Coal Structure Recognition Method Based on LSTM Neural Network

Abstract

Coal structures exhibit considerable differences in rock properties and adsorption capacities. The physical properties of coal rocks are fundamental to understanding oil and gas reservoirs, while adsorption capacity directly impacts the gas content in coal seams. The accurate recognition of coal structures is essential for evaluating productivity and guiding coalbed methane well development. This study examines coal rocks of Benxi Formation in Ordos Basin. Using core photographs and logging curves, we classified the coal structures into undeformed coal, cataclastic coal, and granulated-mylonitized coal. AC, DEN, CAL, GR, and CN15 logging curves were selected to build a coal structure recognition model utilizing a long short-term memory (LSTM) neural network. This approach addresses the gradient vanishing and exploding issues often encountered in traditional neural networks, enhancing the model’s capacity to handle nonlinear relationships. After numerous iterations of learning and parameter adjustments, the model achieved a recognition accuracy of over 85%, with 32 hidden units, a minimum batch size of 28, and up to 150 iterations. Validation with independent well data not involved in the model building process confirmed the model’s effectiveness, meeting the practical needs of the study area. The results suggest that the study area is predominantly characterized by undeformed coal, with cataclastic coal and granulated-mylonitized coal more developed along fault trends.

1. Introduction

Coal-bearing basins commonly undergo multi-stage, complex tectonic evolution, which results in varying degrees of deformation and damage to coal measure strata, leading to the formation of different coal structures [1,2,3]. The Ordos Basin, one of China’s most significant coal basins, spans an area of approximately 37 × 10⁴ km² and is renowned for its substantial coal and natural gas reserves. In addition, the basin ranks among the top regions in the country in terms of oil resources and production, further emphasizing its geological importance. The basin hosts two major coal-bearing sequences from the Carboniferous–Permian and Jurassic periods, both of which are characterized by well-developed coal seams of considerable thickness. The Carboniferous–Permian coal seams are high-rank coals, mainly composed of gas coal and anthracite, with gas contents ranging from 2.46 to 23.25 m³/t. In contrast, the Jurassic coal seams are lower rank, primarily consisting of long flame coal, with gas contents ranging from 0.01 to 6.29 m³/t. The total coalbed methane (CBM) resource volume in the Ordos Basin is approximately 107,235.7 × 10⁸ m³, representing about one-third of China’s total CBM resources, underscoring the basin’s vast exploration and development potential [4].

The coal structure within these seams is the result of a series of geological processes that have acted upon the coal seams over time. It encompasses a wide range of characteristics, such as particle size distribution, coal particle shape, and the relationships between the different components within the coal seam. In coal seams that are significantly influenced by tectonic forces, coal structure directly impacts key petrophysical properties such as permeability, adsorption capacity, and rock mechanics, which in turn affect the effectiveness of production-enhancing techniques during coalbed methane exploration and development [5]. Hence, coal structure plays a crucial role in the economic feasibility and safe production of CBM, making its accurate identification a critical task in the exploration of coal seams and CBM resources [6,7,8,9].

Existing methods for coal structure identification can be broadly categorized into three groups: (1) direct calibration and observation of underground coal core samples; (2) correlation index-based methods; and (3) geophysical methods, primarily involving logging data [10,11]. While direct core sampling offers high accuracy in determining coal structure, it is limited by the fragile nature of coal and the high cost associated with core recovery. Additionally, it cannot provide continuous or real-time coal structure data. In contrast, geophysical methods, particularly logging-based techniques, offer a more cost-effective solution and can be employed for continuous coal structure recognition. Among these, seismic inversion methods have been widely applied, although they are technically demanding and expensive. Logging-based methods, such as those using gamma ray, resistivity, and density logs, are more commonly utilized due to their relatively lower cost and ease of implementation [12].

Recent studies have employed machine learning algorithms to improve the accuracy of coal structure identification from logging data. For example, Shi et al. [13] proposed a semi-supervised learning method based on Laplacian Support Vector Machine (LapSVM) to identify coal structure using a small amount of labeled logging data. LapSVM improves model performance and reduces the over-reliance on labeled data. In practical applications, it has been found that when labeled logging data are limited, LapSVM can serve as a reliable tool for coal structure identification. Thinesh et al. [14], using gamma ray, density, and resistivity logging data from four boreholes in the Talcher coalfield, Eastern India, applied supervised machine learning techniques, including Support Vector Machine (SVM), Decision Tree (DT), Random Forest (RF), Multi-Layer Perceptron (MLP), and Extreme Gradient Boosting (XGBoost). The results of model training showed that all prediction models achieved an accuracy of over 88% in classifying carbonaceous and non-coal lithofacies. Despite the promising results, challenges remain, particularly in overcoming the multi-solution problem often encountered in traditional methods [15,16,17,18]. The key challenge in coal structure identification is the ability to resolve the ambiguities inherent in logging curve morphology, especially when the same curve morphology can correspond to multiple coal structures.

In this context, the integration of artificial neural networks (ANNs), particularly long short-term memory (LSTM) networks, offers a promising solution. LSTM networks are capable of learning complex nonlinear relationships between input data (such as logging curves) and output data (coal structure) due to their self-learning capabilities and the ability to capture long-term dependencies in data sequences. Currently, artificial neural networks (ANN), due to their strong nonlinear mapping ability, adaptability, function fitting, and self-learning capabilities, have achieved ideal results in fields such as coal texture prediction [19], coal structure identification [20], reservoir facies classification [21], and lithology recognition [22]. These applications not only improve accuracy but also significantly enhance efficiency. The core of coal structure identification based on logging curves is to achieve a nonlinear mapping between logging curve values and coal structure identification results [23].

Traditional neural networks, such as recurrent neural networks (RNNs) and backpropagation (BP) networks, effectively reduce human interference and increase recognition speed [24]. However, when learning and processing long sequences, these networks often encounter gradient vanishing and exploding issues during error minimization, which weakens their ability to handle nonlinear relationships. To address this limitation, a gating structure was introduced in each recurrent unit, leading to the development of the long short-term memory (LSTM) neural network [25]. With its distinctive hidden layer neuron connections, LSTM has been widely applied across various disciplines [26], enabling the establishment of complex, high-dimensional mapping relationships and making it especially suitable for solving intricate nonlinear geological problems [27].

This study aims to develop a quantitative coal structure recognition model based on LSTM networks, using conventional logging data and coal core data to address the nonlinearity and multi-solution challenges in coal structure recognition. The model is built on data from seven wells in the Ordos Basin and is trained and optimized within MATLAB R2024a to ensurWe have checked the full text and there is no such probleme reliable and cost-effective coal structure identification. The findings are expected to provide significant support for CBM resource exploration and development in the Ordos Basin and other similar coal-bearing regions.

2. Logging Response Characteristics of Coal Structure

2.1. Coal Structure Division

Coal structure refers to the structural characteristics of coal seams that have been altered by complex tectonic forces during their formation and evolutionary processes. These structural variations significantly influence the physical and chemical properties of the coal. Key factors such as porosity [28], maceral composition [29], and pore structure [30] play crucial roles in determining the characteristics of coal structures. Various coal structure recognition methods have been developed, based on fragmentation degree, genetic type, and structural features, with the three-division and four-division methods being the most commonly used [31].

In the study area, analysis of coal core data reveals that the predominant coal structures are undeformed and cataclastic in nature. Undeformed coal is more prone to mechanical damage during core sampling, complicating the differentiation between granulated coal and mylonitized coal. Both granulated and mylonitized coals exhibit poor mechanical properties, which can lead to substantial coal powder generation under external fracturing conditions. This can obstruct the seepage channels within coal seams and potentially damage mining equipment, thus presenting significant safety risks in coalbed methane extraction [32].

Based on these observations, the coal structures in this study are classified into three types: (1) undeformed coal seam, (2) cataclastic coal seam, and (3) granulated-mylonitized coal seam (

Table 1. Core characteristics of different coal structure types in the study area.

Coal Structure Type	Degree of Coal Fragmentation	Physical Properties	Fracture Development Characteristics	Typical Core Photograph
Undeformed coal	The structure of the coal body is complete, and there is no obvious displacement	Appear black columnar, with a metallic luster, hard texture	Cleft development, rare ectogenesis cracks
Cataclastic coal	The structure of coal body is relatively complete, showing lenticular block and obvious relative displacement	Appear black columnar, with a metallic luster, hard texture	Cleft is developed and cleaved by exogenous fractures
Granulated-mylonitized coal	The structure of the coal body was seriously damaged, which was lumpy or broken, and had unconformable contact with the upper and lower layers	Black lump, dim luster, low hardness, easy to twist powder	The coal body is loose and develops a large number of ectogenesis cracks. It is impossible to distinguish the primary cleat or present broken grain and powder

).

(1)

Undeformed coal: This coal structure remains largely intact, characterized by angular, massive forms with no significant displacement. The coal and rock exhibit well-developed bedding, cleating, and high hardness, with no substantial signs of structural deformation.

(2)

Cataclastic coal: This type has undergone moderate structural damage, exhibiting lenticular block formations with noticeable relative displacement. The original layered structure is largely disrupted, with clear exogenous fractures crossing through the coal and rock clefts, leading to a local lenticular appearance.

(3)

Granulated-mylonitized coal: This type shows more severe structural damage compared to cataclastic coal, often appearing lumpy or fragmented into granulated pieces. The bedding structure is minimal, and the coal layers exhibit unconformity contacts with significant secondary fracturing. The coal is highly friable, frequently breaking into granular fragments and powder.

2.2. Logging Response Characteristics of Different Coal Structures

Under the influence of complex geological structures, coal seam structures have undergone varying degrees of deformation, resulting in distinct differences in the geophysical properties of the coal. This study provides a comprehensive analysis by integrating logging data with coal core samples to investigate the variations in logging responses across different coal structures. The objective is to accurately identify and characterize these structural differences.

Through the analysis of coal core and logging data from the study area, the following characteristics of various coal structures, as observed on logging curves, are summarized (Figure 1). As the degree of damage to the coal seam increases, the development of external fissures intensifies, leading to greater looseness in the coal. As a result, undeformed coal is typically characterized by low density, low radioactivity, and high acoustic time differences.

Furthermore, the increased damage to the coal seam raises the likelihood of wellbore collapse during drilling, which can significantly enlarge the wellbore logging. When the coal seam fractures, the underground pressure balance is disrupted, potentially allowing formation water to infiltrate through certain fractures. This infiltration leads to a marked increase in the hydrogen content index, as measured by compensated acoustic logging.

Based on the analysis of the logging response characteristics of different coal body structures, significant differences were observed in the conventional logging response curves across various coal types. However, a considerable amount of overlapping or stacking was also observed in the logging curve data for different coal body structures (

Table 2. The logging response distribution range of different coal structures in the study area.

Coal Structure Type	GR/(API)	DEN/(g/cm3)	AC/(μs/m)	CAL/(cm)	CNL/(%)
Undeformed structure	109.2~208.8	1.96~2.62	210.9~316.8	21.5~25.2	18.8~51.3
Cataclastic structure	38.6~118.8	1.23~1.98	328.1~416.2	21.2~30.1	40.1~64.1
Granulated-mylonitized structure	16.2~82.2	1.18~1.67	366.5~480.6	24.3~35.5	36.6~68.8

Note: Data 109.2 to 208.8 indicate the minimum to maximum values.

and Figure 2). As a result, establishing linear relationships between the logging curve data and the different coal structures often leads to multiple solutions.

To address this challenge, the application of neural networks provides an effective solution for handling nonlinear mapping problems, mitigating the impact of curve stacking between different coal structures. This method significantly enhances the accuracy of coal structure recognition, improving the reliability of the results.

3. LSTM Neural Network Model for Coal Structure Recognition

3.1. Basic Principle of LSTM Neural Network

Conventional neural networks, such as Recurrent Neural Networks (RNNs) and Backpropagation Neural Networks (BPNNs), employ gradient-based techniques to minimize errors. However, these networks face challenges with “long-term memory” issues, where the process of error reduction can lead to problems like gradient vanishing or explosion, resulting in unsatisfactory recognition accuracy [33]. Hinton et al. [34] emphasized that the main advantage of neural networks lies in their ability to automatically extract and abstract features. This capability eliminates the cumbersome task of manual feature extraction and facilitates the automatic identification of complex and effective high-order features. In recent years, fully connected deep neural networks have been successfully applied to predict clay content [35] and porosity [36], achieving promising results.

Recurrent Neural Networks (RNNs) are a key architecture within deep neural networks. Unlike fully connected deep neural networks (DNNs), where units in hidden layers are independent, RNNs feature interconnected hidden layers that depend on sequential inputs received prior to the current time step. This temporal dependency makes RNNs particularly effective for handling time-dependent data. Long short-term memory (LSTM) networks improve upon traditional RNNs by addressing issues like gradient vanishing through the careful design of recurrent units. As a result, LSTMs have become one of the most successful types of recurrent neural networks in practical applications.

The LSTM network enhances recognition results by incorporating forgetting gates, input gates, and output gates within each neuron (Figure 3). At each time step, the LSTM receives three components as input: the cell memory C_t−1 from the previous time step, the cell state h_t−1 from the previous time step, and the current input information x_t from the current time step.

The information processing within the LSTM begins by passing the previous cell state h_t−1 and the current input x_t through the forgetting and input gates. This allows the network to discard irrelevant information and integrate the current input data to update the cell memory. The updated cell memory is then combined with the previous cell state h_t−1 via the output gate, resulting in the final updated cell state h_t.

Forgetting gate ${\alpha }_1$: Input the cell state h_t−1 at the last moment and the input information x_t at the current moment to output the forgotten information;

$${\mathrm{f}}_{\mathrm{t}}={\alpha }_1({\omega }_{xf}·x_t+{\omega }_{hf}·h_{t-1}+b_f)$$

Input gate ${\mathsf{\alpha }}_2$: Input the current input information x_t and the cell state h_t−1 at the last time, output the required new information, and add it to the cell memory to update the cell memory;

$$i_t={\mathsf{\alpha }}_2({\omega }_{xi}·x_t+{\omega }_{hi}·h_{t-1}+b_i)$$

$${\mathrm{g}}_{\mathrm{t}}=\tan \mathrm{h}({\omega }_{xg}·x_t+{\omega }_{hg}·h_{t-1}+b_g)$$

C_t = f_t·C_t−1 + i_t·g_t

Output gate ${\mathsf{\alpha }}_3$: Input information x_t at the current time, cell state h_t−1 at the last time, and updated cell memory C_t, and output the updated cell state h_t;

$$O_t={\alpha }_3({\omega }_{xO}·x_t+{\omega }_{hO}·h_{t-1}+b_O)$$

$$h_t=O_t\mathit{tanh}\left(C_t\right)$$

In the formula: f_t, i_t, and O_t are the outputs at the time of forgetting gate, input gate, and output gate t, respectively; ${\mathsf{\omega }}_{\mathrm{x}\mathrm{f}}, {\mathsf{\omega }}_{\mathrm{x}\mathrm{i}}, \mathrm{a}\mathrm{n}\mathrm{d} {\mathsf{\omega }}_{\mathrm{x}\mathrm{O}}$ represents the weights of the forgetting gate, input gate, and output gate, respectively; b_f, b_i, and b_o are biases for forgetting gates, input gates, and output gates, respectively; g_t is the output of tanh layer.

3.2. Logging Data Preprocessing

The analysis of the logging response characteristics for different coal body structures reveals that, despite the stacking phenomenon observed in the five logging curves—namely CAL, GR, AC, CNL, and DEN—there remains a significant correlation among them. Therefore, this study selects these five logging curves as characteristic variables for coal body structure recognition.

To mitigate the impact of varying dimensions across the logging curves on coal structure recognition, it is essential to standardize the input data prior to model training. Standardization not only improves the accuracy of the results but also reduces the computational load. In this study, the data from each logging curve are normalized using standardization formulas, ensuring that the values are scaled to a range between 0 and 1, as shown in Equation (1).

$$\mathsf{\Delta }X={\displaystyle \frac{X-Xmin}{Xmax-Xmin}}$$

(1)

In the equation, ΔX represents the standardized logging curve data; X refers to the raw logging curve data; X_min is the minimum value for each logging curve data set; and X_max is the maximum value for each logging curve data set.

This study selected 21 coal core samples from 7 core wells in Area A. The samples include 7 undisturbed coal core samples, corresponding to 210 logging data points; 9 cataclastic coal core samples, with 285 logging data points; and 5 granulated-mylonitized coal core samples, with 145 logging data points. To improve the accuracy and reliability of the LSTM neural network model in identifying coal structures, 70% of the total logging data samples for each coal structure type were randomly selected as training samples, while the remaining 30% of the data samples were set aside for validation.

3.3. Construction and Training of Coal Structure Recognition Model Based on LSTM Neural Network

To extract context from multidimensional information, we utilize the forget gate of the LSTM unit, enabling the network to effectively access multi-directional context.

At each input x in the array, the network receives N hidden vectors h₁,…, h_N and N memory vectors m₁,…, m_N. It calculates a new hidden vector h and a memory vector m, which are passed to the next state in vector form. The network concatenates the transformed input I*N with the N hidden vectors h₁,…, h_N to form the vector H and computes the memory vector using the gating mechanisms as follows:

$$m={\sum }_i^N{g}_i^f\odot m_i+g^u\odot g^c$$

Since LSTM can capture contextual information in multiple directions, this method performs well in coal structure prediction. However, high-intensity information processing increases computational costs. To enhance efficiency, Dropout layers are incorporated into the model to mitigate overfitting and improve operational efficiency.

The proposed prediction model, as shown in Figure 4, comprises the following components: an input layer, three LSTM layers, two Dropout layers, and an output layer combining CTC and Softmax. The model is optimized through the Backpropagation Through Time (BPTT) algorithm, which adjusts parameters based on the error between actual and expected outputs. The steepest descent method is employed to modify network weights and achieve the optimal structure.

High-dimensional logging data are fed into three LSTM layers, each representing a scanning direction. The internal state and output of the LSTM units are computed from the previous positions’ states and outputs along horizontal and vertical directions, as follows:

$$\left(h_{i,j},q_{i,j}\right)=LSTM\left(x_{i,j},h_{i,j\pm 1},\right.q_{i,j\pm 1},\left.q_{i\pm 1,j}\right)$$

Here, $x_{i,j}$ represents the input feature vector at position (i, j), and h and q denote the unit’s output state and internal state, respectively. The ±1 symbol depends on the scanning direction under consideration.

For the optimized model, the workflow is summarized into the following seven steps:

• Normalize the three-dimensional seismic attribute data and sample a result set for backup;
• Perform feature learning on the sample data set;
• Train the LSTM model, using the lower layer output as input for the upper layer;
• After feature learning and training, obtain a trained feature representation model;
• Fine-tune the parameters in the top layer using the BPTT algorithm from front to back;
• Use Softmax as a regression classifier to output the probability of the prediction results;
• Compute the cross-entropy loss of the model and optimize it iteratively using a gradient-based algorithm to achieve the best model.

3.4. Verification of Coal Structure Recognition Results and Accuracy

The coal structure recognition model achieved an accuracy of 85.5% on the validation sample set, as summarized in

Table 3. Coal structure recognition results of LSTM neural network model in the study area.

Coal Structure Type	Sample Size of Validation Set	Accurate Recognition Number	The Recognition Accuracy
Undeformed coal	63	57	90.5%
Cataclastic coal	86	72	83.4%
Granulated-mylonitized coal	44	36	81.8%
Total	193	165	85.5%

. Specifically, the recognition accuracies for undeformed coal, cataclastic coal, and granulated-mylonitized coal were 90.5%, 83.4%, and 81.8%, respectively. Overall, the model’s performance meets the requirements for practical applications.

A deeper analysis of the model’s recognition accuracy indicates that undeformed coal exhibits the highest accuracy, while cataclastic coal and granulated-mylonitized coal show relatively lower recognition rates. This discrepancy is primarily due to the mechanical crushing of cataclastic coal during the coring process, which resulted in core samples being misclassified as granulated-mylonitized coal. Furthermore, the transitional nature of the coal body structure between cataclastic coal and granulated-mylonitized coal introduces additional challenges in accurately distinguishing these two coal types during model training.

4. Discussion

4.1. Parameter Analysis of LSTM Network Model

In this study, key network parameters influencing the accuracy of coal structure recognition using the LSTM neural network include the number of training iterations (Epoch), the number of hidden units (Num Hidden Units), and the minimum batch size (Min Batch Size).

The epoch parameter defines the maximum number of training iterations for the LSTM neural network. If the number of epochs is excessively high, the network may experience poor generalization, as it becomes overly sensitive to individual training samples, leading to overfitting. This, in turn, can decrease the model’s accuracy when applied to unseen data. To optimize the selection of the epoch parameter, the control variable method was used, wherein only the epoch parameter was varied while observing changes in the cross-entropy loss function.

As shown in Figure 5, when the number of iterations reached 150, the loss curve stabilized, and the model accuracy approached 90%. Therefore, 150 epochs were selected for training.

The number of hidden units in a neural network is a critical parameter that significantly impacts the accuracy of the model’s recognition results. If the number of hidden units is too small, the network may lack sufficient capacity for learning and processing information. On the other hand, an excessive number of hidden units can result in an overly complex network structure, leading to slower learning rates and potentially causing the network to get stuck in local minima during the training process.

In this study, the control variable method was employed to optimize the number of hidden units by observing changes in the accuracy of the training set while varying only the number of hidden units. As shown in Figure 6, when the number of hidden units is set to 32, the recognition accuracy for the training set samples reaches 96.4%. Therefore, the number of hidden units was set to 32 for this study.

The minimum batch size refers to the number of output samples processed in a single batch during the training of a network model. In deep learning, to address the slow training speed that can occur when large volumes of data are inputted at once, a small-batch approach is often used. However, if the minimum batch size is too small, the randomness of the network’s learning process may become excessive, which could hinder model performance.

In this study, the control variable method was applied to optimize the minimum batch size by observing changes in the accuracy of the training set while varying only the batch size. As shown in Figure 7, when the minimum batch size was set to 28, the recognition accuracy of the training set samples reached a peak of 97.2%. Therefore, the minimum batch size was set to 28.

4.2. Model Reliability Analysis

To further validate the reliability of the coal structure recognition model built using the LSTM neural network, verification wells located outside the model construction area were selected. In one such verification well, an 8 m coal seam was identified at a depth of 2168 to 2176 m. The coal structure characteristics in the verification well were assessed using the model. From bottom to top, the identified coal types were as follows: undeformed coal (3.5 m), cataclastic coal (2.5 m), and granulated-mylonitized coal (3 m). Three coal cores were identified in the verification well:

(1)

The No. 1 coal core, located at a burial depth of 2170 to 2171.875 m with a length of 1.875 m, is classified as undeformed coal.

(2)

The No. 2 coal core, situated at a depth of 2172.625 to 2173.625 m and measuring 1 m in length, is identified as cataclastic coal.

(3)

The No. 3 coal core, buried between 2174.625 and 2175.875 m, has a length of 1.2 m and is classified as granulated-mylonitized coal.

The comparison between the model’s recognition results and the coal core data demonstrates strong consistency (Figure 8), thereby validating the reliability of the coal body structure recognition model developed using the LSTM neural network.

4.3. Distribution Characteristics of Coal Structures

An analysis of the planar distribution characteristics of coal structures in the study area was conducted to clarify the distribution patterns, thereby providing a more solid theoretical foundation for coalbed methane development.

Based on the coal structure identification results from the LSTM neural network, an analysis of the planar distribution characteristics of coal structures in the study area was conducted (Figure 9). The findings reveal that undeformed coal predominates in the area, followed by cataclastic coal, while granulated-mylonitized coal shows the lowest degree of development.

Additionally, a northeast-trending fault is primarily developed in the study area, with a structural high point located in the central region and relatively lower structural elevations on both sides [37]. Structural analysis indicates that the central part of the study area is subjected to strong geological stress, leading to uplift in this region, which has resulted in the formation of numerous faults and corresponding changes in coal structure. Along the fault trend, cataclastic coal progressively appears from south to north, while granulated-mylonitized coal is sparsely distributed and observed only in the northern and central regions. This distribution pattern suggests that fault activity in the central area caused coal structure fragmentation, while the relatively stable structures on both sides were less affected by fault activity, resulting in a wider distribution of undeformed coal on either side.

5. Conclusions

(1)

Based on the analysis of coal core data from the Benxi Formation in Area A of the Ordos Basin and the observed challenges posed by granulated and mylonitized coal on coalbed methane extraction, the coal structure in Area A is classified into three distinct types: (1) undeformed coal, which displays minimal structural disruption, characterized by angular blocky masses; (2) cataclastic coal, which shows moderate structural damage, forming lenticular blocks with visible fractures; (3) granulated-mylonitized coal, which has undergone significant deformation, resulting in granular fragments with unconformable contacts with adjacent layers and extensive secondary fracturing.

(2)

Research indicates that undeformed coal is characterized by a low density, low radioactivity, and high acoustic transit time. Specifically, the density (DEN) ranges from 1.96 to 2.62 g/cm³, the gamma ray (GR) value is typically over 100 API, and the acoustic transit time (AC) ranges from 210.9 to 316.8 μs/m. In comparison, the density (DEN) of the cataclastic structure ranges from 1.23 to 1.98 g/cm³, with a gamma ray (GR) of 38.6 to 118.8 API, and an acoustic transit time (AC) of 328.1 to 416.2 μs/m. For the granulated-mylonitized structure, the density (DEN) ranges from 1.18 to 1.67 g/cm³, the gamma ray (GR) ranges from 16.2 to 82.2 API, and the acoustic transit time (AC) is 366.5 to 480.6 μs/m.

(3)

This study integrates forgetting gates, input gates, and output gates within the framework of an LSTM neural network to develop a coal structure recognition model. By utilizing five logging curves—CAL, GR, AC, CNL, and DEN—the model significantly improves the accuracy of coal structure identification, achieving an overall accuracy of 85.5%.

(4)

The undeformed coal is more developed in the study area, while the development of cataclastic coal and granulated-mylonitized coal is less pronounced. Along the fault trend, the development of cataclastic coal and granulated-mylonitized coal is more prominent, indicating that fault activity influenced the coal structure.