Sample expansion and intelligent phase judgment method based on VAE and SMOTE
By using VAE model dimensionality reduction and SMOTE method to expand geological constraint samples, combined with a dual-branch neural network, the problems of sample imbalance and loss of geophysical correlation in intelligent facies identification were solved, achieving high-precision sedimentary facies identification.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-02-03
- Publication Date
- 2026-04-03
AI Technical Summary
Existing intelligent phase identification methods have not effectively solved the problem of sample imbalance when processing multidimensional time series curve data, and directly splicing well logging curve features and sandstone parameters leads to the loss of geophysical correlation information, affecting classification accuracy.
The VAE model was used to reduce the dimensionality of the logging curve data, and the correlation between the curve feature vector and sandstone parameters was constructed. The SMOTE method was used to generate new samples, which were trained by a two-branch neural network model. Geological constraint weights were introduced to ensure that the new samples conform to the geological and physical laws.
It improved the quality and reliability of the samples, enhanced the model's ability to identify complex geological conditions, improved the accuracy of sedimentary facies identification, and solved the problem of sample imbalance.
Smart Images
Figure CN121786631A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of data processing technology, specifically to a sample augmentation and intelligent phase determination method based on VAE and SMOTE. Background Technology
[0002] Current intelligent phase identification methods, whether using CNN or RNN models, rely on input samples consisting of multiple well logging curves within a fixed time window or multiple well logging curves constrained by layer depth. Furthermore, to increase classification accuracy, it is often necessary to digitize and parameterize expert domain knowledge, inputting it into the model as domain-driven information to form a multi-input deep learning model. However, regardless of whether CNN, RNN, or hybrid models such as TCN are used, the problem of sample balance remains.
[0003] Commonly used sample equalization methods mainly include undersampling and oversampling. For phase-determining model samples, there is currently no mature oversampling method based on multidimensional time series curves. For example, the SMOTE method only supports one-dimensional vectors and not tensors (multidimensional vectors). A VAE (variable autoencoder) can be used to extract features and compress data from the logging curves, and then combined with sandstone parameters to meet the input requirements of the SMOTE algorithm. However, simply concatenating the curve feature vector and sandstone parameters for expansion does not consider the geophysical correlation between the two.
[0004] Well logging curves reflect the electrical and radioactive physical properties of formations and have an inherent geophysical relationship with parameters such as porosity and permeability of sandstone. For example, well logging curves for high-porosity sandstone should exhibit specific morphological characteristics. If this relationship is ignored during sample expansion, the generated new samples may show mismatches between well logging curve characteristics and sandstone parameters, violating geological laws. This can lead to the introduction of noisy data during model training, reducing classification accuracy. Summary of the Invention
[0005] In view of the above, it is necessary to provide a sample expansion and intelligent phase determination method based on VAE and SMOTE to solve the technical problem of intelligent phase determination under sample imbalance conditions.
[0006] One embodiment of this application provides a sample augmentation and intelligent phase determination method based on VAE and SMOTE, the method comprising: Based on the well logging curve data, sandstone parameter data, and sedimentary facies category of each sandstone interval, samples were obtained. The VAE model was used to reduce the dimensionality of the well logging curve data to obtain curve feature vectors. The geophysical correlation between the curve feature vectors and sandstone parameter data in each sample was analyzed to construct the correlation degree. The curve feature vector of each sample is concatenated with the sandstone parameter data to obtain the joint feature vector of each sample; based on the sedimentary facies category label of the sample and combined with the correlation degree, the samples to be expanded are screened; interpolation coefficients with correlation degree constraints are introduced between the joint feature vectors of the samples to be expanded, and new samples are generated using the SMOTE method to obtain the expanded curve feature vector and sandstone parameter data. A dual-branch neural network model is trained based on the expanded curve feature vector and sandstone parameter data. The trained dual-branch neural network model is then used to obtain sedimentary facies classification results based on the well logging curve data and sandstone parameter data of the sandstone interval to be detected.
[0007] Preferably, obtaining the curve feature vector and the well logging curve image specifically involves: First, the logging curve segments are normalized, and then converted into image format to form logging curve sample images; The well logging curve sample image is used as the input to the encoder of the VAE model. The encoder outputs the mean vector and standard deviation vector. The compressed curve feature vector is obtained through reparameterized sampling.
[0008] Preferably, the construction of the correlation degree specifically includes: For each sample, a vector composed of the normalized values of all sandstone parameters is taken as the sandstone parameter vector; the normalization bias is determined by comprehensively considering the numerical deviations between the curve feature vector after dimensionality reduction mapping and all dimensions of the sandstone parameter vector. The geological consistency weights are determined by comparing the sedimentary facies labels of each sandstone interval with those of its nearest neighboring sandstone intervals. The correlation degree is determined based on the normalized deviation and geological consistency weight corresponding to each sample; wherein the correlation degree is negatively correlated with the normalized deviation and positively correlated with the geological consistency weight.
[0009] Preferably, determining the normalization bias specifically involves: For each curve feature vector, the dimension is reduced to the same dimension as the number of sandstone parameter data types. The root mean square of the numerical deviation between the reduced curve feature vector and the sandstone parameter vector in all the same dimensions is used as the normalized deviation.
[0010] Preferably, the determination of the geological consistency weight is specifically the proportion of samples in the nearest neighbor sample set of each sample that have the same sedimentary facies label as each sample; wherein, the nearest neighbor sample set of each sample is obtained by the KNN algorithm.
[0011] Preferably, the screening of samples to be expanded specifically involves: Based on the sedimentary facies labels, the number of samples in each category is counted. All samples in all categories except the category with the largest number of samples are taken as samples of each minority class. Based on the overall distribution of the correlation degree of all samples in each minority class, the correlation degree threshold is obtained. For each minority class, samples with a correlation degree higher than the corresponding correlation degree threshold are taken as samples to be expanded.
[0012] Preferably, the interpolation coefficients used to introduce correlation constraints among the samples to be expanded are specifically: For each minority class, all samples to be expanded are arranged in the order in which they were obtained; for any two adjacent samples to be expanded, the proportion of the correlation degree of the latter sample to the sum of the correlation degrees of the two samples to be expanded is obtained, and multiplied by a random number that follows a standard normal distribution to obtain the interpolation coefficient between any two adjacent samples to be expanded.
[0013] Preferably, the number of samples in each minority class after expansion will be equal to the number of samples in the largest class in the original sample size.
[0014] Preferably, the dual-branch neural network includes a curve feature branch, a sandstone parameter branch, and a feature fusion classification layer.
[0015] Preferably, the sedimentary facies category identification result is the sedimentary facies category with the highest probability output by the dual-branch neural network.
[0016] This application has at least the following beneficial effects: This application first acquires well logging curve data and sandstone parameter data, which can comprehensively describe the characteristics of each sandstone interval, laying the foundation for subsequent feature extraction and model training. A VAE model is used to reduce the dimensionality of the well logging curve data. This process not only reduces the data dimensionality but also retains important information, making subsequent analysis and modeling more efficient. The dimensionality-reduced curve feature vectors can focus on the most representative features, helping to reveal the geophysical correlations between different samples. The relationship between the curve feature vectors and sandstone parameter data is analyzed to construct the correlation degree. This process helps identify which features are most critical for determining sedimentary facies categories, quantifies the physical correlation between well logging data and geological parameters, provides an objective basis for sample quality assessment, and solves the problem of lost correlation information caused by the simple splicing of multimodal data in traditional methods.
[0017] Secondly, by concatenating the curve feature vector with sandstone parameter data to form a joint feature vector, the information from both can be comprehensively considered, enriching the feature expression of the samples and enhancing the model's ability to identify complex geological conditions. The correlation coefficient is then used to screen the samples to be expanded, ensuring that the expanded samples can effectively represent minority class features and avoiding the introduction of noise and irrelevant data. Furthermore, correlation-constrained interpolation coefficients are introduced between the joint feature vectors of the samples to be expanded, and geological constraint weights are introduced during the SMOTE sample expansion process. This ensures that the generated new samples conform to geological and physical laws, avoiding the noise sample problem caused by blind expansion in traditional methods, and improving the quality and reliability of the expanded samples.
[0018] Finally, the dual-branch neural network model can simultaneously process curve feature vectors and sandstone parameter data, making full use of these two types of information for in-depth learning. This avoids the problem of unbalanced parameter weights caused by direct splicing, improves the classification accuracy of minority class samples, and achieves efficient and intelligent processing of imbalanced big data. The model can learn more representative features when processing complex data, thereby effectively improving the identification accuracy of sedimentary facies categories. Attached Figure Description
[0019] Figure 1 A flowchart of the sample augmentation and intelligent phase determination method based on VAE and SMOTE provided for this application; Figure 2 This is a schematic diagram of the original sample distribution provided in this application; Figure 3 A statistical information graph of the balanced sample provided for this application; Figure 4 A schematic diagram of the CNN model results provided in this application; Figure 5 A schematic diagram of the results of the dual-branch neural network model provided in this application. Detailed Implementation
[0020] In the description of the embodiments in this application, the words "exemplary," "or," and "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design scheme described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design schemes. Specifically, the use of the words "exemplary," "or," and "for example" is intended to present the relevant concepts in a specific manner.
[0021] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in this application's specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the application.
[0022] It should also be noted that the terms "first" and "second" in this application and its accompanying drawings are used to distinguish similar objects, rather than to describe a specific order or sequence. The methods disclosed in the embodiments of this application or the methods shown in the flowcharts include one or more steps for implementing the method. Without departing from the scope of protection of this application, the execution order of multiple steps can be interchanged, and some steps can also be deleted.
[0023] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.
[0024] This application proposes a sample augmentation and intelligent phase determination method based on VAE and SMOTE, which is applied to the field of data processing technology. (See attached document.) Figure 1 The method includes the following steps: S1: Obtain well logging curve data and sandstone parameter data for sandstone strata. Use the VAE model to perform dimensionality reduction on the well logging curve data to obtain curve feature vectors and well logging curve images. Analyze the geophysical correlation between curve feature vectors and sandstone parameters in each sample and construct the correlation degree.
[0025] First, training samples were collected for a known sandstone interval, including well logging data such as the microgradient curve (RMN), micropotential curve (RMG), natural gamma curve (GR), and spontaneous potential curve (SP), with a sampling interval of 0.05 meters. Based on the top and bottom depth information in the sandstone data table, well logging curve segments corresponding to the depth range were extracted. Simultaneously, sandstone parameter data, including sandstone thickness, were collected. Effective thickness Porosity and penetration rate .
[0026] First, the logging curve segments are normalized to 0-1, and then the curve data is converted into an image format with a pixel value range of 0-255 to form a logging curve sample image. In this embodiment, the normalization method used is the maximum-minimum normalization method. Z-score normalization is applied to the sandstone parameters to form sandstone parameter vectors. Each sample includes a well logging curve sample image, a sandstone parameter vector, and its corresponding sedimentary facies category label.
[0027] Construct a variational autoencoder (VAE) model, consisting of two parts: an encoder and a decoder. The encoder processes the input well logging curve sample images. Compressed into a 128-dimensional curve feature vector The encoder outputs the mean vector. and standard deviation vector The curve feature vector is obtained through reparameterized sampling, as shown in the formula. .in, This indicates sampling from a standard normal distribution. This represents element-wise multiplication; the decoder extracts the curve feature vector. Reconstructed into a well logging curve image.
[0028] The loss function of VAE includes reconstruction loss and KL divergence regularization term. Reconstruction loss is used to measure the difference between the reconstructed image and the original image; KL divergence regularization term is used to constrain the latent spatial distribution to be close to the standard normal distribution.
[0029] The VAE model was trained using all well logging curve sample images. After training, only the encoder was used to process each well logging curve sample image. Convert to 128-dimensional curve eigenvectors Let be denoted as the curve eigenvector set. ,in Indicates the total number of samples; , , These represent the curve feature vectors of the 1st sample, the 2nd sample, and the Nth sample, respectively.
[0030] To quantify the geophysical correlation between the curve eigenvectors and sandstone parameters, this application constructs a correlation degree. First, the normalized deviation is calculated to measure the consistency between the curve characteristics and the parameters; the formula is as follows: .in, Indicates the first Normalization bias of each sample Indicates the first The curve feature vector of each sample After dimensionality reduction mapping to 4-dimensional space, the first... Each component is obtained by averaging the data of each of the 32 dimensions of the 128-dimensional vector. This mapping preserves the main information of the curve features while reducing the dimensionality to correspond to the 4-dimensional sandstone parameters. Indicates the first The first sample Sandstone parameters (after standardization) max() represents the maximum value function; This represents a small constant to prevent the denominator from being zero; in this embodiment, it is set to 0.01. This represents the numerical deviation between the eigenvector of the curve after dimensionality reduction and the sandstone parameter vector in all the same dimensions.
[0031] The principle behind the above formula is as follows: by calculating the root mean square of the relative deviation in the corresponding dimension, the normalized deviation value is obtained. Its physical meaning represents the degree of inconsistency between the curve characteristics and the sandstone parameters, and its value range is... The smaller the value, the more closely the two match; when the two are completely identical... .
[0032] Geological consistency weights are constructed based on the degree of clustering of similar samples, specifically as follows: ;in, This represents the geological consistency weight of the i-th sample. Indicates sample In feature space The set consists of 3 nearest neighbor samples, which are obtained using the KNN algorithm. In this embodiment, the number of nearest neighbor samples is... This value is a common configuration for the KNN algorithm, which can capture local features without introducing too much noise; Indicates sample Category tags; This indicates an indicator function; its value is 1 when the condition within the parentheses is true, and 0 otherwise.
[0033] Geological consistency weight primarily measures the proportion of samples of the same class among nearest neighbors. Its physical meaning is the class purity of the sample; a higher value indicates that the sample is surrounded by similar samples in the feature space and is a typical representative of that class. When all nearest neighbors are of the same class... .
[0034] Finally, the correlation degree is constructed, and the formula is: .in, Indicates the first Curve-parameter correlation for each sample; This represents the geological consistency weight of the i-th sample; Indicates the first The normalization bias of each sample. The principle of this formula is: when When the denominator reaches its minimum value of 1, the correlation is at its maximum, indicating that the sample has both high class purity and good curve-parameter matching, making it a high-quality representative sample; when (All neighbors are of the same kind) and When the curve and parameters are perfectly matched, the correlation reaches its maximum value of 1.
[0035] Calculate the correlation degree for all samples to form a correlation degree vector. The final set of curve feature vectors is obtained. Sandstone parameter vector set and correlation vector .
[0036] In one embodiment, logging data was collected from 612 wells, with a logging curve depth of approximately 300 meters and a sampling interval of 0.05 meters. After data cleaning and sample creation, a total of 1370 samples of channel facies (H type), 1181 samples of sandbar facies (SB type), and 4268 samples of interfluvial sands (Z type) were obtained. A schematic diagram of the original sample distribution is shown below. Figure 2 As shown, there is a significant imbalance in sample size, with the number of interfluvial sand samples being 3.6 times that of sandbar facies and 3.1 times that of channel facies.
[0037] S2: Concatenate the curve feature vector of each sample with the sandstone parameter data to obtain the joint feature vector of each sample; based on the sedimentary facies category label of the sample and combined with the correlation degree, screen the samples to be expanded; introduce interpolation coefficients with correlation degree constraints between the joint feature vectors of the samples to be expanded, and use the SMOTE method to generate new samples to obtain the expanded curve feature vector and sandstone parameter data.
[0038] Traditional SMOTE methods uniformly expand all minority class samples, which can also expand low-quality samples, introducing samples that do not conform to geological patterns. In contrast, this application identifies high-quality samples based on correlation vectors and introduces geological constraint weights during the SMOTE interpolation process to ensure that the generated new samples maintain a reasonable geological correlation between curve features and sandstone parameters. Specifically: curve eigenvectors Tail and sandstone parameter vector The first elements are concatenated to form a 132-dimensional joint feature vector. The number of samples in each category is counted to identify the minority class. For example, the number of river channel samples is... The number of sand dam phase samples is The number of river sand samples is The minority categories are channel facies and sandbar facies.
[0039] For each minority class, based on the degree of association Perform sample screening. Set a correlation threshold. Only samples with a correlation score higher than the correlation score threshold are retained for expansion. The correlation score threshold is... The median of the association degree for each category is used to select samples with association degrees in the upper half, thus avoiding the expansion of low-quality samples. The set of all samples to be expanded for each minority class is denoted as... And all the samples to be expanded are arranged in the order in which they were obtained.
[0040] For each sample to be expanded New samples are generated using the SMOTE method. The basic principle of SMOTE is to generate new samples from the samples to be expanded. and the next sample to be expanded Linear interpolation is performed between the two samples to obtain new samples. The formula can be expressed as: .in, This represents the random interpolation coefficients. The traditional SMOTE algorithm uses uniformly distributed random sampling.
[0041] This application introduces weighted interpolation coefficients with correlation constraints, causing the distribution of interpolation coefficients to become non-uniform, but rather skewed towards higher correlation. The specific formula for the interpolation coefficients is as follows: ;in, Represents a distributed random number. , Let represent the correlation between the i-th sample to be expanded and the j-th sample to be expanded, respectively.
[0042] Interpolation coefficients of traditional SMOTE obey The new samples are uniformly distributed with an expected value of 0.5, and are uniformly distributed between the two original samples. This application adjusts the distribution of interpolation coefficients by weighting based on correlation. When nearest neighbor samples... The correlation is higher than that of the current sample At that time, that is ,but interpolation coefficients The expected value is greater than 0.5, making the generated new sample Closer to highly correlated samples This allows for the inheritance of better geological characteristics; that is, the generation of new samples is no longer random interpolation without direction, but rather, under the guidance of geological and physical laws, it is biased towards samples with better matching characteristics and parameters and more reasonable geological significance, thereby achieving geological constraints; the weighting coefficient maintains the basic characteristics of SMOTE interpolation.
[0043] Calculate the number of samples needed to augment each minority class to balance the sample size across classes. In one embodiment, the target sample size is set to the number of samples in the most majority class. For riverine facies, it is necessary to generate... A new sample needs to be generated; for sandbar facies, a new sample needs to be generated. A new sample.
[0044] Repeat the weighted SMOTE interpolation process until a sufficient number of new samples are generated. Combine the original samples and the newly generated samples to form the expanded curve feature vector set. and the expanded sandstone parameter vector set The sample size for each category is 4268. The statistical infographic after sample equilibrium is shown below. Figure 3 As shown.
[0045] Finally, the expanded 132-dimensional joint vector... Decomposed into curve feature part and sandstone parameter part: through vector slicing operation, the first 128 dimensions are extracted as curve feature vector. The extracted 4-dimensional vector is used as the sandstone parameter vector. .
[0046] S3: Train a dual-branch neural network model based on the expanded curve feature vector and sandstone parameter data. Use the trained dual-branch neural network model to obtain sedimentary facies identification results based on the well logging curve data and sandstone parameter data of the sandstone interval to be detected.
[0047] The purpose of this step is to construct a dual-branch neural network classification model, inputting the expanded curve feature vector and sandstone parameter vector into two independent branches respectively, thereby avoiding the problem of unbalanced parameter weights caused by direct splicing and achieving high-precision intelligent classification of sedimentary facies.
[0048] The expanded curve feature vector and sandstone parameter vector are input into two independent branches of a dual-branch neural network, respectively. Classification is performed through a feature fusion layer, and the sedimentary facies category identification result is output, specifically: A dual-branch neural network is constructed, comprising a curve feature branch, a sandstone parameter branch, and a feature fusion classification layer. This structure aims to address the parameter weight imbalance problem caused by direct concatenation. If a 132-dimensional vector is directly input into a single network, its feature weights are easily overwhelmed by the 128-dimensional curve features due to the sandstone parameters' 4-dimensional nature. The dual-branch structure assigns an independent feature extraction path to each modality, ensuring sufficient learning and balanced fusion of the two types of features.
[0049] The curve feature branch receives a 128-dimensional curve feature vector. As input, it contains three fully connected layers: the first layer maps 128 dimensions to 64 dimensions using the ReLU activation function; the second layer maps 64 dimensions to 32 dimensions using the ReLU activation function; and the third layer maps 32 dimensions to 16 dimensions using the ReLU activation function. The output of the curve feature branch is a 16-dimensional feature vector. .
[0050] The sandstone parameter branch receives a 4-dimensional sandstone parameter vector. As input, it contains two fully connected layers: the first layer maps 4 dimensions to 8 dimensions using the ReLU activation function; the second layer maps 8 dimensions to 16 dimensions using the ReLU activation function. The output of the sandstone parameter branch is a 16-dimensional feature vector. .
[0051] The feature fusion classification layer concatenates the outputs of the two branches into a 32-dimensional vector. The process is mapped to 3D (corresponding to three categories: channel facies, sandbar facies, and interchannel sand) through a fully connected layer, and then the Softmax activation function is used to output the category probability distribution.
[0052] The model was trained using cross-entropy loss as the loss function, and the Adam optimizer was used for parameter updates. The learning rate was set to 0.0001, the batch size to 32, and the number of training epochs to 60. The augmented samples were divided into training and validation sets in an 8:2 ratio. The model was trained using the training set and evaluated using the validation set.
[0053] After 60 rounds of training, the model achieved an overall accuracy of 92.97% on the validation set. For the most critical minority class samples, the classification performance for sandbar facies was: precision 90.21%, recall 93.25%, and F1 score 91.7%; the classification performance for channel facies was: precision 95.92%, recall 94.8%, and F1 score 95.36%.
[0054] To verify the validity of this application, a comparison was made with a traditional scheme that does not employ geological constraints. The comparison scheme used the same dataset and directly processed well logging curve images using a CNN model without sample augmentation or correlation constraints. A schematic diagram of the traditional CNN model results is shown below. Figure 4 As shown in the diagram, the result of the two-branch neural network model is illustrated below. Figure 5 As shown.
[0055] from Figure 4 and Figure 5 As can be seen, the traditional CNN model (vae_text1) performs as follows on the sandbar facies: precision 55.65%, recall 68.91%, and F1 score 61.57%. Compared with the traditional approach, the dual-branch neural network model (vae_text2) in this application improves the F1 score by 30.13 percentage points (from 61.57% to 91.7%), precision by 34.56 percentage points, and recall by 24.34 percentage points on the sandbar facies, demonstrating that the geologically constrained sample augmentation method can significantly improve the classification accuracy of minority class samples and effectively solve the problem of imbalanced samples.
[0056] After training, for new samples, the VAE encoder first extracts the curve feature vector and obtains the corresponding sandstone parameter vector. Then, both are input into the trained dual-branch network model to obtain the class probability distribution. The category with the highest probability is selected as the prediction result.
[0057] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions marked in the blocks may occur in a different order than that shown in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. In the descriptions corresponding to the flowcharts and block diagrams in the accompanying drawings, the operations or steps corresponding to different blocks may also occur in a different order than disclosed in the description; sometimes there is no specific order between different operations or steps. For example, two consecutive operations or steps may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. Each block in a block diagram and / or flowchart, and combinations of blocks in a block diagram and / or flowchart, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.
[0058] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.
Claims
1. A sample augmentation and intelligent phase determination method based on VAE and SMOTE, characterized in that, The method includes the following steps: Based on the well logging curve data, sandstone parameter data, and sedimentary facies category of each sandstone interval, samples were obtained. The VAE model was used to reduce the dimensionality of the well logging curve data to obtain curve feature vectors. The geophysical correlation between the curve feature vectors and sandstone parameter data in each sample was analyzed to construct the correlation degree. The curve feature vector of each sample is concatenated with the sandstone parameter data to obtain the joint feature vector of each sample; based on the sedimentary facies category label of the sample and combined with the correlation degree, the samples to be expanded are screened; interpolation coefficients with correlation degree constraints are introduced between the joint feature vectors of the samples to be expanded, and new samples are generated using the SMOTE method to obtain the expanded curve feature vector and sandstone parameter data. A dual-branch neural network model is trained based on the expanded curve feature vector and sandstone parameter data. The trained dual-branch neural network model is then used to obtain sedimentary facies classification results based on the well logging curve data and sandstone parameter data of the sandstone interval to be detected.
2. The sample augmentation and intelligent phase determination method based on VAE and SMOTE as described in claim 1, characterized in that, The specific steps for obtaining the curve feature vector and the well logging curve image are as follows: First, the logging curve segments are normalized, and then converted into image format to form logging curve sample images; The well logging curve sample image is used as the input to the encoder of the VAE model. The encoder outputs the mean vector and standard deviation vector. The compressed curve feature vector is obtained through reparameterized sampling.
3. The sample augmentation and intelligent phase determination method based on VAE and SMOTE as described in claim 1, characterized in that, The construction of the correlation degree specifically involves: For each sample, a vector composed of the normalized values of all sandstone parameters is taken as the sandstone parameter vector; the normalization bias is determined by comprehensively considering the numerical deviations between the curve feature vector after dimensionality reduction mapping and all dimensions of the sandstone parameter vector. The geological consistency weights are determined by comparing the sedimentary facies labels of each sandstone interval with those of its nearest neighboring sandstone intervals. The correlation degree is determined based on the normalized deviation and geological consistency weight corresponding to each sample; wherein the correlation degree is negatively correlated with the normalized deviation and positively correlated with the geological consistency weight.
4. The sample augmentation and intelligent phase determination method based on VAE and SMOTE as described in claim 3, characterized in that, The determination of the normalization bias specifically involves: For each curve feature vector, the dimension is reduced to the same dimension as the number of sandstone parameter data types. The root mean square of the numerical deviation between the reduced curve feature vector and the sandstone parameter vector in all the same dimensions is used as the normalized deviation.
5. The sample augmentation and intelligent phase determination method based on VAE and SMOTE as described in claim 3, characterized in that, The determination of the geological consistency weight is specifically the proportion of samples in the nearest neighbor sample set of each sample that have the same sedimentary facies label as each sample; wherein, the nearest neighbor sample set of each sample is obtained by the KNN algorithm.
6. The sample augmentation and intelligent phase determination method based on VAE and SMOTE as described in claim 1, characterized in that, The screening of samples to be expanded specifically involves: Based on the sedimentary facies labels, the number of samples in each category is counted. All samples in all categories except the category with the largest number of samples are taken as samples of each minority class. Based on the overall distribution of the correlation degree of all samples in each minority class, the correlation degree threshold is obtained. For each minority class, samples with a correlation degree higher than the corresponding correlation degree threshold are taken as samples to be expanded.
7. The sample augmentation and intelligent phase determination method based on VAE and SMOTE as described in claim 6, characterized in that, The interpolation coefficients that introduce correlation constraints among the samples to be expanded are specifically: For each minority class, all samples to be expanded are arranged in the order in which they were obtained; for any two adjacent samples to be expanded, the proportion of the correlation degree of the latter sample to the sum of the correlation degrees of the two samples to be expanded is obtained, and multiplied by a random number that follows a standard normal distribution to obtain the interpolation coefficient between any two adjacent samples to be expanded.
8. The sample augmentation and intelligent phase determination method based on VAE and SMOTE as described in claim 6, characterized in that, The number of samples in each minority class after augmentation will be equal to the number of samples in the largest class of the original sample size.
9. The sample augmentation and intelligent phase determination method based on VAE and SMOTE as described in claim 1, characterized in that, The dual-branch neural network includes a curve feature branch, a sandstone parameter branch, and a feature fusion classification layer.
10. The sample augmentation and intelligent phase determination method based on VAE and SMOTE as described in claim 9, characterized in that, The sedimentary facies category identification result is specifically the sedimentary facies category with the highest probability output by the dual-branch neural network.