A coal rock data generation and permeability prediction method based on deep learning

Abstract

The application discloses a coal rock data generation and permeability prediction method based on deep learning, carries out pretreatment on a coal rock two-dimensional slice image, splices images to obtain three-dimensional binary data; uses a CGAN model to generate image data, uses a Stokes equation and a Darcy law to calculate permeability; parallel feature extraction of CNN and Transform structure is constructed to obtain fusion features for permeability prediction. The application uses a Transform model+CNN model to establish spatial feature selection and sequence feature supplement to predict permeability, ensures the correlation features of sequence graphs, and obtains higher-precision prediction results. Through deep learning, the problem of model underfitting caused by insufficient data is made up, and for the training difficulty of large-size images in the convolution process, a feature based on fusion sampling image slices and non-sampling slices is provided, higher feasibility is brought to model training, model training is facilitated and the demand for training equipment is reduced, and the deployment possibility of the scheme is enhanced.

Description

Technical Field

[0001] This invention belongs to the field of coal and rock development data mining technology, and in particular relates to a method for generating coal and rock data and predicting permeability based on deep learning. Background Technology

[0002] Coal permeability is of profound significance for related development and research in the field of coal petrology. In real-world scenarios, coal seams often undergo multiple phases of tectonic movement after deposition, experiencing repeated compression from tectonic stress fields in different directions. This leads to the destruction of the original structure, resulting in destructive deformation and the formation of numerous natural fractures. During coalbed methane extraction, the drainage and depressurization process alters the original stress state of the coal reservoir, causing deformation of the pores and fractures within the reservoir. This, in turn, affects the dynamic changes in permeability. The combined influence of these multiple factors increases the difficulty of rapidly and effectively predicting the permeability of coal petrological data.

[0003] Coal and rock permeability prediction methods can be divided into two categories: physical simulation-based and deep learning-based approaches. Physical simulation mainly involves extracting relevant structures and simulating streaming data from existing 2D or 3D data to estimate permeability, which often requires significant processing and analysis time. In real-world scenarios, faster, more real-time prediction methods are needed. However, research on deep learning-based coal and rock permeability prediction is limited. Methods for predicting rock permeability primarily rely on random Gaussian data generation, using 2D slices to construct a 2D-to-3D permeability correspondence, and predicting permeability based on relevant physical features of image data. These approaches often significantly damage data information, making it difficult to effectively explain the structure and transform features of image data. Furthermore, due to data scarcity, their prediction models tend to have weak correlation and reliability, hindering their widespread application in real-world scenarios.

[0004] Currently, most methods for predicting coal seam permeability using machine learning are based on a single machine learning model. This approach is simple and fast, building models using large amounts of data. Theoretically, the more training data, the more stable the prediction model. Deep learning methods, through reliable training on large datasets, can quickly and accurately predict coal permeability. However, the accuracy and generalization of these prediction models are poor, and they are highly dependent on the number of samples. Especially when the training sample size is small, the model may exhibit significant errors. Summary of the Invention

[0005] The problem this invention aims to solve is to provide a deep learning-based method for generating coal and rock data and predicting permeability. In particular, by processing coal and rock data, deep learning is introduced to generate data and establish a fast and feasible serialization scheme for permeability prediction, which can avoid the situation where the whole image is used as input, resulting in the inability to train.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by this invention is: a method for generating coal and rock data and predicting permeability based on deep learning, comprising the following steps:

[0007] S1: Data processing, which generates three-dimensional binary data from the collected coal and rock data;

[0008] S2: Data generation. Existing real data is used to generate coal and rock data using the CGAN model. Permeability is calculated using the Stokes equation and Darcy's law to obtain similar and reliable coal and rock generation data.

[0009] S3: Penetration rate prediction. By constructing parallel feature extraction using CNN and Transformer structures, fused features are obtained for penetration rate prediction.

[0010] Furthermore, S1 includes the following steps:

[0011] S11: Crop the edges of the collected coal and rock data to make all slice images cropped to the same size;

[0012] S12: The slice data is filtered and denoised, and adaptive thresholding and threshold segmentation work together to obtain the pore structure and generate its corresponding binarized image.

[0013] S13: Concatenate the slices along their sequential dimensions to generate three-dimensional binary data.

[0014] Furthermore, S12 includes the following steps:

[0015] S121: Divide the cropped slice image into regions with specific features and select and extract slice images based on their pore-crack structure;

[0016] S122: Due to the differences between different images, the threshold settings for segmentation vary. An adaptive threshold segmentation method is used to perform reasonable pore segmentation on the slices and generate corresponding binary slice data.

[0017] S123: Enhance the segmentation details by combining adaptive segmentation and threshold segmentation. Adaptive segmentation adjusts the threshold by averaging local features, and segments more closely with respect to details than a given threshold. Threshold segmentation, on the other hand, can greatly preserve the porosity information of the overall structure by dividing the data by a threshold.

[0018] Furthermore, the adaptive segmentation formula is as follows:

[0019]

[0020]

[0021] In the formula: The segmentation threshold is generated adaptively. It is the average area of ​​the neighborhood. It is used for fine-tuning The defined constant value, For the original image at pixel location grayscale value, The classification number is generated based on adaptive threshold segmentation. In digital images of coal and rock, 1 represents solid matrix and 0 represents pore structure.

[0022] The threshold segmentation formula is:

[0023]

[0024] In the formula: For threshold segmentation, the threshold is set, where f(x,y) is the pixel position of the original image. grayscale value, The classification number is generated based on threshold segmentation. In digital images of coal and rock, 1 represents solid matrix and 0 represents pore structure.

[0025] The final slice segmentation formula is:

[0026]

[0027] In the formula: The classification number is generated by fusing adaptive threshold segmentation and threshold segmentation results.

[0028] Furthermore, step S2 includes the following steps:

[0029] S21: Divide the existing real data into training set and test set in a 6:4 ratio. Use the training set to extract relevant features for data generation condition information. Compress the training set by k times to a specified dimension as real data for training. Use the CGAN model to generate coal and rock data using the above-processed data.

[0030] S22: After upsampling the generated data by k times to restore it to the required size, LBM is used to simulate the fluid movement in the pore structure. The permeability is calculated using Stokes' equation and Darcy's law to obtain similar and reliable coal and rock generation data.

[0031] Furthermore, S21 includes the following steps:

[0032] S211: Calculate the corresponding pore area of ​​the processed data based on its individual slices, and extract its pore area ratio and curvature information as conditional information for data generation.

[0033] S212: Compress the 3D data proportionally by downsampling the data dimensions by a factor of k to provide feasible input for CGAN.

[0034] Furthermore, the objective function of CGAN is:

[0035]

[0036]

[0037] In the formula: For generator, Let x be the discriminator and x be the generated data. The data distribution of the generator is as follows: The noise distribution is as follows ，y The conditional information, noise z, and the result of the conditional information can be denoted as... The result of data x on discriminator D is .

[0038] Furthermore, step S3 includes the following steps:

[0039] S31: The generated data and a portion of the real dataset are fused together to form the training dataset for the penetration rate prediction model;

[0040] S32: Extract relevant features from the input data using selected interval slicing and non-selected interval slicing;

[0041] S33: By constructing parallel feature extraction using CNN and Transformer structures, fused features are obtained for penetration rate prediction.

[0042] Furthermore, S32 includes the following steps:

[0043] S321: Interval image selection is performed by selecting slices according to the slice order information based on t steps;

[0044] S322: Extract relevant feature information from the non-selected slice data, mainly including the extraction of the pore area ratio of the slice data, the extraction of the pore distribution of the slice data, and the construction of relevant feature vectors with sequence characteristics from the relative sequence position information of the slices.

[0045] Furthermore, S33 includes the following steps:

[0046] S331: Continuous convolution of image data is performed by CNN. Continuous features on the image structure are constructed by the spatial extraction capability of CNN. The relative sequence property is preserved by the discontinuous selection of slices, thus achieving dual feature preservation in both time and space.

[0047] S332: Based on the Transformer, the relevant feature vectors of the sequence slices without CNN convolution are extracted by Transformer. The feature extraction of the sequence data is used to make up for the lack of image feature extraction on the discontinuously selected sequence images.

[0048] S333: This method fuses sequence-level features with spatial features from CNN convolutions. A linear weighting strategy is used to select weights for different features, applying them at corresponding positions. The size allocation of different features is adaptively adjusted. A simple penetration rate prediction module is used to construct the correspondence between the fused features and the penetration rate. A multi-layer MLP is built within this module for layer-by-layer feature descent, gradually reducing the error. A residual module is introduced to enhance the contextual association of feature information.

[0049] The loss function for permeability prediction can be expressed as:

[0050]

[0051] Where N is the number of samples, For the true penetration rate, To predict penetration rate.

[0052] The advantages and positive effects of this invention are:

[0053] This invention uses CGAN to generate data to obtain similar images. Then, it employs a Transform model combined with a CNN model to predict penetration rate by selecting spatial features and supplementing sequential features, ensuring the relevance of the sequence images and achieving higher accuracy prediction results. Deep learning is used to generate training data to compensate for underfitting caused by insufficient data. Furthermore, to address the difficulty of training large images during convolution, a feature-based approach that fuses sampled and unsampled image slices is provided, increasing the feasibility of model training, facilitating model training, reducing the need for training equipment, and enhancing the deployment possibilities of the solution. Attached Figure Description

[0054] Figure 1 This is an overall flowchart of an embodiment of the present invention.

[0055] Figure 2 This is a schematic diagram of the model structure of an embodiment of the present invention. Detailed Implementation

[0056] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0058] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation", "connection" and "linking" should be interpreted broadly, and those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0059] The embodiments of the present invention will be further described below with reference to the accompanying drawings:

[0060] like Figure 1 As shown, a method for generating coal and rock data and predicting permeability based on deep learning includes the following steps.

[0061] S1: Data Processing. CT microscopic imaging technology is used to photograph real coal and rock, obtaining a digital sequence of two-dimensional microstructure image slices. The data is then processed to generate three-dimensional binary data from the collected coal and rock data. Specifically, S1 includes the following steps:

[0062] S11: Data downsampling. Considering the edge effect of core scanning, pixels will be cropped from each core scan result as the base data. Specifically, the edges of the collected coal and rock data are cropped to ensure that all slice images are cropped to the same size, avoiding the surrounding black edges acting as strong interference information in constructing the correspondence between slice features and permeability.

[0063] S12: The sliced ​​data undergoes filtering and noise reduction, adaptive thresholding, and threshold segmentation. Segmentation acquires the pore structure and generates its corresponding binarized image. Data noise reduction does not eliminate all noise, but rather minimizes it as much as possible. It modifies the processed data state without altering the original state, transforming it into a state that conforms to the regular relationships between the original data attributes and extracts the original state and outlier information. The sampling process reduces the impact of noise, thus ensuring the accuracy of the measurement results. Data filtering removes interference information from one or more data sets while maintaining the information of the acquisition system. Data filtering is crucial; it can improve computational efficiency and achieve better processing results in certain situations. Data binarization applies contrast enhancement and noise reduction filters to obtain a filtered grayscale image from the original grayscale image, and then obtains binary image data from the filtered grayscale image data. Specifically, S12 includes the following steps:

[0064] S121: Divide the cropped slice image into regions with specific characteristics and select and extract slice images based on their differences in pore-crack structure.

[0065] S122: Due to the differences between different images, the threshold settings for segmentation vary. An adaptive threshold segmentation method is used to perform reasonable pore segmentation on the slices and generate corresponding binary slice data.

[0066] S123: Enhance the segmentation details by combining adaptive segmentation and threshold segmentation. Adaptive segmentation adjusts the threshold by averaging local features, and segments more closely with respect to details than a given threshold. Threshold segmentation, on the other hand, can greatly preserve the porosity information of the overall structure by dividing the data by a threshold.

[0067] Specifically, the adaptive segmentation formula is as follows:

[0068]

[0069]

[0070] In the formula: The segmentation threshold is generated adaptively. It is the average area of ​​the neighborhood. It is used for fine-tuning The defined constant value, For the original image at pixel location grayscale value, The classification number is generated based on adaptive threshold segmentation. In digital images of coal and rock, 1 represents solid matrix and 0 represents pore structure.

[0071] The threshold segmentation formula is:

[0072]

[0073] In the formula: For threshold segmentation, the threshold is set, where f(x,y) is the pixel position of the original image. grayscale value, The classification number is generated based on threshold segmentation. In digital coal and rock images, 1 represents solid matrix and 0 represents pore structure.

[0074] The final slice segmentation formula is:

[0075]

[0076] In the formula: The classification number is generated by fusing adaptive threshold segmentation and threshold segmentation results.

[0077] S13: Concatenate the slices along their sequential dimensions to generate three-dimensional binary data.

[0078] S2: Data Generation. Existing real data is used to generate coal and rock data using the CGAN model. Permeability is calculated using the Stokes equation and Darcy's law to obtain similar and reliable coal and rock generation data. Specifically, S2 includes the following steps:

[0079] S21: Divide the existing real data into training and test sets in a 6:4 ratio. Use the training set to extract relevant features for data generation condition information. Compress the training set by a factor of k to a specified dimension to serve as real data for training. Use the CGAN model to generate coal and rock data using the data processed above.

[0080] S211: Calculate the corresponding pore area of ​​the processed data based on its individual slices, and extract its pore area ratio, curvature information, etc. as conditional information for data generation.

[0081] S212: The 3D data is compressed proportionally by downsampling, compressing the data dimension by a factor of k to provide a feasible input for CGAN. Specifically, considering that the size of the synthesized data would be too large for CGAN, downsampling is used to process the data, which is then used as one of the inputs to the CGAN discriminator. Random noise and conditional labels refer to the relevant features of the 3D image after cropping and processing, such as pore size, pore number, curvature information, etc., which are concatenated as input to the generator. The input to the discriminator is a vector obtained by concatenating the generated image and the real image with the conditional vector. The result is used to train the generator. Through continuous game training, a generator model that can specify the generated result is obtained. Finally, the generated image is upsampled to restore the original image size.

[0082] The main modules for generating CGAN-based data include:

[0083] The generation module: CGAN uses noise as input and adds conditional vectors for supervision and guidance, using the conditional vectors to supervise the generation of highly reliable data with higher confidence.

[0084] Identification Module: CGAN uses the images and conditional vectors generated by the generation module, as well as the real images and conditional vectors, as inputs to generate coal and rock data that are more similar to the real samples.

[0085] Specifically, the objective function of CGAN is:

[0086]

[0087]

[0088] In the formula: For generator, Let x be the discriminator and x be the generated data. The data distribution of the generator is as follows: The noise distribution is as follows ，y The conditional information, noise z, and the result of the conditional information can be denoted as... The result of data x on discriminator D is .

[0089] S22: After upsampling the generated data k times to restore it to the required size, LBM is used to simulate fluid motion in the pore structure. Permeability is calculated using Stokes' equations and Darcy's law to obtain similar and reliable coal and rock generation data. The Lattice Boltzmann Method (LBM) is a flow simulation method suitable for real pore structures. Compared to other traditional CFD calculation methods, this method possesses the characteristics of a mesoscopic model, falling between microscopic molecular dynamics models and macroscopic continuous models. Therefore, it has advantages such as simple description of fluid interactions, easy setting of complex boundaries, ease of parallel computation, and ease of program implementation.

[0090] S3: Permeability Prediction. Through parallel feature extraction using a CNN and Transformer architecture, fused features are obtained for permeability prediction. This provides a feasible model for constructing permeability prediction from coal and rock slices. The data input is divided into interval image slices with structural information and relevant feature extraction from unselected image slices to supplement the sequence information of the selected image slices. Feature extraction refers to the process where the selected image information is fed into a CNN neural network to learn the overall structural information of the coal and rock, extracting the overall structural information. Unselected image information is processed using feature extraction methods according to the overall sequence information, and integrated into a sequence information that is then fed into the Transformer. The discontinuity between unselected image information provides additional continuous information, thereby enhancing the reliability of the features from the selected image information. Specifically, S3 includes the following steps:

[0091] S31: The generated data and a portion of the real dataset are fused together to form the training dataset for the penetration rate prediction model. For the new synthetic training dataset, considering its size and memory consumption in CNNs, a method based on interval feature extraction is proposed. This method predicts penetration rate by fusing intermittently selected slice images for CNN convolution and non-intermittently selected images for Transformer feature extraction.

[0092] S32: Extract relevant features from the input data using selected interval slicing and non-selected interval slicing. Specifically, S32 includes the following steps:

[0093] S321: Interval image selection is performed by selecting slices according to the slice order information based on t steps. Interval data selection refers to selecting images according to a specified step size to extract the overall 3D structural features. Unselected images are processed using feature extraction methods to extract and transform relevant features, including image porosity, pore area, coordination number, average pore diameter, average pore cross-sectional area, and relative numbering of the image sequence.

[0094] S322: Extract relevant feature information from the non-selected slice data, mainly including the extraction of the pore area ratio of the slice data, the extraction of the pore distribution of the slice data, and the construction of relevant feature vectors with sequence characteristics from the relative sequence position information of the slices.

[0095] S33: By constructing parallel feature extraction using CNN and Transformer structures, fused features are obtained for penetration rate prediction.

[0096] Specifically, S33 includes the following steps:

[0097] S331: By performing continuous convolution on image data using CNN, continuous features are constructed on the image structure through the spatial extraction capability of CNN, and the discontinuous selection through slicing preserves its relative sequence properties to a certain extent, thus achieving dual feature preservation in both time and space.

[0098] S332: Based on the Transformer, feature extraction of the sequence data is performed on the sequence slices that have not undergone CNN convolution by extracting relevant feature vectors using Transformer. This preserves the feature extraction of the sequence data to compensate for the lack of image feature extraction on discontinuously selected sequence images.

[0099] S333: By fusing feature information in the sequence dimension with spatial features in the CNN convolution, a linear weighting strategy is used to select weights for different features and apply them to corresponding positions. The size distribution of different features is adaptively adjusted. Through the fused features, a simple penetration prediction module is used to construct the correspondence between the dual fused features and the penetration rate. Multi-layer MLP is constructed in the module to perform layer-by-layer feature descent, slowly reducing the error. A residual module is introduced to enhance the contextual association of feature information.

[0100] The loss functions for permeability prediction are the L2 loss function and the L1 loss function. The loss function for permeability prediction can be expressed as:

[0101]

[0102] Where N is the number of samples, For the true penetration rate, To predict penetration rate.

[0103] In summary, this invention uses CGAN to generate data to obtain similar images. Then, it employs a Transform model combined with a CNN model to predict penetration rate by selecting spatial features and supplementing sequential features, ensuring the relevance of the sequence images and achieving higher accuracy prediction results. Deep learning is used to generate training data to compensate for underfitting caused by insufficient data. Furthermore, it addresses the difficulty of training large images during convolution by providing a feature-based approach that fuses sampled and unsampled image slices, making model training more feasible, facilitating model training, reducing the need for training equipment, and enhancing the deployment possibilities of the solution. Compared to traditional manual data collection and physical methods for penetration rate prediction, deep learning-based data generation and penetration rate prediction require less manual labor, less time, and have higher penetration efficiency.

[0104] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A deep learning-based coal rock data generation and permeability prediction method, characterized in that: Comprise the following steps, S1: data processing, making the collected coal rock data generate three-dimensional binary data; S2: data generation, using the existing real data to generate coal rock data by the model of CGAN, calculating the permeability by the Stokes equation and Darcy's law, and obtaining similar and reliable coal rock generated data; S3: permeability prediction, by constructing the parallel feature extraction of CNN and Transformer structure, obtaining the fusion features for permeability prediction, the S3 comprising the following steps, S31: fuse the generated data and part of the real data set as the training data set of the permeability prediction model; S32: select interval slicing and non-select interval slicing related feature extraction on the input data; S33: by constructing the parallel feature extraction of CNN and Transformer structure, obtaining the fusion features for permeability prediction, the S33 comprising the following steps, S331: by CNN, continuous convolution is performed on image data, continuous features are constructed on the image structure by the spatial extraction capability of CNN, and the relative sequence properties are retained by the discontinuous selection of slicing, while achieving double feature reservation in space and time; S332: according to the Transformer, the Transformer feature extraction of related feature vectors is performed on the sequence slicing without CNN convolution, and the feature extraction of sequence data is retained to make up for the insufficient image feature extraction on the sequence image of discontinuous selection; S333: by fusing the feature information in sequence dimension and the spatial features on CNN convolution, a linear weighting strategy is adopted to select the weight of different features, which is used for weighting in the corresponding position, adaptively adjusting the size allocation of different features, and by fusing the features, a simple permeability prediction module is used to construct the corresponding relationship between the double fusion features and the permeability, a multi-layer MLP is constructed therein to gradually reduce the error, a residual module is introduced to enhance the context association of feature information, The loss function of permeability prediction can be expressed as: where N is the number of samples, is the true permeability, is the predicted permeability.

2. The coal rock data generation and permeability prediction method based on deep learning according to claim 1, characterized in that: The S1 comprises the following steps, S11: cropping the edges of the collected coal rock data, so that all slice images are cropped to the same size; S12: filtering and denoising the slice data, and the adaptive threshold and threshold segmentation jointly act on the slice data, the segmentation is used to obtain the pore structure, and the corresponding binary image is generated; S13: the slice is spliced in sequence dimension to generate three-dimensional binary data.

3. The coal rock data generation and permeability prediction method based on deep learning according to claim 2, characterized in that: The S12 comprises the following steps, S121: the cropped slice image is divided into a region with specific features, and the slice image is selected and extracted according to the difference of its pore-crack structure; S122: due to the difference between different images, the threshold setting of segmentation is different, the adaptive threshold segmentation method is used to reasonably segment the pores of the slice, and the corresponding binary slice data is generated; S123: Enhance the details of the segmentation by adaptive segmentation and threshold segmentation combined segmentation method, adaptive segmentation is through the mean value of local features to adaptively adjust the threshold, relative to the given threshold segmentation pays more attention to the details of information, while threshold segmentation can greatly retain the overall structure of the pore information by threshold division.

4. The coal rock data generation and permeability prediction method based on deep learning according to claim 3, characterized in that: The adaptive segmentation formula is: wherein: is the adaptive generated segmentation threshold, is the average value of the neighborhood area, is the fine-tuning factor, is the constant value defined by, is the gray value of the original image at pixel position , is the classification number generated according to the adaptive threshold segmentation, in which 1 represents the solid matrix and 0 represents the pore structure in the coal digital image. The threshold segmentation formula is: In the formula: is a threshold value of threshold segmentation, f(x, y) is a gray value of a raw image at a pixel position , is a classification number generated according to threshold segmentation, wherein 1 represents a solid matrix and 0 represents a pore structure in a coal rock digital image. The final slice segmentation formula is: In the formula: is a classification number generated according to adaptive threshold segmentation and fusion of threshold segmentation results.

5. The coal rock data generation and permeability prediction method based on deep learning according to any one of claims 1 to 4, characterized in that: The S2 includes the following steps, S21: Divide the existing real data into training set and test set according to the proportion of 6:4, use the training set to extract the relevant features for the condition information of data generation, and compress the training set by k times to the specified dimension as the training real data, use the CGAN model to generate coal rock data using the above processed data; S22: After the generated data is up-sampled k times to recover to the required size, the fluid motion in the pore structure is simulated by LBM, the permeability is calculated by Stokes equation and Darcy law, and similar and reliable coal rock generated data is obtained.

6. The coal rock data generation and permeability prediction method based on deep learning according to claim 5, characterized in that: The S21 includes the following steps, S211: Calculate the corresponding pore area according to the single slice of the data after processing, and extract the pore area ratio and curvature information as the condition information of data generation; S212: Compress the three-dimensional data according to the ratio, and compress the data dimension by k times through down-sampling, so as to provide feasible input for CGAN.

7. The coal rock data generation and permeability prediction method based on deep learning according to claim 6, characterized in that: The objective function of the CGAN is: where: is a generator, is a discriminator, is generated data, the data distribution of the generator is , the noise distribution is , is conditional information, the result of the noise z and the conditional information can be written as , the result of the data x on the discriminator D is .

8. The coal rock data generation and permeability prediction method based on deep learning according to any one of claims 1 to 4, characterized in that: The S32 includes the following steps, S321: Interval image selection is selected by slice according to the sequence information according to t step; S322: Extract the relevant feature information of the non-selected slice data, mainly including the extraction of the pore area ratio of the slice data, the extraction of the pore distribution of the slice data, and the construction of the sequence characteristic related feature vector according to the relative sequence position information of the slice.

Images