Complex carbonate rock logging lithology identification method based on diffusion model

Abstract

The invention belongs to the technical field of carbonate rock oil-gas exploration, and particularly discloses a complex carbonate rock logging lithology identification method based on a diffusion model, and the method comprises the following steps: determining the lithology types of a plurality of observation wells based on the on-site rock core observation and slice analysis, and synchronously obtaining the logging data of the corresponding observation wells, constructing a lithology training data set in combination with a depth corresponding relationship between the lithology category and the logging data; a diffusion model is adopted to generate and supplement lithology categories with insufficient samples; logging data is adopted as an input feature, the lithology category is adopted as an output label, and a convolutional neural network fused with a Bayesian optimization algorithm and a Transformer hybrid model are utilized to train a lithology identification model; and inputting to-be-identified logging data into the trained lithology identification model, and outputting a lithology identification result. According to the method, high-precision lithology identification can be realized under the conditions of deep layers and complex stratums.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of carbonate rock oil and gas exploration, and more particularly relates to a complex carbonate rock logging lithology identification method based on a diffusion model. BACKGROUND

[0002] Carbonate reservoir lithology prediction is a core technical link in oil and gas exploration, especially in complex structural areas. Accurate identification and fine classification of reservoirs are of great significance for subsequent oil and gas development deployment. Traditional lithology identification methods mainly rely on qualitative interpretation of conventional logging curves, rock physical parameter analysis, and core logging and thin section observation. However, these methods are heavily dependent on the professional experience and subjective judgment of geologists, and require high data integrity, making it difficult to achieve efficient and automated processing.

[0003] In particular, in complex carbonate reservoirs, there are various lithology types, irregular spatial distribution, frequent lithofacies combination, and complex pore structure, showing strong heterogeneity and nonlinear response characteristics. Under this background, some lithology types are extremely rare due to special geological causes or sampling conditions, forming a significant small sample problem. Traditional methods often cannot accurately identify rare lithology types when facing such data imbalance or sample scarcity, resulting in decreased classification accuracy and insufficient model generalization ability, making it difficult to meet the technical needs of accurate identification of complex reservoirs in actual exploration.

[0004] Although existing lithology prediction methods have improved identification efficiency and accuracy to some extent, their performance and stability still face serious challenges when faced with small samples and low-quality logging data in complex carbonate rock formations. Carbonate reservoirs have complex lithology combinations, strong heterogeneity, and frequent interlayer mutations, resulting in uneven natural distribution of sample quantities for different lithology types, with some rare lithology classes showing significant small sample characteristics in training data. In deep exploration, logging data is often of poor quality due to factors such as instrument noise, signal attenuation, resolution limitations, and stratigraphic interference, with a large number of outliers and missing points.

[0005] Under such circumstances, traditional lithology prediction methods such as polynomial regression, support vector machine (SVM), and decision tree are usually based on ideal data assumptions and are difficult to cope with high noise, low resolution, or small sample scenarios, resulting in significant decreases in prediction accuracy and robustness. Even current mainstream deep learning methods are prone to overfitting and class bias when faced with extremely unbalanced small sample classes, degraded data quality, and high geological feature variability, making it difficult to fully extract key logging response features corresponding to rare lithology classes, ultimately leading to large prediction errors and weak generalization ability.

[0006] Therefore, in the lithology identification task of complex carbonate reservoirs, it is urgent to break through the high dependence of existing methods on data quality and sample size, and to propose an intelligent prediction framework with strong robustness and the ability to effectively deal with small sample distribution and low-quality logging data, so as to improve the lithology identification accuracy under deep and complex formation conditions. SUMMARY

[0007] In view of the defects of the prior art, the purpose of the present application is to provide a complex carbonate logging lithology identification method based on a diffusion model, which can realize high-precision lithology identification under deep and complex formation conditions.

[0008] To achieve the above purpose, in a first aspect, the present application provides a complex carbonate logging lithology identification method based on a diffusion model, comprising the following steps: S10, determining the lithology categories of a plurality of observation wells based on field core observation and thin section analysis, and synchronously acquiring the logging data of the corresponding observation wells, and then combining the depth correspondence relationship between the lithology categories and the logging data to construct a lithology training data set; S20, for the long-tail distribution of lithology categories in the lithology training data set, a diffusion model is used to generate and supplement the insufficient lithology categories; S30, on the supplemented lithology training data set, using logging data as input features and lithology categories as output labels, a convolutional neural network combined with a Bayesian optimization algorithm and a Transformer hybrid model are used to train a lithology identification model; S40, inputting the logging data to be identified into the trained lithology identification model to output the lithology identification result.

[0009] As a further optimization, step S10 further comprises: calculating the Pearson correlation coefficients between the features of each logging curve, and analyzing the correlation degree between each logging parameter and the known lithology category; According to the correlation strength, the logging feature parameters highly correlated with lithology prediction are selected, and redundant or weakly correlated variables are removed.

[0010] As a further optimization, step S20 specifically comprises: using Label Encoder to convert the text type lithology label into a recognizable numerical type identifier, and using IQR method to remove outliers for the input feature column, and normalizing different features; constructing an autoencoder neural network to train the normalized logging features, compressing and mapping the input features into latent vectors, and the decoder part is used to reversely restore the generated latent vectors to the original feature space; A forward diffusion noise adding strategy based on cosine scheduling is adopted, and noise is gradually added to the latent vector through the cumulative signal preservation coefficient to generate noise-added latent vectors of different degrees as the training input of the denoiser network. Using the DDIM reverse sampling algorithm, starting from a random noise vector, combining the trained denoising model and the forward scheduling coefficient, recursively calculating the new latent sample conforming to the minority class distribution, and then restoring it to the normalized feature space data through the decoder, and then using the maximum and minimum values consistent with the initial training to perform inverse normalization processing to restore the generated data to the actual physical dimension.

[0011] As a further optimization, step S20 further comprises: calculating the Wasserstein distance and KS test value of the real sample and the generated sample in each feature dimension, and outputting the distribution comparison chart, PCA or t-SNE dimension reduction visualization result, and the accuracy and F1 value comparison of the classifier before and after enhancement, forming a complete verification report.

[0012] As a further optimization, in step S30, the construction method of the convolutional neural network and the Transformer hybrid model comprises: The logging feature vector is regarded as a one-dimensional sequence, which is input into a multi-layer one-dimensional convolution module for local pattern capture. The convolution layers are sequentially stacked with batch normalization and ReLU activation functions, and the number of feature channels is gradually increased to match the input dimension of the Transformer encoder. The output of the convolution module is converted into a tensor that meets the input requirements of the Transformer, which is input into the Transformer encoder implemented based on PyTorch. The long-distance dependency in the logging data is captured through the multi-head self-attention mechanism. The output of the Transformer encoder is output through global average pooling and a fully connected layer to output the lithology class probability.

[0013] As a further optimization, in step S30, during the model training phase, a hyperparameter search method based on Bayesian optimization is introduced, and a Focal Loss loss function is introduced together with a class weight to alleviate the impact of class imbalance.

[0014] In a second aspect, the application provides a complex carbonate logging lithology identification system based on a diffusion model, comprising: A data set construction module is used to determine the lithology class of a plurality of observation wells based on field core observation and thin section analysis, and simultaneously acquire logging data of the corresponding observation wells. Then, combined with the depth correspondence relationship between the lithology class and the logging data, a lithology training data set is constructed. The data supplementing module is configured to generate and supplement the lithology categories with insufficient samples by using a diffusion model in view of the long-tail distribution of the lithology categories in the lithology training data set. The training module is configured to train the lithology identification model by using a convolutional neural network combined with a Transformer mixed model based on the fusion of a Bayesian optimization algorithm, and taking the logging data as input features and the lithology categories as output labels. The identification module is configured to input the logging data to be identified into the trained lithology identification model and output a lithology identification result.

[0015] In a third aspect, the present application provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the diffusion model-based complex carbonate rock logging lithology identification method according to any one of the above aspects when executing the program.

[0016] In a fourth aspect, the present application provides a computer-readable storage medium having computer instructions stored thereon, wherein the computer instructions are executable on a processor to implement the diffusion model-based complex carbonate rock logging lithology identification method according to any one of the above aspects.

[0017] Compared with the prior art, the present application has the following remarkable beneficial effects: The present application innovatively combines the lithology prediction task with the small sample data enhancement technology, breaking through the limitation of traditional lithology prediction which only relies on limited real samples. By introducing a conditional diffusion generation model, the lithology categories with insufficient samples are synthesized and supplemented in a targeted manner, which significantly expands the training sample size, improves the representativeness of the minority class samples in model learning, and effectively alleviates the influence of the uneven distribution of lithology categories on the prediction performance.

[0018] Unlike the traditional scheme of using only a single convolutional network or a sequence model, the lithology prediction network structure designed by the present application combines a convolutional neural network (CNN) with a Transformer encoder, fully combining the ability of convolution in extracting local features of logging curves and the advantage of Transformer in capturing long-distance dependencies and global context relationships, so that the model can more comprehensively capture feature patterns when facing logging data with coexistence of local mutations and spatial structures under complex geological backgrounds, thereby improving the accuracy and robustness of lithology identification.

[0019] Through the joint workflow of the conditional diffusion model and the CNN-Transformer network, the generalization performance and learning efficiency of the model are significantly improved in the network training stage. Moreover, the verification results on the actual lithology data in the field show that the generated supplementary samples combined with the manually labeled data can effectively improve the matching degree of the prediction results and the actual lithology, the boundary is clearer and more reasonable, and the stratum distribution characteristics of the complex carbonate reservoir can be more truly reflected.

[0020] According to the comprehensive comparison of the experimental results, the prediction accuracy and F1 value of the method provided in the application are better than those of the conventional method in the case of unbalanced lithology categories and limited sample size, and the method provided in the application has stronger capturing ability for local details and global correlation, has significant application value and promotion prospect, and can provide more reliable data support for subsequent reservoir modeling and resource evaluation. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 is a flowchart of the method for identifying lithology of a complex carbonate reservoir provided in the embodiments of the application based on a diffusion model; Figure 2 is a class number distribution diagram of original lithology data provided in the embodiments of the application; Figure 3 is a general architecture diagram of the diffusion model provided in the embodiments of the application; Figure 4 is a partial lithology training loss and verification loss diagram of the training process provided in the embodiments of the application; Figure 5 is a classification model architecture diagram provided in the embodiments of the application; Figure 6 is a lithology prediction loss diagram and an accuracy diagram provided in the embodiments of the application; Figure 7 is a verification result confusion matrix diagram provided in the embodiments of the application; Figure 8 is a complete flowchart provided in the embodiments of the application. DETAILED DESCRIPTION

[0022] In order to make the purpose, technical scheme and advantages of the application clearer, the application will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the application and do not limit the application.

[0023] To improve the lithology identification accuracy under deep and complex formation conditions, the application provides a complex carbonate rock logging lithology identification method based on a diffusion model. The method first constructs a lithology data generation network based on a diffusion probability model, uses diffusion disturbance and reverse recovery mechanism to generate and supplement a small number of lithology samples, expands the number of samples of rare categories, and effectively alleviates the model deviation problem caused by uneven sample distribution.

[0024] Meanwhile, the application designs a hybrid deep learning model combining convolutional neural network (CNN) and Transformer structure, which fully utilizes the advantages of CNN in extracting local spatial features of logging data and the ability of Transformer in capturing global dependencies of logging sequences, thereby improving the expression and identification ability of different scale lithology features. To further improve the training efficiency and prediction stability of the hybrid network under small sample and low quality logging data, a Bayesian optimization algorithm is introduced to automatically search and configure the key hyperparameters of the network, avoiding the problem of local optimum of traditional manual parameter tuning.

[0025] Among them, the CNN module can extract spatial features from original and generated logging data to identify local lithology changes; the Transformer module captures long-distance dependencies in logging data through self-attention mechanism to enhance the understanding of complex layers and geological background.

[0026] By connecting the small sample based on the diffusion model to the CNN-Transformer network for modeling, an end-to-end lithology prediction process is formed, and the Bayesian optimization is combined in the training stage to automatically adjust the hyperparameters, which significantly improves the prediction accuracy and robustness of the model under small sample scenarios. The scheme can provide more reliable lithology identification results for subsequent reservoir geological modeling and resource evaluation.

[0027] Compared with the traditional method of directly predicting lithology on the original logging data, the application generates and supplements small sample lithology data by introducing a diffusion model, and realizes adaptive adjustment of hyperparameters by combining Bayesian optimization, which not only effectively improves the quality and diversity of training data, but also enhances the adaptability of the model structure to complex inputs. The method not only maintains the integrity of the data features, but also significantly improves the accuracy of lithology prediction, showing stronger robustness and generalization ability.

[0028] To achieve the above purpose, as Figure 1 shown, the technical scheme of the application includes the following steps: S10, based on field core observation and thin section analysis, determine the lithology information of each well, and combine the logging data and depth matching relationship to construct the original lithology training data set.

[0029] S20, for the original data set of lithology category exists long tail distribution, using diffusion model to generate supplement for the sample insufficient category, improve data balance.

[0030] S3, in the supplemented training data set, with logging data as input, lithology as output, using the CNN-Transformer hybrid model fused with Bayesian optimization for training, and introducing Focal Loss to further improve the recognition ability of minority class.

[0031] S40, input the logging data to be identified into the trained classification model, output the corresponding lithology prediction result and verify its accuracy.

[0032] Specifically, in S10, first, through macroscopic observation and thin section analysis of the field core, combined with the judgment of geology experts on the lithology characteristics of each observation well, the lithology types of multiple typical observation wells are determined as the logging data labeling basis. On this basis, the multi-channel logging curve data of each observation well is collected, and the determined lithology type is matched point by point according to the drilling depth, and the original training data set containing complete logging characteristics and accurate lithology label is constructed, ensuring that the sample source is real and reliable, and the spatial distribution is reasonable. Further, the Pearson correlation coefficient of logging characteristics is calculated, the parameters highly related to lithology prediction are selected, the redundant or weakly related variables are removed, and the number of samples of each lithology category is counted to clearly identify the minority class samples that need to be supplemented, providing the basis for subsequent data generation and model training. Figure 2 The number distribution diagram of the original lithology data categories.

[0033] In S20, first, for the convenience of subsequent neural network processing, the text type lithology label is converted into recognizable numerical type identification using Label Encoder, and the IQR method is used for outlier removal on the input feature column to ensure the stability and quality of the training data. At the same time, different features are normalized to map the data to the range of [0, 1], reducing the influence of dimension difference.

[0034] In terms of feature space representation, the normalized 9-dimensional logging features are trained by constructing an autoencoder (Autoencoder) neural network. The encoder part compresses and maps the input features to 64-dimensional latent vectors to preserve the main feature information and reduce the difficulty of generation modeling. The decoder part is used to subsequently reverse the generated latent vectors to the original feature space to ensure the explainability and usability of the generated data. The mean square error between the input and output is used as the loss function for autoencoder training, and the Adam optimizer is selected, combined with the exponential moving average (EMA) to improve the stability of the model.

[0035] To further generate the noisy latent vector for the conditional diffusion model learning, a forward diffusion noise adding strategy based on cosine scheduling is adopted, the total diffusion step number T = 2000 is set, and the cumulative signal preservation coefficient The noise is gradually added to the latent vector to generate different degrees of samples as the training input of the denoiser network. The conditional diffusion denoising model is based on the Transformer encoder structure, the input includes a 64-dimensional noisy latent vector, the current diffusion step t and the category label c, the accurate estimation of the noise term is realized through the multi-head self-attention mechanism combined with the position, time and category embedding, and the loss function is the mean square error of the predicted noise and the real noise.

[0036] In the generation phase, the DDIM inverse sampling algorithm is used to start from a random noise vector, combine the trained denoising model and the forward scheduling coefficient, recursively calculate and , realize the step-by-step denoising of the latent vector until a new latent sample conforming to the minority class distribution is generated. The generated latent vector is restored to the normalized feature space data through the Decoder module, and then the maximum and minimum consistent with the initial training are used for reverse normalization processing to restore the generated data to the actual physical dimension, ensuring that it can be used for subsequent lithology identification and geological analysis. Figure 3 The model overall architecture is Figure 4 part of the lithology training result graph.

[0037] To verify the effectiveness of the generated samples, the Wasserstein distance and KS test value of the real samples and the generated samples in each feature dimension are calculated, and the distribution comparison graph, PCA or t-SNE dimension reduction visualization result, and the accuracy and F1 value comparison of the classifier before and after enhancement are output to form a complete verification report to prove the effectiveness and applicability of the proposed method in minority class sample identification.

[0038] In S30, in further embodiments of the application, for the balanced lithology prediction training data, first, data preprocessing and sample selection are performed, that is, the logging curve data is taken as the model input, the corresponding lithology label is taken as the output, and the input features are normalized to speed up the subsequent network convergence and reduce the influence of the dimension difference of different logging parameters. On this basis, a joint model structure combining convolutional neural network (CNN) and Transformer is designed, such as Figure 5As shown, the CNN and the Transformer complement each other in local feature extraction and global dependency modeling. Specifically, the well logging feature vector is regarded as a one-dimensional sequence, which is first input into a multi-layer one-dimensional convolution module for local pattern capture. The convolution layers are sequentially stacked with batch normalization and ReLU activation functions, and the number of feature channels is gradually increased to match the input dimension of the Transformer encoder. Then, the convolution module output is converted into a tensor that meets the Transformer input requirements and is input into the Transformer encoder implemented based on PyTorch. The number of multi-head attention, the number of encoder layers, the dimension of the feedforward network, and the Dropout parameter are set to effectively capture the long-distance dependency relationship in the well logging data through the multi-head self-attention mechanism. The output of the Transformer encoder is then passed through global average pooling and a fully connected layer to output the lithology class probability, ensuring that both local mutations and global structure information in the well logging data are recognized to improve the accuracy and generalization ability of lithology classification.

[0039] In the model training phase, a hyperparameter search method based on Bayesian optimization is introduced. To further enhance the recognition ability of the minority class of lithology, the Focal Loss loss function is introduced in the training combined with the class weight to alleviate the influence of class imbalance and effectively improve the sensitivity of the model to the rare lithology class and the overall recognition effect. Figure 6 The training loss and accuracy are Figure 7 The confusion matrix is

[0040] In S40, to verify the effectiveness and advancement of the small sample recognition method for complex carbonate lithology in the present application, a comparative experiment is designed based on the process shown in Figure 8 The performance of the present application and the traditional lithology prediction process is evaluated. Specifically, first, the original lithology training data without amplification is directly input into the trained neural network to obtain the prediction result under the traditional scheme; then, the step S20 of the present application is used to generate the diffusion model for directed synthesis and supplement of the minority class samples, to construct a balanced training data set, and input into the CNN-Transformer joint network in step S30 for lithology prediction to obtain the prediction result under the method of the present application. By comparing the prediction accuracy, F1 value and other indicators of the two schemes, the significant advantages of the present application in alleviating class imbalance and improving prediction accuracy are verified.

[0041] By comparing the lithology prediction results obtained by the comparative method and the method of the present application, it is further proved that the proposed lithology classification method based on conditional diffusion generation combined with CNN-Transformer network can significantly improve the identification accuracy of minority classes and the overall prediction robustness in carbonate reservoirs with complex lithology combinations and strong heterogeneity. Experimental results show that compared with traditional methods, the present application can achieve more accurate discrimination of different lithology classes under complex geological conditions, has high practical application value and promotion potential, and can provide reliable data support for geological modeling and development plan of carbonate reservoirs.

[0042] Compared with the prior art, the present application has the following remarkable beneficial effects: The conditional diffusion generation model and the CNN-Transformer lithology prediction network are innovatively integrated, effectively improving the prediction accuracy and stability under the conditions of insufficient data samples, class imbalance and complex local feature details of complex carbonate lithology. The process includes two core neural network modules, wherein the diffusion model is used for sample generation and quality enhancement of the original lithology training data, can perform directional synthesis and supplement for lithology classes with fewer samples, significantly expand the size of the original data set, while retaining the diversity and real distribution characteristics of lithology features, overcoming the problems of pattern collapse and insufficient diversity of generated samples that may be caused by traditional generative adversarial networks (GAN).

[0043] Based on the generated and expanded training data, a lithology prediction network combining convolutional neural network and Transformer encoder is used to learn and predict the supplemented training samples, fully utilizing the advantages of convolutional structure in capturing local mutation features of well logging curves, and the ability of Transformer in long distance dependence modeling and global context information integration, significantly improving the recognizability of lithology details in complex carbonate reservoirs. Field real lithology data experimental results show that compared with traditional methods without sample generation, the data generated by the conditional diffusion model proposed in the present application can effectively improve the identification accuracy of small sample lithology classes, and the prediction results have higher matching degree with the actual lithology, especially in complex areas with subtle lithology boundaries and strong heterogeneity, which can more accurately reflect small-scale lithology changes and lithology transition zones.

[0044] In addition, the experiment comparison shows that the diffusion generation and CNN-Transformer combined method proposed in the application has more accurate prediction results, clearer boundaries, better representation and separation of local features and global correlation, and greatly improves the reliability and accuracy of complex carbonate reservoir geological modeling and resource potential analysis when dealing with low signal-to-noise ratio and more missing sample logging data compared with traditional GAN or single model. Therefore, the application has significant technical advantages and promotion value in small sample amplification of lithology, class balancing, and identification and modeling of complex carbonate reservoirs.

[0045] Those skilled in the art will readily understand that the above description is only the preferred embodiment of the present application, and is not intended to limit the present application, and any modifications, equivalent replacements and improvements made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A complex carbonate lithology identification method based on a diffusion model, characterized in that, The method comprises the following steps: S10, determining the lithology categories of a plurality of observation wells based on field core observation and thin section analysis, and synchronously acquiring logging data of the corresponding observation wells, and then combining the depth correspondence relationship between the lithology categories and the logging data to construct a lithology training data set; S20, for the long-tail distribution of the lithology categories in the lithology training data set, a diffusion model is used to generate and supplement the lithology categories with insufficient samples; S30, on the supplemented lithology training data set, using logging data as input features and lithology categories as output labels, a convolutional neural network combined with a Bayesian optimization algorithm and a Transformer hybrid model are used to train a lithology identification model; S40, inputting the logging data to be identified into the trained lithology identification model to output a lithology identification result.

2. The complex carbonate lithology identification method based on diffusion model according to claim 1, characterized in that, Step S10 further comprises: calculating the Pearson correlation coefficients between the features of each logging curve, and analyzing the correlation degree between the logging parameters and the known lithology categories; According to the correlation degree, the logging feature parameters highly correlated with the lithology prediction are selected, and the redundant or weakly correlated variables are removed.

3. The complex carbonate lithology identification method based on diffusion model according to claim 1, characterized in that, Step S20 specifically comprises: using Label Encoder to convert the text type lithology label into a recognizable numerical type identifier, and using IQR method to remove outliers for the input feature column, and simultaneously normalizing different features; constructing an autoencoder neural network to train the normalized logging features, compressing and mapping the input features into latent vectors, and the decoder part is used to inversely restore the generated latent vectors to the original feature space; using a forward diffusion noise adding strategy based on cosine scheduling, gradually adding noise to the latent vector through a cumulative signal retention coefficient, generating different degrees of noise-added latent vectors as the training input of the denoiser network; using the DDIM inverse sampling algorithm, starting from a random noise vector, combining the trained denoiser model and the forward scheduling coefficient, recursively calculating new latent samples that meet the minority class distribution, restoring them to the normalized feature space data through the decoder, and then using the maximum and minimum values consistent with the initial training to perform inverse normalization processing, restoring the generated data to the actual physical dimension.

4. The method of claim 1, wherein the method is a diffusion model based complex carbonate lithology identification method. Step S20 further comprises: calculating the Wasserstein distance and KS test value of the real samples and the generated samples in each feature dimension, and outputting the distribution comparison graph, PCA or t-SNE dimension reduction visualization result, and the accuracy and F1 value comparison of the classifier before and after enhancement, forming a complete verification report.

5. The method for complex carbonate lithology identification based on diffusion model according to claim 1, characterized in that, In step S30, the construction method of the convolutional neural network combined with the Transformer hybrid model comprises: The logging feature vector is regarded as a one-dimensional sequence, which is input into a multi-layer one-dimensional convolution module for local pattern capture, and the convolution layers are sequentially stacked with batch normalization and ReLU activation function, and the number of feature channels is gradually increased to match the input dimension of the Transformer encoder; The convolution module output is converted into a tensor conforming to the requirements of the Transformer input, and is input into a Transformer encoder implemented based on PyTorch, so as to capture long-distance dependencies in the logging data through a multi-head self-attention mechanism; The output of the Transformer encoder is output through a global average pooling and a fully connected layer, and a lithology category probability is output.

6. The method of claim 1, wherein the method is a diffusion model based complex carbonate lithology identification method. In step S30, a model training phase, a hyperparameter search method based on Bayesian optimization is introduced, and a Focal Loss loss function is introduced and combined with a category weight, so as to alleviate the influence of category imbalance.

7. A complex carbonate lithology identification system based on diffusion model, characterized in that, Comprise: A data set construction module is configured to determine the lithology categories of a plurality of observation wells based on field core observation and thin section analysis, and simultaneously acquire logging data of the corresponding observation wells, and then combine the lithology categories and the depth corresponding relationship of the logging data to construct a lithology training data set; A data supplement module is configured to generate and supplement the sample-deficient lithology categories by using a diffusion model in view of the long-tail distribution of the lithology categories in the lithology training data set; A training module is configured to use the logging data as input features and the lithology categories as output labels to train a lithology identification model by using a convolutional neural network combined with Bayesian optimization algorithm and a Transformer hybrid model on the supplemented lithology training data set; An identification module is configured to input logging data to be identified into the trained lithology identification model, and output a lithology identification result.

8. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the program to implement the complex carbonate logging lithology identification method based on the diffusion model according to any one of claims 1-6.

9. A computer readable storage medium having stored thereon computer instructions, wherein, The computer instructions are executed by the processor to implement the complex carbonate logging lithology identification method based on the diffusion model according to any one of claims 1-6.

Images