to obtain a fourth feature map, and the fourth feature map is input into the fifth encoding convolutional module after performing maximum average pooling.

[0018] The third convolutional layer of the fifth encoding convolutional module is used to perform a fifth feature extraction on the input data to obtain a fifth feature map.

[0019] Preferably, the method for optimizing the fifth feature map using the AC-ASPP module includes:

[0020] The fifth feature map is processed in parallel using the basic convolution block, asymmetric hole convolution block, and pooling block of the AC-ASPP module, and the first layer processing result is output;

[0021] The splicing block of the AC-ASPP module is used to splice the first layer processing results to obtain the fusion processing results;

[0022] The optimization is ended after convolution of the fusion processing result.

[0023] Preferably, the specific method of using four decoding block modules to perform feature recognition on the feature map includes:

[0024] The fourth decoding block module is used to perform feature recognition on the optimized fifth feature map and the fourth feature map, and the recognition results are upsampled to the third decoding block module;

[0025] The third decoding block module performs feature recognition on the third feature map and the data obtained from the fourth decoding block module, and upsamples the recognition result to the second decoding block module;

[0026] The second decoding block module performs feature recognition on the second feature map and the data obtained from the third decoding block module, and upsamples the recognition results to the first decoding block module;

[0027] The first decoding block module performs feature recognition on the first feature map and the data obtained from the second decoding block module, and outputs the three-dimensional data of the crack.

[0028] Preferably, the decoding block module includes a Concatenate module, two decoding convolutional layers, a ResCBAM block and an upsampling block, wherein the Concatenate module is used to splice the received data to obtain spliced ​​data, the decoding convolutional layer is used to sequentially perform feature recognition on the spliced ​​data output by the Concatenate module to obtain feature data, the ResCBAM block is used to perform attention optimization on the feature data, and the upsampling block is used to upsample the attention-optimized feature data to the previous decoding block module.

[0029] Preferably, the method for optimizing the attention of the feature data by the ResCBAM block includes:

[0030] Divide the spliced ​​data into two parts;

[0031] Perform channel attention optimization and spatial attention optimization on one of the spliced ​​data in turn;

[0032] Perform residual connection on another piece of spliced ​​data and the spliced ​​data that has been optimized twice to complete the optimization of the feature data.

[0033] Preferably, the method for training the three-dimensional reconstruction model includes:

[0034] Collect sample images and perform normalization and annotation;

[0035] Group the sample images to obtain training set, validation set and test set;

[0036] Initialize the 3D reconstruction model;

[0037] Iteratively train the 3D reconstruction model using the training set;

[0038] After the iterative training is completed, the validation set is used to optimize the parameters of the 3D reconstruction model;

[0039] The test set is used to perform performance tests on the 3D reconstruction model, and the training is completed if the test passes.

[0040] A three-dimensional reconstruction device for coal-rock combination fractures based on deep learning, including:

[0041] A data acquisition module, used to acquire the CT scan image and perform standardization processing to obtain an analysis image;

[0042] A model operation module is used to process the analysis image using a pre-trained three-dimensional reconstruction model to generate three-dimensional fracture data;

[0043] A data storage module, used for storing three-dimensional fracture data;

[0044] The data output module is used to output visualized three-dimensional fracture data.

[0045] Computer equipment, including:

[0046] Memory for storing computer programs;

[0047] A processor is used to execute the computer program to implement the above-mentioned three-dimensional reconstruction method of coal-rock combination fractures based on deep learning.

[0048] A storage medium stores a computer program, which, when executed, implements the above-mentioned method for three-dimensional reconstruction of coal-rock combination fractures based on deep learning.

[0049] Beneficial effects:

[0050] 1. The present invention adopts the VGG16 network with the fully connected layer removed as the backbone feature extraction network to construct a 3D reconstruction model, and introduces the AC-ASPP module in the decoder module. The size of the convolution kernel in a certain dimension is different from that in other dimensions. This convolution method usually reduces the number of model parameters and reduces the computational complexity. On the other hand, different convolution kernel sizes can also flexibly capture features of different scales and enhance the model's ability to recognize complex shapes and patterns.

[0051] 2. The present invention has pixel-level recognition capability, which avoids the problem that the threshold is difficult to determine when the traditional threshold segmentation method processes coal-rock combinations. In addition, the present invention overcomes the shortcomings of the traditional semantic segmentation model that can only identify simple background cracks and the problem that the traditional threshold segmentation method cannot accurately define the crack threshold through the multi-scale feature extraction capability;

[0052] 3. The present invention introduces the ResCBAM attention module into the three-dimensional reconstruction model, which can ignore irrelevant information and amplify required information, thereby more effectively reflowing the gradient, thereby alleviating the problem of gradient disappearance in the network. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0054] Figure 1 is a schematic diagram of the sample image data augmentation and expansion method;

[0055] Figure 2 It is a structural schematic diagram of the three-dimensional reconstruction model in the present invention;

[0056] Figure 3 It is the VGG16 network with the fully connected layers removed;

[0057] Figure 4 It is a schematic diagram of the structure of the AC-ASPP module;

[0058] Figure 5 It is a schematic diagram of the structure of the ResCBAM module;

[0059] Figure 6 It is a structural diagram of the UNet module;

[0060] Figure 7 It is a comparison chart of the experimental results of the present invention and several prior arts. DETAILED DESCRIPTION

[0061] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0062] like Figure 2-6 As shown, the present invention first provides a three-dimensional reconstruction method of coal-rock combination fractures based on deep learning, including S1 and S2.

[0063] S1. Collect CT scan images of coal-rock combinations and perform standardized processing to obtain analysis images.

[0064] S2. Use the pre-trained 3D reconstruction model to process the analysis image to generate crack 3D data. The 3D reconstruction model is based on UNet, which is mainly composed of a feature extraction backbone network and an enhanced feature extraction network, such as Figure 6 As shown in the figure, the feature extraction backbone network contains five preliminary effective feature layers, each of which consists of two 33 convolutional layers, a ReLU activation function, and a 22 pooling layer, aiming to gradually reduce and abstract image features. The ReLU activation function formula is f(x) = max(0, x), where x is the input value and f(x) is the output value.

[0065] The three-dimensional reconstruction model includes an encoder and a decoder, wherein the encoder includes five sequentially arranged encoding convolution modules, and the decoder includes an AC-ASPP module and four sequentially arranged decoding block modules.

[0066] The encoder uses the VGG16 network with the fully connected layer removed as the backbone feature extraction network. VGG16 uses a smaller 33 convolution kernel and a neural network with more layers, and is designed to improve the model's feature extraction capabilities.

[0067] The encoder has 13 convolutional layers, which are divided into five encoding convolutional modules. Specifically, the first and second convolutional layers form the first encoding convolutional module, the third and fourth convolutional layers form the second encoding convolutional module, the fifth, sixth and seventh convolutional layers form the third encoding convolutional module, the eighth, ninth and tenth convolutional layers form the fourth encoding convolutional module, and the eleventh, twelfth and thirteenth convolutional layers form the fifth encoding convolutional module. The number of convolutional kernels in the first encoding convolutional module, the second encoding convolutional module, the third encoding convolutional module, the fourth encoding convolutional module and the fifth encoding convolutional module are 64, 128, 256, 512 and 512 respectively. The outputs of the first four convolutional modules are divided into two identical copies, one of which is saved as feature1, feature2, feature3 and feature4, and the other is passed to the next encoding convolutional module after a maximum pooling operation. The output of the fifth encoding convolutional module is only saved as feature5.

[0068] The decoder consists of an AC-ASPP module and four decoding blocks. The model runs in the order of the fourth, third, second, and first decoding blocks. Feature 1, feature 2, feature 3, and feature 4 are input to the first and fourth decoding blocks respectively.

[0069] The AC-ASPP module consists of two parts. The first part consists of five parallel modules, including an 11 convolution, three asymmetric dilated convolutions with dilation rates of 6, 12, and 18, and an ASPP pooling layer. The ASPP pooling layer includes an 11 pooling, 11 convolution, and upsampling operations. The formula for dilated convolution is (f*r·g)(x,y)=Σ i ∑ j f(i,j)*g(xr·i,yr·j), where f is the convolution kernel, g is the input image, (x,y) is the coordinate of the output image, i and j are the indexes of the convolution kernel, and r is the void ratio. The second part concatenates the outputs of the five parallel modules in the first part and then performs an 11 convolution operation.

[0070] The decoding block module includes a Concatenate module, two decoding convolutional layers, a ResCBAM block and an upsampling block, wherein the Concatenate module is used to concatenate the received data to obtain concatenated data, the decoding convolutional layer is used to perform feature recognition on the concatenated data output by the Concatenate module in turn to obtain feature data, the ResCBAM block is used to optimize the feature data for attention, and the upsampling block is used to upsample the attention-optimized feature data to the previous decoding block module.

[0071] The ResCBAM module consists of channel attention, spatial attention and residual connection. The input data x is divided into two identical parts, one of which does not perform any operation, and the other obtains the weighted feature map f(x) through the channel attention mechanism and the spatial attention mechanism. Then, the saved input data x is added to the weighted feature map f(x) through the residual connection. The role of the spatial attention mechanism is to emphasize the importance of different positions in the image. Its working principle is: first, the input data is subjected to maximum pooling and average pooling operations, and then the two feature maps are spliced, and the feature map is processed through the convolution layer to generate the spatial attention weight; then the obtained spatial attention weight is applied to the input data to weight the features of each spatial position. The core idea of ​​the residual connection is to add direct connections between one or more layers of the network so that the network can learn the residual between the input and output, which helps the model to more effectively reflux the gradient during the training process, thereby alleviating the problem of gradient vanishing in deep networks. Therefore, by setting the ResCBAM module, the limitations of traditional convolutional neural networks in processing information of different scales, shapes and directions are overcome, and the problem of model gradient vanishing is alleviated. The formulas of the channel attention mechanism and the spatial attention mechanism are:

[0072] M c (F) = σ(W1(W0(F avg ))+W1(W0(F max )));

[0073] M s (F) = σ(f 7×7 ([F avg ; F max ]));

[0074] Where Favg and Fmax are global average pooling and global maximum pooling of feature map F, respectively. W0 and W1 are weight matrices of the fully connected layer. Mc(F) and Ms(F) represent the channel attention map and spatial attention map, respectively. σ represents the Sigmoid activation function.

[0075] Based on the above network structure, the method for processing the analysis image by the three-dimensional reconstruction model includes S21 to S23.

[0076] S21, using five coding convolution modules of the decoder to perform step-by-step feature extraction on the analysis image, and generate five sequentially arranged feature maps. The method of using five coding convolution modules of the decoder to perform step-by-step feature extraction on the analysis image includes S211 to S215.

[0077] S211, using the two first convolutional layers of the first encoding convolutional module to perform a first feature extraction on the analysis image to obtain a first feature map, and performing maximum average pooling on the first feature map and then inputting it into the second encoding convolutional module.

[0078] S212, using the two second convolutional layers of the second encoding convolutional module to perform a second feature extraction on the input data to obtain a second feature map, and performing maximum average pooling on the second feature map and then inputting it into the third encoding convolutional module.

[0079] S213, using the three third convolutional layers of the third encoding convolutional module to perform a third feature extraction on the input data to obtain a third feature map, and performing maximum average pooling on the third feature map and then inputting it into the fourth encoding convolutional module.

[0080] S214, using the three fourth convolutional layers of the fourth encoding convolutional module to perform a fourth feature extraction on the input data to obtain a fourth feature map, and performing maximum average pooling on the fourth feature map and then inputting it into the fifth encoding convolutional module.

[0081] S215. Perform a fifth feature extraction on the input data using the three fifth convolutional layers of the fifth encoding convolutional module to obtain a fifth feature map.

[0082] S22, optimizing the fifth feature map using the AC-ASPP module and then upsampling the optimized feature map to the fourth decoding block module. The method of optimizing the fifth feature map using the AC-ASPP module includes:

[0083] S221. Use the basic convolution block, asymmetric hole convolution block and pooling block of the AC-ASPP module to process the fifth feature map in parallel and output the first layer processing result.

[0084] S222. Use the splicing block of the AC-ASPP module to splice the first layer processing results to obtain a fusion processing result.

[0085] S223, convolve the fusion processing result and then end the optimization.

[0086] S23, using four decoding block modules to perform feature recognition on the feature map and output three-dimensional crack data. The specific method of using four decoding block modules to perform feature recognition on the feature map includes S231 to S234.

[0087] S231, using the fourth decoding block module to perform feature recognition on the optimized fifth feature map and the fourth feature map, and upsampling the recognition results to the third decoding block module.

[0088] S232. The third decoding block module performs feature recognition on the third feature map and the data obtained from the fourth decoding block module, and upsamples the recognition result to the second decoding block module.

[0089] S233, the second decoding block module performs feature recognition on the second feature map and the data obtained from the third decoding block module, and upsamples the recognition result to the first decoding block module.

[0090] S234, the first decoding block module performs feature recognition on the first feature map and the data obtained from the second decoding block module, and outputs the three-dimensional crack data.

[0091] Furthermore, the ResCBAM block is used to perform attention optimization on the feature data. The specific method includes: dividing the spliced ​​data into two parts; performing channel attention optimization and spatial attention optimization on one of the spliced ​​data in turn; and performing residual connection on the other spliced ​​data and the spliced ​​data that has been optimized twice to complete the optimization of the feature data.

[0092] The method for training the 3D reconstruction model is as follows.

[0093] First, sample images are collected and normalized and labeled. Specifically, the pixel values ​​of all pixels in the sample image are divided by 255 to scale to the interval [0, 1] to obtain the true label, and the number of iterations is set to N times, N = [100, 300]. In this embodiment, a total of 300 original sample images are collected, and 1200 sample images are formed after data enhancement and expansion, thereby increasing the number and diversity of sample images and improving the generalization ability and robustness of the trained model. Figure 1 As shown, the methods of data enhancement and expansion include brightness enhancement, noise blurring, left-right flipping, and up-down flipping.

[0094] Second, the sample images are grouped to obtain a training set, a validation set, and a test set. The division ratio of the training set, the validation set, and the test set is 6:2:2, that is, the training set has 720 sample images, the validation set has 210 sample images, and the test set has 240 sample images.

[0095] Third, initialize the 3D reconstruction model.

[0096] Fourth, the 3D reconstruction model is iteratively trained using the training set. During the iteration process, after the first iteration, the model parameters of the current 3D reconstruction model are taken as the best model parameters and saved; except for the first iteration, after other iterations, the model parameters of the current 3D reconstruction model are compared with the best model parameters stored previously, and the model parameters of the 3D reconstruction model with the smallest verified loss function are taken as the new best model parameters and saved, so as to realize the continuous correction of the model parameters.

[0097] Fifth, after the iterative training is completed, the validation set is used to optimize the parameters of the 3D reconstruction model.

[0098] Sixth, use the test set to perform performance test on the 3D reconstruction model. If the test passes, the training is completed.

[0099] In order to more intuitively understand the segmentation performance of the model, this paper evaluates the performance of the classification model based on the confusion matrix. By comparing the model's predicted results with the actual category labels, the model's performance indicators such as mean intersection over union (mIOU), pixel average (mPA) and precision can be calculated. The indicator definition formula is as follows:

[0100]

[0101] Where k represents the number of foregrounds, TP represents the true category of the sample is crack, and the model correctly identifies it; TN represents the true category of the sample is background, and the model identifies it as crack; FP represents the true category of the sample is background, but the model identifies it as crack; FN represents the true category of the sample is crack, but the model misidentifies it.

[0102] Taking UNet, PSPNet, DeeplabV3+, FCN and SegNet in the prior art as comparison, the evaluation results of the present invention and several prior arts are shown in Table 1. In Table 1, the present invention is marked as VRA-UNet, that is, the 3D reconstruction model of the present invention is abbreviated as VRA-UNet.

[0103] Table 1 Evaluation results of the present invention and several prior arts

[0104] Model mIOU mPA Precision UNet 79.17 83.67 88.13 PSPNet 68.52 77.51 77.50 DeeplabV3+ 74.45 77.96 84.55 FCN 78.35 83.04 86.37 SegNet 78.82 83.27 87.64 VRA-UNet 85.22 90.80 91.95

[0105] As shown in Table 1, the VRA-UNet model of the present invention has better comprehensive performance than other network models. When the model converges, the evaluation indicators such as mIOU, mPA and Precision reach 85.22%, 90.80% and 91.95% respectively. Compared with UNet, PSPNet, DeepLabv3+, FCN and SegNet models, the mIOU of the VRA-UNet model of the present invention is increased by 6.05%, 16.7%, 10.77%, 6.87% and 6.4% respectively, the mPA is increased by 7.13%, 13.29%, 12.84%, 7.4% and 7.53% respectively, and the Precision is increased by 3.82%, 14.45%, 7.4%, 5.58% and 4.31% respectively.

[0106] Furthermore, the present invention and the above-mentioned prior arts are applied to the actual identification of cracks in coal-rock assemblages. The results are as follows: Figure 7 As shown. Figure 7 It can be seen that the VRA-UNet model of the present invention achieves higher accuracy in the segmentation of coal-rock combination images, and the edge segmentation is smoother, which effectively overcomes the problem of misrecognition caused by the similarity between the background color of the coal matrix and the cracks in the threshold segmentation method, as well as the problem of threshold setting that cannot take into account both coal and rock in the threshold segmentation process.

[0107] The present invention further provides a three-dimensional reconstruction device for coal-rock combination fractures based on deep learning, including a data acquisition module, a model operation module, a data storage module and a data output module.

[0108] The data acquisition module is used to acquire CT scan images and perform standardization processing to obtain analysis images.

[0109] The model running module is used to process the analysis image using the pre-trained three-dimensional reconstruction model to generate three-dimensional fracture data.

[0110] The data storage module is used to store three-dimensional fracture data.

[0111] The data output module is used to output visualized three-dimensional fracture data.

[0112] Those skilled in the art can clearly understand that, for the convenience and simplicity of description, the specific working process of the system and device described above can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here. In the several embodiments provided in the present disclosure, it should be understood that the disclosed system, device and method can be implemented in other ways. The device embodiment described above is only schematic. For example, the division of the module is only a logical function division. There may be other division methods in actual implementation. For example, multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some communication interfaces, indirect coupling or communication connection of devices or modules, which can be electrical, mechanical or other forms. The module described as a separate component may or may not be physically separated, and the component displayed as a module may or may not be a physical module, that is, it may be located in one place, or it may be distributed on multiple network modules. Some or all of the modules can be selected according to actual needs to achieve the purpose of the scheme of this embodiment.

[0113] The present invention also provides a computer device, including a memory and a processor.

[0114] Memory, used to store computer programs.

[0115] A processor is used to execute a computer program to implement the above-mentioned three-dimensional reconstruction method of coal-rock combination fractures based on deep learning.

[0116] Finally, the present invention provides a storage medium storing a computer program, which, when executed, implements the above-mentioned method for three-dimensional reconstruction of coal-rock combination fractures based on deep learning.

[0117] Based on the storage medium, the technical solution of the present disclosure, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present disclosure. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc., various media that can store program codes.

[0118] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0119] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A three-dimensional reconstruction method of coal-rock combination fractures based on deep learning, characterized in that: The steps include: Collect CT scan images of coal-rock assemblages and perform standardized processing to obtain analytical images; The pre-trained 3D reconstruction model is used to process the analysis image to generate the 3D data of the crack. The 3D reconstruction model includes an encoder and a decoder, wherein the encoder includes five sequentially arranged encoding convolution modules, and the decoder includes an AC-ASPP module and four sequentially arranged decoding block modules. The method for the 3D reconstruction model to process the analysis image includes: The five encoding convolution modules of the decoder are used to extract features of the analysis image step by step, generating five feature maps arranged in order; the fifth feature map is optimized using the AC-ASPP module and then upsampled to the fourth decoding block module; Four decoding block modules are used to perform feature recognition on the feature map and output the three-dimensional data of the crack.

2. A method for three-dimensional reconstruction of coal-rock combination fractures based on deep learning as claimed in claim 1, characterized in that: The method of performing step-by-step feature extraction on the analysis image using the five coded convolutional modules of the decoder includes: The first feature extraction is performed on the analysis image using the two first convolution layers of the first coding convolution module to obtain the first feature map, and the first feature map is subjected to maximum average pooling and then input into the second coding convolution module; The two second convolutional layers of the second encoding convolutional module are used to perform a second feature extraction on the input data to obtain a second feature map, and the second feature map is subjected to maximum average pooling and then input into the third encoding convolutional module; The third convolution layers of the third encoding convolution module are used to perform a third feature extraction on the input data to obtain a third feature map, and the third feature map is subjected to maximum average pooling and then input into the fourth encoding convolution module; The three fourth convolutional layers of the fourth encoding convolutional module are used to perform a fourth feature extraction on the input data to obtain a fourth feature map, and the fourth feature map is input into the fifth encoding convolutional module after performing maximum average pooling. The third convolutional layer of the fifth encoding convolutional module is used to perform a fifth feature extraction on the input data to obtain a fifth feature map.

3. A method for three-dimensional reconstruction of coal-rock combination fractures based on deep learning as claimed in claim 2, characterized in that: The method of optimizing the fifth feature map using the AC-ASPP module includes: The fifth feature map is processed in parallel using the basic convolution block, asymmetric hole convolution block, and pooling block of the AC-ASPP module, and the first layer processing result is output; The splicing block of the AC-ASPP module is used to splice the first layer processing results to obtain the fusion processing results; The optimization is ended after convolution of the fusion processing result.

4. A method for three-dimensional reconstruction of coal-rock combination fractures based on deep learning as claimed in claim 2, characterized in that: The specific method of using four decoding block modules to perform feature recognition on feature maps includes: The fourth decoding block module is used to perform feature recognition on the optimized fifth feature map and the fourth feature map, and the recognition results are upsampled to the third decoding block module; The third decoding block module performs feature recognition on the third feature map and the data obtained from the fourth decoding block module, and upsamples the recognition result to the second decoding block module; The second decoding block module performs feature recognition on the second feature map and the data obtained from the third decoding block module, and upsamples the recognition results to the first decoding block module; The first decoding block module performs feature recognition on the first feature map and the data obtained from the second decoding block module, and outputs the three-dimensional data of the crack.

5. A method for three-dimensional reconstruction of coal-rock combination fractures based on deep learning as claimed in claim 4, characterized in that: The decoding block module includes a Concatenate module, two decoding convolutional layers, a ResCBAM block and an upsampling block, wherein the Concatenate module is used to concatenate the received data to obtain concatenated data, the decoding convolutional layer is used to perform feature recognition on the concatenated data output by the Concatenate module in turn to obtain feature data, the ResCBAM block is used to optimize the feature data for attention, and the upsampling block is used to upsample the attention-optimized feature data to the previous decoding block module.

6. A method for three-dimensional reconstruction of coal-rock combination fractures based on deep learning as claimed in claim 5, characterized in that: The methods by which the ResCBAM block optimizes attention on feature data include: Divide the spliced ​​data into two parts; Perform channel attention optimization and spatial attention optimization on one of the spliced ​​data in turn; Perform residual connection on another piece of spliced ​​data and the spliced ​​data that has been optimized twice to complete the optimization of the feature data.

7. The method for three-dimensional reconstruction of coal-rock combination fractures based on deep learning according to claim 1, characterized in that: Methods for training 3D reconstruction models include: Collect sample images and perform normalization and annotation; Group the sample images to obtain training set, validation set and test set; Initialize the 3D reconstruction model; Iteratively train the 3D reconstruction model using the training set; After the iterative training is completed, the validation set is used to optimize the parameters of the 3D reconstruction model; The test set is used to perform performance tests on the 3D reconstruction model, and the training is completed if the test passes.

8. A three-dimensional reconstruction device for coal-rock combination fractures based on deep learning, characterized in that: include: A data acquisition module, used to acquire the CT scan image and perform standardization processing to obtain an analysis image; A model operation module is used to process the analysis image using a pre-trained three-dimensional reconstruction model to generate three-dimensional fracture data; A data storage module, used for storing three-dimensional fracture data; The data output module is used to output visualized three-dimensional fracture data.

9. Computer device, characterized in that include: Memory for storing computer programs; A processor, used to execute the computer program to implement a three-dimensional reconstruction method of coal-rock combination fractures based on deep learning as described in any one of claims 1 to 7.

10. A storage medium, characterized in that A computer program is stored, which, when executed, implements a three-dimensional reconstruction method of coal-rock combination fractures based on deep learning as described in any one of claims 1 to 7.

Images