A MS-BFI framework and method for identifying shale bedded fractures

Abstract

The present application relates to a kind of MS-BFI framework for identifying shale bedding seam and the method for identifying bedding seam, conditional generation confrontation network includes generator and discriminator, generator generates the pseudo sample data simulating the balanced distribution of various logging data based on conditional information and random noise, discriminator is used to distinguish the true probability of pseudo sample data;Conditional information is prior constraint information;Isolation forest model is used to eliminate abnormal data in pseudo sample data;Mixed classifier is trained based on the mixed data set of sample data set containing logging data and pseudo sample data, and the classification result of bedding seam is output;CGAN is responsible for generating sample, to expand minority sample, improve the quality of generated data, so as to solve the problem of uneven distribution and lack of diversity;Noise filtering scheme is designed based on iForest algorithm to improve the reliability of generated sample;It can realize multi-step accurate prediction of bedding seam to unevenly distributed logging data.

Description

Technical Field

[0001] This invention relates to the field of well logging technology, and in particular to an MS-BFI framework for identifying shale bedding fractures and a method for bedding fracture identification. Background Technology

[0002] Stratification fractures are rock fractures formed during diagenesis and tectonic processes, parallel or nearly parallel to bedding interfaces. They are a common type of natural fracture in fine-grained sedimentary rocks. These fractures are widely distributed in shale and are one of the main fracture types, having a crucial impact on shale reservoir quality, hydrocarbon accumulation, single-well productivity, and fracturing effectiveness. Therefore, accurate identification of stratification fractures is essential for shale reservoir evaluation.

[0003] With the development of artificial intelligence technology, machine learning has become an important tool in the petroleum industry, providing a new perspective for fracture identification using geophysical logging data. Compared with traditional fracture identification methods (such as statistical analysis and cross-plot template methods), machine learning, with its advantages in multidimensional data processing, automatic feature learning, and nonlinear modeling, can more effectively capture the inherent mapping relationship between target features and latent features. As a data-driven technology, the accuracy of machine learning is highly dependent on the quantity and quality of input data. Typically, standard machine learning algorithms assume the existence of a large-scale and balanced training dataset. However, in practical applications, the vast majority of the collected logging data represents non-fracture samples, while the corresponding fracture samples are extremely scarce. This imbalance causes the machine learning classifier to be severely biased towards the majority class (i.e., non-fracture) during training, thus ignoring the minority class (i.e., fracture) and significantly reducing the model's fracture identification performance.

[0004] To date, various methods have been proposed to address the class imbalance problem. These methods can be broadly categorized into data-level approaches (such as oversampling and undersampling) and algorithm-level approaches (such as cost-sensitive learning and ensemble learning). While these methods have made significant contributions to solving the data imbalance problem, they each exhibit distinct drawbacks and limitations when applied to extremely imbalanced data. Specifically, data resampling, though simple to implement, can lead to information loss, introduce noise, and distort the original data structure. Cost-sensitive learning is highly compatible with base classifiers and computationally efficient, but its effectiveness largely depends on preset cost parameters—adjusting these parameters requires extensive prior knowledge and empirical parameter tuning, often failing to achieve optimal results when class imbalance is severe. Ensemble learning, by integrating multiple classifiers, possesses strong robustness and generalization capabilities and can reduce the risk of overfitting; however, its implementation is complex, costly, and time-consuming, and the effective combination of multiple classification models still faces significant challenges.

[0005] Generative Adversarial Networks (GANs), as generative models based on deep learning algorithms, have attracted widespread attention in the field of imbalanced classification due to their powerful data augmentation capabilities. In recent years, GANs have been introduced into the petroleum geology industry, providing an effective solution to the data imbalance problem in well logging identification of bedding fractures. However, directly applying the original GAN ​​model for fracture identification still faces many challenges: 1) The standard GAN architecture is an unsupervised generative algorithm, and the random data it generates lacks specific meaning and cannot cover the overall distribution characteristics of real data; these generated data can only represent a part of the original sample dataset. This may lead to insufficient diversity in model generation and potential bias, i.e., mode collapse. 2) Low-quality samples or anomalous samples inevitably exist in the generated data (for example, some generated samples located on the decision boundary are easily misclassified into other categories), which will reduce the contribution of the generated data to the classifier training; this effect will gradually intensify with the increase of model iterations, eventually damaging the model's performance. 3) The formation of bedding fractures is a continuous geological process, so the well logging response of fractures not only exhibits numerical anomalies but also shows spatiotemporal correlations. Summary of the Invention

[0006] This invention addresses the technical problems existing in the prior art by providing an MS-BFI framework and a method for identifying shale bedding fractures. By comprehensively utilizing the data augmentation capabilities of Conditional Generative Adversarial Network (CGAN), the anomaly detection capabilities of Isolation Forest (iForest), the local feature extraction capabilities of Convolutional Neural Network (CNN), and the sequence modeling capabilities of Bidirectional Long Short-Term Memory Network (BiLSTM), it achieves a dual improvement in the accuracy and stability of bedding fracture identification under complex geological conditions.

[0007] According to a first aspect of the present invention, an MS-BFI framework for identifying shale bedding fractures is provided, comprising: a conditional generative adversarial network, an isolated forest model, and a hybrid classifier; The conditional generative adversarial network includes a generator and a discriminator. The generator generates pseudo-sample data that simulates the balanced distribution of various well logging data based on conditional information and random noise. The discriminator is used to determine the true probability of the pseudo-sample data. The conditional information is prior constraint information. The isolated forest model is used to remove abnormal data from the pseudo-sample data; The hybrid classifier is trained on a mixed dataset containing both well logging data and pseudo-sample data, and outputs the classification result of the bedding fracture.

[0008] Based on the above technical solution, the present invention can also be improved as follows.

[0009] Optionally, the objective function of the conditional generative adversarial network is:

[0010] in, This represents the discriminator network. Represents a generator network; Represents the value function; The distribution of the real data is represented by x; x represents a distribution that follows a certain distribution. Real data; This represents the condition information embedded in the input terminals of the generator and discriminator; z represents a random noise distribution; z represents a distribution that follows a random noise distribution. Random noise; and Representing condition information The outputs of the generator and discriminator described below; It represents the mathematical expectation of the distribution function.

[0011] Optionally, the condition information may take the form of: category labels, attribute parameters, or multi-source modal data; The category labels include: crack density labels.

[0012] Optionally, the generator takes the concatenated vector as input data and outputs the pseudo-sample data; the concatenated vector consists of uniformly distributed random noise and the conditional information. The discriminator takes the pseudo-sample data and real data as input and outputs a scalar prediction score of the probability of classifying the pseudo-sample data as real data. The optimal value of the objective function of the conditional generative adversarial network is solved by alternately executing the network optimization of the generator in step 1 and the network optimization of the discriminator in step k; k is a preset positive integer greater than 1.

[0013] Optionally, the training process of the isolated forest model includes: Constructing an isolation tree from the sample dataset Randomly select a subset of samples , and send to the root node of the isolated tree; Randomly select an attribute of a sample In this attribute Select the splitting threshold within the range of the maximum and minimum values. According to the splitting threshold The sample subsets are sequentially Each sample is divided into the left or right subtree until each node contains only a single sample point, or the binary tree reaches the preset height limit.

[0014] Optionally, the constructed isolated tree uses anomaly scores to perform anomaly detection on the pseudo-sample data, removing abnormal data from the pseudo-sample data; The formula for calculating the abnormal score is: ; in, Let represent any i-th input data, and n represent the number of samples; This represents the path length, which is defined as the number of edges traversed from the root node to an external node in the isolated tree. This represents the average path length for all isolated trees. Indicates the normalization factor; ; For the harmonic operator.

[0015] Optionally, the hybrid classifier includes: a convolutional neural network, a bidirectional long short-term memory network, and a classification head; The convolutional neural network automatically extracts local logging features induced by bedding fractures in the logging data based on a local receptive field and weight sharing mechanism. The bidirectional long short-term memory network processes the input sequence data bidirectionally, and fuses local logging features and long-term dependencies between logging features to obtain a fused feature vector. The input to the classification head is the fused feature vector. A mapping relationship is established between the fused feature vector and the bedding seam category, and the classification probability of the bedding seam is output.

[0016] According to a second aspect of the present invention, a method for identifying layered seams based on the MS-BFI framework is provided, comprising: Step 1: Obtain raw logging data and label the raw logging data according to the core observation and imaging logging interpretation results. After preprocessing the labeled raw logging data, a sample dataset is constructed. Step 2: Input the sample dataset into the conditional generative adversarial network, which generates each pseudo-sample data with the goal of balancing the distribution of samples with different labels. Step 3: Remove outlier data from the pseudo-sample data based on the isolated forest model; Step 4: Train the hybrid classifier using the hybrid dataset, input the logging data to be tested into the trained hybrid classifier, and the hybrid classifier outputs the classification result of the bedding fracture.

[0017] Optionally, the preprocessing of the logging data in step 1 includes: Outlier removal; the original logging data are normalized to a uniform range using a min-max normalization function; tensor sample data are generated using a sliding window method according to a set window height and sliding step size.

[0018] Optionally, step 4 may be followed by: Step 5: Test the hybrid classifier based on the test data and blind well data, and evaluate the performance of the hybrid classifier by combining the confusion matrix.

[0019] This invention provides an MS-BFI framework and method for identifying shale bedding fractures. Addressing the uneven class distribution and nonlinear characteristics of conventional well logging data, a novel MS-BFI fusion framework is proposed. This MS-BFI framework integrates Conditional Generative Adversarial Network (CGAN), Isolation Forest (iForest), Convolutional Neural Network (CNN), and Bidirectional Long Short-Term Memory (BiLSTM). CGAN generates samples to expand minority samples and improve the quality of generated data, thereby solving the problems of uneven distribution and insufficient diversity. A noise filtering scheme based on the iForest algorithm is designed to improve the reliability of generated samples. A hybrid classifier is constructed by combining CNN and BiLSTM to capture the sequential features of well logging data and improve the model's accuracy and generalization ability. The collaborative work of each unit enables the MS-BFI model to accurately predict bedding fractures in multiple steps from imbalanced well logging data. Experiments on real datasets and comparisons with various classic data augmentation techniques demonstrate the superior performance of the proposed method. Application to conventional well logging datasets of Silurian marine shale in the southern Sichuan Basin shows that the proposed MS-BFI model achieves identification accuracies of 94.59% and 94.66% on test data and blind well data, respectively, validating the effectiveness of the method in identifying bedding fractures. This method provides a robust and effective solution for identifying bedding fractures under conditions of data imbalance and complex geological conditions, and is expected to find practical applications in the petroleum industry. Attached Figure Description

[0020] Figure 1 A flowchart illustrating the identification process of an embodiment of the MS-BFI framework for identifying shale bedding fractures provided by the present invention; Figure 2 This is a schematic diagram of the frequency distribution of the linear density of bedding joints provided in an embodiment of the present invention; Figure 3 Cross-plot of conventional logging data for bedding fractures of different density categories provided in embodiments of the present invention; Figure 4 This is a schematic diagram showing the density distribution comparison between actual data and generated data, and the KS test results provided in this embodiment of the invention. Figure 5This is a schematic diagram comparing the Spearman correlation coefficient between the actual data and the generated data provided in this embodiment of the invention. Figure 6 This is a schematic diagram illustrating the result of using the Adam optimizer to optimize the hyperparameters of CNN-BiLSTM through grid search, as provided in an embodiment of the present invention. Figure 7 A schematic diagram of the training and testing curves of the CNN-BiLSTM model provided in this embodiment of the invention; Figure 8 A schematic diagram of the confusion matrix of the training set, test set, and blind test set provided in an embodiment of the present invention; Figure 9 This is a schematic diagram of the identification results of the bedding fractures in blind well B1 provided in an embodiment of the present invention. Detailed Implementation

[0021] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0022] Figure 1 A flowchart illustrating an embodiment of the MS-BFI framework for identifying shale bedding fractures provided by the present invention is shown below. Figure 1 As shown, the MS-BFI framework includes: conditional generative adversarial networks, isolated forest models, and hybrid classifiers.

[0023] Conditional generative adversarial networks consist of a generator and a discriminator. The generator generates pseudo-sample data that simulates the equilibrium distribution of various well logging data based on conditional information and random noise. The discriminator is used to determine the true probability of the pseudo-sample data. The conditional information is prior constraint information.

[0024] The Isolation Forest model is used to remove outlier data from spurious sample data.

[0025] The hybrid classifier is trained on a mixed dataset containing both sample datasets and pseudo-sample datasets, and outputs the classification results of bedding fractures.

[0026] This invention provides an MS-BFI framework for identifying shale bedding fractures. Addressing the uneven class distribution and nonlinear characteristics of conventional well logging data, this MS-BFI framework integrates Conditional Generative Adversarial Network (CGAN), Isolation Forest (iForest), and a hybrid classifier. CGAN generates samples to expand minority samples and improve the quality of generated data, thus solving the problems of uneven distribution and insufficient diversity. A noise filtering scheme based on the iForest algorithm is designed to improve the reliability of generated samples. The hybrid classifier captures the sequential features of well logging data and improves the model's accuracy and generalization ability. The collaborative work of each unit enables the MS-BFI model to accurately predict bedding fractures in multiple steps from imbalanced well logging data. Experiments on real datasets and comparisons with various classic data augmentation techniques demonstrate the superior performance of the proposed method.

[0027] Example 1

[0028] Embodiment 1 provided by this invention is an embodiment of an MS-BFI framework for identifying shale bedding fractures provided by this invention, combined with... Figure 1 It is known that the implementations of the MS-BFI framework include: conditional generative adversarial networks, isolated forest models, and hybrid classifiers.

[0029] Conditional generative adversarial networks consist of a generator and a discriminator. The generator generates pseudo-sample data that simulates the equilibrium distribution of various well logging data based on conditional information and random noise. The discriminator is used to determine the true probability of the pseudo-sample data. The conditional information is prior constraint information.

[0030] In one possible implementation, the core feature of CGAN lies in its unique conditional input mechanism, which allows the model to precisely control the generator and discriminator networks in a supervised manner. This information-guided approach enables the CGAN model to learn unique features from various conditional inputs, thus giving the generation process the flexibility to reflect diverse input trends. Guided by conditional information, CGAN can generate data specific to particular categories; therefore, the data generated by CGAN is more structured and diverse, effectively solving the model collapse problem.

[0031] The goal of the generator network (G) is to capture the distribution characteristics of the real data as accurately as possible, and to minimize the objective function:

[0032] The goal of the discriminator network (D) is to evaluate the probability (ranging from 0 to 1) of whether the generated data is real, and to maximize the objective function:

[0033] CGAN modifies the generator and discriminator networks to integrate additional information. The generator's input data includes a random noise vector and conditional information, while the discriminator's input data contains a combination of generated data and conditional information. The objective function of CGAN is:

[0034] in, This represents the discriminator network. Represents a generator network; Represents the value function; The distribution of the real data is represented by x; x represents a distribution that follows a certain distribution. Real data; This represents the conditional information at the input of the embedded generator and discriminator; z represents a random noise distribution; z represents a distribution that follows a random noise distribution. Random noise; and Representing condition information The outputs of the generator and discriminator; It represents the mathematical expectation of the distribution function.

[0035] In one possible implementation, the conditional information may take the form of category labels, attribute parameters, or multi-source modal data.

[0036] Category labels include: crack density label.

[0037] In one possible implementation, the generator takes a concatenated vector as input data and outputs pseudo-sample data; the concatenated vector consists of uniformly distributed random noise and conditional information.

[0038] The discriminator takes pseudo-sample data and real data as input and outputs a scalar prediction score which measures the probability of classifying pseudo-sample data as real data.

[0039] The optimal value of the objective function of the conditional generative adversarial network is solved by alternately executing the network optimization of the 1-step generator and the network optimization of the k-step discriminator; k is a preset positive integer greater than 1.

[0040] The generator network produces synthetic samples from random noise to simulate the distribution of real samples, while the discriminator network strives to distinguish real data from the fake data generated by the generator. During CGAN training, the generator and discriminator compete with each other like opponents in a zero-sum game: the generator aims to generate data sufficient to deceive the discriminator, while the discriminator strives to accurately identify fake data. This dynamic adversarial mechanism prompts both sides to continuously optimize their performance through competition until the generated data and real data are indistinguishable—a state known as Nash equilibrium in game theory. The globally optimal solution to this adversarial game is achieved by alternately performing one step of optimization for the generator network and k steps of optimization for the discriminator network.

[0041] Solving the objective function of CGAN involves the implicit Jensen-Shannon divergence and Kullback-Leibler divergence within the CGAN framework. The unique solution corresponds to , i.e., for any input, where represents the final data distribution of the generator. ,ie, .

[0042] The CGAN model is constructed to balance the distribution characteristics of well logging data based on the aforementioned basic network architecture. Both the generator and discriminator contain input, hidden, and output layers. The generator takes a concatenated vector consisting of uniformly distributed random noise and fracture density labels as input. This initial vector is first upsampled and reshaped through a fully connected layer, then processed by two transposed convolutional layers (Convtranspose1D), using different phase lengths for upsampling to generate hidden features. To improve model stability and ensure the integrity of feature representation during upsampling, each transposed convolutional layer employs both batch normalization (BN) and ReLU activation functions. The generator's output is generated by the last transposed convolutional layer, which uses hyperbolic tangent (Tanh) as the activation function. The discriminator receives generated and real data as input, concatenates them along the channel dimension during forward propagation, and then passes them sequentially through two convolutional layers (Conv1D) and downsamples them to generate feature maps. To promote efficient gradient flow, each convolutional layer uses a Leaky_ReLU activation function with a gradient value of 0.2. The final output is a scalar prediction score, which represents the probability of distinguishing the input data.

[0043] The CGAN model is used to generate pseudo-sample data to balance the skewed distribution of the original data. Although CGAN can approximately reflect the true distribution of the entire data, low-quality or outlier samples inevitably exist in the generated data, which can significantly affect the final classification performance. To avoid introducing noisy data that may interfere with classifier training, an Isolation Forest model is introduced to filter out potential outliers.

[0044] iForest is a tree-based ensemble learning algorithm for anomaly detection. The core idea is that anomalous samples are sparse and significantly different from normal samples, making them easier to isolate. Data isolation is achieved by constructing a set of isolation trees (iTrees). Each iTree is a binary tree that recursively partitions the data space using randomly selected features and splitting thresholds until every data point is isolated in a leaf node. The length of the isolation path from the root node to a leaf node reflects the degree of data isolation. Because anomalous data has "rare and unique" characteristics, it is usually located in leaf nodes closer to the root node than other normal data, meaning it can be isolated through shorter paths. This path-length-based isolation mechanism enables iForest to effectively detect anomalous data.

[0045] The iForest algorithm comprises two phases: training and evaluation. In one possible implementation, the training process of the isolated forest model includes: Constructing an isolation tree from the sample dataset Randomly select a subset of samples , and send it to the root node of the isolated tree.

[0046] Randomly select an attribute of a sample In this attribute Select the splitting threshold within the range of the maximum and minimum values. Based on the splitting threshold Sample subsets in sequence Each sample is divided into the left or right subtree until each node contains only a single sample point, or the binary tree reaches the preset height limit.

[0047] Once the iTree is fully constructed, anomaly scores are used for anomaly detection during the evaluation phase. Anomaly scores are defined as the normalized result of the data sample path length. In one possible implementation, the constructed isolated tree uses these anomaly scores to detect anomalies in pseudo-sample data, removing abnormal data from the pseudo-sample data.

[0048] The formula for calculating the outlier score is: .

[0049] in, Let represent any i-th input data, and n represent the number of samples.

[0050] This represents the path length, which is defined as the number of edges traversed from the root node to an external node in an isolated tree. This represents the average path length for all isolated trees.

[0051] represents the normalization factor, and represents the expected average path length when the sample size is n; ; For the harmonic operator.

[0052] In practical implementation, the harmonic operator Can be adopted An approximate solution is obtained using Euler's constant.

[0053] In the specific implementation process, the outlier score of each pseudo-sample data is evaluated. If the outlier score is close to 1, the pseudo-sample data is considered an outlier; otherwise, if the outlier score is close to 0, the pseudo-sample data is considered normal data.

[0054] In one possible implementation, the hybrid classifier includes a convolutional neural network, a bidirectional long short-term memory network, and a classification head.

[0055] Stratified fractures typically form under the continuous action of multiple geological processes, resulting in well logging data rich in spatiotemporal information. In fracture identification, this continuity requires comprehensive consideration of the spatiotemporal correlations between different geophysical parameters, rather than relying solely on well logging data from a single depth. Existing mainstream fracture identification methods often treat well logging data as independent samples, neglecting its continuity characteristics, thus limiting identification accuracy and model generalization ability. In reality, due to significant heterogeneity within reservoirs, the well logging response of stratified fractures typically exhibits complex and nonlinear characteristics. Given its obvious nonlinearity and spatiotemporal correlation, this invention proposes a hybrid classifier combining convolutional neural networks (CNNs) and bidirectional long short-term memory networks (BiLSTMs), which can more fully capture the local and continuous features of stratified fractures in well logging data.

[0056] Convolutional Neural Networks (CNNs) are deep neural networks with powerful local receptive capabilities. Their core idea is to extract local features from input data through convolution operations and then construct hierarchical representations using a multi-layered architecture. In this process, key sequence information of the input data is preserved while redundant features are effectively suppressed, thus significantly improving the ability to model local correlations. The core structure of a CNN consists of convolutional layers and pooling layers: convolutional layers transform the input vector into a feature vector through convolution operations; by sliding the convolutional kernel to cover different local regions of the input data, CNNs can accurately capture key features within the local receptive field. A convolution operation can be represented as:

[0057] In the formula, , These are the feature maps of the current layer and the previous layer, respectively. Let be the weights of the i-th convolutional kernel in the l-th network layer; For bias terms; It is an activation function used to introduce nonlinearity into the network and capture complex relationships.

[0058] Pooling layers, located after convolutional layers, control the dimensionality of extracted features. Through downsampling, the feature dimensionality is reduced while key information is preserved, thus lowering computational complexity and reducing the risk of overfitting. The most common pooling technique is max pooling, which selects the maximum value within a specified window. The mathematical expression for max pooling can be represented as:

[0059] In the formula, This is the output of the nth feature in the (l+1)th layer after max pooling; This is the input for the nth feature in the l-th layer; This is the j-th pooling region.

[0060] Bi-Layer Long Short-Term Memory (BiLSTM) is an enhanced recurrent neural network (RNN) that combines Long Short-Term Memory (LSTM) units with a bidirectional network for processing sequential data. By introducing memory units and a bidirectional training strategy, BiLSTM effectively alleviates the gradient problem in RNNs and can simultaneously extract sequential information from the input data both forward and backward. This enables BiLSTM to learn bidirectional dependencies in long sequence data, significantly improving the model's ability to extract rich feature information.

[0061] Bidirectional Long Short-Term Memory (BiLSTM) networks consist of two parallel sets of Long Short-Term Memory (LSTM) units: one sets process input features forward, and the other sets process them backward. The LSTM, the core component of the BiLSTM model, has a chain-like structure, composed of multiple circularly connected memory units. LSTM relies on gating mechanisms to control the discarding and retention of information, thereby capturing the long-term dependencies hidden in temporal data. The flow and updating of information are regulated by three key gate structures: the forget gate, the input gate, and the output gate. Specifically: the forget gate determines whether to retain input information in the current memory unit; the input gate determines whether to write new information to the memory unit; and the output gate controls whether the information in the memory unit can be used for the output at the current time step. BiLSTM can parse temporal information from both forward and backward directions; the model fuses the forward hidden state at each time step. Backward hidden state To obtain the final output .

[0062]

[0063]

[0064]

[0065] in, It is in a hidden state; , This is the weight matrix; It is time-series data; For bias terms; f and This is the activation function.

[0066] This invention constructs a hybrid framework by concatenating a Convolutional Neural Network (CNN) and a Bidirectional Long Short-Term Memory (BiLSTM) network to build a CNN-BiLSTM classifier for bedding fracture identification. In this hybrid classification model, the CNN, relying on a local receptive field and weight sharing mechanism, can automatically extract local logging features induced by bedding fractures; the BiLSTM, through bidirectional temporal operations, can effectively characterize long-term dependencies in logging data. This method can synergistically mine local features and long-term information, maintaining high recognition accuracy and robustness under complex geological conditions.

[0067] The CNN-BiLSTM model architecture consists of two stacked convolutional blocks, each containing a convolutional layer (Conv1D) and a max-pooling layer. The convolutional kernels slide across the logging data to extract local features, and the output is further compressed after processing by the pooling layer. Each convolutional layer is configured with batch normalization and ReLU activation functions to accelerate convergence and improve nonlinear adaptability, respectively. Subsequently, a BiLSTM network containing 64 LSTM units receives these outputs to further extract long-term dependencies; a fully connected (Dense Layer) fuses the information; a dropout layer (dropout rate = 0.3) is then used to prevent overfitting; finally, a Softmax layer calculates the probability of classifying the fractures and bedding planes as the target output.

[0068] Convolutional neural networks automatically extract local logging features caused by bedding fractures in logging data based on a local receptive field and weight sharing mechanism.

[0069] Bidirectional long short-term memory networks process the input sequence data bidirectionally, fusing local logging features and long-term dependencies between logging features to obtain a fused feature vector.

[0070] The input to the classification head is the fused feature vector. It establishes a mapping relationship between the fused feature vector and the bedding seam category, and outputs the classification probability of the bedding seam.

[0071] The hybrid classifier is trained on a mixed dataset containing both sample datasets and pseudo-sample datasets, and outputs the classification results of bedding fractures.

[0072] Example 2

[0073] Embodiment 2 of this invention provides a method for identifying layer seams based on an MS-BFI framework provided in this embodiment of the invention, combined with... Figure 1 It can be seen that embodiments of this method include: Step 1: Obtain raw logging data and label the raw logging data according to the core observation and imaging logging interpretation results. After preprocessing the labeled raw logging data, construct a sample dataset.

[0074] In one possible embodiment, the process of preprocessing the logging data in step 1 includes: Outlier removal; the original logging data is normalized to a uniform range using a min-max normalization function; tensor sample data is generated using the sliding window method according to the set window height and sliding step size.

[0075] Step 2: Input the sample dataset into the conditional generative adversarial network. The conditional generative adversarial network generates pseudo-sample data with the goal of balancing the distribution of samples with different labels.

[0076] Step 3: Remove outlier data from the spurious sample data based on the isolated forest model.

[0077] Step 4: Train the hybrid classifier using the hybrid dataset. Input the logging data to be tested into the trained hybrid classifier, and the hybrid classifier outputs the classification results of the bedding fractures.

[0078] In one possible embodiment, step 4 is followed by: Step 5: Test the hybrid classifier based on the test data and blind well data, and evaluate the performance of the hybrid classifier based on the confusion matrix.

[0079] It is understood that the method for identifying bedding fractures based on the MS-BFI framework provided by the present invention corresponds to the MS-BFI framework for identifying shale bedding fractures provided in the foregoing embodiments. The relevant technical features of the method for identifying bedding fractures based on the MS-BFI framework can be referred to the relevant technical features of the MS-BFI framework for identifying shale bedding fractures, and will not be repeated here.

[0080] Example 3

[0081] Embodiment 3 of this invention is a specific application embodiment of the MS-BFI framework for identifying shale bedding fractures and the method for identifying shale bedding fractures based on the MS-BFI framework provided by this invention. The dataset used in this specific application embodiment was collected from Silurian marine shale in the Luzhou area of ​​southern Sichuan Basin, China. As the core target layer of this invention embodiment, the Lower Silurian Longmaxi Formation (S1l) is currently the main stratum for shale gas exploration and development in the Luzhou area. This stratum was formed in a deep-sea to semi-deep-sea shelf environment and is mainly composed of organic-rich siliceous shale, carbonaceous shale, silty shale, and siltstone.

[0082] like Figure 2 The diagram shows the frequency distribution of linear density of bedding fractures provided in an embodiment of the present invention. This embodiment uses macroscopic fracture density as the basis for identifying bedding fractures in well logging. Linear density (P10, i.e., the number of fractures per unit length) is used to assess the degree of fracture development in the wellbore. Given the limitations of well logging resolution and the frequent changes in fracture development, to more accurately describe fracture distribution characteristics, this embodiment counts the number of bedding fractures in core samples and imaging logs in units of 10 cm depth, and calculates the fracture linear density (unit: dm³). -1 Based on the linear density distribution, the development degree of bedding fractures is divided into five categories: Category I (linear density = 0), Category II (linear density 1-3 dm²), and Category III (linear density 1-3 dm²). -1 Class III (linear density 4-8 dm) -1 Class IV (linear density 9-12 dm) -1 Class V (linear density ≥ 13 dm) and Class V (linear density ≥ 13 dm) -1 It should be noted that the embodiments of the present invention only focus on open bedding joints (i.e., unfilled and partially filled types), while filled cracks are considered as non-cracks.

[0083] Conventional logging data, including natural gamma (GR), caliper (CAL), sonic transit time (AC), density (DEN), compensated neutron (CNL), shallow resistivity (RS), and deep resistivity (RD) logging data, were collected from the five core and imaging logging wells mentioned above. After depth matching and outlier removal (using the Pauta criterion), a total of 2268 logging datasets with corresponding labels were finally obtained from the raw data. The number of samples in classes I to V were 1544, 315, 141, 175, and 93, respectively.

[0084] like Figure 3 The image shown is a cross-plot of conventional logging data for bedding fractures of different density categories provided in an embodiment of the present invention. Combined with... Figure 3It can be observed that the distribution of logging data for different types of bedding fractures varies, but some samples with different types still exhibit similar logging response characteristics. There is significant overlap in logging values ​​between different types of bedding fractures, and the category boundaries are blurred. This is attributed to the fact that logging response characteristics are not only related to the bedding fractures themselves, but also influenced by other geological factors such as mineral composition, pore fluids, and hydrocarbon content. For example, high levels of natural gas and clay minerals in shale reservoirs can lead to increased AC and GR values, resulting in logging responses similar to those of bedding fractures. These interfering factors make it difficult to distinguish logging responses originating from the fractures themselves. Therefore, logging identification of bedding fractures is a complex classification problem involving nonlinear characteristics.

[0085] The proposed MS-BFI framework is used to identify bedding fractures, integrating CGAN, iForest, and CNN-BiLSTM methods for data balancing, noise filtering, and fracture classification, respectively. Each module is designed to systematically address the various challenges in the fracture identification process. The identification process based on this framework includes the following key steps.

[0086] Step 1: Data Preprocessing. First, the raw logging data is labeled based on core observations and imaging logging interpretations. After removing outliers, the labeled dataset is normalized to [0,1] using a min-max function to eliminate the impact of differences in logging data magnitude on model training. To avoid overfitting, the logging data is divided into training and test sets based on depth (approximately 75% is selected as the training set, and the remaining approximately 25% as the test set). Then, a sliding window method is applied to the training and test sets to generate tensors with a window height of 5 and a sliding step size of 1.

[0087] Step 2: Data Balancing Based on CGAN. Training data is allocated to the CGAN, and data generation is used to balance the distribution of samples with different labels. During this process, crack labels are integrated into the CGAN to constrain the data generation process. Both the generator and discriminator are optimized using the Adam optimizer. After multiple training trials, the learning rates for the generator and discriminator are set to 0.0002 and 0.0001, respectively; the input noise dimension is set to 100, and the batch size is 64. To effectively guide the generator's optimization direction, each generator update is accompanied by five discriminator updates. The entire training process includes 50 iterations of pre-training and 5000 iterations of formal training.

[0088] Step 3: Noise Filtering Based on iForest. iForest is used as a noise filter to improve the quality of the CGAN-generated data and reduce the impact of noisy data on classifier performance. To balance filter performance and efficiency, the initial parameters of iForest are set as follows: n_estimators (number of iTrees) = 100, max_samples (sample size) = 256, contamination = 0.1, random_state = 30.

[0089] Step 4: Layer seam classification based on CNN-BiLSTM. This CNN-BiLSTM hybrid classifier is trained on a mixed dataset containing both the training dataset and filtered generated samples. During training, a grid search method is used to optimize the model hyperparameters; simultaneously, an early stopping strategy is employed to prevent overfitting—training is stopped if the validation loss does not improve within five consecutive training epochs.

[0090] Step 5: Model Validation. Apply the constructed recognition model to test data and blind well data to validate model performance.

[0091] Given the significant class imbalance between non-crack and crack data, data generation was limited to a few classes (i.e., bedding fracture data). To avoid diminishing the representativeness of the original dataset and the model's over-reliance on the generated samples, a total of 2285 synthetic datasets were ultimately generated based on CGAN. These generated datasets provided the data foundation for subsequent model development. Therefore, this embodiment of the invention quantifies the statistical fidelity of the generated data from multiple dimensions through comparative analysis.

[0092] In this embodiment of the invention, probability density curves obtained from kernel density estimation (KDE) are used to visualize the distribution of the original data and the generated data. For example... Figure 4 The diagram shown illustrates a comparison of the density distribution of actual data and generated data, along with the KS test results, provided in an embodiment of the present invention. Figure 4The results demonstrate that the generated data generally fits the distribution characteristics of the original data well, with only minor differences in local intervals. Particularly noteworthy is the complete preservation of complex distribution patterns in certain categories, indicating CGAN's strong ability to maintain the variability of the original data. Furthermore, the cumulative distribution functions of the original and generated data were compared using the Kolmogorov-Smirnov (KS) test: the KS statistics for all eight variables ranged from 0.053 to 0.166 (mean 0.096), indicating no statistically significant difference in the distributions of the two groups; simultaneously, the p-values ​​for all eight variables ranged from 0.158 to 0.739 and were all greater than 0.05, meaning the null hypothesis could not be rejected at the 95% confidence level. The KS test results strongly demonstrate that CGAN can generate samples that preserve the statistical characteristics of the original data. In addition, Spearman correlation analysis was used to quantitatively compare the consistency of the correlation strengths of the variables in the original and generated data. The correlation heatmap shows that the color distribution is centered on the diagonal. Figure 5 The diagram shown is a comparison of the Spearman correlation coefficient between the actual data and the generated data provided in this embodiment of the invention. Figure 5 The results show that the overall correlation patterns of the variables remain largely unchanged, with diagonal autocorrelation and associations among key features being maintained. Notably, some original correlations have weakened, which may be related to the introduction of new correlation factors into the generated data. Nevertheless, CGAN still captures most of the correlation structure in the original data. These results confirm that the CGAN model can effectively augment the training data sample, thereby alleviating the data imbalance problem.

[0093] After iForest noise filtering, a total of 2076 generated data sets were retained and combined with the original data to construct a hybrid dataset for training the CNN-BiLSTM classifier, with a class ratio of 1:0.69:0.50:0.64:0.43. Compared with the original data, the class imbalance problem in the hybrid data was significantly alleviated.

[0094] The ability of the CNN-BiLSTM model to capture the spatiotemporal features of well logging data largely depends on the optimal configuration of its hyperparameters. To achieve optimal performance in the bedding fracture classification task, this embodiment of the invention employs a grid search tool to systematically tune key hyperparameters (including learning rate, batch size, and weight decay). The learning rate determines the frequency of parameter updates in each iteration and directly affects the model's stability and convergence speed; the batch size determines the number of samples used for parameter updates, and an appropriate batch size helps avoid overfitting or underfitting; weight decay controls the decay rate of the model weights and has a significant impact on the model's complexity and generalization ability. The search range for hyperparameter tuning is set to the learning rate. Batch size Weight decay The model was trained under different parameter combinations, and the results were recorded. Finally, the optimal hyperparameter configuration was determined through accuracy evaluation.

[0095] like Figure 6 The diagram shown illustrates the results of hyperparameter optimization of CNN-BiLSTM using the Adam optimizer via grid search, as provided in an embodiment of the present invention. Figure 6 This demonstrates the grid search process used to find the globally optimal hyperparameters of a CNN-BiLSTM model; the optimal hyperparameters are marked with black boxes. Combined with... Figure 6 It can be seen that the model performance fluctuates continuously with changes in the learning rate, batch size, and weight decay value, and these three parameters are 0.001, 32, and 10, respectively. -4 When the optimal state is reached, the optimal hyperparameter combination of the CNN-BiLSTM model is determined as follows: learning rate = 0.001, batch size = 32, weight decay = 10. -4 .

[0096] The trained model is then applied to test data to verify the reliability of the training results. For example... Figure 7 The figure shown is a schematic diagram of the training curve and test curve of the CNN-BiLSTM model provided in an embodiment of the present invention. Figure 7 This demonstrates the classification accuracy trends of the CNN-BiLSTM model on the training and test sets. It can be seen that the trend of test accuracy is completely consistent with the training data: the training accuracy rapidly increases from 75% to 95% in the initial stage, then stabilizes after about 50 training epochs, eventually settling at approximately 96%; the test accuracy shows an overall upward trend with slight fluctuations over 200 training epochs, and later closely matches the training accuracy. Figure 8 A schematic diagram of the confusion matrix of the training set, test set, and blind test set provided in an embodiment of the present invention; Figure 8 Medium green represents the number of correctly identified samples and its proportion of the total sample; the bottom row (light gray) represents precision; the rightmost column represents recall; and dark gray represents overall precision. Figure 8 The confusion matrix analysis shows that the CNN-BiLSTM classifier consistently maintains high training accuracy across all five categories. More importantly, the confusion matrix of the test data demonstrates highly stable classification performance, with accuracy only slightly decreasing compared to the training results. These results indicate that the optimized CNN-BiLSTM model performs excellently on both the training and test sets, exhibiting high accuracy and good generalization ability in the layered seam classification task.

[0097] To further verify the effectiveness of the constructed MS-BFI model in identifying bedding fractures, researchers selected another representative shale gas well (Well B1) in the Luzhou area for a blind logging experiment. The logging data from this well was not used for modeling analysis. The test section of Well B1 covers the Longmaxi Formation (depth range 4040.5-4126.7 meters), and a total of 862 datasets were obtained from the logging data. Figure 9 This is a schematic diagram illustrating the identification results of bedding fractures in blind well B1 provided in an embodiment of the present invention, combined with... Figure 9 It can be seen from the annotation based on the imaging logging results that there is a significant distribution imbalance among different fracture density categories: the sample sizes for categories I to V are 552, 112, 73, 82 and 43, respectively.

[0098] The blind logging results from well B1 show a high degree of consistency between the MS-BFI model's identification results and the imaging logging interpretation results. Although there was a slight performance degradation, the overall accuracy of the blind logging (94.66%) was largely consistent with the training data (96.10%) and the test data (94.59%). The blind logging confusion matrix indicates that the model accurately identifies most categories (Class I) and also demonstrates excellent performance for a few categories (Classes II, III, IV, and V). The accuracy, recall, and precision for each category in the blind logging all exceeded 88%, maintaining a high degree of agreement with the modeling data. Furthermore, misclassification mainly occurred in the non-fractured sections (Class I) of well B1: for example, intervals ① (depth 4058.10-4058.50 meters) and ② (depth 4119.20-4119.50 meters) were incorrectly classified as fractured sections. Figure 9 These intervals likely correspond to low-angle structural fractures—whose logging response signals are too similar to bedding fractures to be easily distinguished. Nevertheless, the blind test results fully demonstrate the high effectiveness and reliability of the MS-BFI model in identifying bedding fractures under unbalanced data and complex geological conditions.

[0099] This invention provides an MS-BFI framework and method for identifying shale bedding fractures. Addressing the uneven distribution and nonlinear characteristics of well logging data, a novel MS-BFI fusion framework is proposed. This MS-BFI framework integrates Conditional Generative Adversarial Network (CGAN), Isolation Forest (iForest), and a hybrid classifier. CGAN generates samples to expand minority samples and improve the quality of generated data, thereby solving the problems of uneven distribution and insufficient diversity. A noise filtering scheme is designed based on the iForest algorithm to improve the reliability of generated samples. A hybrid classifier is constructed by combining CNN and BiLSTM to capture the sequential features of well logging data and improve the model's accuracy and generalization ability. The collaborative work of each unit enables the MS-BFI model to accurately predict bedding fractures in multiple steps from imbalanced well logging data. Experiments on real datasets and comparisons with various classic data augmentation techniques demonstrate the superior performance of the proposed method. Application to well logging datasets of Silurian marine shale in the southern Sichuan Basin demonstrates that the proposed MS-BFI model achieves identification accuracies of 94.59% and 94.66% on test data and blind well data, respectively, validating the effectiveness of this method in identifying bedding fractures. This method provides a robust and effective solution for identifying bedding fractures under conditions of data imbalance and complex geological conditions, and holds promise for practical application in the petroleum industry.

[0100] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0101] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0102] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations 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, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0103] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0104] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0105] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0106] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. An MS-BFI framework for identifying shale bedding fractures, characterized in that, The MS-BFI framework includes: Conditional Generative Adversarial Networks, Isolated Forest Model, and Hybrid Classifiers; The conditional generative adversarial network includes a generator and a discriminator. The generator generates pseudo-sample data that simulates the balanced distribution of various well logging data based on conditional information and random noise. The discriminator is used to determine the true probability of the pseudo-sample data. The conditional information is prior constraint information. The isolated forest model is used to remove abnormal data from the pseudo-sample data; The hybrid classifier is trained on a mixed dataset containing both well logging data and pseudo-sample data, and outputs the classification result of the bedding fracture.

2. The MS-BFI framework according to claim 1, characterized in that, The objective function of the conditional generative adversarial network is: in, This represents the discriminator network. Represents a generator network; Represents the value function; The distribution of the real data is represented by x; x represents a distribution that follows a certain distribution. Real data; This represents the condition information embedded in the input terminals of the generator and discriminator; z represents a random noise distribution; z represents a distribution that follows a random noise distribution. Random noise; and Representing condition information The outputs of the generator and discriminator described below; It represents the mathematical expectation of the distribution function.

3. The MS-BFI framework according to claim 1, characterized in that, The conditional information may take the form of: category labels, attribute parameters, or multi-source modal data; The category labels include: crack density labels.

4. The MS-BFI framework according to claim 1, characterized in that, The generator takes the concatenated vector as input data and outputs the pseudo-sample data; the concatenated vector is composed of uniformly distributed random noise and the conditional information. The discriminator takes the pseudo-sample data and real data as input and outputs a scalar prediction score of the probability of classifying the pseudo-sample data as real data. The optimal value of the objective function of the conditional generative adversarial network is solved by alternately executing the network optimization of the generator in step 1 and the network optimization of the discriminator in step k; k is a preset positive integer greater than 1.

5. The MS-BFI framework according to claim 1, characterized in that, The training process of the isolated forest model includes: Constructing an isolation tree from the sample dataset Randomly select a subset of samples , and send to the root node of the isolated tree; Randomly select an attribute of a sample In this attribute Select the splitting threshold within the range of the maximum and minimum values. According to the splitting threshold The sample subsets are sequentially Each sample is divided into the left or right subtree until each node contains only a single sample point, or the binary tree reaches the preset height limit.

6. The MS-BFI framework according to claim 5, characterized in that, The constructed isolated tree uses anomaly scores to detect anomalies in the pseudo-sample data and removes abnormal data from the pseudo-sample data. The formula for calculating the abnormal score is: ; in, Let represent any i-th input data, and n represent the number of samples; This represents the path length, which is defined as the number of edges traversed from the root node to an external node in the isolated tree. This represents the average path length for all isolated trees. Indicates the normalization factor; ; For the harmonic operator.

7. The MS-BFI framework according to claim 1, characterized in that, The hybrid classifier includes: a convolutional neural network, a bidirectional long short-term memory network, and a classification head; The convolutional neural network automatically extracts local logging features induced by bedding fractures in the logging data based on a local receptive field and weight sharing mechanism. The bidirectional long short-term memory network processes the input sequence data bidirectionally, and fuses local logging features and long-term dependencies between logging features to obtain a fused feature vector. The input to the classification head is the fused feature vector, which establishes a mapping relationship between the fused feature vector and the bedding seam category, and outputs the probability of the bedding seam classification.

8. A method for identifying layered seams based on the MS-BFI framework described in claims 1-7, characterized in that, The method includes: Step 1: Obtain raw logging data and label the raw logging data according to the core observation and imaging logging interpretation results. After preprocessing the labeled raw logging data, a sample dataset is constructed. Step 2: Input the sample dataset into the conditional generative adversarial network, which generates each pseudo-sample data with the goal of balancing the distribution of samples with different labels. Step 3: Remove outlier data from the pseudo-sample data based on the isolated forest model; Step 4: Train the hybrid classifier using the hybrid dataset, input the logging data to be tested into the trained hybrid classifier, and the hybrid classifier outputs the classification result of the bedding fracture.

9. The method according to claim 7, characterized in that, The preprocessing of the well logging data in step 1 includes: Outlier removal; the original logging data are normalized to a uniform range using a min-max normalization function; tensor sample data are generated using a sliding window method according to a set window height and sliding step size.

10. The method according to claim 7, characterized in that, Step 4 is followed by: Step 5: Test the hybrid classifier based on the test data and blind well data, and evaluate the performance of the hybrid classifier by combining the confusion matrix.

Images