Training method, recognition method and device of coal rock maceral recognition neural network model

Abstract

The invention discloses a coal rock maceral recognition neural network model training method, recognition method and device, and the method comprises the steps: obtaining a training sample set, and enabling each sample in the training sample set to comprise a coal rock microscopic image and a maceral sub-classification label marked by the coal rock microscopic image; using samples in the training sample set to train a deep neural network model containing one or more dynamic polymerization layers, after the coal rock microscopic image is input through an input layer of the deep neural network model, extracting microscopic component features of vitrinite and / or inertinite in the coal rock microscopic image through the dynamic polymerization layers, comparing a maceral sub-classification prediction label of a coal rock microscopic image output by an output layer of the deep neural network model with a maceral sub-classification label in a sample so as to estimate model parameters in the deep neural network model and then obtain a coal rock maceral identification neural network model; the problems that an existing method is high in subjectivity, low in efficiency, poor in repeatability and the like are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of microscopic image data processing, and in particular to a training method, an identification method and a device for a neural network model for identifying microscopic components of coal and rock. Background Art

[0002] The identification and classification of coal rock microscopic components is an important research topic in fields such as coal petrology, sedimentology, and energy geology. It plays a key role in coal genesis analysis, reservoir evaluation, combustion performance prediction, and the development of coalbed methane and shale gas resources. Traditionally, the identification of coal rock microscopic components relies mainly on manual microscopic identification. Researchers observe coal rock thin sections using equipment such as polarizing microscopes and fluorescence microscopes, and make judgments based on experience. However, this method has drawbacks such as strong subjectivity, long time consumption, and poor repeatability. Especially when dealing with large-scale samples or deep coal rocks with complex microstructures, it is easy to lead to a decrease in identification accuracy, which ultimately affects the stability and reliability of the analysis results.

[0003] In recent years, with the development of computer vision and artificial intelligence technologies, automated maceral identification based on deep learning has become a new trend in coal petrology research. By learning features from microscopic images using neural network models, vitrinite, inertinite, and their subcomponents can be automatically identified, reducing manual intervention and improving the objectivity and consistency of analysis. In particular, in the evaluation of deep coal gas and unconventional oil and gas reservoirs, efficient and accurate maceral analysis is crucial for predicting reservoir properties, assessing coalbed methane content, and optimizing coal utilization strategies. Therefore, developing intelligent, high-precision, and automated maceral identification methods will not only significantly improve research efficiency but also promote the intelligent and data-driven development of coal petrology, energy geology, and mineralogy.

[0004] With the continuous development of microscopy and image processing technologies, the dimensionality and complexity of microscopic image data have shown an exponential growth. Especially in fields such as mineralogy, materials science, and biology, microscopic image data often contain a large amount of detailed information. Traditional microscopic image data processing methods usually rely on fixed-structure neural networks or traditional feature extraction algorithms. However, when faced with highly complex and diverse image data, these methods often exhibit insufficient feature extraction capabilities, over-reliance on local features, or difficulty in capturing nonlinear relationships, resulting in poor accuracy and stability of the processing results. Summary of the Invention

[0005] The present invention is made in order to solve the problems of strong subjectivity, low efficiency and poor repeatability in the traditional method of identifying microscopic components of coal and rock.

[0006] As one aspect of the present invention, an embodiment of the present invention provides a method for training a neural network model for identifying microscopic components of coal and rock, which may include:

[0007] Acquire a training sample set, where each sample in the training sample set includes a coal rock microscopic image and a microscopic component sub-classification label annotated on the coal rock microscopic image;

[0008] A deep neural network model comprising one or more dynamic aggregation layers is trained using samples in the training sample set, wherein, after the coal rock microscopic image is input through the input layer of the deep neural network model, the microscopic component characteristics of the vitrinite and / or inertinite in the coal rock microscopic image are extracted through the dynamic aggregation layer, and the microscopic component sub-classification prediction labels of the coal rock microscopic image output by the output layer of the deep neural network model are compared with the microscopic component sub-classification labels in the sample, so as to estimate the model parameters in the deep neural network model and obtain a coal rock microscopic component identification neural network model.

[0009] In one embodiment, obtaining the training sample set may include:

[0010] Acquire coal rock microscopic images in multiple imaging modes at the same location of the coal rock sample; wherein the multiple imaging modes include: single polarization imaging mode, reflected light imaging mode and fluorescence spectrum imaging mode;

[0011] Classifying and labeling the microscopic component characteristics of the vitrinite and / or inertinite in the coal rock microscopic image, and cross-validating the single polarization labeling results using the reflected light characteristics and fluorescence spectrum characteristics of the coal rock microscopic image to determine the sub-classification labels of the microscopic components annotated in the coal rock microscopic image;

[0012] Each coal rock microscopic image and its annotated microscopic component sub-classification label are taken as a sample to construct a training sample set.

[0013] In another embodiment, determining the microscopic component subclassification label annotated on the coal rock microscopic image may include: annotating the microscopic component subclassification in each coal rock microscopic image that accounts for the largest proportion of the area of ​​the coal rock microscopic image as the unique microscopic component subclassification label of the coal rock microscopic image.

[0014] In another embodiment, each coal rock microscopic image and its annotated microscopic component sub-classification label is taken as a sample. Before constructing the training sample set, the following steps may be further included:

[0015] The image areas other than the coal rock microscopic image corresponding to the unique microscopic component subclassification in each coal rock microscopic image are eliminated.

[0016] In another embodiment, the deep neural network model includes a dynamic aggregation layer, in which the forward propagation dynamic aggregation layer extracts the microscopic component features of the vitrinite and / or inertinite in the coal rock microscopic image to extract the microscopic component feature vector of the coal rock microscopic image, and performs feature aggregation based on the microscopic component feature vector to obtain an aggregated microscopic combination feature vector, and determines a holographic projection output based on the microscopic component feature vector; the aggregated microscopic combination feature vector and the holographic projection output serve as inputs for the next iteration; the backward propagation dynamic aggregation layer calculates the gradient of the weight matrix and the dynamic aggregation weight coefficient based on the loss function; the neural network model determines the loss function based on the aggregated microscopic component feature vector, the holographic projection output, the microscopic component sub-classification prediction label output by the model, and the microscopic component sub-classification label in the sample to train the dynamic aggregation weight coefficient and weight matrix in the dynamic aggregation layer to obtain a coal rock microscopic component identification neural network model;

[0017] When the deep neural network model includes multiple dynamic aggregation layers, the first dynamic aggregation layer is forward propagated to extract the microscopic component features of the vitrinite and / or inertinite group in the coal rock microscopic image to extract the microscopic component feature vector of the coal rock microscopic image, the first dynamic aggregation layer performs feature aggregation based on the microscopic component feature vector to obtain the aggregated microscopic combination feature vector, and determines the holographic projection output based on the microscopic component feature vector; the microscopic component feature vector output by the first dynamic aggregation layer is used as the input of the next dynamic aggregation layer, and the aggregated microscopic combination feature vector and the holographic projection output output by the first dynamic aggregation layer are used as the input of the next iteration; each dynamic aggregation layer is back-propagated to calculate the gradient of the weight matrix and the dynamic aggregation weight coefficient based on the loss function; the neural network model determines the loss function based on the aggregated microscopic component feature vector, the holographic projection output, the microscopic component sub-classification prediction label output by the model and the microscopic component sub-classification label in the sample to train the dynamic aggregation weight coefficient and weight matrix in the dynamic aggregation layer to obtain a coal rock microscopic component identification neural network model.

[0018] As another aspect of the present invention, an embodiment of the present invention provides a method for identifying coal rock microscopic components based on a dynamic aggregation strategy, which may include:

[0019] Inputting the coal rock microscopic image into a pre-trained coal rock microscopic component recognition neural network model to identify the microscopic component subclassification of the coal rock microscopic image;

[0020] The coal rock micro-component identification neural network model is pre-trained according to the above-mentioned coal rock micro-component identification neural network model training method.

[0021] As another aspect of the present invention, an embodiment of the present invention provides a training device for a neural network model for identifying microscopic components of coal and rock, which may include:

[0022] An acquisition module is used to acquire a training sample set, where each sample in the training sample set includes a coal rock microscopic image and a microscopic component sub-classification label annotated on the coal rock microscopic image;

[0023] A training module is used to train a deep neural network model containing one or more dynamic aggregation layers using samples in the training sample set, wherein, after the coal rock microscopic image is input through the input layer of the deep neural network model, the microscopic component characteristics of the vitrinite and / or inertinite in the coal rock microscopic image are extracted through the dynamic aggregation layer, and the microscopic component sub-classification prediction label of the coal rock microscopic image output by the output layer of the deep neural network model is compared with the microscopic component sub-classification label in the sample, so as to estimate the model parameters in the deep neural network model and obtain a coal rock microscopic component identification neural network model.

[0024] As another aspect of the present invention, an embodiment of the present invention provides a device for identifying coal rock microscopic components based on a dynamic polymerization strategy, which may include:

[0025] A recognition module is used to input the coal rock microscopic image into a pre-trained coal rock microscopic component recognition neural network model to identify the microscopic component subclassification of the coal rock microscopic image;

[0026] The coal rock micro-component identification neural network model is pre-trained according to the above-mentioned coal rock micro-component identification neural network model training method.

[0027] As another aspect of the present invention, an embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-mentioned training method of the neural network model for identifying the microscopic components of coal and rock, or implements the above-mentioned method for identifying the microscopic components of coal and rock based on the dynamic aggregation strategy.

[0028] As another aspect of the present invention, an embodiment of the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and runnable on the processor. When the processor executes the program, it implements the above-mentioned training method of the coal rock micro-component identification neural network model, or implements the above-mentioned coal rock micro-component identification method based on the dynamic aggregation strategy.

[0029] The beneficial effects of the above technical solutions provided by the embodiments of the present invention include at least:

[0030] Embodiments of the present invention provide a training method, identification method, and apparatus for a neural network model for identifying coal microscopic components. The training method obtains a training sample set, each sample in the training sample set including a coal microscopic image and a microscopic component sub-classification label annotated with the coal microscopic image. Then, the samples in the training sample set are used to train a deep neural network model comprising one or more dynamic aggregation layers. The coal microscopic image is input into the input layer of the deep neural network model, and the dynamic aggregation layer extracts microscopic component features of the vitrinite and / or inertinite in the coal microscopic image. The predicted microscopic component sub-classification labels of the coal microscopic image output by the output layer of the deep neural network model are compared with the microscopic component sub-classification labels in the sample to estimate the model parameters in the deep neural network model and obtain the coal microscopic component identification neural network model. The training method utilizes a neural network algorithm based on a dynamic aggregation strategy. Unlike traditional fixed-structure neural networks, the dynamic aggregation layer can adaptively adjust the weights and activation strengths of neurons based on the features of the input microscopic image data, thereby improving the accuracy of feature extraction. This addresses the problems of traditional neural networks' fixed structure and insufficient feature extraction capabilities, enabling the neural network to flexibly adapt to microscopic differences between different samples. By building an intelligent microscopic component identification model, real-time detection and automatic classification of microscopic components can be achieved, reducing the workload of manual identification and improving research efficiency.

[0031] Other features and advantages of the present invention will be described in the following description, and in part will become apparent from the description, or will be understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures particularly pointed out in the written description and the accompanying drawings.

[0032] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] 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:

[0034] Figure 1 This is a flow chart of a training method for a neural network model for identifying coal rock microscopic components provided in an embodiment of the present invention;

[0035] Figure 2 for Figure 1 Step S11 is specifically implemented in the flowchart;

[0036] Figure 3a-Figure 3l Schematic diagram of microscopic images of subclassification of microscopic components provided in an embodiment of the present invention;

[0037] Figure 4 for Figure 1 Step S12 is specifically implemented in the flowchart;

[0038] Figure 5 This is a structural diagram of a training device for a neural network model for identifying coal and rock microscopic components provided in an embodiment of the present invention;

[0039] Figure 6 This is a schematic diagram showing the impact of different hyperparameters on model training time provided in an embodiment of the present invention;

[0040] Figure 7 This is a graph showing the relationship between the accuracy of different models and the number of training rounds provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0041] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0042] The traditional neural network mentioned in the present invention is a neural network with a fixed layer structure and weights. For example, when compared with a specific example, the traditional neural network is the neural network used in "BP Neural Network Model and Learning Algorithm_Fan Zhenyu" in "Software Guide" Volume 10, Issue 7, July 2011.

[0043] After analyzing the existing automated microscopic component identification methods based on deep learning, the inventors found that, on the one hand, traditional neural networks usually use fixed layer structures and weights, which may lead to insufficient accuracy when processing data with diverse features and cannot flexibly adapt to the characteristics of different data; on the other hand, in traditional neural networks, feature extraction often relies on the preset structure of the network, which can easily lead to the loss of important information or excessive reliance on certain features, and cannot fully capture the potential patterns of the data; on the other hand, traditional neural networks are prone to dimensionality disasters when processing high-dimensional data, and cannot effectively retain and utilize local and global information in microscopic images; at the same time, because traditional methods may over-rely on certain features when processing complex data, the model overfits. Furthermore, when processing multimodal data, the existing technology is usually unable to effectively fuse information from different data sources, and it is difficult to achieve a balance between morphological features and spectral features. Based on the above technical defects, the present invention is proposed.

[0044] The present invention provides a training method, an identification method, and an apparatus for a neural network model for identifying coal rock microscopic components. The training method obtains a training sample set, each sample in the training sample set including a coal rock microscopic image and a microscopic component sub-classification label annotated with the coal rock microscopic image. Then, the samples in the training sample set are used to train a deep neural network model comprising one or more dynamic aggregation layers. The coal rock microscopic image is input into the input layer of the deep neural network model, and the dynamic aggregation layer extracts microscopic component features of the vitrinite and / or inertinite in the coal rock microscopic image. The predicted microscopic component sub-classification labels of the coal rock microscopic image output by the output layer of the deep neural network model are compared with the microscopic component sub-classification labels in the sample to estimate the model parameters in the deep neural network model and obtain the coal rock microscopic component identification neural network model. The training method utilizes a neural network algorithm based on a dynamic aggregation strategy. Unlike traditional fixed-structure neural networks, the dynamic aggregation layer can adaptively adjust the weights and activation strengths of neurons based on the features of the input microscopic image data, thereby improving the accuracy of feature extraction. This solves the problems of traditional neural networks' fixed structure and insufficient feature extraction capabilities, enabling the neural network to flexibly adapt to microscopic differences between different samples. By building an intelligent microscopic component identification model, real-time detection and automatic classification of microscopic components can be achieved, reducing the workload of manual identification and improving research efficiency.

[0045] Example 1

[0046] In the first embodiment of the present invention, a method for training a neural network model for identifying microscopic components of coal and rock is provided. Figure 1 As shown, the training method may include the following steps:

[0047] Step S11: Obtain a training sample set, where each sample in the training sample set includes a coal rock microscopic image and a microscopic component sub-classification label annotated on the coal rock microscopic image.

[0048] This step involves obtaining a constructed training sample set. Each sample in the training sample set consists of a microscopic image of coal rock. The image sources can be a variety of natural mineral specimens or laboratory-synthesized samples. These images, captured using specialized microscope equipment, contain the distinct types and morphological features of vitrinite and inertinite, enabling annotation of microscopic component subclassifications.

[0049] Step S12: Use samples in the training sample set to train a deep neural network model containing one or more dynamic aggregation layers, wherein after the coal rock microscopic image is input through the input layer of the deep neural network model, the microscopic component characteristics of the vitrinite and / or inertinite in the coal rock microscopic image are extracted through the dynamic aggregation layer, and the microscopic component sub-classification prediction labels of the coal rock microscopic image output by the output layer of the deep neural network model are compared with the microscopic component sub-classification labels in the sample, so as to estimate the model parameters in the deep neural network model and obtain a coal rock microscopic component recognition neural network model.

[0050] In this step, feature extraction is performed on high-dimensional coal and rock microscopic images based on the deep neural network model. The dynamic aggregation layer can effectively extract potential structures and enhance the generalization ability of the deep neural network model in the face of complex microscopic image data. Based on the dynamic aggregation strategy and feature mapping combined with holographic projection, it can not only extract potential microscopic component characteristics in coal and rock microscopic image data, but also improve the nonlinear relationship capture ability of the deep neural network model, effectively avoiding the problems of fixed structure and insufficient feature extraction ability in traditional neural networks, thereby enhancing the ability of feature learning and microscopic image data modeling.

[0051] The training method for the coal and rock microscopic component identification neural network model provided in the embodiments of the present invention utilizes a neural network algorithm based on a dynamic aggregation strategy. Unlike traditional fixed-structure neural networks, the dynamic aggregation layer can adaptively adjust the weights and activation strengths of neurons based on the characteristics of the input microscopic image data, thereby improving the accuracy of feature extraction. This addresses the issues of fixed structure and insufficient feature extraction capabilities of traditional neural networks, enabling the neural network to flexibly adapt to microscopic differences between different samples. By constructing an intelligent microscopic component identification model, real-time detection and automatic classification of microscopic components are achieved, reducing the workload of manual identification and improving research efficiency.

[0052] In the embodiment of the present invention, the above step S11 obtains the training sample set, referring to Figure 2 As shown, the following steps may be specifically included:

[0053] Step S111: Acquire coal rock microscopic images in multiple imaging modes at the same position of the coal rock sample; wherein the multiple imaging modes include: single polarization imaging mode, reflected light imaging mode and fluorescence spectrum imaging mode.

[0054] In this step, high-resolution optical microscopy and high-performance imaging are used to capture microscopic details of the vitrinite and inertinite layers, such as boundary morphology, fracture state, and texture characteristics, using different imaging modes (e.g., single polarization mode). By acquiring multiple imaging modes at the same location on the coal sample, a multimodal comparative dataset can be generated, providing rich information for subclassification of microscopic components.

[0055] It should be noted that the coal and rock microscopic images in the embodiments of the present invention are stored in lossless compressed TIFF or high-quality PNG format to ensure information integrity. The resolution is set to 3000x3000 pixels or higher, with a magnification of 400x or 1000x to capture microscopic details. Single-channel grayscale images or multi-channel RGB images are selected as needed to highlight optical reflectance or fluorescence spectral features. Metadata (such as sample number, shooting conditions, and microscope parameters) is attached to each image to facilitate subsequent tracking and data organization.

[0056] Step S112: divide and label the microscopic component characteristics of the vitrinite and / or inertinite in the coal rock microscopic image, and cross-validate the single polarization labeling results through the reflected light characteristics and fluorescence spectrum characteristics of the coal rock microscopic image to determine the sub-classification labels of the microscopic components annotated in the coal rock microscopic image.

[0057] In this step, specialized tools (e.g., Labelbox, LabelImg, etc.) are used to annotate the microscopic features of the vitrinite and / or inertinite components (e.g., boundaries, texture orientation, crack distribution, etc.) in detail. These features are then divided into subclasses according to the hierarchical annotation principle to improve classification accuracy. This step cross-validates the single-polarization annotation results by combining reflected light and fluorescence spectral features, and performs multimodal feature comparison and correction for each imaging position to ensure classification consistency and reliability. Furthermore, domain experts can be invited to review the annotation results to reduce subjective errors and ensure data quality.

[0058] The differences in the microscopic characteristics of the vitrinite and inertinite in this step are summarized in Table 1:

[0059] Table 1 Summary of the differences in microscopic characteristics between vitrinite and inertinite

[0060]

[0061]

[0062]

[0063] In this step, vitrinite refers to the organic component of coal with a glassy structure and no crystalline features. It is primarily formed by the degradation and transformation of plant tissue through coalification. Its optical properties and morphological characteristics are clearly recognizable under a microscope, often exhibiting a strong luster and varying degrees of reflectivity under reflected light. Based on its structure, optical properties, and microscopic morphology, it can be further divided into the following categories:

[0064] (1) Structured vitrinite

[0065] Structural vitrinite is characterized by the preservation of obvious plant tissue structure. It is commonly found in the early stages of coalification and usually presents a strip, fiber or tubular morphology. Some of the original xylem, vessels and fiber structures can be seen. Under a microscope, its tissue arrangement is clear, with a certain directionality along the fiber direction. It exhibits high gloss and high reflectivity under reflected light. Because it highly retains the characteristics of plant origin, it is an important indicator for determining the type of parent material and the sedimentary environment. Figure 3a As shown, the structural vitrinite cell cavity is severely deformed and microcracks are visible.

[0066] (2) Homogeneous vitrinite

[0067] Homogeneous vitrinite appears as a dense mass or evenly distributed form, with no obvious plant tissue residues, representing a more advanced stage of coalification. Its microstructure is uniform and optically isotropic, meaning that when the microscope stage is rotated, the brightness and darkness do not change significantly. It has a uniform luster under reflected light and a lower reflectivity than structural vitrinite. It is usually the most common type of vitrinite in coal, see Figure 3b As shown, microcracks develop in the homogeneous vitrinite.

[0068] (3) Matrix vitrinite

[0069] The main characteristic of matrix vitrinite is its uneven distribution, appearing diffuse or mixed with other vitrinites. Its morphological boundaries are relatively fuzzy, and there may be fine cracks or transition zones in some areas. Under a microscope, the optical anisotropy is relatively weak, but local directionality can still be seen. Under reflected light, it appears to have a medium luster and medium reflectivity. Due to its strong deformability, it is often associated with inertinite or minerals. Figure 3c As shown, the matrix vitrinite cements the fungal body and the inertinid body.

[0070] (4) Clumped vitrinite

[0071] Agglomerated vitrinite presents irregular blocky shapes and is commonly found in crack filling areas or compressed and deformed areas in coal. There may be micro cracks or inclusions inside it, which may cause its optical properties to vary in local areas. Figure 3d As shown, the gloss under reflected light is average, the reflectivity is slightly lower than that of homogeneous vitrinite, and the overall optical heterogeneity is low, but slight changes in brightness and darkness may occur locally due to the presence of cracks.

[0072] (5) Detrital vitrinite

[0073] Vitreous bodies are characterized by tiny fragments with different particle sizes and clear boundaries. They may be derived from the mechanical crushing of vitrinite or the fragments formed during coalification. Figure 3e As shown, it has strong optical anisotropy, with distinct changes in brightness and darkness observed when the stage is rotated, and exhibits high gloss and reflectivity under reflected light. Due to its small particle size, it is often used as a microparticle component in coal petrographic analysis, often mixed with other coal petrographic components.

[0074] The main subclasses of the inertinite group in the present invention: According to the international classification of coal rock microscopic components, the inertinite group can be further divided into several types, each with its own characteristics in morphology and optical characteristics. They mainly include silken bodies, semi-silky bodies, fungal bodies, secretory bodies, coarse-grained bodies, microscopic bodies and inertinite bodies, etc.

[0075] (1) Fusinite

[0076] The fibrous body is characterized by a woody fiber or grid-like structure, and often retains the morphology of the original plant cell wall, so under the microscope, holes and outlines similar to the wood grain can be seen. This component is actually fossil charcoal formed by ancient fires of buried peat, and its cell walls have been preserved after carbonization. Under a reflected light microscope, the fibrous body shows extremely high reflectivity and strong luster, almost appearing as white crystals, and is one of the brightest reflective components in coal. The optical anisotropy is significant. When the stage is rotated, the intensity of the reflected light from the fibrous body changes dramatically, often showing bright and dark flashes. This shows that the carbon arrangement inside the fibrous body is highly directional, which is consistent with its structural characteristics of being derived from the carbonization of fibrous tissue. In addition, the fibrous body has high mechanical strength and is brittle. It often appears as sharp edges or long strips in polished coal samples. Its high carbon and low volatile content make it difficult to gasify or coke in the industrial use of coal, which is basically equivalent to the "coke" component that pre-existed in the coal. Reference Figure 3f As shown, homogeneous vitrinite and silken body are produced in stripe shape and micro cracks are developed.

[0077] (2) Semifusinite

[0078] Semi-silkenite is formed by partially carbonized plant tissue during the coal formation process, and its characteristics are between vitrinite and fusinite. Morphologically, it is often strip-like or fibrous, and sometimes residual plant cell structures can still be identified, but compared to fusinite, its structure is not completely preserved and the surface is slightly rough. Under a microscope, the optical anisotropy of semi-silkenite is moderate, and the light and dark changes when the stage is rotated are weaker than those of fusinite, but still detectable. Under reflected light, the brightness and reflectivity of semi-silkenite are lower than those of fusinite but higher than those of ordinary vitrinite. Usually, its reflected light color is light gray and the luster is relatively soft. Under a fluorescence microscope, semi-silkenite in low-rank coal sometimes emits a dim brown-yellow fluorescence, which may be related to the partial uncarbonized matter it retains. In general, semi-silkenite is regarded as a product of the transition from fusinite to vitrinite, representing a stage in which the organic matter in the coal undergoes moderate oxidation or incomplete pyrolysis. Its existence indicates that the local environment during coal formation may have undergone weak fire or dry oxidation, which caused partial transformation of plant tissue but not complete carbonization. Reference Figure 3g As shown, the semi-filamentous body cell cavity structure is intact as a whole, but some cells are flattened, arranged irregularly, and cracks are seen.

[0079] (3)Funginite

[0080] Fungal bodies are microscopic components of the inertinite group formed by fungal remnants, and were previously referred to as "mycelium" or "sclerotium" in the literature. Their morphology is mostly spherical, elliptical or branched, similar to the morphology of fungal sclerotia or fungal communities, with a relatively loose structure and lacking the cell wall characteristics of higher plants. Fungal bodies are believed to be the result of carbonization or lignification of large fungal entities (such as sclerotia) in peat bogs. Under a reflected light microscope, the reflectivity of fungal bodies is relatively low, and they appear as dark gray to black particles with a weaker luster than mycelium and semi-mycelium. However, under ultraviolet excitation, they often emit bright yellow-white to orange-yellow fluorescence, which is one of the significant features for identifying fungal bodies. Since fungal bodies are mainly found in low-metamorphic coal, their abundance can be used as an indicator of the intensity of fungal action in coal - a high content of fungal bodies often suggests that the peat bog experienced strong fungal erosion and decomposition during the coal formation process. This biogenic inertinite component is indicative for restoring information such as fungal activity and dry oxidative conditions in ancient environments ( Figure 3d ).

[0081] (4) Secretinite

[0082] Secretions are a special type of component within the inertinite group derived from plant secretions (such as resins and exudates). Their name derives from the concept of "secretion." Under a microscope, secretions often appear as irregular masses or droplets, with smooth outlines or one side rounded and the other side irregular. Their overall appearance is relatively homogeneous, lacking internal cellular structure. Originating from carbonized plant resins or secretory glandular tissue, secretions have a gelatinous texture and often exhibit a clear demarcation from the surrounding matrix in coal samples. Under reflected light, secretions appear as very bright particles with high reflectivity, approaching or even reaching the brightness of silky bodies. Correspondingly, their optical anisotropy is weak, often exhibiting nearly isotropic optical properties, meaning that changes in brightness are not noticeable when the stage is rotated. This is likely due to the formation of homogeneous amorphous carbon after the carbonization of resinous materials. Secretions typically exhibit a bright yellow-orange fluorescence under ultraviolet light, which is related to the resin's rich composition of aromatic and polycyclic compounds and is an important clue for identifying this component. It is worth noting that previous literature has referred to this type of structureless, highly reflective inert group as "resin-sclerotinite", but modern classification has clearly divided it into two types: secretory bodies and fungal bodies. The existence of secretory bodies indicates that the coal-forming plants are rich in resin or secretory substances, and that they have undergone an oxidative hardening process in the early stages of coal formation, preserving them as high-carbon inert residues. Figure 3h As shown, it has oxidized edges and pores and is irregular in shape.

[0083] (5) Macrinite

[0084] Coarse particles are amorphous, massive inertinite components, typically occurring as relatively large, irregular particles or massive fillings with fuzzy edges and a lack of discernible internal structure. Coarse particles may represent inert organic colloids remaining after the complete decomposition of plant tissues in peat, or they may be formed by the filling of plant cell cavities with inert organic matter. Biochemical humification plays a significant role in their formation: extensive microbial decomposition degrades the organic matter, enriching it with carbon, thus forming coarse particles. Under a microscope, coarse particles typically appear uniformly gray or dark gray, with a poorly defined boundary from the surrounding vitrinite matrix. However, due to their high degree of carbonization, their reflected light brightness remains slightly brighter than the surrounding matrix. Coarse particles generally exhibit weak optical anisotropy, with no significant shift in reflected light or dark upon rotation, making them approximately isotropic. Coarse particles have a moderate luster in reflected light, less bright than silky particles. Under fluorescence, a faint yellow-green fluorescence may sometimes be observed, suggesting the presence of a small amount of unagglomerated organic structure. The presence of coarse particles indicates that the organic matter has undergone sufficient humification and late thermal evolution during coal formation, which is a manifestation of highly mature inert organic matter in coal. Figure 3i Images of coarse, filamentous, inerticulate, and microscopic bodies, along with references Figure 3j Microscopic image of coarse grains cemented by vitrinite in the intermediate matrix.

[0085] (6) Microsomes

[0086] Microbodies are the smallest particle size within the inertinite group. They are often uniformly dispersed within the coal matrix as extremely fine particles, typically measuring only a few microns. Microbodies may originate from secondary carbonaceous materials generated during the thermal evolution of peat, such as small particles formed by the thermal decomposition and polymerization of organic colloids at high temperatures. Due to their extremely small size, they are almost completely opaque under transmitted light, and their presence can only be observed under reflected light. Under a reflected light microscope, microbodies appear as small, dark gray to black specks with a very weak luster and the lowest reflectivity among the inertinite group (but still higher than most vitrinite and exinite components). Optical anisotropy is not pronounced, and there is almost no change in brightness when rotated, indicating a highly disordered or isotropic internal structure. Under ultraviolet fluorescence, microbodies typically do not fluoresce or only exhibit extremely weak, dark fluorescence, indicating that they contain few readily luminescent functional groups. Microbodies often fill pores or cracks in coal or are uniformly dispersed as part of the matrix, indicating that organic matter underwent intense thermal evolution or decomposition and polymerization during the late stages of coal formation. The presence of a large number of microparticles is often one of the characteristics of high-rank coal (such as anthracite) or abnormally heated coal seams ( Figure 3k ).

[0087] (7) Inertodetrinite

[0088] Inerticolites can be considered the detrital counterpart of the inertinite, similar to vitrinites. They are composed of small fragments of various inertinite components, with irregular and variable morphologies, ranging from angular to granular, often present in a mixed distribution. Inerticolites typically form when plant debris in peat is broken into small fragments by intense oxidation, bioturbation, or fire, and then incorporated into coal seams through sedimentation. Under a microscope, inerticolites exhibit high reflectivity and strong anisotropy. When the stage is rotated, the sharp contrast between light and dark within these fragments reveals that they originate from carbonaceous planes with varying orientations. Inerticolites often exhibit varying brightness in reflected light, ranging from a white-brightness similar to that of fusinites to a grayish brightness similar to that of semifusinites, but overall they are brighter than vitrinites. Due to their small size and diverse morphology, the boundaries between inerticolites and the surrounding matrix are unclear, resembling a mass of broken glass scattered within the matrix. When observed by fluorescence, inerticulate bodies typically lack significant fluorescence because their components are mostly highly carbonized and lack fluorescent chromophores. The abundance of inerticulate bodies can reflect the intensity of oxidative damage in peat bogs: the stronger the oxygen content or the more frequent the fire, the more inertic debris is produced. Therefore, in coal rock analysis, inerticulate bodies are often used as indicators of paleoenvironmental drought, fire events, or peat transport disturbances ( Figure 3l ).

[0089] Step S113: Take each coal rock microscopic image and its annotated microscopic component sub-classification label as a sample to construct a training sample set.

[0090] In an optional embodiment, determining the microscopic component sub-classification labels annotated to the coal microscopic images in step S112 may specifically include labeling the microscopic component sub-classification with the largest proportion of the area of ​​the coal microscopic image in each coal microscopic image as the unique microscopic component sub-classification label for that coal microscopic image. In order to improve the computing power during the training of the neural network model, in this embodiment of the present invention, the microscopic component sub-classification in each coal microscopic image is no longer labeled by image segmentation. Instead, each coal microscopic image is labeled with a unique microscopic component sub-classification label, thereby ensuring classification consistency and reliability.

[0091] In another optional embodiment, in the above step S113, each coal rock microscopic image and its annotated microscopic component subclassification label are taken as a sample. Before constructing the training sample set, the following steps may also be included, namely, step S114, removing other image areas outside the coal rock microscopic image corresponding to the unique microscopic component subclassification in each coal rock microscopic image. In this embodiment, by removing other image areas outside the microscopic image corresponding to the unique microscopic component subclassification, all microscopic components in a coal rock microscopic image correspond to the unique microscopic component subclassification, which makes the accuracy more accurate during the model training process, improves the training computing power, and avoids the negative impact of other image areas that do not correspond to the unique microscopic component subclassification on the neural network model training process.

[0092] In a specific embodiment, when the above step S12 is implemented, the deep neural network model includes a dynamic aggregation layer. The forward propagation dynamic aggregation layer extracts the microscopic component features of the vitrinite and / or inertinite group in the coal rock microscopic image to extract the microscopic component feature vector of the coal rock microscopic image, and performs feature aggregation based on the microscopic component feature vector to obtain the aggregated microscopic combination feature vector, and determines the holographic projection output based on the microscopic component feature vector; the aggregated microscopic combination feature vector and the holographic projection output are used as inputs for the next iteration; the back propagation dynamic aggregation layer calculates the gradient of the weight matrix and the dynamic aggregation weight coefficient based on the loss function; the neural network model determines the loss function based on the aggregated microscopic component feature vector, the holographic projection output, the microscopic component sub-classification prediction label output by the model and the microscopic component sub-classification label in the sample to train the dynamic aggregation weight coefficient and weight matrix in the dynamic aggregation layer to obtain a coal rock microscopic component recognition neural network model; the deep neural network model includes multiple dynamic aggregation layers. In the state of the combined layers, the first layer of dynamic aggregation layer is forward propagated to extract the microscopic component features of the vitrinite and / or inertinite in the coal rock microscopic image to extract the microscopic component feature vector of the coal rock microscopic image. The first layer of dynamic aggregation layer performs feature aggregation based on the microscopic component feature vector to obtain the aggregated microscopic combination feature vector, and determines the holographic projection output based on the microscopic component feature vector; the microscopic component feature vector output by the first layer of dynamic aggregation layer is used as the input of the next layer of dynamic aggregation layer, and the aggregated microscopic combination feature vector and the holographic projection output output by the first layer of dynamic aggregation layer are used as the input of the next iteration; each dynamic aggregation layer is back-propagated to calculate the gradient of the weight matrix and the dynamic aggregation weight coefficient based on the loss function; the neural network model determines the loss function based on the aggregated microscopic component feature vector, the holographic projection output, the microscopic component sub-classification prediction label output by the model and the microscopic component sub-classification label in the sample to train the dynamic aggregation weight coefficient and weight matrix in the dynamic aggregation layer to obtain a coal rock microscopic component recognition neural network model.

[0093] In this embodiment, when implementing Figure 4 As shown in Figure 2, the training process of the deep neural network is as follows:

[0094] Step S21, initialize the structure of the deep neural network; the deep neural network in the embodiment of the present invention is illustrated by taking the architecture of the input layer, 5 dynamic aggregation layers (hidden layers), and the output layer as an example. The number of neurons in each hidden layer is set according to the dimension and complexity of the input coal rock microscopic image, for example, 100, 200, 200, 200, and 100 neurons respectively. In the embodiment of the present invention, the deep neural network initialization adopts a random weight initialization strategy to avoid the problem of gradient disappearance or gradient explosion.

[0095] Step S22, forward propagation, extracting the microscopic component feature vectors in the coal rock microscopic image; in this step, the dynamic aggregation layer can adaptively adjust the weight (dynamic aggregation weight coefficient) and activation strength of each neuron according to the characteristics of the currently input coal rock microscopic image. Compared with the weights and activation functions of neurons in traditional neural networks, which usually remain fixed during the training process, the dynamic aggregation layer in the embodiment of the present invention can adjust the contribution of each neuron by adopting an aggregation strategy based on objective function optimization, thereby improving the accuracy of feature extraction. In mineral specimens and synthetic samples, since their microstructures may vary greatly, dynamic aggregation can help the deep neural network model better capture the microscopic differences between different samples, thereby improving the accuracy of feature extraction. For example, in the coal rock microscopic image data of vitrinite and inertinite, the vitrinite may show more uniform texture characteristics, while the inertinite may show a more complex structure. By dynamically adjusting the weights of the neural network, the deep neural network model can flexibly learn these different features. The output calculation method of the dynamic aggregation layer is expressed as:

[0096]

[0097] Where, is the output feature vector of the neural network layer l, which represents the aggregated features calculated from the neurons in the neural network layer l; p () is the activation function of the neural network, such as the ReLU activation function or the Sigmoid activation function, which is used for nonlinear transformation and solves the problem of fixed layer structure in traditional neural networks, so that the neural network can extract potential features more accurately during training; is the dynamic aggregation weight coefficient of the lth layer of the neural network, which represents the contribution of the neurons in the lth layer of the neural network to the final output; is the weight matrix of the lth layer of the neural network; is the i-th input sample of the neural network (the first layer is the microscopic image of coal and rock, and the second and subsequent layers are the microscopic component feature vectors output by the first layer); is the bias term of the neural network; n p is the number of neurons in the lth layer of the neural network; i is a positive integer; l is a positive integer, 1-5.

[0098] Step S23: Determine the holographic projection output. In this step, to enhance the expressive power of features, the neural network uses holographic projection theory to map the input microscopic image data. Holographic projection increases the feature dimension through a dimensionality-increasing operation, mapping the features of each input microscopic image data sample into a high-dimensional space. This preserves the local information of the microscopic image data while utilizing global information. During this process, the features of the input coal and rock microscopic image are projected through the holographic mapping matrix, resulting in a representation in a high-dimensional space. The holographic mapping matrix in the embodiment of the present invention is optimized using the Euler-Lagrange equations to ensure that the potential patterns of the microscopic image data are captured to the maximum extent. Microscopic image data typically has a high dimensionality, where each feature point may contain a large amount of detailed information. Holographic projection maps the image data into a high-dimensional space, allowing each input sample to better retain local information and combine global information to capture potential patterns. For the analysis of natural mineral specimens, the microscopic features of these samples may exhibit highly complex and diverse patterns. Through high-dimensional mapping, the deep neural network model can find patterns in a wider feature space. For example, the surface of some minerals may exhibit complex spectral reflectance characteristics, which are difficult to capture in low-dimensional space. However, through holographic projection, these characteristics can be fully expressed in high-dimensional space, as expressed by:

[0099]

[0100] Where, is the holographic projection output of the lth layer of the neural network; is the holographic projection matrix of the lth layer of the neural network, which maps the input features to a high-dimensional space.

[0101] Furthermore, the holographic projection matrix is ​​a training parameter, which is solved through optimization and regularization terms are used to maintain the orthogonality of the matrix, thereby ensuring that key information is not lost during the projection process. The gradient vector method update method is expressed as:

[0102]

[0103] Where ← is the parameter update operation; I is the identity matrix, ensuring the orthogonality of the projection matrix; is the transpose of the holographic projection matrix of the neural network layer l; L p is the loss function of the neural network; is the symbol of partial derivative; p is the regularization coefficient of the neural network, which controls the norm of the projection matrix. Preferably, λ p Set to 0.2; η p is the learning rate of the neural network, preferably, η p Set to 0.01.

[0104] Step S24, back propagation, calculates the gradient of the weight matrix and the dynamic aggregation weight coefficient based on the loss function; in the feature extraction process, the neural network can converge quickly by dynamically adjusting the connection weights of each layer. At each iteration, the neural network calculates the activation value generated after the input microscopic image data is propagated through each layer, and performs back propagation based on the loss function to calculate the gradient of each weight and dynamic aggregation coefficient. When processing multimodal data including polarized light microscopy data and fluorescence spectra, the synchronous gradient calculation of each weight and aggregation coefficient during the back propagation process enables the network to automatically balance the contribution ratio of morphological features and spectral features. The gradient calculation method of the loss function for each weight and aggregation coefficient is expressed as:

[0105]

[0106] Where, is the output feature vector of the lth layer of the neural network.

[0107] Furthermore, the gradient of each weight and aggregation coefficient is updated based on the loss function. By jointly optimizing the weight matrix and dynamic aggregation coefficient, the neural network can more accurately adjust the weight and contribution of each neuron, thereby improving the feature extraction capability. The update method of the weight matrix and dynamic aggregation weight coefficient is expressed as:

[0108]

[0109] Where η p is the learning rate of the neural network, preferably, η p Set to 0.01.

[0110] Step S25: Perform feature aggregation based on the microscopic component feature vectors to obtain an aggregated microscopic composite feature vector. This step is an adaptive optimization of feature aggregation. Based on the contribution of each feature, the feature aggregation method is dynamically adjusted to better reflect the essence of the microscopic image data. For example, when analyzing the vitrinite reflectance, the bottom-level network captures the microscopic texture of the cell wall (resolution of approximately 0.5 μm), and the top-level network extracts the macroscopic reflectance distribution pattern. The dynamic allocation of the dynamic aggregation weight coefficient of the neural network allows the final feature to contain both microstructural and optical property information. The calculation method of the aggregation operation is expressed as:

[0111]

[0112] Where, is the feature vector after aggregation; L pe is the number of layers of the neural network; is the dynamic aggregation weight coefficient of the lth layer of the neural network, which indicates the importance of the feature vector of the lth layer of the neural network in the final fusion.

[0113] Step S26, determining the loss function based on the aggregated microscopic component feature vector, the holographic projection output, the microscopic component sub-classification prediction label output by the model, and the microscopic component sub-classification label in the sample; during the feature extraction process, the loss function of the deep neural network in the embodiment of the present invention needs to consider the cross-entropy loss and the parameters of the dynamic aggregation layer to ensure the stability and generalization ability of the entire network. The cross-entropy loss function measures the difference between the model prediction result and the true label. In the process of processing coal rock microscopic image data, the goal is to classify the types of different mineral samples. Different types of minerals in microscopic images (such as vitrinite and inertinite) have different morphological and fluorescence spectral characteristics, and these characteristics are extremely complex in high-dimensional space. The cross-entropy loss function prompts the model to optimize its parameters during training by comparing the predicted labels and true labels output by the model, thereby improving the classification accuracy of the deep neural network model. In addition, the dynamic aggregation weight coefficient of the dynamic aggregation layer directly affects the contribution of each neuron. In order to prevent the deep neural network model from over-relying on certain specific features, the loss function adopts a regularization term. By penalizing the dynamic aggregation weight coefficient and controlling the weight distribution of neurons in each layer, it can avoid the weight of some neurons being too large during training, which leads to overfitting of certain local features of the model. For microscopic image data, especially when processing mineral specimens with complex texture and morphological characteristics, regularization can help the model extract global information in the image more evenly, rather than focusing too much on certain local features. The calculation method of the neural network loss function is expressed as:

[0114]

[0115] Where, is the cross entropy loss, which represents the error between the model output label and the true label. The model output label is calculated by the preset Softmax function on the aggregated feature vector; λ pe1 To adjust the impact of the dynamic aggregation coefficient on the loss function, you can customize it; λ pe2 In order to adjust the influencing parameters of the holographic projection matrix on the loss function, it can be customized; is the dynamic aggregation weight coefficient of the lth layer of the neural network; is the holographic projection matrix of the lth layer of the neural network.

[0116] In the embodiment of the present invention, the model training is completed by repeatedly iterating the above steps S22 to S26 until a preset stopping condition is met. For example, the preset stopping condition is reaching a preset maximum number of iterations, preferably, the preset maximum number of iterations is set to 1000.

[0117] The beneficial effects of the training method for the coal rock micro-component identification neural network model provided in the embodiments of the present invention may include:

[0118] (1) The neural network based on dynamic aggregation strategy is different from the traditional fixed structure neural network. The dynamic aggregation layer can adaptively adjust the weights and activation strengths of neurons according to the characteristics of the input microscopic image data, thereby improving the accuracy of feature extraction. It solves the problems of fixed structure and insufficient feature extraction ability of traditional neural networks, and enables the neural network to flexibly adapt to the microscopic differences between different samples.

[0119] (2) The microscopic image data is mapped through holographic projection theory, and the input data is mapped to a high-dimensional space using a dimensionality increase operation, thereby retaining more local and global information. For feature extraction of high-dimensional microscopic image data, the potential patterns of microscopic images can be better captured in high-dimensional space, solving the problem of information loss in low-dimensional space.

[0120] (3) When processing multimodal data including polarized light microscopy data and fluorescence spectral data, the model can simultaneously optimize the contribution ratio of morphological features and spectral features, and improve the modeling effect of different types of microscopic image data through adaptive optimization.

[0121] (4) A loss function that comprehensively considers cross entropy loss and dynamic aggregation layer parameters is adopted. By optimizing the model, not only can the classification accuracy be improved, but also the overfitting problem can be avoided. In the processing of complex microscopic image data, the generalization ability of the model is guaranteed by the regularization term.

[0122] (5) Dynamically adjust the feature aggregation method according to the contribution of each feature, which enhances the model's ability to express microscopic image data and can balance the extraction of microstructure and optical property information.

[0123] Based on the same inventive concept, the present invention also provides a training device for a neural network model for identifying microscopic components of coal and rock, referring to Figure 5 As shown, the training device may include: an acquisition module 51 and a training module 52, and its working principle is as follows:

[0124] The acquisition module 51 is used to acquire a training sample set, where each sample in the training sample set includes a coal rock microscopic image and a microscopic component sub-classification label annotated on the coal rock microscopic image;

[0125] The training module 52 is used to train a deep neural network model containing one or more dynamic aggregation layers using samples in the training sample set, wherein, after the coal rock microscopic image is input through the input layer of the deep neural network model, the microscopic component characteristics of the vitrinite and / or inertinite in the coal rock microscopic image are extracted through the dynamic aggregation layer, and the microscopic component sub-classification prediction labels of the coal rock microscopic image output by the output layer of the deep neural network model are compared with the microscopic component sub-classification labels in the sample, so as to estimate the model parameters in the deep neural network model and obtain a coal rock microscopic component recognition neural network model.

[0126] In one embodiment, the acquisition module 51 is specifically configured to:

[0127] Acquire coal rock microscopic images in multiple imaging modes at the same location of the coal rock sample; wherein the multiple imaging modes include: single polarization imaging mode, reflected light imaging mode and fluorescence spectrum imaging mode;

[0128] Classifying and labeling the microscopic component characteristics of the vitrinite and / or inertinite in the coal rock microscopic image, and cross-validating the single polarization labeling results using the reflected light characteristics and fluorescence spectrum characteristics of the coal rock microscopic image to determine the sub-classification labels of the microscopic components annotated in the coal rock microscopic image;

[0129] Each coal rock microscopic image and its annotated microscopic component sub-classification label are taken as a sample to construct a training sample set.

[0130] In another embodiment, the acquisition module 51 is further specifically used to: mark the microscopic component sub-classification in each coal rock microscopic image that accounts for the largest proportion of the area of ​​the coal rock microscopic image as the unique microscopic component sub-classification label of the coal rock microscopic image.

[0131] In another embodiment, the acquisition module 51 is further specifically configured to remove image regions other than the coal rock microscopic image corresponding to the unique microscopic component subclassification in each coal rock microscopic image.

[0132] In another embodiment, the training module 52 is specifically configured to:

[0133] In a state where the deep neural network model includes a dynamic aggregation layer, the forward propagation dynamic aggregation layer extracts the microscopic component features of the vitrinite and / or inertinite group in the coal rock microscopic image to extract the microscopic component feature vector of the coal rock microscopic image, and performs feature aggregation based on the microscopic component feature vector to obtain the aggregated microscopic combination feature vector, and determines the holographic projection output based on the microscopic component feature vector; the aggregated microscopic combination feature vector and the holographic projection output serve as inputs for the next iteration; the backward propagation dynamic aggregation layer calculates the gradient of the weight matrix and the dynamic aggregation weight coefficient based on the loss function; the neural network model determines the loss function based on the aggregated microscopic component feature vector, the holographic projection output, the microscopic component sub-classification prediction label output by the model and the microscopic component sub-classification label in the sample to train the dynamic aggregation weight coefficient and the weight matrix in the dynamic aggregation layer to obtain a coal rock microscopic component identification neural network model;

[0134] When the deep neural network model includes multiple dynamic aggregation layers, the first dynamic aggregation layer is forward propagated to extract the microscopic component features of the vitrinite and / or inertinite group in the coal rock microscopic image to extract the microscopic component feature vector of the coal rock microscopic image, the first dynamic aggregation layer performs feature aggregation based on the microscopic component feature vector to obtain the aggregated microscopic combination feature vector, and determines the holographic projection output based on the microscopic component feature vector; the microscopic component feature vector output by the first dynamic aggregation layer is used as the input of the next dynamic aggregation layer, and the aggregated microscopic combination feature vector and the holographic projection output output by the first dynamic aggregation layer are used as the input of the next iteration; each dynamic aggregation layer is back-propagated to calculate the gradient of the weight matrix and the dynamic aggregation weight coefficient based on the loss function; the neural network model determines the loss function based on the aggregated microscopic component feature vector, the holographic projection output, the microscopic component sub-classification prediction label output by the model and the microscopic component sub-classification label in the sample to train the dynamic aggregation weight coefficient and weight matrix in the dynamic aggregation layer to obtain a coal rock microscopic component identification neural network model.

[0135] Based on the same inventive concept, an embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the training method of the above-mentioned neural network model for identifying the microscopic components of coal and rock is implemented.

[0136] Based on the same inventive concept, an embodiment of the present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and runnable on the processor. When the processor executes the program, the training method of the above-mentioned neural network model for identifying the microscopic components of coal and rock is implemented.

[0137] The principles of the problems solved by the above-mentioned devices, media, and related equipment in the embodiments of the present invention are similar to the training method of the aforementioned coal rock micro-component identification neural network model. Therefore, its implementation can refer to the implementation of the aforementioned method, and the repeated parts will not be repeated.

[0138] Example 2

[0139] A second embodiment of the present invention provides a method for identifying coal rock microscopic components based on a dynamic aggregation strategy, comprising: inputting a coal rock microscopic image into a pretrained coal rock microscopic component identification neural network model to identify the microscopic component subclassifications of the coal rock microscopic image; the coal rock microscopic component identification neural network model is pretrained according to the training method for the coal rock microscopic component identification neural network model described in the first embodiment. This identification method addresses the issues of high subjectivity, low efficiency, and poor repeatability inherent in traditional coal rock microscopic component identification methods, enabling real-time detection and automatic classification of microscopic components, reducing the workload of manual identification and improving research efficiency.

[0140] In a specific example, this study used deep coal gas sample X from Well A1. The sample, obtained from a deep coal gas reservoir, is primarily composed of vitrinite and inertinite, and includes multiple sub-classes of macerals. The research objective is to automatically identify 12 typical macerals of the vitrinite and inertinite using a dynamic aggregation strategy, thereby improving the accuracy and automation of coal microscopic analysis. The specific classifications include five sub-classes of the vitrinite (structural vitrinite, homogeneous vitrinite, matrix vitrinite, agglomerated vitrinite, and vitrinite fragments) and seven sub-classes of the inertinite (filamentous bodies, semi-filamentous bodies, fungal bodies, secretory bodies, coarse-grained bodies, microscopic bodies, and inertinite fragments).

[0141] Data acquisition was performed using a high-resolution optical microscope, with single-polarized light, reflected light, and fluorescence imaging performed on the samples to obtain multimodal microscopic images, ensuring the complete preservation of the morphological and optical characteristics of the vitrinite and inertinite groups. Image resolution was set to 3000×3000 pixels, and lossless TIFF format was used for storage to preserve detailed information. During the data annotation process, professionals manually annotated the boundary morphology, fracture status, and textural characteristics of each microscopic component, and the annotation quality was verified by experts. A minimum of 50 images were collected for each microscopic subclass to ensure sufficient data for subsequent model training. Holographic projection mapping was used to increase the dimensionality of the input data to preserve more complete microscopic component characteristics. Ultimately, the model effectively identified the major subclasses of the vitrinite and inertinite groups. Comparison with manual annotation revealed that the automatic identification error was below 5%, significantly improving the efficiency and accuracy of microscopic component analysis. The model is suitable for practical applications in coal petrology research, reservoir evaluation, and coalbed methane resource development.

[0142] In this example, in order to verify the performance of the coal rock micro-component identification neural network model (dynamic aggregation-based neural network) provided in the embodiment of the present invention compared with the traditional neural network model, this example uses the same traditional neural network model with 5 hidden layers and the deep neural network model with 5 dynamic aggregation layers in this embodiment for comparison and verification, as follows:

[0143] Reference Figure 6 As shown in Figure 2, the performance of dynamic aggregation and traditional neural networks are compared. By comparing the training time of the two with different numbers of layers and neurons, the experimental results show that the training method proposed in this invention has a shorter training time. Figure 7 As shown, in order to compare the accuracy of the deep neural network model of the dynamic aggregation layer with the traditional neural network, the accuracy of the traditional neural network and the dynamic aggregation neural network are recorded respectively under different training rounds to ensure that the data truly reflects the performance of the model in actual tasks. The experimental results show that the dynamic aggregation-based neural network used in the present invention has a higher accuracy.

[0144] Based on the same inventive concept, an embodiment of the present invention also provides a coal rock microscopic component identification device based on a dynamic aggregation strategy, including: an identification module, used to input the coal rock microscopic image into a pre-trained coal rock microscopic component identification neural network model to identify the microscopic component sub-classification of the coal rock microscopic image; wherein, the coal rock microscopic component identification neural network model is pre-trained according to the training method of the coal rock microscopic component identification neural network model in the above-mentioned embodiment one.

[0145] Based on the same inventive concept, an embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the above-mentioned method for identifying coal rock microscopic components based on the dynamic aggregation strategy is implemented.

[0146] Based on the same inventive concept, an embodiment of the present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and runnable on the processor. When the processor executes the program, the above-mentioned method for identifying coal rock microscopic components based on the dynamic aggregation strategy is implemented.

[0147] The principles of the problems solved by the above-mentioned devices, media, and related equipment in the embodiments of the present invention are similar to the aforementioned method for identifying coal rock microscopic components based on dynamic polymerization strategy. Therefore, its implementation can refer to the implementation of the aforementioned method, and the repeated parts will not be repeated.

[0148] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage and optical storage, etc.) containing computer-usable program code.

[0149] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0150] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0151] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0152] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

Claims

1. A method for training a neural network model for identifying coal rock microscopic components, characterized in that: include: Acquire a training sample set, where each sample in the training sample set includes a coal rock microscopic image and a microscopic component sub-classification label annotated on the coal rock microscopic image; A deep neural network model comprising one or more dynamic aggregation layers is trained using samples in the training sample set, wherein, after the coal rock microscopic image is input through the input layer of the deep neural network model, the microscopic component characteristics of the vitrinite and / or inertinite in the coal rock microscopic image are extracted through the dynamic aggregation layer, and the microscopic component sub-classification prediction labels of the coal rock microscopic image output by the output layer of the deep neural network model are compared with the microscopic component sub-classification labels in the sample, so as to estimate the model parameters in the deep neural network model and obtain a coal rock microscopic component identification neural network model.

2. The method according to claim 1, characterized in that Acquiring the training sample set includes: Acquire coal rock microscopic images in multiple imaging modes at the same location of the coal rock sample; wherein the multiple imaging modes include: single polarization imaging mode, reflected light imaging mode and fluorescence spectrum imaging mode; Classifying and labeling the microscopic component characteristics of the vitrinite and / or inertinite in the coal rock microscopic image, and cross-validating the single polarization labeling results using the reflected light characteristics and fluorescence spectrum characteristics of the coal rock microscopic image to determine the sub-classification labels of the microscopic components annotated in the coal rock microscopic image; Each coal rock microscopic image and its annotated microscopic component sub-classification label are taken as a sample to construct a training sample set.

3. The method according to claim 2, characterized in that Determining the microscopic component subclassification label annotated on the coal rock microscopic image includes: marking the microscopic component subclassification in each coal rock microscopic image that accounts for the largest proportion of the area of ​​the coal rock microscopic image as the unique microscopic component subclassification label of the coal rock microscopic image.

4. The method according to claim 3, characterized in that Each coal rock microscopic image and its annotated microscopic component sub-classification label is taken as a sample. Before constructing the training sample set, the following steps are also included: The image areas other than the coal rock microscopic image corresponding to the unique microscopic component subclassification in each coal rock microscopic image are eliminated.

5. The method according to any one of claims 1 to 4, characterized in that In a state where the deep neural network model includes a dynamic aggregation layer, the forward propagation dynamic aggregation layer extracts the microscopic component features of the vitrinite and / or inertinite group in the coal rock microscopic image to extract the microscopic component feature vector of the coal rock microscopic image, and performs feature aggregation based on the microscopic component feature vector to obtain the aggregated microscopic combination feature vector, and determines the holographic projection output based on the microscopic component feature vector; the aggregated microscopic combination feature vector and the holographic projection output serve as inputs for the next iteration; the backward propagation dynamic aggregation layer calculates the gradient of the weight matrix and the dynamic aggregation weight coefficient based on the loss function; the neural network model determines the loss function based on the aggregated microscopic component feature vector, the holographic projection output, the microscopic component sub-classification prediction label output by the model and the microscopic component sub-classification label in the sample to train the dynamic aggregation weight coefficient and the weight matrix in the dynamic aggregation layer to obtain a coal rock microscopic component identification neural network model; When the deep neural network model includes multiple dynamic aggregation layers, the first dynamic aggregation layer is forward propagated to extract the microscopic component features of the vitrinite and / or inertinite group in the coal rock microscopic image to extract the microscopic component feature vector of the coal rock microscopic image, the first dynamic aggregation layer performs feature aggregation based on the microscopic component feature vector to obtain the aggregated microscopic combination feature vector, and determines the holographic projection output based on the microscopic component feature vector; the microscopic component feature vector output by the first dynamic aggregation layer is used as the input of the next dynamic aggregation layer, and the aggregated microscopic combination feature vector and the holographic projection output output by the first dynamic aggregation layer are used as the input of the next iteration; each dynamic aggregation layer is back-propagated to calculate the gradient of the weight matrix and the dynamic aggregation weight coefficient based on the loss function; the neural network model determines the loss function based on the aggregated microscopic component feature vector, the holographic projection output, the microscopic component sub-classification prediction label output by the model and the microscopic component sub-classification label in the sample to train the dynamic aggregation weight coefficient and weight matrix in the dynamic aggregation layer to obtain a coal rock microscopic component identification neural network model.

6. A method for identifying coal rock microscopic components based on dynamic aggregation strategy, characterized in that: include: Inputting the coal rock microscopic image into a pre-trained coal rock microscopic component recognition neural network model to identify the microscopic component subclassification of the coal rock microscopic image; The coal rock micro-component identification neural network model is pre-trained according to the training method for the coal rock micro-component identification neural network model according to any one of claims 1 to 5.

7. A training device for a neural network model for identifying coal and rock microscopic components, characterized in that: include: An acquisition module is used to acquire a training sample set, where each sample in the training sample set includes a coal rock microscopic image and a microscopic component sub-classification label annotated on the coal rock microscopic image; A training module is used to train a deep neural network model containing one or more dynamic aggregation layers using samples in the training sample set, wherein, after the coal rock microscopic image is input through the input layer of the deep neural network model, the microscopic component characteristics of the vitrinite and / or inertinite in the coal rock microscopic image are extracted through the dynamic aggregation layer, and the microscopic component sub-classification prediction label of the coal rock microscopic image output by the output layer of the deep neural network model is compared with the microscopic component sub-classification label in the sample, so as to estimate the model parameters in the deep neural network model and obtain a coal rock microscopic component identification neural network model.

8. A device for identifying coal rock microscopic components based on dynamic polymerization strategy, characterized in that: include: A recognition module is used to input the coal rock microscopic image into a pre-trained coal rock microscopic component recognition neural network model to identify the microscopic component subclassification of the coal rock microscopic image; The coal rock micro-component identification neural network model is pre-trained according to the training method for the coal rock micro-component identification neural network model according to any one of claims 1 to 5.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, it implements the training method of the neural network model for identifying the microscopic components of coal and rock as described in any one of claims 1 to 5, or implements the method for identifying the microscopic components of coal and rock based on the dynamic aggregation strategy as described in claim 6.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, it implements the training method of the neural network model for identifying the microscopic components of coal and rock as described in any one of claims 1 to 5, or implements the method for identifying the microscopic components of coal and rock based on the dynamic aggregation strategy as described in claim 6.

Images