A pore type classification method and classification model for coal rock scanning electron microscope images

Abstract

This invention discloses a method and model for classifying pore types in coal and rock scanning electron microscope (SEM) images. The method includes: training a generative adversarial network (GAN) using a dual manifold optimization mechanism on a first sample set to obtain an image generation model; using the image generation model to augment a second sample set to establish a third sample set, where both the first and second sample sets contain SEM images of primary, metamorphic, mineral, and strain-induced pores in coal and rock; training an extreme learning machine (ELM) using a dynamic boundary smoothing mechanism on the third sample set to obtain a classification model; and determining the pore type of the coal and rock SEM images based on the classification model. This method introduces adversarial network training data to expand the classification model, solving the problems of insufficient training data and poor model generalization ability, while the dual manifold optimization mechanism maintains the topological structure of the image data; the use of an ELM with a dynamic boundary smoothing mechanism for pore type identification further improves classification accuracy.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and model for classifying pore types in coal and rock scanning electron microscope images. Background Technology

[0002] Identifying pores of different origins in coalbed methane reservoirs is of great significance. Currently, many scholars tend to classify pore types based on pore size. While this method provides some classification criteria, it neglects the structural characteristics of pore formation, making it difficult to reveal geological information about the coal-forming environment and its evolutionary history. The origin of pores in coal reservoirs is not only closely related to the coal-forming environment but also directly affects the storage and migration performance of coalbed methane. However, due to the complexity and diversity of pore types, efficient and automated identification of pores of different origins still faces technical bottlenecks. If advanced automatic identification technology can accurately classify the origin of pores, it can not only improve identification efficiency but also provide more comprehensive and accurate geological data for the exploration and evaluation of coalbed methane reservoirs, thereby significantly improving the scientific and economic efficiency of reservoir modeling and gas reservoir development. Therefore, accurately identifying the types and characteristics of pores is crucial for coalbed methane development and reservoir stimulation. In traditional coal reservoir porosity analysis, scanning electron microscopy (SEM) images, as a high-resolution characterization method, can provide detailed pore morphology characteristics. However, due to the diverse types of pores in coal reservoirs and the massive amount of data in scanning electron microscope (SEM) images, which are difficult to annotate, manually annotating pore types is extremely labor-intensive and prone to errors. Therefore, how to efficiently and accurately automate the identification of pore types has become a pressing technical challenge in this field.

[0003] Chinese invention patent CN119021675A proposes an automatic identification method for concealed karst pores based on downhole television cameras. Using high-definition downhole television camera data as the main source, and combining multi-parameter hydrological logging curves, core samples, and logging data cross-plots, the method employs image processing and computer vision techniques that integrate edge detection and corner detection algorithms to extract key features (three-dimensional spatial features such as the width, height, and distribution of karst caves or fissures) from the images. This method can statistically analyze the karst development at different depths and orientations of the borehole, significantly improving the accuracy and efficiency of karst development feature characterization and statistical analysis.

[0004] Chinese invention patent CN118097220A discloses a method and system for identifying voids in mineral-rock particle accumulation systems, relating to the field of image recognition technology. The method includes: acquiring original slice images of the mineral-rock particle accumulation system using CT scanning and preprocessing them; selecting different data parameters such as quantity, size, and content to create four types of datasets and using them to train a neural network; building an AttentionR2U-net and training the network using the datasets to obtain a model; qualitatively and quantitatively evaluating the model's recognition performance under the four datasets based on labeled and recognized images, and further determining the optimal dataset parameters by comparing binary confusion matrix indices; finally, using the optimal dataset parameters as a reference, creating the optimal dataset and training the neural network with it; repeatedly training the neural network and adjusting hyperparameters to obtain the optimal model; and verifying the convergence and accuracy of the model using qualitative and quantitative methods. This method improves the efficiency and accuracy of the neural network training and recognition process in the field of void identification in mineral-rock particle accumulation systems. Summary of the Invention

[0005] The inventors discovered in their work that existing technologies for identifying pore types and automatically recognizing image types in coal and gas reservoirs have many shortcomings:

[0006] 1. The main problem with existing technologies in coal-rock gas reservoir porosity identification is the lack of efficient automatic identification techniques for porosity of multiple origins. Current research mainly focuses on porosity classification based on pore size or morphology, but these methods cannot accurately reveal the genetic characteristics of porosity and are insufficient to provide key information related to coal-forming environments and reservoir gas migration capabilities. Furthermore, traditional identification methods rely on manual annotation or simple algorithmic processing, which is not only inefficient but also susceptible to subjective bias.

[0007] 2. Traditional image data augmentation methods often fail to fully preserve the topological structure of images, and the generated images may lack diversity and fail to effectively cover the various shapes and sizes of pores. This leads to excessive reliance on limited original data during model training, which can easily result in overfitting or underfitting.

[0008] 3. Existing dimensionality reduction methods often fail to fully consider the local structure and boundary effects of the data, which may result in the loss of important features or the introduction of outliers during the dimensionality reduction process, leading to unstable features after dimensionality reduction and affecting subsequent classification results.

[0009] 4. Traditional classification algorithms may not be able to effectively adjust the decision boundaries between categories when dealing with high-dimensional, imbalanced datasets, resulting in lower recognition accuracy for some categories. Furthermore, classification performance is easily affected when faced with uneven data distribution.

[0010] To at least partially solve the above problems, embodiments of the present invention provide a method and model for classifying pore types in coal and rock scanning electron microscope images. A generative adversarial network based on dual manifold optimization is used to expand the training data, solving the problems of insufficient data and poor model generalization ability. At the same time, an extreme learning machine classification algorithm based on dynamic boundary smoothing mechanism is used to identify pore types, further improving the model's ability to process complex geological data and its classification accuracy.

[0011] In a first aspect, embodiments of the present invention provide a method for classifying pore types in coal and rock scanning electron microscope images, including:

[0012] The pore type of coal and rock scanning electron microscope images is determined based on a classification model, which is established through the following steps:

[0013] Using the first sample set, the generator and discriminator of the generative adversarial network are trained by the dual manifold optimization mechanism to obtain the image generation model;

[0014] A second sample set is obtained, and the image is augmented using the image generation model to establish a third sample set. Both the first and second sample sets contain scanning electron microscope images of primary pores, metamorphic pores, mineral pores, and strain pores in coal and rock reservoirs.

[0015] Based on the third sample set, an extreme learning machine is trained using a dynamic boundary smoothing mechanism to obtain a classification model.

[0016] Optionally, the step of training the generator and discriminator of the generative adversarial network using the first sample set and a dual manifold optimization mechanism includes:

[0017] Initialize the weight parameters of the generator and discriminator of the generative adversarial network respectively;

[0018] A random noise vector is input into the generator, and the image generated by the generator and the first sample set are input into the discriminator to obtain the authenticity judgment result of the input image.

[0019] The objective function of the generator is determined by using a dual manifold optimization mechanism. The generator loss function is determined by maximizing the misclassification probability of the discriminator and minimizing the objective function. The discriminator loss function is determined by maximizing the sum of the correct classification of the real image and the misclassification of the generated image.

[0020] The generator's weight parameters are updated based on the gradient of the generator's loss function, and the discriminator's weight parameters are updated based on the gradient of the discriminator's loss function.

[0021] If the iteration termination condition is not met, return to the step of inputting a random noise vector into the generator.

[0022] Optionally, the step of determining the objective function of the generator using a dual manifold optimization mechanism includes:

[0023] Determine the difference between the random noise vector and the generated image, and use the square of the L2 norm of this difference as the first objective term;

[0024] The square of the F-norm of the product of this difference and the gradient of the random noise vector, multiplied by a coefficient that controls the importance of higher-order derivative terms, is used as the second objective term.

[0025] The sum of the first objective term and the second objective term, multiplied by the weight adjustment coefficient of the dual manifold, is used as the objective function of the generator.

[0026] Optionally, training the extreme learning machine based on the third sample set using a dynamic boundary smoothing mechanism includes:

[0027] Input the third sample set into the Extreme Learning Machine and initialize the weight parameters and biases of the Extreme Learning Machine;

[0028] The average within-class variance and average between-class distance are determined based on the initial classification results to obtain the initial smoothing parameters;

[0029] The boundary smoothing function is determined based on the current smoothing parameters and classification results. The new classification result is obtained from the boundary smoothing function, the weight parameters and biases of the extreme learning machine.

[0030] If the objective function is determined, and the iteration termination condition is not met, the weight parameters and biases are updated through backpropagation, the smoothing parameters are updated based on the current iteration number, and the process returns to the step of determining the boundary smoothing function based on the current smoothing parameters and classification results.

[0031] Optionally, obtaining the initial smoothing parameters includes:

[0032] Determine the ratio of the average within-class variance to the average between-class distance, and use the product of this ratio and the smoothing adjustment factor as the initial smoothing parameter.

[0033] Optionally, determining the boundary smoothing function based on the current smoothing parameters and classification results includes:

[0034] Based on the current smoothing parameters and classification results, the boundary smoothing function is determined using the following formula:

[0035]

[0036] Where, For boundary smoothing functions, For the input data of the Extreme Learning Machine, For the The smoothing parameters after the next iteration For the The mean vector of the class. For the The mean vector of the class. For the number of categories, It is an L2 norm.

[0037] Optionally, determining the optimization objective function includes:

[0038]

[0039]

[0040] Where, To optimize the objective function, The output data of the Extreme Learning Machine The target output of the Extreme Learning Machine For the regularization parameters of the extreme learning machine, These are the weight parameters of the Extreme Learning Machine. For multi-objective optimization terms, and These are the weighting factors for intra-class variance and inter-class distance, respectively. For the Class variance For the The mean vector of the class. It is the first The mean vector of the class. It is the number of categories. It is an L2 norm.

[0041] Optionally, updating the smoothing parameters based on the current iteration number includes:

[0042] Update the smoothing parameters based on the current iteration number using the following formula:

[0043]

[0044] Where, It is the first The smoothing parameters after the next iteration These are the initial smoothing parameters. t is the smooth decay rate, and t is the number of iterations.

[0045] Optionally, after establishing the third sample set, the method further includes:

[0046] A feature extraction model is obtained by training an autoencoder based on boundary smoothing using the third sample set; correspondingly,

[0047] The training of the extreme learning machine based on the third sample set using a dynamic boundary smoothing mechanism includes:

[0048] The fourth sample set is obtained from the output data of the feature extraction model, and the extreme learning machine is trained using a dynamic boundary smoothing mechanism.

[0049] Optionally, the autoencoder includes an encoder and a decoder, wherein the encoder is used to perform boundary smoothing on the decoder output data using the following smoothing function:

[0050]

[0051] in, It is a smooth function. It is the output data of the decoder. It is a smoothing intensity parameter; It is a smoothness threshold parameter.

[0052] Optionally, the smoothing intensity parameter is determined using an adaptive method based on local data density.

[0053] Optionally, training the boundary-smoothing-based autoencoder using the third sample set includes:

[0054] Input the third sample set into the autoencoder and initialize the weight parameters and biases of the encoder and decoder of the autoencoder respectively;

[0055] Based on the current weight parameters and biases of the encoder, the output data of the encoder is obtained through the activation function. Based on the current weight parameters and biases of the decoder, the output data of the decoder is obtained through the activation function. The encoder performs boundary smoothing processing on the output data of the decoder to obtain the final smooth output.

[0056] The autoencoder loss function is determined based on the final smoothed output and the weight coefficients of the autoencoder.

[0057] The weight parameters and biases of the encoder and decoder are updated through backpropagation of the loss function;

[0058] If the iteration termination condition is not met, return to the current weight parameters and biases of the encoder, and obtain the encoder output data through the activation function.

[0059] Optionally, the images include: primary pore SEM images, including intercellular pores and interstitial pores; metamorphic pore SEM images, including pores, shrinkage pores and interchain pores; and mineral pore SEM images, including dissolution pores, intergranular pores and mold pores.

[0060] Secondly, embodiments of the present invention provide a method for establishing a classification model of pore types in coal and rock scanning electron microscope images, including:

[0061] Using the first sample set, the generator and discriminator of the generative adversarial network are trained by the dual manifold optimization mechanism to obtain the image generation model;

[0062] A second sample set is obtained, and the image is augmented using the image generation model to establish a third sample set. Both the first and second sample sets contain scanning electron microscope images of primary pores, metamorphic pores, mineral pores, and strain pores in coal and rock reservoirs.

[0063] Based on the third sample set, an extreme learning machine is trained using a dynamic boundary smoothing mechanism to obtain a classification model, which is used to determine the pore type of the input coal and rock scanning electron microscope image.

[0064] Thirdly, embodiments of the present invention provide a pore type classification model for coal and rock scanning electron microscope (SEM) images, used to determine the pore type of input coal and rock SEM images. The classification model is established through the following steps:

[0065] Using the first sample set, the generator and discriminator of the generative adversarial network are trained by the dual manifold optimization mechanism to obtain the image generation model;

[0066] A second sample set is obtained, and the image is augmented using the image generation model to establish a third sample set. Both the first and second sample sets contain scanning electron microscope images of primary pores, metamorphic pores, mineral pores, and strain pores in coal and rock reservoirs.

[0067] Based on the third sample set, an extreme learning machine is trained using a dynamic boundary smoothing mechanism to obtain a classification model.

[0068] Fourthly, embodiments of the present invention provide a computer storage medium storing computer-executable instructions, which, when executed by a processor, implement any of the methods described above.

[0069] Fifthly, embodiments of this disclosure provide a server, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement any of the methods described above.

[0070] The beneficial effects of the above-described technical solutions provided in the embodiments of the present invention include at least the following:

[0071] (1) The coal and rock scanning electron microscope image pore type classification method provided in the embodiments of the present invention greatly reduces the time and error required for manual analysis through automated classification, improves the working efficiency of deep coal and rock reservoir exploration, and provides more scientific and efficient technical support for coal and rock reservoir exploration and reservoir gas migration characteristic analysis.

[0072] (2) The coal and rock scanning electron microscope image pore type classification method provided in the embodiments of the present invention uses a dual manifold optimized generative adversarial network to expand the sample of coal and rock reservoir scanning electron microscope images. Through the dual manifold optimization mechanism, the topological structure of the image data is maintained during the generation process, making the generated pore images more realistic in visual effect, and the generated data is closer to the original data in distribution. The dual manifold optimization adjusts the training process of the generator and the discriminator, making the generated images more diverse and consistent in category distribution and morphology.

[0073] (3) The coal and rock scanning electron microscope image pore type classification method provided in this embodiment of the invention adopts the extreme learning machine classification algorithm based on dynamic boundary smoothing mechanism to identify pore types. By dynamically adjusting the classification boundary, it solves the sensitivity of traditional classification methods to uneven distribution between categories when processing imbalanced data. It can adaptively adjust the classification boundary according to the characteristics of sample distribution, effectively avoid the problem of imbalance between class boundaries, and thus improve the accuracy of pore type identification.

[0074] (4) The coal and rock scanning electron microscope image pore type classification method provided in the embodiments of the present invention uses an autoencoder based on boundary smoothing to extract pore features. The autoencoder compresses high-dimensional data into low-dimensional space through the structure of encoder and decoder, and avoids outliers and boundary effects in the dimensionality reduction process through boundary smoothing strategy. Through adaptive adjustment of smoothing intensity, the model's adaptability to local data structure is enhanced, and the stability and accuracy of feature extraction are improved.

[0075] (5) The coal and rock scanning electron microscope image pore type classification method provided in the embodiments of the present invention can not only accurately distinguish multiple genetic types of pores such as primary pores, metamorphic pores, mineral pores and strain pores, but also further realize the identification of nine sub-types of genetic pores, providing a data basis for the correlation analysis between pore type and coal formation environment.

[0076] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings.

[0077] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0078] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0079] Figure 1 This is a flowchart of the coal and rock scanning electron microscope image pore type classification method in Embodiment 1 of the present invention;

[0080] Figure 2 for Figure 1 The detailed implementation flowchart of step S11 is shown below;

[0081] Figure 3 for Figure 1 The detailed implementation flowchart of step S13 is shown below;

[0082] Figure 4 This is a flowchart of the training process of the autoencoder in Embodiment 1 of the present invention;

[0083] Figures 5a-5i This is a typical pore type diagram of a coal and rock scanning electron microscope image in Embodiment 2 of the present invention. Detailed Implementation

[0084] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0085] It should be understood that the terminology used in this invention is merely for describing particular embodiments and is not intended to limit the invention. Furthermore, with respect to numerical ranges in this invention, it should be understood that each intermediate value between the upper and lower limits of the range is also specifically disclosed. Every smaller range between any stated value or intermediate value within a stated range, and any other stated value or intermediate value within said range, is also included in this invention. The upper and lower limits of these smaller ranges may be independently included or excluded from the range.

[0086] Unless otherwise stated, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. While only preferred methods and materials have been described herein, any methods and materials similar or equivalent to those described herein may be used in the implementation or testing of this invention. All references to this specification are incorporated by way of citation to disclose and describe methods and / or materials associated with those references. In the event of any conflict with any incorporated reference, the content of this specification shall prevail.

[0087] Example 1

[0088] Embodiment 1 of the present invention provides a method for classifying pore types in coal and rock scanning electron microscope images, the process of which is as follows: Figure 1 As shown, it includes the following steps:

[0089] Step S11: Using the first sample set, train the generator and discriminator of the generative adversarial network using the dual manifold optimization mechanism to obtain the image generation model.

[0090] A generative adversarial network with dual manifold optimization is used to augment the samples of scanning electron microscope images of coal and rock reservoirs. Through the dual manifold optimization mechanism, the topological structure of the image data is maintained during the generation process, making the generated pore images more visually realistic and the generated data more similar to the original data in distribution. The dual manifold optimization adjusts the training process of the generator and discriminator, making the generated images more diverse and consistent in terms of category distribution and morphology.

[0091] For details, see Figure 2 As shown, the training process for a generative adversarial network is as follows:

[0092] Step S111: Initialize the weight parameters of the generator and discriminator of the generative adversarial network respectively.

[0093] The method for initializing the weight parameters of the generator and discriminator is expressed as follows:

[0094]

[0095]

[0096] Where, and These represent the weight parameters of the generator and discriminator, respectively. This represents a normal distribution with a mean of 0 and a variance of 0.01.

[0097] Step S112: Input a random noise vector into the generator, and input the image generated by the generator and the first sample set into the discriminator to obtain the authenticity judgment result of the input image.

[0098] The initial input for generator training is a random noise vector generated by a uniform or Gaussian distribution. The image generation process is represented as follows:

[0099]

[0100] Where, Let be the input random noise vector, and let the dimension of the input noise vector be . ; For generator functions, based on weights Map noise to the image space; The generated image.

[0101] The image generated by the generator is input into the discriminator along with the real images in the first sample set to obtain the realism judgment result of the input image.

[0102] Step S113: Determine the objective function of the generator using a dual manifold optimization mechanism. Determine the generator loss function by maximizing the misclassification probability of the discriminator and minimizing the objective function. Determine the discriminator loss function by maximizing the sum of the correct classification of the real image and the misclassification of the generated image.

[0103] To maintain the inherent topological relationships of image data during the generation process, a dual manifold optimization mechanism is adopted, and an objective function is set to preserve the inherent structure and topological relationships of the generated data.

[0104] The process of determining the objective function includes: determining the difference between the random noise vector and the generated image; taking the square of the L2 norm of this difference as the first objective term; multiplying the square of the F-norm of the product of this difference and the gradient of the random noise vector by a coefficient controlling the importance of higher-order derivative terms as the second objective term; and multiplying the sum of the first and second objective terms by the weight adjustment coefficient of the dual manifold as the objective function of the generator. The formula is expressed as:

[0105]

[0106] Where, Let be the objective function. The weighting adjustment coefficients for the dual manifold. It is a random noise vector. For generator functions; This represents the L2 norm, used to measure the difference between the generated data and the original noise; To control the coefficients of the importance of higher-order derivative terms, The Frobenius norm is used to ensure that, during the data generation process, not only are the positions of the original data points preserved, but their local structure on the manifold is also protected. Represents the random noise vector Gradient operations.

[0107] The training objective of the generator is to minimize its loss function, which is expressed as:

[0108]

[0109] Where, Let represent the discriminator function, indicating that the generator generates realistic images by maximizing the probability of misclassification by the discriminator while minimizing . To preserve the topological structure of the image.

[0110] Furthermore, the training objective of the discriminator is to minimize its loss function, which is expressed as:

[0111]

[0112] Where, Represents real image data.

[0113] Step S114: Update the weight parameters of the generator based on the gradient of the generator loss function, and update the weight parameters of the discriminator based on the gradient of the discriminator loss function.

[0114] Dynamically adjusting the learning rates of the generator and discriminator based on the training progress helps to achieve rapid convergence in the early stages of training and fine-tune the generation effect in the later stages of training, as shown below:

[0115]

[0116]

[0117] Where, and The learning rates for the generator and discriminator are respectively. The initial learning rate, and These are the learning rate decay coefficients for the generator and discriminator, respectively. This represents the current training iteration number; and These are the feedback adjustment coefficients in the learning rate adjustment of the generator and discriminator, respectively. It evaluates the average generation quality of the generator in the most recent batch of samples. This quality is measured by the distance between the generated samples and the real samples, allowing the model to adaptively adjust the learning rate during training and optimize based on the quality of the generated samples.

[0118] The gradients calculated through backpropagation, combined with the updated learning rate, are used to optimize the weights of the generator and discriminator, allowing the network weights to gradually converge after multiple rounds of training. This enables the generator to produce more realistic images and the discriminator to more accurately distinguish between real and generated images.

[0119] Step S115: Determine whether the iteration termination condition is met.

[0120] The preset iteration termination condition can be reaching a preset maximum number of iterations; preferably, the preset maximum number of iterations is set to 1000.

[0121] If step S115 determines no, return to step S112; if step S115 determines yes, end the training of the generative adversarial network and use the trained generator as the image generation model.

[0122] Step S12: Obtain the second sample set, use the image generation model to augment the images, and establish the third sample set.

[0123] The first and second sample sets mentioned above both contain scanning electron microscope images of primary pores, metamorphic pores, mineral pores, and strain pores in coal reservoirs, which will be described in detail later.

[0124] After the image generation model is trained, the number of samples is expanded using the trained image generation model.

[0125] In one embodiment, if the original second sample set contains 800 samples, and after image augmentation using an image generation model, 200 samples are obtained, then the augmented third dataset contains 1000 samples.

[0126] Step S13: Based on the third sample set, the extreme learning machine is trained using a dynamic boundary smoothing mechanism to obtain the classification model.

[0127] An extreme learning machine classification algorithm based on dynamic boundary smoothing mechanism is used to identify pore categories. By dynamically adjusting the classification boundary, the sensitivity of traditional classification methods to uneven distribution between categories when dealing with imbalanced data is solved. It can adaptively adjust the classification boundary according to the characteristics of sample distribution, effectively avoiding the imbalance problem between class boundaries, thereby improving the accuracy of pore category identification.

[0128] For details, see Figure 3 As shown, the training process of the Extreme Learning Machine classification algorithm based on the dynamic boundary smoothing mechanism is as follows:

[0129] Step S131: Input the third sample set into the Extreme Learning Machine and initialize the weight parameters and biases of the Extreme Learning Machine.

[0130] The weights and biases of the extreme learning machine are initialized based on the dimensionality of the input data and the complexity of the classification task:

[0131]

[0132]

[0133] In the formula, These are the weights of the Extreme Learning Machine. It is the bias of the Extreme Learning Machine. The mean is 0 and the standard deviation is The normal distribution The number of input units, and Indicates the interval Uniform distribution on This indicates that it conforms to a specific distribution.

[0134] Step S132: Determine the average intra-class variance and average inter-class distance based on the initial classification results to obtain the initial smoothing parameters.

[0135] Determine the ratio of the average within-class variance to the average between-class distance, and use the product of this ratio and the smoothing adjustment factor as the initial smoothing parameter.

[0136] The initial smoothing parameter takes into account the dynamic changes of within-class variance and between-class distance, and is calculated as follows:

[0137]

[0138]

[0139]

[0140] In the formula, These are the initial smoothing parameters. It is a smoothing adjustment factor. It is the average within-class variance of all categories. It is the average inter-class distance. For the first Class variance For the number of categories, For the first The mean vector of the class. For the first The mean vector of the class.

[0141] Step S133: Determine the boundary smoothing function based on the current smoothing parameters and classification results. Obtain the new classification result from the boundary smoothing function, the weight parameters of the extreme learning machine, and the bias.

[0142] A dynamic boundary smoothing mechanism is employed to adjust the decision boundary through an adaptive function to handle the distribution characteristics of different classes. Based on the distribution density and overlap of each class, the classification boundary is dynamically adjusted to enhance the model's performance when handling multi-class, imbalanced data. The boundary smoothing function is expressed as:

[0143]

[0144] In the formula, It is a boundary smoothing function; It is The specific update method for the smoothing parameters after the next iteration will be described in detail later. It is the input data for the Extreme Learning Machine.

[0145] The Extreme Learning Machine (ELM) is trained using optimized weights and biases, along with adjusted classification boundaries. The classification mechanism achieved by the ELM is represented as follows:

[0146]

[0147] In the formula, It is the output of the Extreme Learning Machine. It is the input data matrix. It is an activation function, specifically the Sigmoid activation function; It follows a normal distribution; This represents the noise variance.

[0148] Furthermore, the mean vector represents the class center of the sample distribution, and is calculated based on the weighted average of samples within each class, expressed as:

[0149]

[0150] Where, Indicates the first Sample set of the class, It is a sample The weights are obtained based on the contribution rate of sample features in the previous classification.

[0151] Furthermore, the calculation method for within-class variance is expressed as follows:

[0152]

[0153] Where, For the Variance of the class.

[0154] Step S134: Determine the objective function for optimization.

[0155] Aligning the distribution of support vectors ensures a more reasonable and balanced decision boundary between different classes. Multi-objective optimization terms are used to optimize the balance and reasonableness of the decision boundary, reducing intra-class variance and maximizing inter-class distance, thereby improving classification accuracy and stability. The calculation method is expressed as follows:

[0156]

[0157] Where, For multi-objective optimization terms; and These are the weighting factors for intra-class variance and inter-class distance, respectively. For the Variance of the class.

[0158] The model is fine-tuned and optimized using a feedback mechanism to improve generalization ability and reduce overfitting. The goal is to ensure that the model maintains high recognition accuracy when faced with new and unseen data. The objective function for optimization is:

[0159]

[0160] Where, It is the total error of the Extreme Learning Machine, and the training objective is to make it... Minimize as much as possible;

[0161] It is the target output of the Extreme Learning Machine. This is the regularization parameter of the Extreme Learning Machine (ELM), designed to balance the model's fitting error and complexity, thereby improving the model's generalization performance. Preferably, Set it to 0.3.

[0162] Step S135: Determine whether the objective function satisfies the iteration termination condition.

[0163] If step S135 is determined to be negative, proceed to step S136; if step S135 is determined to be positive, end the training of the extreme learning machine and obtain a classification model for determining the pore type of the input coal and rock scanning electron microscope image.

[0164] Step S136: Update the weight parameters and biases through backpropagation, and update the smoothing parameters based on the current iteration number.

[0165] The smoothing parameter is updated as follows:

[0166]

[0167] It is the smooth decay rate, preferably set to 0.3.

[0168] After step S136, return to step S133.

[0169] In one embodiment, an adaptive noise injection mechanism is employed to dynamically inject noise into the input data during training, simulating various interferences that may be encountered in actual operation, thereby improving the model's adaptability to noise. The noise is calculated as follows:

[0170]

[0171] in, It is the noise standard deviation applied to each sample. It represents the baseline noise level, which is related to the overall noise characteristics of the dataset. It is a noise level that is dynamically adjusted based on the characteristics of the input samples.

[0172] Furthermore, the noise level is calculated based on the local density or variability of the sample, and is expressed as:

[0173]

[0174] In the formula, It is a noise modulation factor. It is data Local variance in its neighborhood; It is local density, calculated as data. The number of samples within the neighborhood. Preferably, Set to 2.

[0175] After adding noise to the input data, the classification result of the Extreme Learning Machine is represented as follows:

[0176]

[0177] The coal and rock scanning electron microscope image pore type classification method provided in Embodiment 1 of the present invention significantly reduces the time and error required for manual analysis through automated classification, improves the working efficiency of deep coal and rock reservoir exploration, and provides more scientific and efficient technical support for coal and rock reservoir exploration and reservoir gas migration characteristic analysis.

[0178] In one embodiment, the Extreme Learning Machine does not include a feature extraction layer. Instead, it uses a third sample set obtained after image augmentation to train an autoencoder based on boundary smoothing to obtain a feature extraction model. The output data of the feature extraction model is used to obtain a fourth sample set, which is then input into the Extreme Learning Machine and trained using a dynamic boundary smoothing mechanism to obtain a classification model.

[0179] An autoencoder consists of two parts: an encoder and a decoder. The encoder compresses the input data into a low-dimensional representation, and the decoder reconstructs this low-dimensional representation into a feature vector that satisfies the similarity condition with the input data, that is, it is as similar as possible to the input data.

[0180] For details, see Figure 4 As shown, the training process of the autoencoder is as follows:

[0181] Step S41: Input the third sample set into the autoencoder and initialize the weight parameters and biases of the encoder and decoder of the autoencoder respectively.

[0182] In one embodiment, the initialization method is represented as follows:

[0183]

[0184]

[0185]

[0186]

[0187] Where, Represents the encoder weights, Indicates the encoder bias; Indicates the weights of the decoder. Indicates the decoder bias; Represents a random function; It is the dimension of the encoder input layer. It is the dimension of the encoder's intermediate layer. It is the dimension of the encoder output layer; This is a function that generates a vector of all zeros.

[0188] Step S42: Based on the current weight parameters and bias of the encoder, the output data of the encoder is obtained through the activation function. Based on the current weight parameters and bias of the decoder, the output data of the decoder is obtained through the activation function. The encoder performs boundary smoothing processing on the output data of the decoder to obtain the final smooth output.

[0189] To handle outliers and boundary effects during dimensionality reduction, an automatic boundary smoothing strategy is adopted. This strategy enhances model robustness by dynamically adjusting boundary value processing. Specifically, it is implemented by smoothing the encoder output using a smoothing function, as shown below:

[0190]

[0191]

[0192] Where, It is the final smooth output of the encoder. It is a smooth function. It is a smoothing intensity parameter; It is an automatically adjusted smoothing threshold parameter used to control the degree of smoothing and improve adaptability to input data boundaries.

[0193] In one embodiment, the smoothing intensity parameter is determined using an adaptive method based on local data density, and the calculation method is expressed as follows:

[0194]

[0195] Where, calculate In its neighborhood Kernel density estimation in the model can accurately handle data in both high-density and sparse regions, improving the robustness and adaptability of the overall model. Indicates the neighborhood; Here is the kernel density estimation function; This is a function for taking the average value; This is a preset small constant. Preferably, Set to 0.001.

[0196] Step S43: Determine the autoencoder loss function based on the final smoothed output and the weight coefficients of the autoencoder.

[0197] During the training of the autoencoder, data is processed through forward propagation, and then the autoencoder parameters are updated using gradient descent via error backpropagation. The loss function of the autoencoder is calculated as follows:

[0198]

[0199]

[0200] Where, The loss function of the autoencoder; It is the input data of the self-encoder. It is the output reconstructed by the decoder. It is the Sigmoid activation function; It is an L2 norm; It is the output of the smoothed encoder; It is the first Weight coefficients of the dimensional autoencoder; This is the number of samples input into the autoencoder in the current batch.

[0201] In one embodiment, the weight coefficients of the autoencoder are used to adjust the contribution of each dimension of features to the error. These weights are dynamically assigned according to the importance of the features, and the calculation method is expressed as follows:

[0202]

[0203] Where, It is the first The feature variance of the input data of the 3D autoencoder; It is the first Characteristics of the input data of a 3D autoencoder; It is a small constant to avoid division by zero; It is the variance function.

[0204] Step S44: Update the weight parameters and biases of the encoder and decoder through backpropagation of the loss function.

[0205] The calculation method for the parameter update amount in each iteration is expressed as follows:

[0206]

[0207]

[0208]

[0209]

[0210] Where, This represents the amount of weight parameter update for the encoder; This is the amount of time the encoder's bias parameters are updated. This represents the amount of weight parameter update for the decoder; This represents the amount of time the decoder's bias parameters are updated. It is the first The learning rate of the autoencoder in the next iteration.

[0211] Furthermore, the autoencoder parameters are updated, and the update method is expressed as follows:

[0212]

[0213]

[0214]

[0215]

[0216] Where, This indicates a parameter update operation.

[0217] The training process of the autoencoder is optimized using a convex hull convergence strategy. By adjusting the learning step size and weight update strategy, the network converges quickly. The learning rate is dynamically adjusted based on the data distribution characteristics in the current iteration.

[0218]

[0219] Where, It is the learning rate of the initial autoencoder; It is the first The learning rate of the autoencoder in each iteration; It is the learning rate decay factor of the autoencoder. It is the first The average diameter of the convex hull in the next iteration It is the diameter of the initial convex hull. The learning rate decreases as the diameter of the convex hull decreases to prevent instability caused by excessively large step sizes in the later stages of learning.

[0220] Step S45: Determine whether the iteration termination condition is met.

[0221] The preset iteration termination condition can be reaching a preset maximum number of iterations; preferably, the preset maximum number of iterations is set to 1000.

[0222] If step S45 determines no, return to step S42; if step S45 determines yes, end the training of the autoencoder and obtain the feature extraction model for image feature extraction.

[0223] The coal and rock scanning electron microscope image pore type classification method provided in this embodiment of the invention uses an autoencoder based on boundary smoothing to extract pore features. The autoencoder compresses high-dimensional data into a low-dimensional space through the structure of the encoder and decoder, and avoids outliers and boundary effects in the dimensionality reduction process through the boundary smoothing strategy. By adaptively adjusting the smoothing intensity, the model's adaptability to the local structure of the data is enhanced, and the stability and accuracy of feature extraction are improved.

[0224] The above describes scanning electron microscope (SEM) images of primary pores, metamorphic pores, mineral pores, and strain pores in coal reservoirs, included in both the first and second sample sets. Further, the SEM images of primary pores include intercellular and interstitial pores; the SEM images of metamorphic pores include vesicles, shrinkage pores, and interchain pores; and the SEM images of mineral pores include dissolution pores, intergranular pores, and mold pores.

[0225] The data collection and labeling process is described below:

[0226] This invention relates to the acquisition of scanning electron microscope (SEM) images of coal and rock reservoirs, sourced from core samples from multiple coalbed methane wells in the same well area. The acquisition is performed using a field emission scanning electron microscope (FET), enabling the acquisition of high-resolution images. The acquired image data undergoes preprocessing, including noise reduction and contrast enhancement, to ensure data quality. All image data is stored in standard TIFF format. The acquired SEM image data is presented in 512... The image is stored at a resolution of 512 pixels, with each pixel using 16-bit depth to preserve the original spectral characteristics of the image, ensuring that every detail is accurately captured; at least 1,000 images are sampled to ensure that the images are clear and the pore features are typical; the collected data is labeled, and during the data labeling process, the pores are divided into four main genetic categories: primary pores, metamorphic pores, mineral pores, and strain pores.

[0227] Primary pores include intercellular pores and interstitial pores. Intercellular pores are formed by remnants of the cellular structure of coal-derived plants, usually resulting from incomplete decomposition of plant tissues during coalification. These pores retain the original anatomical features of coal-derived plants; they are generally larger than 10 μm, mostly round or elliptical, unevenly distributed, and commonly found in the vitrinite and filamentous tissues of low- to medium-rank coals. These pores are often filled with minerals. In high-rank coals, these pores become scarce and irregular in shape due to compaction and mineral filling during coalification. Interstitial pores are formed by the gaps between detrital particles in coal. These particles may originate from the deposition of detritus, minerals, or other heterogeneous components from coal-derived plants. These pores are typically larger than 1 μm in diameter and irregular in shape. In low- to medium-rank coals, interparticle pores are more developed, with larger diameters and higher concentrations; while in high-rank coals, the interparticle pores are affected by compaction, resulting in smaller diameters and fewer pores.

[0228] Metamorphic pores mainly include vesicles, contraction pores, and interchain pores. Vesicles are mainly formed during the coal metamorphism process by the generation and release of thermogenic gases. As the coal rank increases, thermogenic gases are continuously generated, and their discharge process leaves pores inside the organic matter. Development characteristics: Vesicles in low-rank coal have regular shapes and small diameters; while vesicles in high-rank coal can be divided into two categories: (1) residual vesicles, which are formed by the deformation of primary pores under pressure, and are mostly short linear; (2) secondary vesicles, which are formed by the later gas discharge or fracture action, and are mostly clustered, with round or elliptical shapes. Contraction pores are formed during the coalification process by the volume shrinkage of organic matter due to dehydration, degassing and condensation, as well as the separation of minerals and organic matter. Development characteristics: These pores are mainly concentrated in high-rank coal and distributed in the interface region between organic matter and minerals. The pore size is relatively large, belonging to mesopores (2-50 nm) and macropores (>50 nm). Complex in shape, often clustered and interconnected, these pores are the most important connecting pores in high-rank coals, significantly influencing coalbed methane migration. Interchain pores are formed by the condensation of organic matter molecular chains during coalification. These pores are mainly distributed in the organic matter of coal, reflecting structural changes at the molecular level. The pore size is primarily less than 10 nm, and the shape is irregular. With increasing coal rank, the pore size gradually decreases, while the concentration initially decreases and then increases. These pores make a significant contribution to the gas adsorption properties of coal.

[0229] Strain pores are pores in coal seams caused by relative displacement between particles, shear slip between minerals and organic matter, or localized compressive deformation under tectonic deformation or stress. They typically form in shear zones, compression-shear zones, or tensile fracture zones of coal seams, and are a direct result of external forces adjusting the internal structure of the coal seam. Strain pores exhibit diverse morphologies, primarily irregular shapes such as linear, crack-like, or arc-shaped. Pore diameters range from micrometers to millimeters and are controlled by the deformation intensity. In weakly deformed regions, strain pores are mainly micro-cracks, while in strongly deformed regions, large-scale slip cracks or even interconnected pores appear. Strain pores usually have high connectivity, but intense compaction or slippage can cause some pores to close. Furthermore, the distribution of strain pores shows a clear directionality, usually aligned with the direction of the tectonic stress field, which has a significant impact on fluid transport.

[0230] The identification and labeling of each type of pore were reviewed by domain experts to ensure accuracy. The basic characteristics of each type of pore are shown in Table 1 and... Figures 5a-5i As shown. Among them, Figure 5a Intercellular pores Figure 5b For inter-chip holes, Figure 5c For pores, Figure 5d For shrinkage holes and inter-chain holes, Figure 5e and Figure 5f For dissolution pits, Figure 5g Intercrystalline pores Figure 5h For casting mold holes, Figure 5i It is a strain gauge hole.

[0231] Table 1. Types, origins, and developmental characteristics of pores of different origins in coal and rock.

[0232]

[0233]

[0234] Example 2

[0235] Embodiment 2 of the present invention provides a specific application of a method for classifying pore types in coal and rock scanning electron microscope images.

[0236] Core sample X, with a total porosity of 6.91%, was obtained from the coal-gas reservoir of well A1 in the xx basin. High-resolution SEM images were acquired using field emission scanning electron microscopy (FE-SEM). The acquired images had a resolution of 512×512 pixels and used 16-bit grayscale depth to preserve pore detail. A total of 3212 FE-SEM images were obtained. All samples underwent preprocessing, including noise reduction and contrast enhancement, to ensure image quality met analytical requirements. The acquired images were annotated by domain experts, classifying porosity into four main genetic types: primary pores (intercellular and interclastic pores), metamorphic pores (vesicle pores, shrinkage pores, and interchain pores), mineral pores (dissolution pores, intergranular pores, and molding pores), and strain pores. Annotation referenced pore morphology, size, and development characteristics to ensure accurate classification. Due to the limited number of samples, a generative adversarial network (GAN) based on dual manifold optimization was used to expand the sample size. The expanded dataset covers different categories of pore morphology and size. The generated samples are visually consistent with the original data while significantly increasing data diversity. Dimensionality reduction is performed on the expanded dataset using an autoencoder based on boundary smoothing. The encoder compresses high-dimensional features to a low-dimensional representation, and the decoder reconstructs the features to ensure information integrity. A boundary smoothing strategy is also introduced to reduce the impact of outliers on model performance. The dimensionality-reduced features are used for subsequent classification model training. An Extreme Learning Machine (ELM) based on dynamic boundary smoothing is used to classify pore categories. During training, the model dynamically adjusts the classification boundary, optimizing intra-class variance and inter-class distance, significantly improving classification accuracy and generalization ability. The final classification results are shown in Table 2, with an error within ±0.3% compared to human recognition results.

[0237] Table 2. Comparison of automatic and manual porosity identification results for sample X from well A1 in the xx basin.

[0238]

[0239] Based on the above results, it was found that in sample X1, porosity of different genetic types exhibited significant distribution characteristics and proportions. Primary porosity (intercellular and interdetrital pores) in sample X1 had a low porosity (0.06%-0.02%) due to compaction and mineral filling during coalification, and its porosity was relatively small. The porosity identified automatically and manually was close, with small errors (absolute errors of 0.01% and 0.06%, respectively), indicating that it was largely undeveloped in sample X1.

[0240] Metamorphic pores, including vesicles, contraction pores, and interchain pores, are the most prevalent pore types in sample X1. Vesicles exhibit a high porosity (0.42%) and relatively large porosity (2.90% automatically identified, 3.20% manually identified), demonstrating good connectivity, but with a high error rate (0.30% absolute error), reflecting their complexity. Contraction pores and interchain pores show moderate porosity (0.10%), with similar porosity values ​​identified automatically and manually (0.01% absolute error for contraction pores, 0.29% absolute error for interchain pores). Interchain pores, especially within the coal seam, exhibit microporous characteristics and are crucial for adsorbing coalbed methane.

[0241] Mineral porosity (dissolution pores, intergranular pores, and mold pores) was relatively low in sample X1, with a porosity of 0.08%–0.12%. Intergranular pores and dissolution pores exhibited relatively high porosity (0.83% and 0.55%, respectively), indicating that these pores may be well-developed in local reservoir regions and possess some storage and fluid transport capabilities. Mold pores had a porosity and porosity of 0, indicating that they were essentially undeveloped in sample X1.

[0242] Strain porosity (such as structural deformation porosity) is particularly prominent in the tectonic zone, with a porosity of 0.10% and relatively high porosity (0.69%-0.50%). The absolute error between automatic and manual identification is 0.19%, reflecting the modification effect of tectonic stress on the coal reservoir porosity in sample X1. These pores typically have good connectivity and serve as important channels for gas migration.

[0243] Overall, the X1 sample exhibits predominantly metamorphic pores, with some mineral and strain pores developing in localized areas, while primary pores are largely undeveloped. This pore characteristic is of significant importance for the storage and migration of coalbed methane.

[0244] This invention comprehensively utilizes multiple techniques, including data acquisition and annotation, data augmentation, feature extraction, and deep learning classifiers, to achieve automatic identification of porosity in coal-rock gas reservoirs with multiple genetic origins. By introducing generative adversarial networks to expand training data, the problems of insufficient data and poor model generalization ability are addressed. An ecosystem-based neural network parameter optimization method is employed to improve the model's training efficiency and accuracy. Furthermore, the combination of a nonlinear latent mapping autoencoder and an improved quantum-encoded high-order neural network classifier further enhances the model's ability to process complex geological data and its classification accuracy.

[0245] Based on the inventive concept of this invention, embodiments of this invention also provide a method for establishing a classification model of pore types in coal and rock scanning electron microscope images, including:

[0246] Using the first sample set, the generator and discriminator of the generative adversarial network are trained by the dual manifold optimization mechanism to obtain the image generation model;

[0247] A second sample set is obtained, and the image is augmented using the image generation model to establish a third sample set. Both the first and second sample sets contain scanning electron microscope images of primary pores, metamorphic pores, mineral pores, and strain pores in coal and rock reservoirs.

[0248] Based on the third sample set, an extreme learning machine is trained using a dynamic boundary smoothing mechanism to obtain a classification model, which is used to determine the pore type of the input coal and rock scanning electron microscope image.

[0249] Based on the inventive concept of this invention, embodiments of this invention also provide a classification model for pore types in coal and rock scanning electron microscope (SEM) images. This classification model is used to determine the pore type of an input coal and rock SEM image, and is established through the following steps:

[0250] Using the first sample set, the generator and discriminator of the generative adversarial network are trained by the dual manifold optimization mechanism to obtain the image generation model;

[0251] A second sample set is obtained, and the image is augmented using the image generation model to establish a third sample set. Both the first and second sample sets contain scanning electron microscope images of primary pores, metamorphic pores, mineral pores, and strain pores in coal and rock reservoirs.

[0252] Based on the third sample set, an extreme learning machine is trained using a dynamic boundary smoothing mechanism to obtain a classification model.

[0253] The method for establishing the above classification model has been described in detail in Example 1, and will not be elaborated here.

[0254] Based on the inventive concept of the present invention, embodiments of the present invention also provide a computer storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.

[0255] Based on the inventive concept of the present invention, embodiments of the present invention also provide a server, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the above-described method.

[0256] Unless otherwise specifically stated, terms such as processing, calculation, operation, determination, display, etc., may refer to the actions and / or processes of one or more processing or computing systems or similar devices that represent the manipulation and conversion of data representing physical (e.g., electronic) quantities within the registers or memory of the processing system into other data similarly representing physical quantities within the memory, registers, or other such information storage, transmission, or display devices of the processing system. Information and signals can be represented using any of a variety of different techniques and methods. For example, data, instructions, commands, information, signals, bits, symbols, and chips mentioned throughout the above description can be represented by voltage, current, electromagnetic waves, magnetic fields or particles, light fields or particles, or any combination thereof.

[0257] It should be understood that the specific order or hierarchy of steps in the disclosed process is an example of an exemplary method. Based on design preferences, it should be understood that the specific order or hierarchy of steps in the process may be rearranged without departing from the scope of this disclosure. The appended method claims provide elements of various steps in an exemplary order and are not intended to limit the scope to the specific order or hierarchy described.

[0258] In the detailed description above, various features are combined together in a single embodiment to simplify this disclosure. This approach to disclosure should not be construed as reflecting an intention that embodiments of the claimed subject matter require more features than those stated in each claim. Rather, as reflected in the appended claims, the invention is presented with fewer features than all of the features in a single disclosed embodiment. Therefore, the appended claims are hereby clearly incorporated into the detailed description, wherein each claim stands alone as a preferred embodiment of the invention.

[0259] Those skilled in the art will also understand that the various illustrative logic blocks, modules, circuits, and algorithm steps described in conjunction with the embodiments herein can be implemented as electronic hardware, computer software, or a combination thereof. To clearly illustrate the interchangeability between hardware and software, the various illustrative components, blocks, modules, circuits, and steps described above are generally described in terms of their functionality. Whether such functionality is implemented as hardware or software depends on the specific application and the design constraints imposed on the overall system. Those skilled in the art can implement the described functionality in alternative ways for each specific application; however, such implementation decisions should not be construed as departing from the scope of this disclosure.

[0260] The steps of the methods or algorithms described in conjunction with the embodiments herein can be directly embodied in hardware, software modules executed by a processor, or a combination thereof. The software modules can reside in RAM memory, flash memory, ROM memory, EPROM memory, EEPROM memory, registers, hard disks, removable disks, CD-ROMs, or any other form of storage medium well known in the art. An exemplary storage medium is connected to the processor, enabling the processor to read information from and write information to the storage medium. Of course, the storage medium can also be a component of the processor. The processor and storage medium can reside in an ASIC. The ASIC can reside in a user terminal. Alternatively, the processor and storage medium can exist as discrete components in the user terminal.

[0261] For software implementation, the techniques described in this application can be implemented using modules (e.g., procedures, functions, etc.) that perform the functions described in this application. This software code can be stored in memory units and executed by a processor. The memory units can be implemented within the processor or outside the processor; in the latter case, they are communicatively coupled to the processor via various means, as is well known in the art.

[0262] The foregoing description includes examples of one or more embodiments. It is certainly impossible to describe all possible combinations of components or methods in order to describe the above embodiments, but those skilled in the art will recognize that further combinations and arrangements of the various embodiments are possible. Therefore, the embodiments described herein are intended to cover all such changes, modifications, and variations that fall within the scope of the appended claims. Furthermore, the term “comprising” as used in the specification or claims is interpreted in a manner similar to the term “including,” as it is understood when used as a conjunction in the claims. Additionally, the use of any term “or” in the specification of the claims is intended to mean “non-exclusive or.” The terms “first,” “second,” etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

Claims

1. A method for classifying pore types in coal and rock scanning electron microscope images, characterized in that, include: The pore type of coal and rock scanning electron microscope images is determined based on a classification model, which is established through the following steps: Using the first sample set, the generator and discriminator of the generative adversarial network are trained by the dual manifold optimization mechanism to obtain the image generation model; A second sample set is obtained, and the image generation model is used to augment the images to establish a third sample set. Both the first and second sample sets contain scanning electron microscope (SEM) images of primary pores, metamorphic pores, mineral pores, and strain pores of coal and rock reservoir genesis. The primary pore SEM images include intercellular pores and interclastic pores. The metamorphic pore SEM images include vesicles, shrinkage pores, and interchain pores. The mineral pore SEM images include dissolution pores, intergranular pores, and mold pores. Input the third sample set into the Extreme Learning Machine (ELM) to initialize its weight parameters and biases; determine the average intra-class variance and average inter-class distance based on the initialized classification results to obtain the initial smoothing parameters; determine the boundary smoothing function based on the current smoothing parameters and classification results; and obtain the new classification results from the boundary smoothing function, the weight parameters and biases of the ELM. Determine the objective function for optimization. If the objective function does not meet the iteration termination condition, update the weight parameters and biases through backpropagation, update the smoothing parameters based on the current iteration number, and return to execute the step of determining the boundary smoothing function based on the current smoothing parameters and classification results. If the optimization objective function satisfies the iteration termination condition, a classification model is obtained; The step of determining the boundary smoothing function based on the current smoothing parameters and classification results includes: Based on the current smoothing parameters and classification results, the boundary smoothing function is determined using the following formula: ; In the formula, f boundary ( ) represents the boundary smoothing function, X u For the input data of the Extreme Learning Machine, Let be the smoothing parameter after the t-th iteration. Let be the mean vector of the i-th class. Let be the mean vector of the j-th class, and k be the number of classes. It is an L2 norm; The determination of the optimization objective function includes: ; In the formula, E u To optimize the objective function, Y u For the output data of the Extreme Learning Machine, T u The target output of the Extreme Learning Machine W is the regularization parameter for the Extreme Learning Machine. u L represents the weight parameters of the Extreme Learning Machine. align For multi-objective optimization terms, and These are the weighting factors for intra-class variance and inter-class distance, respectively. Let be the variance of the i-th class. Let be the mean vector of the i-th class. Let be the mean vector of the j-th class, and k be the number of classes. It is an L2 norm; The process of obtaining the classification model also includes using an adaptive noise injection mechanism to dynamically inject noise into the input data during training. The noise standard deviation applied to each sample is equal to the sum of the base noise level and the amount of noise dynamically adjusted based on the characteristics of the input sample.

2. The method as described in claim 1, characterized in that, The step of training the generator and discriminator of the generative adversarial network using the first sample set and a dual manifold optimization mechanism includes: Initialize the weight parameters of the generator and discriminator of the generative adversarial network respectively; A random noise vector is input into the generator, and the image generated by the generator and the first sample set are input into the discriminator to obtain the authenticity judgment result of the input image. The objective function of the generator is determined by using a dual manifold optimization mechanism. The generator loss function is determined by maximizing the misclassification probability of the discriminator and minimizing the objective function. The discriminator loss function is determined by maximizing the sum of the correct classification of the real image and the misclassification of the generated image. The generator's weight parameters are updated based on the gradient of the generator's loss function, and the discriminator's weight parameters are updated based on the gradient of the discriminator's loss function. If the iteration termination condition is not met, return to the step of inputting a random noise vector into the generator.

3. The method as described in claim 2, characterized in that, The objective function of the generator is determined using a dual manifold optimization mechanism, including: Determine the difference between the random noise vector and the generated image, and use the square of the L2 norm of this difference as the first objective term; The square of the F-norm of the product of this difference and the gradient of the random noise vector, multiplied by a coefficient that controls the importance of higher-order derivative terms, is used as the second objective term. The sum of the first objective term and the second objective term, multiplied by the weight adjustment coefficient of the dual manifold, is used as the objective function of the generator.

4. The method as described in claim 1, characterized in that, The process of obtaining the initial smoothing parameters includes: Determine the ratio of the average within-class variance to the average between-class distance, and use the product of this ratio and the smoothing adjustment factor as the initial smoothing parameter.

5. The method as described in claim 1, characterized in that, The update of the smoothing parameters based on the current iteration number includes: Update the smoothing parameters based on the current iteration number using the following formula: ; Where, Let be the smoothing parameter after the t-th iteration. These are the initial smoothing parameters. t is the smooth decay rate, and t is the number of iterations.

6. The method as described in claim 1, characterized in that, After establishing the third sample set, the following is also included: A feature extraction model is obtained by training an autoencoder based on boundary smoothing using the third sample set; correspondingly, Based on the third sample set, an extreme learning machine is trained using a dynamic boundary smoothing mechanism, including: The fourth sample set is obtained from the output data of the feature extraction model, and the extreme learning machine is trained using a dynamic boundary smoothing mechanism.

7. The method as described in claim 6, characterized in that, The autoencoder comprises an encoder and a decoder. The encoder is used to perform boundary smoothing on the decoder output data using the following smoothing function: ; in, ( ) is a smoothing function. It is the output data of the decoder. It is a smoothing intensity parameter; It is a smoothness threshold parameter.

8. The method as described in claim 7, characterized in that, The smoothing intensity parameter is determined using an adaptive method based on local data density.

9. The method as described in claim 7, characterized in that, The step of training a boundary-smoothing-based autoencoder using the third sample set includes: Input the third sample set into the autoencoder and initialize the weight parameters and biases of the encoder and decoder of the autoencoder respectively; Based on the current weight parameters and biases of the encoder, the output data of the encoder is obtained through the activation function. Based on the current weight parameters and biases of the decoder, the output data of the decoder is obtained through the activation function. The encoder performs boundary smoothing processing on the output data of the decoder to obtain the final smooth output. The autoencoder loss function is determined based on the final smoothed output and the weight coefficients of the autoencoder. The weight parameters and biases of the encoder and decoder are updated through backpropagation of the loss function; If the iteration termination condition is not met, return to the current weight parameters and biases of the encoder, and obtain the encoder output data through the activation function.

10. A method for establishing a classification model of pore types in coal and rock scanning electron microscope images, characterized in that, include: Using the first sample set, the generator and discriminator of the generative adversarial network are trained by the dual manifold optimization mechanism to obtain the image generation model; A second sample set is obtained, and the image generation model is used to augment the images to establish a third sample set. Both the first and second sample sets contain scanning electron microscope (SEM) images of primary pores, metamorphic pores, mineral pores, and strain pores of coal and rock reservoir genesis. The primary pore SEM images include intercellular pores and interclastic pores. The metamorphic pore SEM images include vesicles, shrinkage pores, and interchain pores. The mineral pore SEM images include dissolution pores, intergranular pores, and mold pores. Input the third sample set into the Extreme Learning Machine (ELM) to initialize its weight parameters and biases; determine the average intra-class variance and average inter-class distance based on the initialized classification results to obtain the initial smoothing parameters; determine the boundary smoothing function based on the current smoothing parameters and classification results; and obtain the new classification results from the boundary smoothing function, the weight parameters and biases of the ELM. The objective function is determined. If the objective function does not meet the iteration termination condition, the weight parameters and biases are updated through backpropagation, and the smoothing parameters are updated based on the current iteration number. The process is then returned to execute the step of determining the boundary smoothing function based on the current smoothing parameters and classification results. If the objective function meets the iteration termination condition, a classification model is obtained, which is used to determine the pore type of the input coal and rock scanning electron microscope image. The step of determining the boundary smoothing function based on the current smoothing parameters and classification results includes: Based on the current smoothing parameters and classification results, the boundary smoothing function is determined using the following formula: ; In the formula, f boundary ( ) represents the boundary smoothing function, X u For the input data of the Extreme Learning Machine, Let be the smoothing parameter after the t-th iteration. Let be the mean vector of the i-th class. Let be the mean vector of the j-th class, and k be the number of classes. It is an L2 norm; The determination of the optimization objective function includes: ; In the formula, E u To optimize the objective function, Y u For the output data of the Extreme Learning Machine, T u The target output of the Extreme Learning Machine W is the regularization parameter for the Extreme Learning Machine. u L represents the weight parameters of the Extreme Learning Machine. align For multi-objective optimization terms, and These are the weighting factors for intra-class variance and inter-class distance, respectively. Let be the variance of the i-th class. Let be the mean vector of the i-th class. Let be the mean vector of the j-th class, and k be the number of classes. It is an L2 norm; The process of obtaining the classification model also includes using an adaptive noise injection mechanism to dynamically inject noise into the input data during training. The noise standard deviation applied to each sample is equal to the sum of the base noise level and the amount of noise dynamically adjusted based on the characteristics of the input sample.

11. A classification model for pore type in coal and rock scanning electron microscopy images, characterized in that, The classification model is used to determine the pore type of the input coal and rock scanning electron microscope image. The classification model is established through the following steps: Using the first sample set, a dual manifold optimization mechanism is employed to train the generator and discriminator of the generative adversarial network, thus obtaining an image generation model. A second sample set is obtained, and the image generation model is used to augment the images to establish a third sample set. Both the first and second sample sets contain scanning electron microscope (SEM) images of primary pores, metamorphic pores, mineral pores, and strain pores of coal and rock reservoir genesis. The primary pore SEM images include intercellular pores and interclastic pores. The metamorphic pore SEM images include vesicles, shrinkage pores, and interchain pores. The mineral pore SEM images include dissolution pores, intergranular pores, and mold pores. Input the third sample set into the Extreme Learning Machine (ELM) to initialize its weight parameters and biases; determine the average intra-class variance and average inter-class distance based on the initialized classification results to obtain the initial smoothing parameters; determine the boundary smoothing function based on the current smoothing parameters and classification results; and obtain the new classification results from the boundary smoothing function, the weight parameters and biases of the ELM. Determine the objective function for optimization. If the objective function does not meet the iteration termination condition, update the weight parameters and biases through backpropagation, update the smoothing parameters based on the current iteration number, and return to execute the step of determining the boundary smoothing function based on the current smoothing parameters and classification results. If the optimization objective function satisfies the iteration termination condition, a classification model is obtained; The step of determining the boundary smoothing function based on the current smoothing parameters and classification results includes: Based on the current smoothing parameters and classification results, the boundary smoothing function is determined using the following formula: ; In the formula, f boundary ( ) represents the boundary smoothing function, X u For the input data of the Extreme Learning Machine, Let be the smoothing parameter after the t-th iteration. Let be the mean vector of the i-th class. Let be the mean vector of the j-th class, and k be the number of classes. It is an L2 norm; The determination of the optimization objective function includes: ; In the formula, E u To optimize the objective function, Y u For the output data of the Extreme Learning Machine, T u The target output of the Extreme Learning Machine W is the regularization parameter for the Extreme Learning Machine. u L represents the weight parameters of the Extreme Learning Machine. align For multi-objective optimization terms, and These are the weighting factors for intra-class variance and inter-class distance, respectively. Let be the variance of the i-th class. Let be the mean vector of the i-th class. Let be the mean vector of the j-th class, and k be the number of classes. It is an L2 norm; The process of obtaining the classification model also includes using an adaptive noise injection mechanism to dynamically inject noise into the input data during training. The noise standard deviation applied to each sample is equal to the sum of the base noise level and the amount of noise dynamically adjusted based on the characteristics of the input sample.

12. A computer storage medium, characterized in that, The computer storage medium stores computer-executable instructions, which, when executed by a processor, implement the method described in any one of claims 1 to 10.

13. A server, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the method of any one of claims 1 to 10.

Images