A smart identification method for distinguishing shale crystals

Abstract

This invention discloses an intelligent identification method for shale and mudstone reservoirs, relating to the field of geophysical logging technology. Based on the mineral composition and structural characteristics of shale and mudstone reservoirs, this invention manually calibrates shale on logging data of the target well section. The logging data is preprocessed, and Pearson correlation analysis is performed on various logging parameters. Redundant curves are removed, and a sample library is constructed. Samples are extracted from the sample library to construct training and validation sets. The smote method is used to expand the shale samples in the training set. A CNN-BiLSTM-Attention combined model based on the attention mechanism is then constructed to build a shale discriminator. The discriminator is trained using the training set for shale identification, and then the validation set is used to verify the discriminator's recognition effect. Finally, the validated discriminator is used to identify shale in the well section to be identified. This invention solves the problem of accurate shale identification in complex shale and mudstone reservoirs, improves the rate of shale and mudstone shale identification, and is beneficial for the exploration and development of complex shale and mudstone reservoirs.

Description

Technical Field

[0001] This invention relates to the field of geophysical logging technology, specifically to an intelligent identification method for distinguishing bright crystals in shale and mudstone. Background Technology

[0002] Bright crystal identification, as an important foundational task in oil exploration and development, provides valuable evidence for targeted layer selection in shale oil exploration from a completely new geological perspective.

[0003] Currently, commonly used methods for identifying bright crystals include core description, charting, and statistical analysis. However, due to the complex geological conditions, shale reservoirs are characterized by strong heterogeneity, rapid sequence changes, and high complexity, resulting in weak bright crystal structure mapping relationships and small sample sizes in their logging responses, posing a significant challenge to bright crystal identification. While core description is precise, it is time-consuming and labor-intensive, and yields limited sample sizes. Conventional charting, while simple and intuitive, is hampered by the indistinct logging responses in shale reservoirs, making accurate bright crystal differentiation difficult. Statistical analysis, while capable of identifying some linearly separable bright crystals to a certain extent, still has limitations in its application to complex reservoirs.

[0004] Therefore, existing conventional methods are insufficient for accurate identification of spar crystal structures. Thus, there is an urgent need to develop an intelligent identification method for spar crystals in shale and mudstone reservoirs, addressing challenges such as small sample size, difficulty in extracting nonlinear features, and rapid sequence changes, to achieve accurate identification of spar crystals in complex shale and mudstone reservoir environments. Summary of the Invention

[0005] To address the challenge of accurately identifying bright crystals in complex shale reservoirs, this invention proposes an intelligent method for identifying bright crystals in shale. This method overcomes the limitations of training due to the small sample size and imbalanced sample categories. By combining the Smote data expansion method with a CNN-BiLSTM-Attention combined model, the method improves the speed and accuracy of bright crystal identification in shale, thus benefiting the exploration and development of complex shale reservoirs.

[0006] The present invention adopts the following technical solution:

[0007] A smart identification method for distinguishing spar crystals in shale and mudstone specifically includes the following steps:

[0008] Step 1: Obtain the logging curves of the target well section in the shale reservoir, obtain the logging data of the target well section, combine with the core data, analyze the mineral composition and structural characteristics in the shale reservoir, manually determine the bright crystalline and non-bright crystalline sections on the logging curves, and manually mark the labels on the logging curves.

[0009] Step 2: After preprocessing the logging data of the target well section, perform Pearson correlation analysis on each logging parameter. Based on the Pearson correlation coefficient of each logging parameter, filter out the redundant curves in the logging curves of the target well section. Then, combine the artificially labeled bright crystal and non-bright crystal labels on the logging curves of the target well section to construct a sample library including bright crystal sample set and non-bright crystal sample set.

[0010] Step 3: Randomly select bright crystal samples and non-bright crystal samples from the sample library as training samples and validation samples to form training set and validation set. Expand the bright crystal samples in the training set based on the smote method so that the number of bright crystal samples in the training set is equal to the number of non-bright crystal samples.

[0011] Step 4: Construct a bright crystal discriminator using a CNN-BiLSTM-Attention combined model based on the attention mechanism;

[0012] Step 5: Use the training set to train the bright crystal discriminator constructed by the CNN-BiLSTM-Attention combined model based on the attention mechanism, and obtain the trained bright crystal discriminator.

[0013] Step 6: Validate the bright crystal discriminator using the validation set to obtain the validated bright crystal discriminator;

[0014] Step 7: Use the verified bright crystal discriminator to perform bright crystal identification on the well section to be identified. Input the logging data extracted from the logging curve of the well section to be identified into the bright crystal discriminator according to the depth sequence to obtain the bright crystal identification result of the well section to be identified.

[0015] Preferably, the logging curves include natural gamma curves, spontaneous potential curves, sonic transit time curves, density curves, compensated neutron curves, and resistivity curves.

[0016] Preferably, step 2 specifically includes the following steps:

[0017] Step 2.1: Preprocess the logging data of the target well section, including depth correction and noise reduction.

[0018] Step 2.2: Perform Pearson correlation analysis on each logging parameter. The formula for calculating the Pearson correlation coefficient of the logging parameters is as follows:

[0019]

[0020] In the formula, ρ X,Y Here, X and Y are the Pearson correlation coefficients, where i is the sequence number and n is the total number of logging parameters. i Let X be the measured value of the i-th logging parameter. Let X be the average value of the logging parameter, and Y be the average value of the logging parameter.i Let Y be the measured value of the i-th logging parameter. This represents the average value of the logging parameter Y;

[0021] The Pearson correlation coefficient ρ X,Y A value less than 0 indicates a negative correlation between logging parameter X and logging parameter Y; the Pearson correlation coefficient ρ X,Y A value greater than 0 indicates a positive correlation between logging parameter X and logging parameter Y, and the Pearson correlation coefficient ρ is greater than 0. X,Y The closer the correlation is to 0, the worse the correlation between logging parameter X and logging parameter Y becomes;

[0022] Step 2.3, based on the Pearson correlation coefficient ρ X,Y Screen the logging curves of the target well section, and |ρ X,Y The logging curves at |≥0.9 are considered redundant curves. After removing redundant curves from the logging curves of the target well section, the logging data on the filtered target well section are used, combined with the bright crystal and non-bright crystal labels on the logging curves. The logging data at the same depth point in the bright crystal section are taken as a bright crystal sample, and multiple bright crystal samples are obtained to construct a bright crystal sample set. Then, the logging data at the same depth point in the non-bright crystal section are taken as a non-bright crystal sample, and multiple non-bright crystal samples are obtained to construct a non-bright crystal sample set. The resulting sample library includes the bright crystal sample set and the non-bright crystal sample set.

[0023] Preferably, the ratio of training samples in the training set to validation samples in the validation set is 9:1.

[0024] Preferably, the bright crystal discriminator is constructed based on a CNN-BiLSTM-Attention combined model with an attention mechanism. The CNN-BiLSTM-Attention combined model with an attention mechanism includes a convolutional neural network model and a BiLSTM network model with an attention mechanism, which are connected in sequence.

[0025] The attention-based convolutional neural network model includes a CNN branch, an attention mechanism branch, and a multiplication layer arranged in parallel.

[0026] The CNN branch is used to extract the feature vector of the input sample, including a first convolutional layer, a first activation layer, a second convolutional layer, and a second activation layer. Both the first and second convolutional layers are 3×1 convolutional layers. The first convolutional layer has 32 convolutional kernels, and the second convolutional layer has 64 convolutional kernels. The convolution operation of the first and second convolutional layers is as shown in formula (2):

[0027]

[0028] In the formula, h i x is the feature vector obtained from the convolution operation. i Let w be the input value of the i-th element in the convolutional layer. i Let b be the weight matrix value corresponding to the i-th element. i Let be the bias value of the i-th element, and N×N be the size of the convolution kernel;

[0029] Both the first and second activation layers use the ReLU function as the activation function, as shown in formula (3):

[0030] y i =max(0,h) i (3)

[0031] In the formula, y i The activation function value is max(0,h). i ) is the maximum value function, used to select 0 and the feature vector h obtained from the convolution operation. i The maximum value in;

[0032] The attention mechanism branch is used to assist the convolutional neural network in extracting feature vectors. It includes a global average pooling layer, a first fully connected layer, a first activation layer, a second fully connected layer, and a second activation layer. The number of channels in the first fully connected layer is set to 16, and the number of channels in the second fully connected layer is set to 64. The first activation layer uses the ReLU function as the activation function, and the second activation layer uses the sigmoid function as the activation function.

[0033] The attention mechanism branch uses the SE module of the soft attention mechanism to assist the convolutional neural network in extracting feature vectors. Weights representing importance are added to focus on specific channels, including the Squeeze part, the Excitation part, and the Reweight part.

[0034] The global average pooling layer of the attention mechanism branch corresponds to the Squeeze part of the SE module. It performs a global pooling operation on the feature vectors in the input samples and performs average pooling on the feature vectors in each channel to obtain the global information vector, as shown in formula (4):

[0035]

[0036] In the formula, z c The global information vector is the output of the global average pooling layer, where W is the maximum width of the input feature vector, H is the maximum height of the input feature vector, C is the number of channels in the global average pooling layer, and u c (I,J) represents the input feature vector value with width I and height J;

[0037] The first fully connected layer, the first activation layer, the second fully connected layer, and the second activation layer of the attention mechanism branch correspond to the Excitation part of the SE module. The global information vector obtained by the global average pooling layer passes through the first fully connected layer, the first activation layer, the second fully connected layer, and the second activation layer in sequence to generate a channel weight vector to represent the channel importance, as shown in formula (5):

[0038] s c =σ(W2δ(W1z) c (5)

[0039] In the formula, s c σ(·) is the channel weight vector, δ(·) is the sigmoid activation function value, δ(·) is the ReLU activation function value, W1 is the fully connected layer parameter used for compression, and W2 is the fully connected layer parameter used for dimension restoration.

[0040] The second activation layer of the CNN branch and the second activation layer of the attention mechanism branch are both connected to the multiplication layer. The multiplication layer corresponds to the Reweight part of the SE module and is used to fuse the feature vector extracted by the CNN branch and the channel weight vector generated by the attention mechanism branch. By multiplying the channel weight vector with the extracted feature vector, the fused data after feature enhancement is obtained, as shown in formula (6):

[0041]

[0042] In the formula, For the fused data after feature enhancement, F scale (·) represents the reweight function, u c The input feature vector values;

[0043] The BiLSTM network model uses the fused data output by the attention-based convolutional neural network model as input data to capture and propagate information from adjacent depths in the input data. The BiLSTM network model includes a forward LSTM network for forward learning and a backward LSTM network for backward learning. Both the forward and backward LSTM networks include the input at time t, cell state, temporary cell state, hidden layer state, forget gate, memory gate, and output gate. After the input data enters the BiLSTM network model, it sequentially enters the forward and backward LSTM networks. The forward LSTM network processes the input data in a forward order, while the backward LSTM network processes it in a reverse order. The forward and backward LSTM networks respectively use the input data and the hidden state from the previous time step to calculate the hidden state at the current time step. The hidden states of the forward and backward LSTM networks are then connected or merged to obtain the output of the BiLSTM network model.

[0044] The forward propagation of the forward LSTM network is shown in Equation (7):

[0045]

[0046] In the formula, x t The input at time t, For the forget gate that propagates forward at time t, Forget gates that propagate forward at time t-1, LSTM(·) is a bidirectional long short-term memory neural network model;

[0047] The backpropagation of the backpropagation LSTM network is shown in Equation (8):

[0048]

[0049] In the formula, For the forget gate that propagates backward at time t, Forget gates that propagate backwards at time t+1;

[0050] The output of the BiLSTM network model is:

[0051]

[0052] In the formula, h t Let t represent the state of the hidden layer of the BiLSTM network model at time t.

[0053] Preferably, the bright crystal discriminator constructed by the CNN-BiLSTM-Attention combined model based on the attention mechanism is trained using the Adam gradient descent algorithm. During the training process, the bright crystal discriminator adaptively adjusts the learning rate and updates the network weights. The cross-entropy loss function is used as the loss function during the training process to evaluate the accuracy of the bright crystal discriminator in recognizing bright crystals, as shown in formula (10):

[0054]

[0055] In the formula, L(·) is the objective function to be minimized, m is the number of samples; y is the true class, which takes the value of 0 or 1; h(x) is the probability of the bright crystal discriminator predicting sample x, and θ is the model parameter of the bright crystal discriminator.

[0056] Preferably, step 5 specifically includes the following steps:

[0057] Step 5.1: Set the maximum number of training iterations and the initial learning rate for the bright crystal discriminator;

[0058] Step 5.2: Randomly select training samples from the training set and input them into the bright crystal discriminator. Use the bright crystal discriminator to identify whether the training samples are bright crystals. The bright crystal discriminator automatically labels the training samples that are identified as bright crystals with a bright crystal label and labels the training samples that are not identified as non-bright crystals with a non-bright crystal label.

[0059] Step 5.3: Compare the labels assigned by the bright crystal discriminator with the manually assigned labels in the training samples. If they match, proceed to step 5.4. Otherwise, adaptively adjust the learning rate of the bright crystal discriminator and adjust the network weights to update the bright crystal discriminator, and return to step 5.2 to continue training the bright crystal discriminator.

[0060] Step 5.4: Determine whether the current training coefficients have reached the preset maximum number of training iterations. If the current number of training iterations has reached the maximum number of training iterations, stop training the bright crystal discriminator and obtain the trained bright crystal discriminator. Otherwise, return to step 5.2 to continue training the bright crystal discriminator.

[0061] Preferably, the maximum number of training iterations is not less than 2000, the initial learning rate is set to 0.001, and the loss function value of the bright crystal discriminator gradually stabilizes at 0.2 during the training process.

[0062] Preferably, step 6 specifically includes the following steps:

[0063] Step 6.1: Input the verification samples in the verification set into the bright crystal discriminator in sequence. Use the bright crystal discriminator to identify whether the verification samples are bright crystals and automatically mark the verification samples that are identified as bright crystals with bright crystal labels and mark the verification samples that are not identified as non-bright crystals with non-bright crystal labels. Compare with the manually labeled labels in the verification samples to determine whether the bright crystal recognition result of the bright crystal discriminator is accurate.

[0064] Step 6.2: After the bright crystal discriminator has completed the bright crystal recognition of all verification samples, obtain the accuracy, precision, and recall of the bright crystal discriminator during the bright crystal recognition process, and compare them with the preset accuracy, precision, and recall. If the accuracy, precision, and recall of the bright crystal discriminator have reached the preset accuracy, precision, and recall, then proceed to step 7; otherwise, return to step 5 to continue training the bright crystal discriminator.

[0065] The present invention has the following beneficial effects:

[0066] This invention proposes an intelligent identification method for shale and mudstone reservoirs. By performing Pearson correlation analysis on various logging parameters to remove redundant curves and extract sample data, and expanding the shale and mudstone sample set in the sample library with the smote method, a shale and mudstone discriminator is constructed based on a CNN-BiLSTM-Attention combined model with an attention mechanism. After training the shale and mudstone discriminator and verifying its accuracy in identifying shale and mudstone, the shale and mudstone discriminator is used to identify shale and mudstone in the well section to be identified, thus achieving accurate identification of shale and mudstone in complex shale and mudstone reservoirs.

[0067] This invention employs the smote method to expand the bright crystal sample data, effectively solving the problems of small quantity and class imbalance of bright crystal samples obtained in the field of shale and mudstone. This provides data support for the training of the bright crystal discriminator. Furthermore, this invention constructs a bright crystal discriminator based on a CNN-BiLSTM-Attention combined model with an attention mechanism, and utilizes a parallel model of a convolutional neural network based on an attention mechanism for feature extraction and recognition, effectively solving the problem of nonlinear feature extraction. Combined with the BiLSTM network model, this fully exploits the depth sequence information of well logging curves to capture the correlation of sequence layers, improving both the speed of shale and mudstone bright crystal identification and achieving accurate identification of shale and mudstone bright crystals. This lays the foundation for guiding the exploration and development of complex shale and mudstone reservoirs. Attached Figure Description

[0068] Figure 1 This is a flowchart of an intelligent identification method for distinguishing bright crystals in mudstone and shale according to the present invention.

[0069] Figure 2 The graph shows the results of the Pearson correlation analysis.

[0070] Figure 3This is a schematic diagram of expanding bright crystal samples based on the smote method; in the figure, Figure (a) is the bright crystal sample set before expansion, and Figure (b) is the bright crystal sample set after expansion using the smote method.

[0071] Figure 4 This is a schematic diagram of the structure of a convolutional neural network model based on the attention mechanism.

[0072] Figure 5 This is a diagram showing the effect of identifying bright crystals in mudstone and shale in well X in the embodiment.

[0073] Figure 6 This is a diagram showing the effect of identifying shale bright crystals in well Y in the embodiment. Detailed Implementation

[0074] The following section uses a specific area containing a shale reservoir as the study area, and further explains the specific embodiments of the present invention in conjunction with the accompanying drawings:

[0075] This invention proposes an intelligent identification method for distinguishing shale crystals, such as... Figure 1 As shown, the specific steps include:

[0076] Step 1: Obtain logging curves for the target well section in the shale reservoir. Obtain logging data for the target well section, including natural gamma curve, spontaneous potential curve, sonic transit time curve, density curve, compensated neutron curve, and resistivity curve. Combined with core data, by observing core samples and core thin section data, and referring to geological data of the study area, analyze the mineral composition and structural characteristics of the shale reservoir. Manually determine the spar crystal and non-spar crystal sections on the logging curves, and manually label them on the logging curves.

[0077] Step 2: After preprocessing the logging data of the target well section, perform Pearson correlation analysis on each logging parameter. Based on the Pearson correlation coefficient of each logging parameter, filter out redundant curves in the logging curves of the target well section. Then, combine the manually labeled bright crystal and non-bright crystal labels on the logging curves of the target well section to construct a sample library including bright crystal sample sets and non-bright crystal sample sets. The specific steps are as follows:

[0078] Step 2.1: Perform preprocessing on the logging data of the target well section, such as depth correction and noise reduction, to obtain preprocessed logging data.

[0079] Step 2.2: Perform Pearson correlation analysis on each logging parameter, such as... Figure 2 As shown, the Pearson correlation coefficients between various logging parameters are obtained, and the Pearson correlation coefficients are used to evaluate the degree of correlation between the logging parameters.

[0080] In this embodiment, the formula for calculating the Pearson correlation coefficient of the logging parameters is as follows:

[0081]

[0082] In the formula, ρ X,Y Here, X and Y are the Pearson correlation coefficients, where i is the sequence number and n is the total number of logging parameters. i Let X be the measured value of the i-th logging parameter. Let X be the average value of the logging parameter, and Y be the average value of the logging parameter. i Let Y be the measured value of the i-th logging parameter. This represents the average value of the logging parameter Y.

[0083] The Pearson correlation coefficient ρ X,Y A value less than 0 indicates a negative correlation between logging parameter X and logging parameter Y; the Pearson correlation coefficient ρ X,Y A value greater than 0 indicates a positive correlation between logging parameter X and logging parameter Y, and the Pearson correlation coefficient ρ is greater than 0. X,Y The closer the correlation is to 0, the worse the correlation between logging parameter X and logging parameter Y becomes.

[0084] Step 2.3, based on the Pearson correlation coefficient ρ X,Y Screen the logging curves of the target well section, and |ρ X,Y The logging curves at |≥0.9 are considered redundant curves. After removing redundant curves from the logging curves of the target well section, the logging data on the filtered target well section are used, combined with the bright crystal and non-bright crystal labels on the logging curves. The logging data at the same depth point in the bright crystal section are taken as a bright crystal sample, and multiple bright crystal samples are obtained to construct a bright crystal sample set. Then, the logging data at the same depth point in the non-bright crystal section are taken as a non-bright crystal sample, and multiple non-bright crystal samples are obtained to construct a non-bright crystal sample set. The resulting sample library includes the bright crystal sample set and the non-bright crystal sample set.

[0085] Step 3: Randomly select bright crystal samples and non-bright crystal samples from the sample library as training samples and validation samples to form training set and validation set. The ratio of training samples in training set to validation samples in validation set is 9:1. Then, expand the bright crystal samples in the training set based on the smote method so that the number of bright crystal samples in the training set is equal to the number of non-bright crystal samples, thus balancing the bright crystal samples and non-bright crystal samples in the training set.

[0086] In this embodiment, the process of expanding the bright crystal sample based on the smote method is as follows: Figure 3 As shown, Figure 3 The squares represent existing bright crystal samples, and the dots represent bright crystal samples expanded using the smote method.

[0087] Step 4: Construct a bright crystal discriminator using a CNN-BiLSTM-Attention combined model based on the attention mechanism.

[0088] In this embodiment, the bright crystal discriminator is constructed based on the CNN-BiLSTM-Attention combined model with attention mechanism. The CNN-BiLSTM-Attention combined model with attention mechanism includes a convolutional neural network (CNN) model and a BiLSTM network model with attention mechanism, which are connected in sequence.

[0089] The attention-based convolutional neural network model described above possesses excellent feature extraction and recognition capabilities, including parallel CNN branches, attention mechanism branches, and multiplication layers, such as... Figure 4 As shown.

[0090] The CNN branch is used to extract the feature vector of the input sample, including a first convolutional layer, a first activation layer, a second convolutional layer, and a second activation layer. Both the first and second convolutional layers are 3×1 convolutional layers. The first convolutional layer has 32 convolutional kernels, and the second convolutional layer has 64 convolutional kernels. The convolution operation is performed on the first and second convolutional layers. Since both bright and non-bright samples are one-dimensional data, one-dimensional convolutional kernels are selected to perform convolution operations on the input samples, as shown in formula (2):

[0091]

[0092] In the formula, h i x is the feature vector obtained from the convolution operation. i Let w be the input value of the i-th element in the convolutional layer. i Let b be the weight matrix value corresponding to the i-th element. i Let be the bias value of the i-th element, and N×N be the size of the convolution kernel;

[0093] To better address the problem of nonlinear feature extraction from input samples, both the first and second activation layers use the ReLU function as the activation function, as shown in formula (3):

[0094] y i =max(0,h) i (3)

[0095] In the formula, y i The activation function value is max(0,h). i ) is the maximum value function, used to select 0 and the feature vector h obtained from the convolution operation. iThe maximum value in.

[0096] To improve the attention of convolutional neural networks to important features in input samples, an attention mechanism branch is set up to extract salient features of input samples in parallel. As a concrete method of human prior in machine learning components, the attention mechanism can focus attention on important features when the network processes input samples, effectively avoiding the shortcomings of traditional neural networks that cannot distinguish the importance of signals when processing information.

[0097] The attention mechanism branch is used to assist the convolutional neural network in extracting feature vectors. It includes a global average pooling layer, a first fully connected layer, a first activation layer, a second fully connected layer, and a second activation layer. The number of channels in the first fully connected layer is set to 16, and the number of channels in the second fully connected layer is set to 64. The first activation layer uses the ReLU function as the activation function, and the second activation layer uses the sigmoid function as the activation function.

[0098] The attention mechanism branch uses a soft attention mechanism SE module to assist the convolutional neural network in extracting feature vectors. The soft attention mechanism improves the accuracy of feature extraction by adding weights that represent the importance of specific channels. Specifically, it includes a Squeeze part, an Excitation part, and a Reweight part.

[0099] The global average pooling layer of the attention mechanism branch corresponds to the Squeeze part of the SE module. It performs a global pooling operation on the feature vectors in the input samples and performs average pooling on the feature vectors in each channel to obtain the global information vector, as shown in formula (4):

[0100]

[0101] In the formula, z c The global information vector is the output of the global average pooling layer, where W is the maximum width of the input feature vector, H is the maximum height of the input feature vector, C is the number of channels in the global average pooling layer, and u c (I,J) represents the input feature vector values ​​with width I and height J.

[0102] The first fully connected layer, the first activation layer, the second fully connected layer, and the second activation layer of the attention mechanism branch correspond to the Excitation part of the SE module. The global information vector obtained by the global average pooling layer passes through the first fully connected layer, the first activation layer, the second fully connected layer, and the second activation layer in sequence to generate a channel weight vector to represent the channel importance, as shown in formula (5):

[0103] s c =σ(W2δ(W1z) c (5)

[0104] In the formula, s c σ is the channel weight vector, σ(·) is the sigmoid activation function value, δ(·) is the ReLU activation function value, W1 is the fully connected layer parameter used for compression, and W2 is the fully connected layer parameter used for dimension restoration.

[0105] The second activation layer of the CNN branch and the second activation layer of the attention mechanism branch are both connected to the multiplication layer. The multiplication layer corresponds to the Reweight part of the SE module and is used to fuse the feature vector extracted by the CNN branch and the channel weight vector generated by the attention mechanism branch. By multiplying the channel weight vector with the extracted feature vector, the fused data after feature enhancement is obtained, as shown in formula (6):

[0106]

[0107] In the formula, For the fused data after feature enhancement, F scale (·) represents the reweight function, u c The input feature vector values.

[0108] Considering that the attention-based convolutional neural network model cannot fully utilize the potential information of sedimentary rocks in the depth domain sequence in bright crystal discrimination, the feature-enhanced fused data is used as the input of the BiLSTM network model to continue deep sequence modeling.

[0109] The BiLSTM network model is used to capture and propagate information from adjacent depths in the input data. It includes a forward LSTM network for forward learning and a backward LSTM network for backward learning. Both the forward and backward LSTM networks include the input at time t, cell state, temporary cell state, hidden layer state, forget gate, memory gate, and output gate.

[0110] In this embodiment, the BiLSTM network model incorporates a backward LSTM network in conjunction with a forward LSTM network, integrating gating architecture and bidirectional characteristics. This fully considers the impact of historical and future information on the current state, resulting in better memory and robustness. Furthermore, the BiLSTM network model treats well logging data along the depth direction as an ordered sequence. This allows for further extraction of feature information from the input samples and capture of features propagating to adjacent depths, fully utilizing the long-term dependency features of the input samples for learning.

[0111] When input data enters the BiLSTM network model, it sequentially enters the forward LSTM network and the backward LSTM network. The forward LSTM network processes the input data in a forward order, while the backward LSTM network processes the input data in a reverse order. The forward and backward LSTM networks respectively use the input data and the hidden state of the previous time step to calculate the hidden state of the current time step. The hidden states of the forward and backward LSTM networks are then connected or merged to obtain the output of the BiLSTM network model.

[0112] in,

[0113] The forward propagation of the forward LSTM network is shown in Equation (7):

[0114]

[0115] In the formula, x t The input at time t, For the forget gate that propagates forward at time t, Forget gates that propagate forward at time t-1, LSTM(·) is a bidirectional long short-term memory neural network model.

[0116] The backpropagation of the backpropagation LSTM network is shown in Equation (8):

[0117]

[0118] In the formula, For the forget gate that propagates backward at time t, Forget gates that propagate backwards at time t+1.

[0119] The output of the BiLSTM network model is:

[0120]

[0121] In the formula, h t Let t represent the state of the hidden layer of the BiLSTM network model at time t.

[0122] In this embodiment, the bright crystal discriminator constructed by the CNN-BiLSTM-Attention combined model based on the attention mechanism is trained using the Adam gradient descent algorithm. Before training, the initial learning rate of the bright crystal discriminator is preset to 0.001. During the training process, the bright crystal discriminator adaptively adjusts the learning rate and updates the network weights. The cross-entropy loss function is used as the loss function during the training process to evaluate the accuracy of the bright crystal discriminator in recognizing bright crystals, as shown in formula (10):

[0123]

[0124] In the formula, L(·) is the objective function to be minimized, m is the number of samples; y is the true class, which takes the value of 0 or 1; h(x) is the probability of the bright crystal discriminator predicting sample x, and θ is the model parameter of the bright crystal discriminator.

[0125] Step 5: Train the bright crystal discriminator constructed using the CNN-BiLSTM-Attention combined model based on the attention mechanism using the training set to obtain the trained bright crystal discriminator. This specifically includes the following steps:

[0126] Step 5.1: Set the maximum number of training iterations for the bright crystal discriminator to 2000, and set the initial learning rate of the bright crystal discriminator to 0.001.

[0127] Step 5.2: Randomly select training samples from the training set and input them into the bright crystal discriminator. Use the bright crystal discriminator to identify whether the training samples contain bright crystals. The bright crystal discriminator automatically labels the training samples that are identified as bright crystals with a bright crystal label and labels the training samples that are not identified as non-bright crystals with a non-bright crystal label.

[0128] Step 5.3: Compare the labels assigned by the bright crystal discriminator with the manually assigned labels in the training samples. If they match, proceed to step 5.4. Otherwise, adaptively adjust the learning rate of the bright crystal discriminator and adjust the network weights to update the bright crystal discriminator, and return to step 5.2 to continue training the bright crystal discriminator.

[0129] Step 5.4: Determine whether the current training coefficients have reached the preset maximum number of training iterations. If the current number of training iterations has reached the maximum number of training iterations, stop training the bright crystal discriminator and obtain the trained bright crystal discriminator. Otherwise, return to step 5.2 to continue training the bright crystal discriminator.

[0130] In this embodiment, during the training process, the loss function of the bright crystal discriminator gradually stabilizes as the number of training iterations increases, and eventually the loss function value of the bright crystal discriminator gradually stabilizes at 0.2, at which point the recognition effect of the bright crystal discriminator is the best.

[0131] Step 6: Validate the bright crystal discriminator using the validation set to obtain the validated bright crystal discriminator. This specifically includes the following steps:

[0132] Step 6.1: Input the verification samples in the verification set into the bright crystal discriminator in sequence. Use the bright crystal discriminator to identify whether the verification samples are bright crystals and automatically mark the verification samples that are identified as bright crystals with bright crystal labels and mark the verification samples that are not identified as non-bright crystals with non-bright crystal labels. Compare with the manually labeled labels in the verification samples to determine whether the bright crystal recognition result of the bright crystal discriminator is accurate.

[0133] Step 6.2: After the bright crystal discriminator has completed the bright crystal recognition of all verification samples, obtain the accuracy, precision, and recall of the bright crystal discriminator during the bright crystal recognition process, and compare them with the preset accuracy, precision, and recall. If the accuracy, precision, and recall of the bright crystal discriminator have reached the preset accuracy, precision, and recall, then proceed to step 7; otherwise, return to step 5 to continue training the bright crystal discriminator.

[0134] Step 7: Use the verified bright crystal discriminator to perform bright crystal identification on the well section to be identified. Input the logging data extracted from the logging curve of the well section to be identified into the bright crystal discriminator according to the depth sequence to obtain the bright crystal identification result of the well section to be identified.

[0135] In this embodiment, to compare the bright crystal recognition performance of CNN networks, BiLSTM networks, and the bright crystal discriminator (a CNN-BiLSTM-Attention combined model based on an attention mechanism) constructed in this invention, single-well logging data and adjacent-well logging data were input into CNN networks, BiLSTM networks, and the bright crystal discriminator constructed in this invention, respectively, to conduct single-well bright crystal recognition experiments and adjacent-well bright crystal recognition experiments. The bright crystal recognition results of CNN networks, BiLSTM networks, and the bright crystal discriminator constructed in this invention in the single-well bright crystal recognition experiments and adjacent-well bright crystal recognition experiments were obtained, as follows: Figure 5 and Figure 6 As shown in Tables 1 and 2, the accuracy, precision, and recall of the CNN network, BiLSTM network, and the bright crystal discriminator constructed in this invention were obtained in the single-well bright crystal recognition experiment and the adjacent-well bright crystal recognition experiment.

[0136] Table 1. Evaluation of the bright crystal recognition performance of each model in the single-well bright crystal recognition experiment.

[0137]

[0138] Table 2 Evaluation table of the bright crystal recognition performance of each model in the adjacent well bright crystal recognition experiment.

[0139]

[0140] By comparing Tables 1 and 2, the accuracy, precision, and recall of each model in the bright crystal identification process were compared. Combined with phase field core photographs, core descriptions, and thin section data, it was found that the bright crystal discriminator constructed by the CNN-BiLSTM-Attention combined model based on the attention mechanism of the present invention was the best in both single-well bright crystal identification experiments and adjacent-well bright crystal identification. Therefore, the method of the present invention effectively improves the identification accuracy of bright crystals in shale and provides technical support for complex shale reservoirs.

[0141] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A smart identification method for distinguishing shale crystals, characterized in that, Specifically, the following steps are included: Step 1: Obtain the logging curves of the target well section in the shale reservoir, obtain the logging data of the target well section, combine with the core data, analyze the mineral composition and structural characteristics in the shale reservoir, manually determine the bright crystalline and non-bright crystalline sections on the logging curves, and manually mark the labels on the logging curves. Step 2: After preprocessing the logging data of the target well section, perform Pearson correlation analysis on each logging parameter. Based on the Pearson correlation coefficient of each logging parameter, filter out the redundant curves in the logging curves of the target well section. Then, combine the artificially labeled bright crystal and non-bright crystal labels on the logging curves of the target well section to construct a sample library including bright crystal sample set and non-bright crystal sample set. Step 3: Randomly select bright crystal samples and non-bright crystal samples from the sample library as training samples and validation samples to form training set and validation set. Expand the bright crystal samples in the training set based on the smote method so that the number of bright crystal samples in the training set is equal to the number of non-bright crystal samples. Step 4: Construct a bright crystal discriminator using a CNN-BiLSTM-Attention combined model based on the attention mechanism; Step 5: Use the training set to train the bright crystal discriminator constructed by the CNN-BiLSTM-Attention combined model based on the attention mechanism, and obtain the trained bright crystal discriminator. Step 6: Validate the bright crystal discriminator using the validation set to obtain the validated bright crystal discriminator; Step 7: Use the verified bright crystal discriminator to perform bright crystal identification on the well section to be identified. Input the logging data extracted from the logging curve of the well section to be identified into the bright crystal discriminator according to the depth sequence to obtain the bright crystal identification result of the well section to be identified.

2. The intelligent identification method for distinguishing shale crystals according to claim 1, characterized in that, The logging curves include natural gamma curve, spontaneous potential curve, sonic transit time curve, density curve, compensated neutron curve, and resistivity curve.

3. The intelligent identification method for distinguishing shale crystals according to claim 1, characterized in that, Step 2 specifically includes the following steps: Step 2.1: Preprocess the logging data of the target well section, including depth correction and noise reduction. Step 2.2: Perform Pearson correlation analysis on each logging parameter. The formula for calculating the Pearson correlation coefficient of the logging parameters is as follows: In the formula, ρ X,Y Here, X and Y are the Pearson correlation coefficients, where i is the sequence number and n is the total number of logging parameters. i Let X be the measured value of the i-th logging parameter. Let X be the average value of the logging parameter, and Y be the average value of the logging parameter. i Let Y be the measured value of the i-th logging parameter. This represents the average value of the logging parameter Y; The Pearson correlation coefficient ρ X,Y A value less than 0 indicates a negative correlation between logging parameter X and logging parameter Y; the Pearson correlation coefficient ρ X,Y A value greater than 0 indicates a positive correlation between logging parameter X and logging parameter Y, and the Pearson correlation coefficient ρ is greater than 0. X,Y The closer the correlation is to 0, the worse the correlation between logging parameter X and logging parameter Y becomes; Step 2.3, based on the Pearson correlation coefficient ρ X,Y Screen the logging curves of the target well section, and |ρ X,Y The logging curves at |≥0.9 are considered redundant curves. After removing redundant curves from the logging curves of the target well section, the logging data on the filtered target well section are used, combined with the bright crystal and non-bright crystal labels on the logging curves. The logging data at the same depth point in the bright crystal section are taken as a bright crystal sample, and multiple bright crystal samples are obtained to construct a bright crystal sample set. Then, the logging data at the same depth point in the non-bright crystal section are taken as a non-bright crystal sample, and multiple non-bright crystal samples are obtained to construct a non-bright crystal sample set. The resulting sample library includes the bright crystal sample set and the non-bright crystal sample set.

4. The intelligent identification method for distinguishing shale crystals according to claim 1, characterized in that, The ratio of training samples in the training set to validation samples in the validation set is 9:

1.

5. The intelligent identification method for distinguishing shale crystals according to claim 1, characterized in that, The bright crystal discriminator is constructed based on a CNN-BiLSTM-Attention combined model with an attention mechanism. The CNN-BiLSTM-Attention combined model with an attention mechanism includes a convolutional neural network model and a BiLSTM network model with an attention mechanism, which are connected in sequence. The attention-based convolutional neural network model includes a CNN branch, an attention mechanism branch, and a multiplication layer arranged in parallel. The CNN branch is used to extract the feature vector of the input sample, including a first convolutional layer, a first activation layer, a second convolutional layer, and a second activation layer. Both the first and second convolutional layers are 3×1 convolutional layers. The first convolutional layer has 32 convolutional kernels, and the second convolutional layer has 64 convolutional kernels. The convolution operation of the first and second convolutional layers is as shown in formula (2): In the formula, h i x is the feature vector obtained from the convolution operation. i Let w be the input value of the i-th element in the convolutional layer. i Let b be the weight matrix value corresponding to the i-th element. i Let be the bias value of the i-th element, and N×N be the size of the convolution kernel; Both the first and second activation layers use the ReLU function as the activation function, as shown in formula (3): y i = max(0,h i ) (3) In the formula, y i The activation function value is max(0,h). i ) is the maximum value function, used to select 0 and the feature vector h obtained from the convolution operation. i The maximum value in; The attention mechanism branch is used to assist the convolutional neural network in extracting feature vectors. It includes a global average pooling layer, a first fully connected layer, a first activation layer, a second fully connected layer, and a second activation layer. The number of channels in the first fully connected layer is set to 16, and the number of channels in the second fully connected layer is set to 64. The first activation layer uses the ReLU function as the activation function, and the second activation layer uses the sigmoid function as the activation function. The attention mechanism branch uses the SE module of the soft attention mechanism to assist the convolutional neural network in extracting feature vectors. Weights representing importance are added to focus on specific channels, including the Squeeze part, the Excitation part, and the Reweight part. The global average pooling layer of the attention mechanism branch corresponds to the Squeeze part of the SE module. It performs a global pooling operation on the feature vectors in the input samples and performs average pooling on the feature vectors in each channel to obtain the global information vector, as shown in formula (4): In the formula, z c The global information vector is the output of the global average pooling layer, where W is the maximum width of the input feature vector, H is the maximum height of the input feature vector, C is the number of channels in the global average pooling layer, and u c (I,J) represents the input feature vector value with width I and height J; The first fully connected layer, the first activation layer, the second fully connected layer, and the second activation layer of the attention mechanism branch correspond to the Excitation part of the SE module. The global information vector obtained by the global average pooling layer passes through the first fully connected layer, the first activation layer, the second fully connected layer, and the second activation layer in sequence to generate a channel weight vector to represent the channel importance, as shown in formula (5): s c =σ(W2δ(W1z c )) (5) In the formula, s c σ(·) is the channel weight vector, δ(·) is the sigmoid activation function value, δ(·) is the ReLU activation function value, W1 is the fully connected layer parameter used for compression, and W2 is the fully connected layer parameter used for dimension restoration. The second activation layer of the CNN branch and the second activation layer of the attention mechanism branch are both connected to the multiplication layer. The multiplication layer corresponds to the Reweight part of the SE module and is used to fuse the feature vector extracted by the CNN branch and the channel weight vector generated by the attention mechanism branch. By multiplying the channel weight vector with the extracted feature vector, the fused data after feature enhancement is obtained, as shown in formula (6): In the formula, For the fused data after feature enhancement, F scale (·) represents the reweight function, u c The input feature vector values; The BiLSTM network model uses the fused data output by the attention-based convolutional neural network model as input data to capture and propagate information from adjacent depths in the input data. The BiLSTM network model includes a forward LSTM network for forward learning and a backward LSTM network for backward learning. Both the forward and backward LSTM networks include the input at time t, cell state, temporary cell state, hidden layer state, forget gate, memory gate, and output gate. After the input data enters the BiLSTM network model, it sequentially enters the forward and backward LSTM networks. The forward LSTM network processes the input data in a forward order, while the backward LSTM network processes it in a reverse order. The forward and backward LSTM networks respectively use the input data and the hidden state from the previous time step to calculate the hidden state at the current time step. The hidden states of the forward and backward LSTM networks are then connected or merged to obtain the output of the BiLSTM network model. The forward propagation of the forward LSTM network is shown in Equation (7): In the formula, x t The input at time t, For the forget gate that propagates forward at time t, Forget gates that propagate forward at time t-1, LSTM(·) is a bidirectional long short-term memory neural network model; The backpropagation of the backpropagation LSTM network is shown in Equation (8): In the formula, For the forget gate that propagates backward at time t, Forget gates that propagate backwards at time t+1; The output of the BiLSTM network model is: In the formula, h t Let t represent the state of the hidden layer of the BiLSTM network model at time t.

6. The intelligent identification method for distinguishing shale crystals according to claim 5, characterized in that, The bright crystal discriminator constructed by the CNN-BiLSTM-Attention combined model based on the attention mechanism is trained using the Adam gradient descent algorithm. During the training process, the bright crystal discriminator adaptively adjusts the learning rate and updates the network weights. The cross-entropy loss function is used as the loss function to evaluate the accuracy of the bright crystal discriminator in recognizing bright crystals, as shown in Equation (10): In the formula, L(·) is the objective function to be minimized, m is the number of samples; y is the true class, which takes the value of 0 or 1; h(x) is the probability of the bright crystal discriminator predicting sample x, and θ is the model parameter of the bright crystal discriminator.

7. The intelligent identification method for distinguishing shale crystals according to claim 1, characterized in that, Step 5 specifically includes the following steps: Step 5.1: Set the maximum number of training iterations and the initial learning rate for the bright crystal discriminator; Step 5.2: Randomly select training samples from the training set and input them into the bright crystal discriminator. Use the bright crystal discriminator to identify whether the training samples are bright crystals. The bright crystal discriminator automatically labels the training samples that are identified as bright crystals with a bright crystal label and labels the training samples that are not identified as non-bright crystals with a non-bright crystal label. Step 5.3: Compare the labels assigned by the bright crystal discriminator with the manually assigned labels in the training samples. If they match, proceed to step 5.

4. Otherwise, adaptively adjust the learning rate of the bright crystal discriminator and adjust the network weights to update the bright crystal discriminator, and return to step 5.2 to continue training the bright crystal discriminator. Step 5.4: Determine whether the current training coefficients have reached the preset maximum number of training iterations. If the current number of training iterations has reached the maximum number of training iterations, stop training the bright crystal discriminator and obtain the trained bright crystal discriminator. Otherwise, return to step 5.2 to continue training the bright crystal discriminator.

8. The intelligent identification method for distinguishing shale crystals according to claim 7, characterized in that, The maximum number of training iterations is no less than 2000, the initial learning rate is set to 0.001, and the loss function value of the bright crystal discriminator gradually stabilizes at 0.2 during the training process.

9. The intelligent identification method for distinguishing shale crystals according to claim 1, characterized in that, Step 6 specifically includes the following steps: Step 6.1: Input the verification samples in the verification set into the bright crystal discriminator in sequence. Use the bright crystal discriminator to identify whether the verification samples are bright crystals and automatically mark the verification samples that are identified as bright crystals with bright crystal labels and mark the verification samples that are not identified as non-bright crystals with non-bright crystal labels. Compare with the manually labeled labels in the verification samples to determine whether the bright crystal recognition result of the bright crystal discriminator is accurate. Step 6.2: After the bright crystal discriminator has completed the bright crystal recognition of all verification samples, obtain the accuracy, precision, and recall of the bright crystal discriminator during the bright crystal recognition process, and compare them with the preset accuracy, precision, and recall. If the accuracy, precision, and recall of the bright crystal discriminator have reached the preset accuracy, precision, and recall, then proceed to step 7; otherwise, return to step 5 to continue training the bright crystal discriminator.

Images