A Multidimensional Well Logging Lithology Identification Method and Device

Abstract

This invention relates to a multi-dimensional well logging lithology identification method and apparatus. The method includes acquiring current well logging data of a well to be tested; performing sequence stratigraphy on the well to obtain sedimentary sequence stratigraphy results based on the current well logging data; dividing the current well logging data into data segments of different sequences based on the sedimentary sequence stratigraphy results; inputting the data segment of each sequence into a lithology identification model for lithology identification to obtain the lithology probability distribution corresponding to each sequence; and performing geological constraint correction on the lithology probability distribution based on the continuous depth on the well to obtain the final lithology identification result. This invention performs sequence stratigraphy on the well to be tested, thereby enabling lithology identification of data segments of different sequences, obtaining well logging performance in different strata, and combining geological information to perform geological constraint correction on the lithology probability distribution, thus combining geological information with well logging data and improving the lithology identification effect.

Description

Technical Field

[0001] This invention relates to the field of formation lithology identification technology, and in particular to a multidimensional logging lithology identification method and apparatus. Background Technology

[0002] In carbonate reservoirs, geophysical logging data is an effective and irreplaceable method for understanding subsurface stratigraphic structure and lithological characteristics. However, due to the prevalence of transitional lithologies such as calcareous dolomite and dolomitic limestone in carbonate reservoirs, these lithologies exhibit significant multiple ambiguities in their logging responses due to the continuous and gradual changes in mineral composition. Firstly, the physical response ranges of different lithologies highly overlap, with sonic transit time (AC) and density (DEN) curves showing continuous distribution in the transition zone. Secondly, lithological boundaries are difficult to define, with conventional threshold discrimination methods exhibiting a misclassification rate exceeding 35%. Thirdly, transitional lithologies constitute a severely insufficient proportion of the cored samples, resulting in a serious sample scarcity problem. These characteristics pose fundamental technical challenges to traditional lithology identification methods in complex diagenetic zones. With the rapid development of machine learning technology, automated lithology identification methods based on machine learning are gradually becoming mainstream.

[0003] Existing well logging lithology identification methods mainly use optimal framework models to predict the probabilities of various lithologies in well logging data. While this method can quickly classify lithologies and improve the efficiency and accuracy of lithology identification, it does not comprehensively consider the geological information inherent in the well logging data itself. Although significant progress has been made in well logging lithology identification combined with machine learning, its effectiveness is often greatly reduced in areas with complex strata and a wide variety of lithologies. This is because the response characteristics of well logging curves for some different types of lithologies are similar and difficult to distinguish. The information in the well logging cannot be fully integrated through any mathematical method to construct a true subsurface lithology sequence. Furthermore, the well logging performance of the same lithology varies in different strata, resulting in poor lithology identification results.

[0004] Therefore, there is an urgent need to propose a multidimensional logging lithology identification method and device to solve the technical problem that existing logging lithology identification methods do not comprehensively consider the geological information contained in the logging data itself, the underground lithological sequence, and the logging performance in different strata, resulting in poor lithology identification effect. Summary of the Invention

[0005] In view of this, it is necessary to provide a multi-dimensional logging lithology identification method and device to solve the technical problem that existing logging lithology identification methods do not comprehensively consider the geological information contained in the logging data itself, the underground lithology sequence, and the logging performance in different strata, resulting in poor lithology identification effect.

[0006] To address the aforementioned problems, in a first aspect, the present invention provides a multidimensional logging lithology identification method, comprising: Obtain the current logging data of the well to be inspected; Based on the current logging data, the well to be tested is sequence-divided to obtain sedimentary sequence division results; Based on the sedimentary sequence division results, the current logging data is divided into data segments of different sequences; The data segments of each sequence are input into the lithology identification model for lithology identification to obtain the lithology probability distribution corresponding to each sequence. The lithology probability distribution is geologically constrained and corrected based on the continuous depth of the well to be tested, and the final lithology identification result is obtained.

[0007] In one possible implementation, the step of performing geological constraint correction on the lithology probability distribution based on continuous depths on the well to be tested to obtain the final lithology identification result includes: The lithology probability distribution is used to predict the lithology, and the lithology prediction results at continuous depths of the well to be tested are obtained. The lithology prediction results are then subjected to geological constraint correction to obtain the final lithology identification results.

[0008] In one possible implementation, the step of performing geological constraint correction on the lithology prediction results to obtain the final lithology identification results includes: The influencing factors are identified, and the lithological transformation rules within each sequence are established based on these factors. The influencing factors include changes in sedimentary environment and tectonics. The lithological transformation rules are the rules governing the transition between transitional lithologies and end-member lithologies. Based on the lithological transformation law, the lithological prediction results are corrected by geological constraints to obtain the final lithological identification results.

[0009] In one possible implementation, the method further includes: Construct a hierarchical partitioning model; The current logging data is input into the sequence stratigraphy model to perform sequence stratigraphy on the well to be tested, and the sedimentary sequence stratigraphy result is obtained.

[0010] In one possible implementation, the training process of the hierarchical partitioning model includes: Obtain raw logging data from multiple wells and corresponding core / cuttings logging lithology, lithological composition identified by microscopic thin sections, and sequence stratigraphy schemes; The lithology of the core / cutting logging and the lithological components are combined to obtain lithological labels, and the sequence stratigraphy scheme is converted into sequence labels; Response feature patterns are extracted based on the lithology label, the sequence label, and the original logging data to obtain the sequence logging response features; The preset sequence partitioning model is trained based on the sequence logging response characteristics to obtain the trained sequence partitioning model.

[0011] In one possible implementation, the step of extracting response feature patterns based on the lithology label, the sequence label, and the original logging data to obtain sequence logging response features includes: The raw logging data is subjected to quality control, depth matching and standardization to obtain preprocessed raw logging data; Based on the depth connection of the multiple wells, the preprocessed raw logging data is curve-fitted to obtain the first logging curve; The first logging curve is mapped one-to-one with the lithology label and the sequence label in terms of depth to obtain a structured dataset; The response feature patterns of the first logging curves within each sequence are extracted according to the sequence division scheme to obtain the sequence logging response features.

[0012] In one possible implementation, the training process of the lithology identification model includes: Based on the depth connection, curve fitting is performed on the original logging data to obtain a second logging curve; The second logging curve is subjected to multi-scale wavelet decomposition to obtain multi-scale features; the multi-scale features include approximation coefficients and detail coefficients of different frequency bands; Based on the NEAT algorithm, the multi-scale features and the original logging data are curve standardized to obtain the third logging curve; The third logging curve is divided into training set, validation set and test set according to a fixed window length by sliding window segmentation and label assignment; The dung beetle optimization algorithm is used to automatically optimize the key hyperparameters of the lithology identification model to obtain the optimal hyperparameter combination; After initializing the lithology identification model using the optimal hyperparameter combination, the lithology identification model is trained using the training set, the validation set, and the test set to obtain the trained lithology identification model.

[0013] In one possible implementation, the lithology identification model is an encoder model built on the Transformer-KAN architecture. The encoder model includes a one-dimensional convolutional layer, a learnable positional encoding module, multiple cascaded Transformer-KAN encoding layers, a global average pooling layer, and a fully connected classification layer. Each Transformer-KAN encoding layer contains a KAN multi-head attention module that uses a radial basis function kernel to calculate the nonlinear similarity between the query and the key, and a feedforward neural network. Each sub-layer is followed by residual connections and layer normalization.

[0014] In one possible implementation, the step of automatically optimizing the key hyperparameters of the lithology identification model using the dung beetle optimization algorithm to obtain the optimal hyperparameter combination includes: The search space is defined by including the number of encoder layers, the number of attention heads, the dimension of hidden layers, the size of the KAN grid, the learning rate, and the batch size; Initialize the population; using the validation set weighted score as the fitness function, iteratively execute rolling ball, reproduction, foraging, and stealing behaviors to update candidate solutions for parameters, and output the optimal hyperparameter combination.

[0015] Secondly, the present invention also provides a multi-dimensional logging lithology identification device, comprising: The data acquisition module is used to acquire the current logging data of the well to be inspected; The sequence division module is used to perform sequence division on the well to be tested based on the current logging data, and obtain the sedimentary sequence division result; The data segmentation module is used to divide the current well logging data into data segments of different sequences based on the sedimentary sequence division results; The lithology identification module is used to input the data segments of each sequence into the lithology identification model for lithology identification, and obtain the lithology probability distribution corresponding to each sequence. The geological constraint module is used to perform geological constraint correction on the lithology probability distribution based on the continuous depth on the well to be tested, so as to obtain the final lithology identification result.

[0016] The beneficial effects of this invention are as follows: It acquires the current logging data of the well to be tested; it performs sequence stratigraphy on the well based on the current logging data to obtain sedimentary sequence stratigraphy results; based on the sedimentary sequence stratigraphy results, it divides the current logging data into data segments of different sequences; it inputs the data segment of each sequence into a lithology identification model for lithology identification, obtaining the lithology probability distribution corresponding to each sequence; and it performs geological constraint correction on the lithology probability distribution based on the continuous depth on the well to obtain the final lithology identification result. This invention performs sequence stratigraphy on the well to be tested, thereby enabling lithology identification of data segments of different sequences, obtaining logging performance in different strata, and combining geological information to perform geological constraint correction on the lithology probability distribution. This allows for the integration of geological information with logging data, improving the lithology identification effect. Attached Figure Description

[0017] Figure 1 This is a schematic flowchart of an embodiment of the multidimensional logging lithology identification method provided by the present invention; Figure 2 A schematic flowchart of an embodiment for determining sequence logging response characteristics provided by the present invention; Figure 3 A schematic flowchart illustrating an embodiment of the training process of the sex recognition model provided by the present invention; Figure 4 This is a schematic diagram of an embodiment of the multidimensional logging lithology identification device provided by the present invention. Detailed Implementation

[0018] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0019] like Figure 1 As shown in the figure, a specific embodiment of the present invention discloses a multi-dimensional logging lithology identification method, comprising: S101. Obtain the current logging data of the well to be tested.

[0020] This invention can be applied to core aspects of oil and gas exploration and development, specifically serving reservoir evaluation and geological modeling. Its application scenarios span multiple stages from regional exploration to oilfield development, and it holds irreplaceable value, particularly in reservoirs with strong heterogeneity and difficult lithological identification, such as carbonate rocks and complex clastic rocks. Specifically, it can read raw logging data from the testing files of the well to be tested. This raw logging data includes, but is not limited to, natural gamma ray (GR), resistivity (RT / RXO), bulk density (DEN), neutron porosity (CNL), sonic transit time (AC), and spontaneous potential (SP). It can also perform quality control processes on the current logging data, including removing anomalous jump values, filling in missing segments (e.g., using adjacent depth interpolation), and Z-score standardizing all curves to eliminate dimensional differences and baseline shifts caused by different logging series or instruments.

[0021] S102. Based on the current logging data, perform sequence division of the well to be tested to obtain the sedimentary sequence division results.

[0022] The embodiments of this invention involve a key step in introducing geological constraints. Specifically, preprocessed well logging data is used as input to perform sequence stratigraphy on the well to be tested, resulting in sedimentary sequence stratigraphy results.

[0023] S103. Based on the sedimentary sequence division results, the current logging data is divided into data segments of different sequences.

[0024] This invention can divide current well logging data into data segments of different sequences based on the sequence boundary depths in the sedimentary sequence delineation results. Well logging curves within each data segment share similar sedimentary backgrounds and environmental energies, providing a stable "sedimentary environmental context" for subsequent lithological identification. For example, all data points belonging to the "open platform facies highstand systems tract" can be grouped into a single processing unit.

[0025] S104. Input the data segment of each sequence into the lithology identification model to perform lithology identification and obtain the lithology probability distribution corresponding to each sequence.

[0026] The embodiments of the present invention can construct a lithology identification model. The input of the lithology identification model can be a data segment for each sequence, and the output can be the lithology probability distribution corresponding to each sequence. The data segment is input into the lithology identification model to perform lithology identification and obtain the corresponding lithology probability distribution.

[0027] S105. Based on the continuous depth on the well to be tested, the lithology probability distribution is corrected by geological constraints to obtain the final lithology identification result.

[0028] The embodiments of the present invention can synthesize the lithology probability distribution by continuous depth on the well to be detected, and then perform geological constraint correction to obtain the final lithology identification result.

[0029] Compared with existing technologies, this embodiment provides the following: acquiring the current logging data of the well to be tested; performing sequence stratigraphy on the well based on the current logging data to obtain sedimentary sequence stratigraphy results; dividing the current logging data into data segments of different sequences based on the sedimentary sequence stratigraphy results; inputting the data segment of each sequence into a lithology identification model for lithology identification to obtain the lithology probability distribution corresponding to each sequence; and performing geological constraint correction on the lithology probability distribution based on the continuous depth on the well to obtain the final lithology identification result. This invention performs sequence stratigraphy on the well to be tested, thereby enabling lithology identification of data segments of different sequences, obtaining logging performance in different strata, and combining geological information to perform geological constraint correction on the lithology probability distribution. This allows for the integration of geological information and logging data, improving the lithology identification effect.

[0030] In some embodiments of the present invention, the method further includes: Construct a hierarchical partitioning model; The current logging data is input into the sequence stratigraphy model to perform sequence stratigraphy on the well to be tested, and the sedimentary sequence stratigraphy results are obtained.

[0031] In this embodiment of the invention, preprocessed well logging data is used as input to load and invoke a pre-trained sequence partitioning model. This model is a machine learning model based on random forest, which performs sequence partitioning on a single well. Thus, the current well logging data can be input into the sequence partitioning model to perform sequence partitioning on the well to be tested, obtaining the sedimentary sequence partitioning results.

[0032] In some embodiments of the present invention, the training process of the hierarchical partitioning model includes: Obtain raw logging data from multiple wells and corresponding core / cuttings logging lithology, lithological composition identified by microscopic thin sections, and sequence stratigraphy schemes.

[0033] This invention can collect raw logging data from multiple wells in the target work area, including natural gamma (GR), resistivity (RT / RXO), bulk density (DEN), neutron porosity (CNL), sonic transit time (AC), and spontaneous potential (SP). Simultaneously, "hard data" serving as monitoring labels is collected: lithology from core / cuttings logging, precise lithological composition identified by microscopic thin sections, and sequence stratigraphy schemes completed by sedimentologists (including sequence boundaries and systems tract labels).

[0034] The lithology and lithological components of core / cutting logging are combined to obtain lithological labels, and the sequence stratigraphy scheme is converted into sequence labels.

[0035] The core / cuttings logging lithology in the embodiments of the present invention can represent most of the lithology, and the lithological components identified by microscopic thin sections can represent more precise lithology. Combining the core / cuttings logging lithology and lithological components can yield complete and more precise lithological labels. To facilitate subsequent processing, the sequence stratigraphy scheme can be converted into sequence labels.

[0036] Response feature patterns are extracted based on lithology labels, sequence labels, and raw logging data to obtain sequence logging response features.

[0037] The embodiments of the present invention can process the original well logging data and then extract the response feature patterns of the original well logging data through lithology tags and sequence tags to obtain sequence logging response features.

[0038] In some embodiments of the present invention, such as Figure 2 As shown, response feature patterns are extracted based on lithological labels, sequence labels, and raw well logging data to obtain sequence logging response features, including: S201. Perform quality control, depth matching, and standardization on the raw logging data to obtain preprocessed raw logging data.

[0039] The embodiments of the present invention can perform quality control, depth matching, and standardization processing on the raw logging data. Quality control may include outlier detection and removal, depth matching may include key layer comparison and correlation analysis, depth correction and resampling, and standardization processing may include mean-variance method. The specific processing procedures can be selected and operated according to the actual situation, and the embodiments of the present invention do not limit them.

[0040] S202. Based on the depth connection of multiple wells, the preprocessed raw logging data is curve fitted to obtain the first logging curve.

[0041] The embodiments of the present invention can determine the depth connection of multiple wells, thereby performing curve fitting on the preprocessed raw logging data based on the depth connection, and thus obtaining the fitted first logging curve.

[0042] S203. The first logging curve is matched one-to-one with the lithology label and sequence label in depth to obtain a structured dataset.

[0043] The embodiments of the present invention can map the first logging curves with lithology labels and sequence labels at depth for the depth of multiple wells, thereby generating a structured dataset.

[0044] S204. Extract the response feature patterns of the first logging curve within each sequence according to the sequence division scheme to obtain the sequence logging response features.

[0045] This invention can analyze and extract the response characteristic patterns of well logging curves within each sequence based on an expert-defined sequence division scheme, thereby obtaining the sequence logging response characteristics. The response characteristic patterns are shown in Table 1. Table 1. Response Feature Patterns

[0046] S205. Train the preset sequence partitioning model based on the sequence logging response characteristics to obtain the trained sequence partitioning model.

[0047] This invention trains a pre-defined sequence stratigraphy model based on sequence logging response characteristics. This integrates traditional geological constraints into the prediction process, establishing a stable "sedimentary environment background" for subsequent lithology identification models. It provides upper and lower boundaries for fine-grained lithology identification across different sedimentary environments within the same well, avoiding baseline shifts in logging responses caused by differences in sedimentary environments across different sequences. For example, a transgressive systems tract might exhibit an upward decreasing GR, while a highstand systems tract might show a stable low GR. Without differentiation, the model might mistakenly identify different logging responses of the same lithology (e.g., limestone) in different sequences as different lithologies. Therefore, performing sequence prediction first provides a "sedimentary environment" context for subsequent lithology identification.

[0048] In some embodiments of the present invention, such as Figure 3 As shown, the training process of the lithology identification model includes: S301. Based on depth connectivity, curve fitting is performed on the original logging data to obtain the second logging curve.

[0049] In this embodiment of the invention, curve fitting can be performed on the original logging data based on the depth connection of multiple wells to obtain a second logging curve.

[0050] S302. Perform multi-scale wavelet decomposition on the second logging curve to obtain multi-scale features; the multi-scale features include approximation coefficients and detail coefficients of different frequency bands.

[0051] In this embodiment of the invention, wavelet transform and NEAT algorithm can be combined for well logging curve processing. Specifically, the second well logging curve can be first decomposed into multi-scale wavelet transform, and the approximation coefficients and detail coefficients of different frequency bands can be extracted as features to obtain multi-scale features.

[0052] S303. Based on the NEAT algorithm, curve standardization is performed on multi-scale features and raw logging data to obtain the third logging curve.

[0053] In this embodiment of the invention, these multi-scale features, along with the original well logging data, constitute the input layer of the NEAT neural network, enabling the network to simultaneously perceive both the macroscopic trend and microscopic details of the curve. The NEAT algorithm, through its unique topological evolution mechanism, automatically searches for and optimizes a deep neural network structure. This network learns how to fuse and weight these multi-scale features, thereby achieving intelligent curve denoising, reconstruction, or anomaly identification. The adaptive function is designed according to the processing objective: for denoising tasks, it minimizes the error between the network output and the expert-labeled "clean" standard curve; for feature enhancement tasks, it maximizes the salience of the target geological response in the output. Through NEAT's species formation and history labeling mechanisms, the network can explore various feature fusion strategies and nonlinear transformation paths, ultimately evolving a highly adaptive intelligent processing model that considers both global background and local details. This intelligent processing model can output a third well logging curve after curve standardization, significantly improving the accuracy and geological interpretability of well logging curve processing.

[0054] S304. Perform sliding window cutting and label assignment on the third logging curve according to a fixed window length, and divide it into training set, validation set and test set according to well.

[0055] In this embodiment of the invention, the standardized continuous curve is cut into sliding windows of a fixed length (e.g., 128 sampling points), with a 50% overlap between windows. Each window is labeled according to the majority of lithology categories. Finally, the data is divided into training, validation, and test sets (7:2:1 ratio) by well, and converted into a tensor format acceptable to the model, i.e., the input shape is [batch size, window length, number of features], and the label is the unique thermal encoding of the corresponding lithology category.

[0056] S305. The dung beetle optimization algorithm is used to automatically optimize the key hyperparameters of the lithology identification model and obtain the optimal hyperparameter combination.

[0057] In order to improve model performance, this embodiment of the invention employs the Dung Beetle Optimization (DBO) algorithm to automatically optimize key hyperparameters and obtain the optimal combination of hyperparameters.

[0058] S306. After initializing the lithology identification model with the optimal hyperparameter combination, the lithology identification model is trained according to the training set, validation set and test set to obtain the trained lithology identification model.

[0059] The lithology identification model in this embodiment of the invention adopts an encoder structure. First, the original well logging features are embedded into a high-dimensional space through a one-dimensional convolutional layer and a learnable positional encoding is added. The lithology identification model is an encoder model built based on the Transformer-KAN architecture. The encoder model includes a one-dimensional convolutional layer, a learnable positional encoding module, 2-4 cascaded Transformer-KAN encoding layers, a global average pooling layer, and a fully connected classification layer. Each Transformer-KAN encoding layer contains a KAN multi-head attention module that uses a radial basis function kernel to calculate the nonlinear similarity between the query and the key (using a radial basis function kernel to calculate the nonlinear similarity between the query and the key) and a feedforward neural network. Each sub-layer is followed by residual connections and layer normalization to stabilize training. The encoder output is fed into the fully connected classification layer after global average pooling. Finally, the probability distribution of each lithology category is output through the Softmax function.

[0060] In some embodiments of the present invention, step S305 includes: The search space is defined by including the number of encoder layers, the number of attention heads, the dimension of hidden layers, the size of the KAN grid, the learning rate, and the batch size; Initialize the population; using the validation set weighted score as the fitness function, iteratively execute rolling ball, reproduction, foraging, and stealing behaviors to update candidate solutions for parameters, and output the optimal hyperparameter combination.

[0061] This invention employs the Dung Beetle Optimization (DBO) algorithm to automatically optimize key hyperparameters: a search space is defined including the number of encoder layers, the number of attention heads, the dimension of hidden layers, the size of the KAN grid, the learning rate, and the batch size; a population of 30 individuals is initialized; the fitness function is obtained by iteratively executing four behaviors—rolling, breeding, foraging, and stealing—to update candidate solutions for parameters, iterating a total of 200 times or until the fitness converges; finally, the optimal combination of hyperparameters is output and the model is reinitialized accordingly.

[0062] Further training of the model using optimal hyperparameter combinations may include using optimal hyperparameter configurations, with cross-entropy loss as the optimization objective, AdamW as the optimizer, and cosine annealing learning rate scheduling to train the model completely, incorporating an early stopping strategy to prevent overfitting; evaluating performance on the validation set and saving the best weights after each cycle; calculating the overall accuracy, precision / recall / F1 score for various lithologies on the independent test set after training, plotting the confusion matrix, and visualizing the comparison between the predicted and actual lithology profiles to comprehensively evaluate model performance. The trained model can then be integrated into the prediction system, and then step S101 can be performed to detect the current logging data of the well to be tested.

[0063] Finally, the validation well dataset was uploaded as input to the algorithm models before and after optimization, and the accuracy and speed of lithology identification and localization were compared before and after model optimization. As shown in Table 2: Table 2. Experimental Results of Model Performance Comparison

[0064] As shown in Table 2, the optimized model significantly improved its accuracy after training, with an average accuracy of over 89%, representing a 10.4% increase in accuracy and a doubling of the speed. This indicates that the geologically constrained DBO-transformer-KAN model is suitable for lithological identification of carbonate reservoirs.

[0065] In some embodiments of the present invention, S105 includes: Lithology probability distribution is used to predict lithology, and lithology prediction results are obtained at continuous depths of the well to be tested. Geological constraint corrections were applied to the lithology prediction results to obtain the final lithology identification results.

[0066] This invention can generate lithology prediction results at continuous depths through sliding window voting or probability-weighted averaging, and output lithology profile maps and confidence curves. The system supports batch processing of multi-well data, result visualization, and export in LAS / CSV format, and also provides a KAN attention weight visualization tool to enhance model interpretability. Then, geological constraint correction is applied to the lithology prediction results to obtain the final lithology identification results.

[0067] In some embodiments of the present invention, geological constraint correction is performed on the lithology prediction results to obtain the final lithology identification results, including: Identify the influencing factors and establish the lithological transformation rules within each sequence based on these factors; the influencing factors include changes in sedimentary environment and tectonics; the lithological transformation rules are the rules governing the transition between transitional lithologies and end-member lithologies; Based on the lithological transformation law, the lithological prediction results are corrected by geological constraints to obtain the final lithological identification results.

[0068] The embodiments of this invention define the lithological transformation rules within each sequence based on factors such as changes in the sedimentary environment and tectonics in the study area, as shown in Table 3. Table 3. Definition of Geophysical Constraint Rules

[0069] This step provides the correct geological constraints for subsequent sequence capture. It can effectively solve the problem that a single well cannot contain all sequence lithologies, enabling the model to effectively capture such sequence transformation relationships in different sequences of different wells. For example, "there are often transitional lithologies (such as dolomitic limestone and calcareous dolomite) between limestone and dolomite" and "at the bottom of the transgressive systems tract, the lithology may suddenly change from coarse to fine", and a series of fixed transformation conditions, outputting the final lithology identification results.

[0070] The advantages of the embodiments of the present invention are as follows: 1. High prediction accuracy and good geological fit: Through a cascaded prediction framework of "sequence stratigraphy first, lithology second," the model first grasps the macroscopic sedimentary environment, avoiding misjudgments such as "seeing the trees but not the forest." Under the constraint of geological regularization, the model's prediction results are no longer simple data fitting, but rational inferences consistent with geological common sense. The final lithological identification profile has a higher degree of agreement with actual core data and the understanding of geological experts.

[0071] 2. KAN's exceptional interpretability breaks down the "black box" barrier in AI, a revolutionary advantage brought about by its introduction. Traditional deep learning models are like a "black box," providing only the result, not the reason. KAN, however, can visualize the quantitative relationship between each logging curve and the final lithology determination (e.g., showing how density values ​​influence the probability of identifying dolomite through a specific function). Geologists and engineers can understand the model's "thinking process," thus trusting its predictions. This may reveal previously unknown hidden relationships between certain logging curves and specific lithologies, driving the development of logging interpretation theory itself.

[0072] 3. Because the model learns general sedimentary patterns and physical relationships (through KAN's interpretable functions and regularization constraints), rather than merely memorizing data characteristics from specific wells, it maintains robust performance when applied to carbonate reservoirs in different regions and geological eras. This significantly reduces the model's dependence on large amounts of labeled data, making it highly valuable for exploration in new areas and regions lacking core data, thus solving the pain point of traditional machine learning models that "fail when changing wells."

[0073] 4. This integrates sequence stratigraphy and lithological identification—two tasks highly dependent on expert experience—into an automated workflow. It frees geologists from arduous manual comparison and interpretation work, reducing tasks from days or even weeks to hours, significantly improving the efficiency of well logging interpretation and supporting rapid evaluation and decision-making.

[0074] 5. This invention does not replace geologists with AI, but rather embeds geologists' knowledge (sequence stratigraphy schemes, lithological transformation laws) as a core component into the model. This is a "teaching a man to fish" approach, allowing AI to learn the thinking methods of geologists. This makes the model a powerful "AI geologist assistant," capable of handling massive amounts of data while adhering to geological laws, achieving an optimal combination of human expert wisdom and computer computing power.

[0075] 6. The advantages of this invention lie in the deep integration of the DBO-transformer-KAN innovative architecture, the sequence constraint framework, and physical regularization, rather than the simple application of a single technology. This multi-technology integrated solution is difficult to replicate simply. In fields such as oil and gas exploration and development, and geological scientific research, it can provide solutions with higher accuracy, more reasonable interpretation, and stronger adaptability, forming a solid technological barrier and core competitiveness.

[0076] To better implement the multidimensional logging lithology identification method in the embodiments of the present invention, based on the multidimensional logging lithology identification method, the embodiments of the present invention also provide a multidimensional logging lithology identification device, such as... Figure 4 As shown, the multidimensional logging lithology identification device 400 includes: The data acquisition module 401 is used to acquire the current logging data of the well to be inspected; Sequence stratigraphy module 402 is used to perform sequence stratigraphy on the well to be inspected based on the current logging data, and obtain sedimentary sequence stratigraphy results; The data segmentation module 403 is used to divide the current logging data into data segments of different sequences based on the sedimentary sequence division results; The lithology identification module 404 is used to input the data segment of each sequence into the lithology identification model for lithology identification, and obtain the lithology probability distribution corresponding to each sequence. The geological constraint module 405 is used to perform geological constraint correction on the lithology probability distribution based on the continuous depth on the well to be tested, so as to obtain the final lithology identification result.

[0077] The multidimensional logging lithology identification device 400 provided in the above embodiments can realize the technical solutions described in the above multidimensional logging lithology identification method embodiments. The specific implementation principles of each module or unit can be found in the corresponding content of the above multidimensional logging lithology identification method embodiments, which will not be repeated here.

[0078] The multidimensional logging lithology identification method and device provided by the present invention have been described in detail above. Specific examples have been used to illustrate the principle and implementation of the present invention. The description of the above embodiments is only for the purpose of helping to understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation and application scope based on the idea of ​​the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A multidimensional logging lithology identification method, characterized in that, include: Obtain the current logging data of the well to be inspected; Based on the current logging data, the well to be tested is sequence-divided to obtain sedimentary sequence division results; Based on the sedimentary sequence division results, the current logging data is divided into data segments of different sequences; The data segments of each sequence are input into the lithology identification model for lithology identification to obtain the lithology probability distribution corresponding to each sequence. The lithology probability distribution is geologically constrained and corrected based on the continuous depth of the well to be tested, and the final lithology identification result is obtained.

2. The multidimensional logging lithology identification method according to claim 1, characterized in that, The step of performing geological constraint correction on the lithology probability distribution based on the continuous depth on the well to be tested, to obtain the final lithology identification result, includes: The lithology probability distribution is used to predict the lithology, and the lithology prediction results at continuous depths of the well to be tested are obtained. The lithology prediction results are then subjected to geological constraint correction to obtain the final lithology identification results.

3. The multidimensional logging lithology identification method according to claim 1, characterized in that, The geological constraint correction of the lithology prediction results to obtain the final lithology identification results includes: The influencing factors are identified, and the lithological transformation rules within each sequence are established based on these factors. The influencing factors include changes in sedimentary environment and tectonics. The lithological transformation rules are the rules governing the transition between transitional lithologies and end-member lithologies. Based on the lithological transformation law, the lithological prediction results are corrected by geological constraints to obtain the final lithological identification results.

4. The multidimensional logging lithology identification method according to claim 1, characterized in that, The method further includes: Construct a hierarchical partitioning model; The current logging data is input into the sequence stratigraphy model to perform sequence stratigraphy on the well to be tested, and the sedimentary sequence stratigraphy result is obtained.

5. The multidimensional logging lithology identification method according to claim 4, characterized in that, The training process of the hierarchical partitioning model includes: Obtain raw logging data from multiple wells and corresponding core / cuttings logging lithology, lithological composition identified by microscopic thin sections, and sequence stratigraphy schemes; The lithology of the core / cutting logging and the lithological components are combined to obtain lithological labels, and the sequence stratigraphy scheme is converted into sequence labels; Response feature patterns are extracted based on the lithology label, the sequence label, and the original logging data to obtain the sequence logging response features; The preset sequence partitioning model is trained based on the sequence logging response characteristics to obtain the trained sequence partitioning model.

6. The multidimensional logging lithology identification method according to claim 1, characterized in that, The step of extracting response feature patterns based on the lithology label, the sequence label, and the original logging data to obtain sequence logging response features includes: The raw logging data is subjected to quality control, depth matching and standardization to obtain preprocessed raw logging data; Based on the depth connection of the multiple wells, the preprocessed raw logging data is curve-fitted to obtain the first logging curve; The first logging curve is mapped one-to-one with the lithology label and the sequence label in terms of depth to obtain a structured dataset; The response feature patterns of the first logging curves within each sequence are extracted according to the sequence division scheme to obtain the sequence logging response features.

7. The multidimensional logging lithology identification method according to claim 5, characterized in that, The training process of the lithology identification model includes: Based on the depth connection, curve fitting is performed on the original logging data to obtain a second logging curve; The second logging curve is subjected to multi-scale wavelet decomposition to obtain multi-scale features; the multi-scale features include approximation coefficients and detail coefficients of different frequency bands; Based on the NEAT algorithm, the multi-scale features and the original logging data are curve standardized to obtain the third logging curve; The third logging curve is divided into training set, validation set and test set according to a fixed window length by sliding window segmentation and label assignment; The dung beetle optimization algorithm is used to automatically optimize the key hyperparameters of the lithology identification model to obtain the optimal hyperparameter combination; After initializing the lithology identification model using the optimal hyperparameter combination, the lithology identification model is trained using the training set, the validation set, and the test set to obtain the trained lithology identification model.

8. The multidimensional logging lithology identification method according to claim 1, characterized in that, The lithology identification model is an encoder model built on the Transformer-KAN architecture. The encoder model includes a one-dimensional convolutional layer, a learnable positional encoding module, multiple cascaded Transformer-KAN encoding layers, a global average pooling layer, and a fully connected classification layer. Each Transformer-KAN encoding layer contains a KAN multi-head attention module that uses a radial basis function kernel to calculate the nonlinear similarity between the query and the key, and a feedforward neural network. Each sub-layer is followed by residual connections and layer normalization.

9. The multidimensional logging lithology identification method according to claim 7, characterized in that, The step of using the dung beetle optimization algorithm to automatically optimize the key hyperparameters of the lithology identification model and obtain the optimal hyperparameter combination includes: The search space is defined by including the number of encoder layers, the number of attention heads, the dimension of hidden layers, the size of the KAN grid, the learning rate, and the batch size; Initialize the population; using the validation set weighted score as the fitness function, iteratively execute rolling ball, reproduction, foraging, and stealing behaviors to update candidate solutions for parameters, and output the optimal hyperparameter combination.

10. A multi-dimensional well logging lithology identification device, characterized in that, include: The data acquisition module is used to acquire the current logging data of the well to be inspected; The sequence division module is used to perform sequence division on the well to be tested based on the current logging data, and obtain the sedimentary sequence division result; The data segmentation module is used to divide the current well logging data into data segments of different sequences based on the sedimentary sequence division results; The lithology identification module is used to input the data segments of each sequence into the lithology identification model for lithology identification, and obtain the lithology probability distribution corresponding to each sequence. The geological constraint module is used to perform geological constraint correction on the lithology probability distribution based on the continuous depth on the well to be tested, so as to obtain the final lithology identification result.

Images