Lithology identification model dynamic optimization training and application method based on three-network collaboration and related device

Abstract

The invention discloses a lithology identification model dynamic optimization training and application method based on three-network collaboration and a related device, and relates to the technical field of lithology identification, the method comprises the following steps: during training, obtaining a data set and constructing a dynamic optimization model, the dynamic optimization model comprises a three-network model and an optimizer, the three-network model comprises a multi-scale ResCNN module, a BiLSTM module, a global optimization attention mechanism module and a multi-layer full-connection module which are connected in sequence, the data set is used as input, an optimizer is used for training the three-network model to obtain a lithology identification model, and during application, logging parameters of a plurality of continuous depth points belonging to the same well are used as input to obtain a lithology identification model; and determining the lithology of the last depth point in the plurality of continuous depth points belonging to the same well by using the lithology identification model, and completing lithology identification. The lithology identification cost can be reduced, and the lithology identification efficiency and accuracy can be improved.

Description

Technical Field

[0001] This application relates to the field of lithology identification technology, and in particular to a method and related apparatus for dynamic optimization training and application of a lithology identification model based on three-network collaboration. Background Technology

[0002] Lithological identification is a crucial aspect of reservoir characterization and resource assessment, and its accuracy directly impacts the efficiency of reservoir delineation and oil and gas reservoir development. As the development of traditional sandstone and carbonate reservoirs gradually reaches its limits, volcanic reservoirs, as an emerging exploration area, are receiving increasing attention. Lithological identification is fundamental to the evaluation of volcanic reservoirs. However, the complexity of volcanic reservoirs presents a significant challenge to accurate lithological identification.

[0003] (1) The mineral composition, rock structure, and rock texture of volcanic rock reservoirs depend on the superposition and modification of different volcanic activities, magma types and later tectonic movements. The logging curve response characteristics of different lithologies may overlap or be redundant.

[0004] (2) Volcanic activity is frequent and has the characteristics of multiple eruptions. The scale, mode (overflow / explosion), material and energy of each eruption are different, thus forming a rock layer combination with great thickness difference. The thinner volcanic clastic rocks produced by the explosion and the thicker lava produced by the overflow alternate to form a typical "thin interbedded" structure, which makes the volcanic rock reservoirs of different lithologies exhibit the characteristics of "rapid lateral changes and multiple vertical cycles", increasing the difficulty of lithology identification.

[0005] (3) The logging curve response characteristics of volcanic rock reservoirs have strong sequence dependence, and the logging curve response characteristics change with depth.

[0006] The aforementioned complex geophysical characteristics make traditional lithology identification methods, which rely on empirical assumptions or single indicators, difficult to accurately characterize lithology. Specifically, traditional lithology identification methods mainly rely on techniques such as core observation, thin section identification, seismic attribute analysis, well logging cross-plot analysis, and fuzzy mathematics. Core observation and thin section identification require manual work. Seismic attribute analysis distinguishes lithology based on seismic attributes, but its accuracy is directly limited by the sensitivity of the selected seismic attribute to the lithology; a single seismic attribute is often insufficient for accurate lithology identification. Well logging cross-plot analysis distinguishes lithology through cross-plots of sensitive well logging parameters, but it is difficult to effectively classify lithologies and achieve accurate lithology identification when lithology is complex or when there is a large overlap in well logging curve response characteristics. Fuzzy mathematics uses the principles of fuzzy mathematics to establish identification patterns to distinguish lithology, but its accuracy is limited by the sensitivity of well logging parameters, making it difficult to achieve accurate lithology identification. Obviously, traditional lithology identification methods require manual intervention, are costly, inefficient, overly reliant on expert experience, are greatly affected by subjective factors, and are difficult to achieve accurate lithology identification, resulting in significant uncertainty in the identification results. Summary of the Invention

[0007] The purpose of this application is to provide a method and related apparatus for dynamic optimization training and application of a lithology identification model based on three-network collaboration, which can reduce the cost of lithology identification and improve the efficiency and accuracy of lithology identification.

[0008] To achieve the above objectives, this application provides the following solution.

[0009] Firstly, this application provides a dynamic optimization training method for a lithology identification model based on three-network collaboration, the method comprising: Obtain a dataset; the dataset includes multiple sample data and label data corresponding to each sample data, the sample data includes logging parameters of multiple consecutive depth points belonging to the same well, and the label data includes the lithology of the last depth point among the multiple consecutive depth points belonging to the same well; A dynamic optimization model is constructed; the dynamic optimization model includes a three-network model and an optimizer. The three-network model includes a multi-scale ResCNN module, a BiLSTM module, a global optimization attention mechanism module, and a multi-layer fully connected module connected in sequence. The optimizer is an Adam-SGD dynamic switching optimizer. Using the dataset as input, the optimizer is used to train the three-network model to obtain the lithology identification model.

[0010] Secondly, this application provides a method for applying a lithology identification model based on three-network collaboration, the method comprising: Obtain logging parameters for multiple consecutive depth points belonging to the same well; Using logging parameters from multiple consecutive depth points belonging to the same well as input, the lithology of the last depth point among the multiple consecutive depth points belonging to the same well is determined using a lithology identification model; the lithology identification model is a model trained using the above-mentioned dynamic optimization training method for lithology identification models based on three-network collaboration.

[0011] Thirdly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the computer program to implement the above-described dynamic optimization training method for lithology identification model based on three-network collaboration or the above-described application method for lithology identification model based on three-network collaboration.

[0012] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described dynamic optimization training method for a lithology identification model based on three-network collaboration or the above-described application method for a lithology identification model based on three-network collaboration.

[0013] According to the specific embodiments provided in this application, this application has the following technical effects.

[0014] This application provides a method and related apparatus for dynamic optimization training and application of a lithology identification model based on three-network collaboration. During training, a dataset is acquired, and a dynamic optimization model is constructed. The dynamic optimization model includes a three-network model and an optimizer. The dataset is used as input, and the optimizer is used to train the three-network model to obtain the lithology identification model. During application, well logging parameters of multiple consecutive depth points belonging to the same well are used as input, and the lithology identification model is used to determine the lithology of the last depth point among the multiple consecutive depth points belonging to the same well, thus completing the lithology identification. This application designs a three-network model comprising a multi-scale ResCNN module, a BiLSTM module, a global optimization attention mechanism module, and a multi-layer fully connected module connected sequentially. Combining the feature extraction capabilities of ResCNN, the bidirectional relationship mining capabilities of BiLSTM for sequence data, and the ability of the global optimization attention mechanism to highlight key information, this model addresses the challenges posed by overlapping or redundant response features of logging curves for different lithologies in volcanic reservoirs, dramatic variations in stratum thickness (i.e., extremely large thickness differences, "thin interlayered" structures, and "rapid lateral changes and multiple vertical cycles"), and strong sequential dependence of lithological vertical changes. This improves the accuracy of lithology identification. Furthermore, the optimized model employs a dynamic switching optimizer between Adam and SGD. The Adam optimizer is used for initial training until the three-network model converges stably, followed by training with the SGD optimizer. This combines the rapid convergence advantage of the Adam optimizer in the early training stages with the robustness and generalization ability of the SGD optimizer in the later training stages, achieving optimal convergence speed and performance, thus improving the accuracy of lithology identification. Compared to core observation and thin section identification, it does not require a large amount of manual labor, solving the problems of high cost, low efficiency, excessive reliance on expert experience, and great influence from subjective factors. Compared to seismic attribute analysis, well logging cross-plot analysis, and fuzzy mathematics, due to the network structure of the designed three-network model and the dynamic switching of the optimizer, it solves the problems of difficulty in achieving accurate lithology identification and the large uncertainty of identification results. Thus, it can reduce the cost of lithology identification, improve the efficiency and accuracy of lithology identification, and is well applicable to lithology identification of volcanic reservoirs. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is an application environment diagram for a dynamic optimization training and application method of a lithology identification model based on three-network collaboration, provided in Embodiment 1 of this application.

[0017] Figure 2This is a flowchart illustrating a dynamic optimization training method for a lithology identification model based on three-network collaboration, as provided in Embodiment 1 of this application.

[0018] Figure 3 This is a technical roadmap for a dynamic optimization training and application method of a lithology identification model based on three-network collaboration, provided in Embodiment 1 of this application.

[0019] Figure 4 A comparison diagram of the correction results of the compensated neutron curve provided in Embodiment 1 of this application; wherein, Figure 4 (a) in the figure represents the wellbore diameter curve. Figure 4 (b) in the figure represents the compensated neutron curve. Figure 4 In the figure, (c) represents the corrected compensated neutron curve, with the vertical axis representing depth and the horizontal axis representing the value of the curve at that depth.

[0020] Figure 5 This is a distribution map of sample data for different types of lithology provided in Embodiment 1 of this application; wherein, Figure 5 (a) in the figure is a histogram of the number of sample data. Figure 5 (b) in the figure is a pie chart showing the number of sample data.

[0021] Figure 6 A flowchart of the SMOTE (Synthetic Minority Oversampling Technique) method provided in Embodiment 1 of this application.

[0022] Figure 7 The network structure diagram of the three-network model based on ResCNN (Residual Convolutional Neural Network), BiLSTM (Bidirectional Long Short-Term Memory), and LGQAM (Learnable Global Query Attention Mechanism) provided in Embodiment 1 of this application is shown.

[0023] Figure 8 The flowchart illustrates the dynamic switching optimization strategy of Adam (Adaptive Moment Estimation) - SGD (Stochastic Gradient Descent) provided in Embodiment 1 of this application.

[0024] Figure 9 The flowchart of a lithology identification model application method based on three-network collaboration is provided for Embodiment 2 of this application.

[0025] Figure 10 This is a schematic diagram of the structure of a computer device provided in Embodiment 3 of this application. Detailed Implementation

[0026] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0027] Example 1.

[0028] The dynamic optimization training method for lithology identification models based on three-network collaboration provided in this application can be applied to, for example... Figure 1 The application environment shown illustrates this. The terminal communicates with the server via a network. A data storage system stores the data the server needs to process. This system can be set up independently, integrated into the server, or located in the cloud or on another server. The terminal can send training requests to the server. Upon receiving the request, the server retrieves the dataset, which includes multiple sample data points and corresponding label data for each sample. The sample data includes logging parameters from multiple consecutive depth points belonging to the same well, and the label data includes the lithology of the last depth point among these consecutive depth points. A dynamic optimization model is constructed, comprising a three-network model and an optimizer. The three-network model consists of a multi-scale ResCNN module, a BiLSTM module, a global optimization attention mechanism module, and a multi-layer fully connected module connected sequentially. The optimizer is an Adam-SGD dynamically switching optimizer. Using the dataset as input, the optimizer trains the three-network model to obtain a lithology identification model. The server can then feed back the training result—the lithology identification model for the training request—to the terminal.

[0029] Furthermore, in some embodiments, the dynamic optimization training method for the lithology identification model based on three-network collaboration can also be implemented by the server or the terminal alone. For example, the terminal can directly process the training request to be processed, or the server can obtain the training request to be processed from the data storage system and process it.

[0030] In one exemplary embodiment, such as Figure 2As shown, a dynamic optimization training method for a lithology identification model based on three-network collaboration is provided. This method is executed by a computer device, specifically by a terminal or server alone, or by both a terminal and a server. In this embodiment, the method is applied to... Figure 1 The following steps are used as an example of a server in the example.

[0031] Step S1: Obtain the dataset; the dataset includes multiple sample data and label data corresponding to each sample data. The sample data includes logging parameters of multiple consecutive depth points belonging to the same well, and the label data includes the lithology of the last depth point among the multiple consecutive depth points belonging to the same well.

[0032] Step S2: Construct a dynamic optimization model; the dynamic optimization model includes a three-network model and an optimizer. The three-network model includes a multi-scale ResCNN module, a BiLSTM module, a global optimization attention mechanism module, and a multi-layer fully connected module connected in sequence. The optimizer is an Adam-SGD dynamic switching optimizer.

[0033] Step S3: Using the dataset as input, train the three-network model using the optimizer to obtain the lithology identification model.

[0034] By implementing steps S1 to S3 as described above, this embodiment can reduce the cost of lithology identification and improve the efficiency and accuracy of lithology identification.

[0035] In recent years, with the continuous improvement of computer hardware performance and the rapid development of big data technology, the application of artificial intelligence algorithms in volcanic reservoir lithology identification has gradually become a research hotspot. Identification methods combining artificial intelligence algorithms with well logging curves have significantly improved the efficiency and accuracy of lithology identification. Currently, commonly used artificial intelligence algorithms include logistic regression (LR), support vector machine (SVM), gradient boosting tree (GBDT), and deep learning models (such as CNN and LSTM). These artificial intelligence algorithms provide new technical approaches for volcanic reservoir lithology identification by exploring the nonlinear relationship between well logging curves and lithology.

[0036] To address the challenges of diverse mineral compositions in volcanic reservoirs, overlapping or redundant response characteristics of logging curves for different lithologies, dramatic variations in stratum thickness, and strong sequence dependence of vertical lithological changes, this embodiment constructs a ResCNN-BiLSTM-LGBAM neural network hybrid model (i.e., a three-network model). This model combines the feature extraction capabilities of ResCNN with the bidirectional relationship mining capabilities of BiLSTM for sequential data. Furthermore, a global optimization attention mechanism is introduced to highlight key information. The constructed neural network hybrid model efficiently learns the nonlinear relationship between logging curves and lithology categories, thereby improving the accuracy of identifying complex lithologies.

[0037] like Figure 3As shown, this study focuses on volcanic reservoirs and uses core observations, thin section photographs, and well logging curves from the study area as data. It analyzes the sensitive response characteristics of different lithologies based on the morphological and numerical features of well logging curves. Addressing issues such as the influence of non-stratigraphic factors on lithological samples, low discriminative power of single curve features, and class imbalance, this embodiment employs a sequential lithology identification training sample construction based on multi-dimensional well logging curve correction, feature enhancement, and hierarchical sliding window sampling. Targeted data preprocessing and enhancement operations, including curve correction, feature construction, data standardization, and SMOTE oversampling, are performed. A hierarchical, well-grouped partitioning strategy is adopted, combined with a sliding window mechanism, to construct sample data for the sequential model, resulting in the dataset. To accurately address the challenges of overlapping or redundant logging response characteristics, dramatic variations in stratum thickness, and strong sequence dependencies in vertical lithological changes within volcanic reservoirs, this embodiment employs a hybrid neural network model design based on a global optimization attention mechanism (RsCNN-BiLSTM). A ResCNN-BiLSTM-LGQAM hybrid model is constructed, with a multi-scale ResCNN module for nonlinear feature extraction, a BiLSTM module learning contextual information from the lithological time series to uncover complex global features and establish vertical dependencies of lithology at depth, and an LGQAM module (i.e., the global optimization attention mechanism module) introducing a learnable global query vector to adaptively calculate the contribution weight of each time step in the BiLSTM module to the final decision, highlighting key information and reducing redundancy and noise. During model training, this embodiment uses an adaptive training method based on dynamic optimization switching and perturbation enhancement to help the model achieve rapid convergence and improve generalization ability. Before each batch of data is input into the model, the sample data is dynamically augmented to improve the model's generalization ability and robustness to minor changes in the sample data. Gradient clipping is applied to effectively prevent gradient explosion during training. An Adam-SGD dynamic switching optimizer is introduced, combining the fast convergence advantage of the Adam optimizer in the early stages of training with the robustness and generalization ability of the SGD optimizer in the later stages of training. The system dynamically switches from the Adam optimizer to the SGD optimizer when necessary during training, thereby achieving the best convergence speed and optimal performance. The validation set accuracy is used as the main monitoring indicator, and a patience value mechanism is used to determine early stopping. The model weights are only saved when the difference between the validation set accuracy of the current round and the historical best validation set accuracy exceeds the minimum improvement threshold, ensuring the preservation of the best model.

[0038] The following, combined with Figure 3 This paper provides a detailed introduction to a dynamic optimization training method for a lithology identification model based on three-network collaboration used in this embodiment.

[0039] (I) Construction of training samples for sequence lithology identification based on multidimensional logging curve correction, feature enhancement and hierarchical sliding window sampling.

[0040] The logging curves in this embodiment include sonic transit time (AC) curves, density (DEN) curves, compensated neutron (CNL) curves, natural gamma (GR) curves, intrusion resistivity (RI) curves, true formation resistivity (RT) curves, and spontaneous potential (SP) curves. The lithological categories in this embodiment include andesite, volcanic breccia, breccia tuff, and tuff. These lithologies are distributed in different volcanic rock formation environments such as volcanic conduits, overflows, eruptions, and volcanic sediments. Andesite is a dense rock mainly composed of amphibole and plagioclase, and its logging curves exhibit response characteristics such as high resistivity, high density, long sonic transit time, low compensated neutron, and low natural gamma. Volcanic breccia is mainly composed of igneous and metamorphic rocks, with coarse-grained, mixed... The grain size structure of the breccia tuff exhibits high resistivity, high compensated neutron, high natural gamma, medium density, and low acoustic propagation time in its logging curves. The breccia tuff has a mixed grain size structure, with smaller gravel particles. Its main components are glass fragments (silica-rich minerals), crystal fragments (plagioclase, partially metamorphic), and volcanic conglomerate. Its logging curves exhibit high acoustic propagation time, high compensated neutron, high natural gamma, low density, and low resistivity. The tuff is dense and composed of volcanic and terrigenous fragments. The volcanic fragments are mainly composed of volcanic ash, glass fragments, crystal fragments, and a small amount of rock fragments. The volcanic ash cements the tuff, and its logging curves exhibit high compensated neutron, medium natural gamma, low acoustic propagation time, low density, and low resistivity.

[0041] (1) The neutron-caliber curve correction method based on empirical formulas is adopted to calibrate the non-geological disturbances in the CNL curve affected by the wellbore scaling.

[0042] In volcanic reservoirs, due to rapid lithological changes and strong heterogeneity, wellbore enlargement or reduction is highly likely during drilling. The compensated neutron curve (CNL), a key parameter for inferring porosity and fluid properties in volcanic reservoir lithology identification, is extremely sensitive to interference from non-formation factors. Uncorrected CNL data may result in false low or high porosity readings due to wellbore enlargement effects, misleading the model's lithology assessment. Therefore, to ensure that the input data accurately reflects the formation's physical parameters, a wellbore environment correction method was applied to the CNL curve. The correction process primarily addresses the impact of wellbore diameter changes. Using the CAL data from the wellbore diameter (CAL) curve, the CNL data in the original CNL curve was corrected according to an empirically fitted correction formula, as shown in Table 1 below.

[0043] Table 1. Correction formulas for compensated neutron curves

[0044] In Table 1, CAL represents the CAL data for a specific data point on the CAL curve, and CNL... 校This refers to the corrected CAL data for a specific data point in the corrected CNL curve. CNL represents the CAL data for a specific data point in the CNL curve. Since the depth of a data point is the same across the three curves, they belong to the same well. For example... Figure 4 As shown, it is a schematic diagram of the correction results.

[0045] (2) Based on multiple logging curves, multi-parameter-physical correlation coupling derivative feature construction is carried out to enhance the lithological characterization dimension.

[0046] Well logging curves of different lithologies may have overlapping numerical features. This embodiment introduces a variety of nonlinear construction methods on the basis of well logging curves to enhance feature discrimination, optimize decision boundaries, and improve the model's ability to represent nonlinear relationships and variable interaction effects.

[0047] 1) Univariate Nonlinear Extension: By performing a square transformation on the logging curve, a quadratic nonlinear characteristic is introduced, as follows: (1) In equation (1), The first well logging curve obtained after square transformation is the... The values ​​of each depth point; The first in the well logging curve The values ​​of each depth point; A depth point is a data point sampled at a certain depth.

[0048] 2) Construction of multivariate interaction terms: For two or more well logging curves that are physically or geologically related, point-by-point addition, subtraction, multiplication, or division transformations are performed to generate new cross-logging curves, enhancing the model's ability to express the interaction relationships between variables and the combination of rock physical properties, as follows: (2) (3) (4) (5) In the above formula, The first well logging curve obtained after the addition transformation is the first... The values ​​of each depth point; To and In another different logging curve, the first The values ​​of each depth point; The first well logging curve obtained after subtraction transformation is the first... The values ​​of each depth point; The first well logging curve obtained after multiplication transformation is the first... The values ​​of each depth point; The first well logging curve obtained after the phase division transformation is the... The values ​​of each depth point.

[0049] 3) Mathematical Transformations: For well logging curves with large numerical ranges or severe skewness, logarithmic transformations are introduced to reduce the influence of extreme values ​​and enhance the distinguishability and linear correlation between variables and lithology categories. Square root transformations are introduced to compress high values ​​and enhance the resolution of low-value portions. Difference transformations and gradient transformations are introduced to capture local variations and trends, as follows: (6) (7) (8) (9) In the above formula, The first logarithmic transformation is used to obtain the new logging curve. The values ​​of each depth point; It is a tiny constant (usually taken as 1). 10 -6 This helps avoid errors when performing logarithmic transformations on logging data with zero or near-zero values. The first well logging curve obtained after square root transformation is the... The values ​​of each depth point; The first well logging curve obtained after differential transformation is the... The values ​​of each depth point; To and In the same logging curve, the first The values ​​of each depth point; The first well logging curve obtained after gradient transformation The values ​​of each depth point; To and In the same logging curve, the first The values ​​of each depth point.

[0050] 4) Statistical Features: To incorporate local contextual information and smooth the original data for noise reduction, a sliding window transformation is used to expand the features vertically, as follows: (10) (11) In the above formula, When the statistic is the mean, the th statistic in the new logging curve obtained after sliding window transformation is... The values ​​of each depth point; This represents the average value of all depth points within the calculation window; Let be the window length, which can be 9. In this case, the window will be 9. Centered on, with a length of The window; When the statistic is the standard deviation, the first statistic in the new logging curve obtained after sliding window transformation is... The values ​​of each depth point; This represents the standard deviation of the values ​​taken at all depth points within the calculation window.

[0051] After the above feature construction, the feature dimension is significantly increased. To remove redundant features and select the feature set that maximizes the representation of lithological information, three complementary feature optimization strategies are adopted while retaining the logging curves: the F-test method (capturing linear correlation), the mutual information method (capturing nonlinear correlation), and the XGBoost feature importance method (capturing model-driven importance). These strategies determine the final feature set for the model input, resulting in the filtered logging curves. A composite feature optimization method with three-domain joint discrimination is employed, combining the variance discrimination capability of the F-test, the nonlinear correlation identification capability of mutual information, and the importance ranking capability of the XGBoost model. This achieves joint feature discrimination across the statistical domain, information domain, and model domain, selecting the key variables with the highest sensitivity to lithological identification.

[0052] (3) The hierarchical sliding window sequence construction method with well isolation constraints is adopted. By using a hierarchical division strategy and well grouping combined with the sliding window mechanism, a serialized training sample that satisfies the temporal continuity and well independence is constructed.

[0053] 1) Dataset partitioning strategy.

[0054] The original dataset is grouped by well number, and then sorted by depth within each well to ensure the correct temporal order. Within each well, the data is strictly divided into an initial training set (70%), an initial validation set (15%), and an initial test set (15%) according to depth order. This method ensures that the data used for training, validation, and testing come from different wells, thus more realistically evaluating the model's generalization ability. During the partitioning process, to prevent sequences from crossing boundaries, buffers are set at the partition boundaries between the initial training set and the initial validation set, as well as between the initial validation set and the initial test set. Specifically, the last L-1 depth points in the initial training set and the initial validation set are removed. Finally, the partitioning results of each well are concatenated to form a complete training set, validation set, and test set.

[0055] 2) Sequence sample construction strategy.

[0056] A sliding window mechanism is used to generate sequence samples within each predefined subset (i.e., training set, validation set, and test set).

[0057] Let the logging parameters of the depth points in the subset be arranged in depth order as { , , ..., }, and its corresponding lithology is { , , ..., }, This represents the number of depth points.

[0058] The input sequence (i.e., sample data) is defined as: for the th The input sequence is constructed from points at depth. The feature values ​​(i.e., logging parameters, including the values ​​of each target logging curve, which includes the target logging curve and the selected logging curve) are included in itself and the L-1 depth points before it, as follows: (12) In equation (12), , , , The first The depth point, the first The depth point, the first The depth point, the first The feature values ​​of each depth point. L can be 10.

[0059] The output label (i.e., label data) is defined as: the output label corresponding to the input sequence. Only the true lithology of the last depth point in the input sequence Input historical information from the sequence { } is used only as a contextual feature to assist the ResCNN-BiLSTM-LGQAM neural network hybrid model in determining the lithology of the last depth point.

[0060] This embodiment incorporates geological continuity constraints: To ensure the geological validity of the input sequence, strict boundary checks are performed during sliding window generation. An input sequence is only retained if all depth points within it belong to the same well. Any input sequence containing depth points from two or more different wells is discarded, thus strictly guaranteeing that each input sequence represents a continuous logging interval within a single well.

[0061] (4) In order to eliminate the difference in dimensionality between different feature dimensions and improve the stability and convergence efficiency of model training, the Z-score standardization method with global scale constraint is used to standardize all features.

[0062] To ensure the stability and efficiency of model training, and to avoid interference from scale differences between different feature dimensions, all feature values ​​were standardized using the Z-score normalization method. This transformed the feature values ​​into a standard normal distribution with a mean of 0 and a standard deviation of 1, as shown in the following formula: (13) In equation (13), For the input sequence Features The standardized eigenvalues; For the input sequence Features eigenvalues; Features The average value; Features The standard deviation.

[0063] It should be noted that the number of well logging curves and the number of filtered well logging curves equals the number of features. To prevent data leakage, the calculation of standardized parameters strictly follows the following: features average and standard deviation The calculations are performed using only the training set data. Subsequent transformations of the training, validation, and test sets ensure that the model does not come into contact with any information from the validation or test sets during training.

[0064] (5) Set up corresponding sampling strategies based on the original number of each lithological sample in the training set, and perform differentiated SMOTE oversampling processing with category ratio constraints.

[0065] In volcanic reservoir lithology identification tasks, due to the complexity of geological structures and the characteristics of drilling sampling, the number of sample data for different types of lithology often varies significantly, such as... Figure 5 As shown, this data imbalance can cause the model to favor the majority class, resulting in reduced accuracy in identifying the minority class. To address this issue, this embodiment sets a corresponding sampling strategy based on the original quantity of sample data for each type of lithology in the training set. After standardization, oversampling is performed on the sample data for each type of lithology, effectively balancing the proportion of lithologies in the training set while preserving the characteristic distribution of the original data.

[0066] like Figure 6 As shown, firstly, each sample data point from the minority class is selected sequentially as the reference sample for synthesizing the new sample. Then, according to the Euclidean distance calculation formula, the reference sample's distance is found among the other sample data in the minority class. k The nearest neighbor: (14) In equation (14), Euclidean distance; The feature dimension is the number of logging curves and the number of filtered logging curves; The first of the benchmark samples 1 eigenvalue, For the other sample data in the minority class samples, the first Each feature value. This distance metric is suitable for continuous numerical data, selecting the nearest Euclidean distance. k Each other sample data is used as the nearest neighbor.

[0067] Then, based on the degree of imbalance in the distribution of sample categories, the sampling factor is set, and the sample is taken from the established benchmark sample. k Randomly select one nearest neighbor from the nearest neighbors as an auxiliary sample, and repeat this sampling operation a certain number of times.

[0068] Finally, in the benchmark sample With randomly selected auxiliary samples Linear interpolation is performed between them to construct new samples. The calculation formula is as follows: (15) In equation (15), It is a random number between [0, 1].

[0069] In this embodiment, a dataset is obtained, which includes multiple sample data and label data corresponding to each sample data. The sample data includes logging parameters of multiple consecutive depth points belonging to the same well, and the label data includes the lithology of the last depth point among multiple consecutive depth points belonging to the same well.

[0070] Obtaining the dataset specifically includes the following steps.

[0071] (1) Obtain the logging curves of each well in multiple wells. The logging curves include sonic transit time curve, density curve, compensated neutron curve, natural gamma curve, resistivity curve of the invaded zone, real formation resistivity curve and spontaneous potential curve.

[0072] (2) Perform feature construction on the logging curve to obtain the constructed logging curve, and perform feature screening on the constructed logging curve to obtain the screened logging curve.

[0073] The logging curves are characterized to obtain the constructed logging curves, and the constructed logging curves are then filtered to obtain the filtered logging curves. The specific steps include the following steps.

[0074] 1) The compensated neutron curve is corrected using the caliper curve (using Table 1) to obtain the corrected compensated neutron curve. The sonic transit time curve, density curve, corrected compensated neutron curve, natural gamma curve, resistivity curve of the invaded zone, real formation resistivity curve and spontaneous potential curve are combined to form the corrected logging curve. The caliper curve and the compensated neutron curve belong to the same well.

[0075] 2) For each corrected logging curve, the corrected logging curve is independently transformed to obtain the first transformed logging curve. The independent transformation includes at least one of the following: square transformation (using equation (1)), logarithmic transformation (using equation (6)), square root transformation (using equation (7)), difference transformation (using equation (8)), gradient transformation (using equation (9)) and sliding window transformation. The sliding window transformation is the statistic within the sliding calculation window. The statistic includes at least one of the following: mean (using equation (10)) and standard deviation (using equation (11)).

[0076] 3) For any two corrected logging curves, perform an interactive transformation on the two corrected logging curves to obtain a second transformed logging curve. The interactive transformation includes at least one of the following: addition transformation (using equation (2)), subtraction transformation (using equation (3)), multiplication transformation (using equation (4)) and division transformation (using equation (5)).

[0077] 4) Combine the first transformed logging curve and the second transformed logging curve to form the constructed logging curve.

[0078] 5) The F-test method, mutual information method and XGBoost feature importance method are used to screen the post-construction logging curves to obtain the first screening set, the second screening set and the third screening set. The post-construction logging curves that belong to the first screening set, the second screening set and the third screening set are selected as the screened post-logging curves.

[0079] After selecting the constructed logging curves that simultaneously belong to the first screening set, the second screening set, and the third screening set as the screened logging curves, the screened logging curves can be further screened again through human experience to obtain the final screened logging curves to be used.

[0080] (3) For each well, sort the data points of the target logging curve in order of depth from small to large or from large to small to obtain multiple depth points, determine the logging parameters and lithology of each depth point, the target logging curve includes the logging curve and the filtered logging curve, and the logging parameters are the values ​​of the target logging curve.

[0081] (4) Construct a dataset based on the logging parameters and lithology at each depth point of each well.

[0082] A dataset is constructed based on the logging parameters and lithology at each depth point of each well, specifically including the following steps.

[0083] 1) For each well, all depth points are divided into an initial training set, an initial validation set, and an initial test set according to a preset ratio.

[0084] 2) Combine the initial training set of each well into a training set, the initial validation set of each well into a validation set, and the initial test set of each well into a test set.

[0085] The initial training set of each well is combined into a training set, the initial validation set of each well is combined into a validation set, and the initial test set of each well is combined into a test set. Specifically, for each well, the last L-1 depth points in the initial training set are removed to obtain an intermediate training set, and the last L-1 depth points in the initial validation set are removed to obtain an intermediate validation set, where L is the number of depth points included in the sample data; the intermediate training set of each well is combined into a training set, the intermediate validation set of each well is combined into a validation set, and the initial test set of each well is combined into a test set.

[0086] 3) For each depth point in each subset, select the logging parameters of several depth points before the depth point and the logging parameters of the depth point as initial sample data, and select the lithology of the depth point as the label data corresponding to the initial sample data to obtain multiple initial sample data of the subset and label data corresponding to each initial sample data. The subset includes training set, validation set and test set, and the depth points included in the initial sample data are continuous.

[0087] 4) For each initial sample data, if the depth points included in the initial sample data belong to the same well, then the initial sample data is used as the sample data, and multiple sample data of each subset and the label data corresponding to each sample data are obtained.

[0088] 5) Combine the multiple sample data of all subsets and the label data corresponding to each sample data to form a dataset.

[0089] The dataset is composed of multiple sample data from all subsets and the label data corresponding to each sample data. Specifically, for each sample data, the sample data is Z-score standardized to obtain standardized sample data; multiple standardized sample data from all subsets and the label data corresponding to each standardized sample data are combined to form an initial dataset; the initial dataset is oversampled using the SMOTE method to obtain the final dataset.

[0090] (II) Design of a neural network hybrid model based on global optimization attention mechanism ResCNN-BiLSTM.

[0091] This embodiment proposes a lithology identification method for volcanic reservoirs based on a ResCNN-BiLSTM neural network hybrid model (i.e., a ResCNN-BiLSTM-LGQAM neural network hybrid model) with a global optimization attention mechanism. Figure 7As shown, the architecture is designed to precisely address the challenges posed by the dramatic changes in stratum thickness, the strong dependence of lithological vertical variation sequences, and the tendency for overlapping or redundancy in the response characteristics of logging curves for different lithologies in volcanic reservoirs.

[0092] The model's processing flow can be divided into the following four stages. First, the multi-scale residual convolution (ResCNN) module is used to extract multi-scale spatial features from the training data. The convolutional layer is responsible for extracting nonlinear local features. The introduction of the residual structure ensures that the model does not lose its ability to capture subtle, multi-scale spatial information such as thin layers when extracting deep features, thus effectively adapting to the characteristics of drastic changes in the thickness of volcanic reservoir strata. Among them, the batch normalization layer is used to stabilize the feature distribution and accelerate model convergence, thereby improving the performance of the subsequent BiLSTM module in processing time-series information, extracting key features and achieving dimensionality reduction. The pooling layer uses the one-dimensional max pooling method to downsample the extracted feature data. Second, the bidirectional long short-term memory (BiLSTM) module learns the past and future contextual information in the lithological time series, iteratively mining more complex global features from local features, thereby effectively establishing the vertical dependence of lithology in depth. Furthermore, building upon this foundation, a Global Optimized Attention (LGQAM) module is introduced. This module receives the output from the BiLSTM module. Unlike traditional attention mechanisms that rely on the last time step, LGQAM introduces a learnable global query vector, which, together with the features from each time step in the sequence, calculates attention weights. This adaptively focuses on the time series most discriminative for the identification task, highlighting key information and effectively reducing the impact of redundant or noisy features on the identification results. This allows the model to automatically learn the temporal patterns most discriminative for lithology identification. Finally, the attention-weighted context vector is passed to a multi-layer fully connected module to generate the final lithology identification result. In addition, the model incorporates a dropout technique, randomly discarding neurons to suppress overfitting in deep networks and improve the model's generalization ability.

[0093] In this embodiment, a dynamic optimization model is constructed, which includes a three-network model and an optimizer. The three-network model includes a multi-scale ResCNN module, a BiLSTM module, a global optimization attention mechanism module, and a multi-layer fully connected module connected in sequence. The optimizer is an Adam-SGD dynamic switching optimizer.

[0094] The multi-scale ResCNN module includes a first branch, a second branch, and a third branch connected in sequence. The input of the first branch is the input of the multi-scale ResCNN module, and the output of the third branch is the output of the multi-scale ResCNN module.

[0095] The first branch includes a first input layer, a first convolutional layer, a first batch normalization layer, a first ReLU activation function layer, a first pooling layer, a first addition layer, and a first dropout layer connected in sequence. The input of the first addition layer is also connected to the output of the first input layer. The input of the first input layer is the input of the first branch, and the output of the first dropout layer is the output of the first branch.

[0096] The second branch includes a second input layer, a second convolutional layer, a second batch normalization layer, a second ReLU activation function layer, a second pooling layer, a second addition layer, and a second dropout layer connected in sequence. The input of the second addition layer is also connected to the output of the second input layer. The input of the second input layer is the input of the second branch, and the output of the second dropout layer is the output of the second branch.

[0097] The third branch consists of a third input layer, a third convolutional layer, a third batch normalization layer, a third ReLU activation function layer, a third addition layer, and a third dropout layer connected in sequence. The input of the third addition layer is also connected to the output of the third input layer. The input of the third input layer is the input of the third branch, and the output of the third dropout layer is the output of the third branch.

[0098] The kernel sizes of the first, second, and third convolutional layers are different: the kernel size of the first convolutional layer is 1, the kernel size of the second convolutional layer is 3, and the kernel size of the third convolutional layer is 5.

[0099] The BiLSTM module consists of a BiLSTM network, a fourth batch normalization layer, and a fourth Dropout layer connected in sequence. The input of the BiLSTM network is the input of the BiLSTM module, and the output of the fourth Dropout layer is the output of the BiLSTM module.

[0100] The global optimization attention mechanism module is used to calculate the relevance score of each time step based on the output vector of the BiLSTM module and the learnable global query vector. The relevance score of each time step is normalized to obtain the attention weight of each time step. Based on the attention weight of each time step, the output vector of each time step is weighted and summed to obtain the context vector.

[0101] Specifically, unlike traditional attention mechanisms that rely on the output vector of the last time step (i.e., the hidden state) as the query vector, the global optimization attention mechanism module introduces a learnable global query vector. This learnable global query vector is automatically updated through backpropagation during training, learning the global temporal pattern that is most discriminative for the recognition task. It can effectively capture key features in the sequence and overcome the long-distance dependency problem.

[0102] (1) Calculation of attention score.

[0103] The relevance score between the output vector at each time step and the learnable global query vector is calculated through a non-linear transformation. The formula for calculating the attention score is: (16) In equation (16), For time steps The relevance score; The projection vector; This is the first learnable weight matrix; A learnable global query vector; This is the learnable second weight matrix; For time steps The output vector.

[0104] (2) Attention score normalization.

[0105] The relevance score at each time step is normalized using the Softmax function to obtain the attention weight for each time step, which measures the relative importance of the information at that time step in the global context. (17) In equation (17), For time step Attention weights; The number of time steps; For time step Attention weights are assigned. To enhance the model's generalization ability, a Dropout mechanism is introduced during the weight allocation stage to prevent the model from becoming overly reliant on specific time steps.

[0106] (3) Context vector aggregation.

[0107] The attention weights obtained after normalization are weighted and summed with the output vector to generate a context vector containing globally significant features. (18) In equation (18), This is the context vector.

[0108] Unlike traditional attention mechanisms that rely on the final state of a sequence, the global attention mechanism proposed in this embodiment introduces a shared, learnable global query vector that evolves continuously during training, learning to identify the temporal patterns most representative of the final decision. This design effectively solves the problem of information loss in long-distance dependencies, enabling the model to flexibly allocate weights according to the feature distribution of the data itself, thereby extracting more discriminative feature patterns.

[0109] The multilayer fully connected module includes a first fully connected layer, a fifth batch normalization layer, a fourth ReLU activation function layer, a fifth Dropout layer, a second fully connected layer, a sixth batch normalization layer, a fifth ReLU activation function layer, a third fully connected layer, a sixth Dropout layer, and an output layer, all connected in sequence. The input of the first fully connected layer is the input of the multilayer fully connected module, and the output of the output layer is the output of the multilayer fully connected module.

[0110] (III) Adaptive training method based on dynamic optimization switching-perturbation enhancement.

[0111] (1) Adam-SGD dynamic switching optimization strategy.

[0112] This embodiment combines the fast convergence advantage of the Adam optimizer in the early stage of training with the robustness and generalization ability of the SGD optimizer in the later stage of training, and proposes an Adam-SGD dynamic switching optimization strategy based on the SWATS idea. This strategy allows the model to adaptively switch from the Adam optimizer to the SGD optimizer during training, thereby improving the final generalization performance of the model while ensuring the convergence speed.

[0113] The core of this Adam-SGD dynamic switching optimization strategy lies in tracking the theoretical learning rate estimate of the Adam optimizer (i.e., the stable and effective learning rate level exhibited by the Adam optimizer over a period of time) and the empirical learning rate estimate (i.e., the equivalent learning rate level of the SGD optimizer). By monitoring the difference between the theoretical and empirical learning rate estimates, it determines whether the adaptability of the Adam optimizer no longer provides additional gains, i.e., whether the model has entered the stationary convergence phase. The Adam-SGD dynamic switching optimization strategy process is as follows: Figure 8 As shown.

[0114] Adam step size estimation: In each iteration of the Adam optimizer, the Adam step size... (That is, the true update direction and scale / step size of the Adam optimizer on the current parameters) can be estimated based on the first moment of the current gradient. and second-order moment estimation Perform the calculation: (19) In equation (19), It is a tiny constant to prevent division by zero.

[0115] Empirical learning rate estimation: Empirical learning rate It measures the Adam step size suggested by the Adam optimizer. With the current gradient The scalar of the relationship reflects the equivalent SGD learning rate level exhibited by the Adam optimizer in the current iteration, and is calculated as follows: (20) (twenty one) In the above formula, and Adam's step size and the current gradient In parameters The step size component and gradient component are calculated. In each iteration, the step size component and gradient component of each parameter are calculated. Then, by truncating the average aggregation and exponential smoothing, the final average empirical learning rate is obtained. .

[0116] Theoretical learning rate estimation: Theoretical learning rate This represents the stable equivalent learning rate level of the Adam optimizer over a certain time scale. It is determined by analyzing the Adam optimizer's learning rate at the [time scale]. k The equivalent SGD learning rate level estimated in the next iteration The exponential moving average is used to remove single-step gradient noise. The calculation formula is as follows: (twenty two) In equation (22), It is the exponential smoothing coefficient; For the first k The theoretical learning rate in -1 iterations.

[0117] Because the exponential moving average tends to be systematically undervalued in the early stages of training, therefore... The bias correction is calculated using the following formula: (twenty three) In equation (23), This is the corrected theoretical learning rate.

[0118] The corrected theoretical learning rate obtained after bias correction As a theoretical learning rate estimate for the Adam optimizer, it is used to compare with the average empirical learning rate. (It is an empirical learning rate estimate) to compare and thus determine whether the adaptability of the Adam optimizer no longer brings significant advantages.

[0119] Switching judgment: When training enters the later stage, if and The long-term convergence indicates that the adaptability of the Adam optimizer has reached saturation, and its optimization performance has degenerated to approximately that of a fixed-learning-rate SGD optimizer. Continuing to use the Adam optimizer will hardly yield any additional gains. Therefore, the following dual switching mechanism is designed, which requires one of the following conditions to be met and at least T training sessions to complete the switching. min After (i.e., the minimum number of switching rounds), you can switch from the Adam optimizer to the SGD optimizer.

[0120] Condition 1: and Difference: When Adam optimization enters a stable phase, its corrected theoretical learning rate... Should be compared with the average empirical learning rate The relative differences between the two tend to converge. The calculation formula is as follows: (twenty four) In equation (24), It is a tiny constant to prevent division by zero and ensure the stability of the calculated values.

[0121] like If the value is less than or equal to a small threshold (e.g., 0.25) and remains stable, it indicates that the adaptability of the Adam optimizer is no longer significant, and the model can be switched to the SGD optimizer.

[0122] Condition 2: Performance stability: Monitor the standard deviation of the validation set loss over the most recent 8 epochs. ， If the standard deviation is less than or equal to a certain preset threshold (e.g., 0.0005), it indicates that the model has converged stably. At this point, switch to the SGD optimizer to seek better generalization.

[0123] The validation set loss is the total loss calculated using the loss function when the validation set is used as input. The loss function is the focus loss, which is used to alleviate class imbalance by dynamically adjusting the weights of individual samples, thereby improving the model's ability to accurately classify the minority class.

[0124] When switching to the SGD optimizer, to ensure that the SGD optimizer still has sufficient exploration capability to escape shallow local optima, the initial learning rate of the SGD optimizer is set to the current Adam optimizer's learning rate current_Adam_lr × 10, and further fine-tuning is performed using the ReduceLROnPlateau scheduler (which is a learning rate adjuster). To ensure... and It can purely reflect the convergence state of the Adam optimizer itself, thereby triggering the switch from the Adam optimizer to the SGD optimizer at the most accurate time. Before the optimizer switch, the ReduceLROnPlateau scheduler is not used, and the learning rate of the Adam optimizer remains unchanged.

[0125] In addition, early stop triggers are designed during the training process.

[0126] (2) A random perturbation enhancement strategy was introduced. Through multi-scale noise injection, feature masking and random scaling, the training samples were dynamically perturbed, which enabled the model to learn a more stable feature distribution.

[0127] To enhance the model's robustness and generalization ability under noisy samples and different logging conditions, this embodiment introduces a random perturbation enhancement strategy during the training phase. This strategy dynamically perturbs the training samples through multi-scale noise injection, random feature masking, and random scaling without changing the sample labels, prompting the model to learn a more stable feature distribution. This strategy is a sequentially cascaded random enhancement mechanism. In each training batch, multi-scale noise injection, random feature masking, and random scaling are executed in a fixed order, but each enhancement is triggered independently with a certain probability. This ensures the controllability of the enhancement strategy while achieving a random combination of multiple perturbation forms, effectively simulating the cumulative effects of noise, missing data, and systematic biases in real logging data, thus improving the model's robustness to complex disturbance conditions.

[0128] Multi-scale noise injection: To simulate random fluctuations and instrument noise in logging signals, small-amplitude random noise related to the sample variance is added to the input features. This enhancement operation is performed with a 30% probability (if the random number is less than 30%), as shown in the following formula: (25) In equation (25), Sample data after multi-scale noise injection; For sample data; For variance; Standard deviation; This is the noise proportionality factor. [0.01, 0.03]; This represents the standard deviation of the sample data along the feature dimension.

[0129] Random Feature Masking: Considering the presence of missing or anomalous readings in actual well logging data, a random masking mechanism is introduced. This involves randomly masking 5%–15% of the feature values ​​with a 15% probability, helping the model to maintain stable predictions even under conditions of localized information loss. The formula is as follows: (26) In equation (26), The sample data after random mask injection; Sample data before random mask injection; This represents element-wise multiplication; The binary mask matrix is ​​generated randomly, and the number of elements is... The number of elements is the same; For the randomly generated binary mask matrix, the first... Each element, which is... The probability is set to 0, at which point the feature is masked, with a probability of 1- The probability is set to 1, and the feature remains unchanged. For the mask ratio, [0.05, 0.15].

[0130] Random scaling: To improve the model's adaptability to changes in feature scale, a uniform random scaling transformation is performed on all features. This augmentation operation is executed with a 20% probability (if the random number is less than 20%), as shown in the following formula: (27) In equation (27), The feature data is randomly scaled. This is the scaling factor.

[0131] In each training iteration, the sample data input to the model is subjected to the above perturbation process with a certain probability.

[0132] (3) Apply gradient clipping to suppress gradient explosion and ensure stable convergence during training.

[0133] After calculating the loss value using the loss function, backpropagation is performed to calculate the gradient, and then the optimizer is used to update the parameters. Gradient clipping is applied during backpropagation. By constraining the L2 norm of the model parameter gradients, when the L2 norm of the gradient exceeds a threshold, the gradient is scaled proportionally, thereby effectively suppressing the gradient explosion problem and ensuring the numerical stability and convergence of the deep network training process.

[0134] (4) Using the validation set accuracy as the monitoring indicator, early stopping is determined by the patience value mechanism. The model weights are saved only when the validation set accuracy in the current round exceeds the historical best and exceeds the minimum improvement threshold.

[0135] To prevent overfitting in the later stages of training and improve training efficiency, an early stopping mechanism is introduced. Validation set accuracy is used as the monitoring metric. Training is terminated early if the validation set accuracy fails to improve significantly over several consecutive training epochs. A patience value of 15 is set. The model performance is considered not to have improved effectively only if the validation set accuracy in the current epoch does not exceed the historical best validation set accuracy and its improvement is less than the minimum improvement threshold of 0.001. In this case, the patience counter is incremented by one. Otherwise, the patience counter is reset. When the patience counter reaches the preset patience value of 15, the model is considered to have converged, training is terminated early, and the model parameters corresponding to the epoch with the highest validation set accuracy are loaded as the final model parameters.

[0136] In this embodiment, the dataset is used as input, and the three-network model is trained using an optimizer to obtain a lithology identification model. The optimizer is the Adam-SGD dynamic switching optimizer, which adopts the Adam-SGD dynamic switching optimization strategy. The Adam-SGD dynamic switching optimization strategy is to first use the Adam optimizer for training until the three-network model is stably converged, and then use the SGD optimizer for training.

[0137] The learning rate of the Adam optimizer remains constant, while the learning rate of the SGD optimizer is adjusted by the learning rate scheduler.

[0138] The switching condition is that the number of iterations reaches a preset number T. min The corrected theoretical learning rate of the Adam optimizer and average empirical learning rate The relative difference is less than or equal to a first preset threshold (e.g., 0.25), and the standard deviation of the relative difference over multiple consecutive iterations is less than or equal to a second preset threshold. Alternatively, the switching condition is that the number of iterations reaches a preset number T. min Furthermore, the standard deviation of the validation set loss over multiple consecutive iterations is less than or equal to the third preset threshold.

[0139] Using the dataset as input, the three-network model is trained using an optimizer to obtain a lithology identification model. Specifically, in each iteration, several sample data are selected from the dataset as target sample data, and the target sample data are randomly noise-added (using Equation (25)), randomly masked (using Equation (26)) and randomly scaled (using Equation (27)) in sequence to obtain enhanced sample data. Using the enhanced sample data as input and the label data corresponding to the enhanced sample data as labels, the three-network model is trained using an optimizer to obtain a lithology identification model.

[0140] (iv) Intelligent identification of complex lithology.

[0141] Intelligent identification of complex lithology encompasses the entire chain from sample construction and hybrid model design to training optimization, aiming to fundamentally improve the accuracy and robustness of complex lithology identification.

[0142] After training is completed and a lithology identification model is obtained, logging parameters of multiple consecutive depth points along the depth direction belonging to the same well are selected to form an input sequence. This input sequence is then input into the lithology identification model for prediction. The lithology identification model uses the logging information of the first L-1 depth points in the input sequence as contextual features to identify the lithology of the last depth point in the input sequence, thereby achieving "sequence-to-point" lithology identification based on contextual information.

[0143] This embodiment, taking into account the diverse lithological composition, complex structure, large thickness differences, and strong vertical temporal dependence of volcanic reservoirs, establishes an integrated volcanic reservoir lithology identification method. This method integrates multi-dimensional feature correction and enhancement with hierarchical sliding window sequence sample construction, targeted hybrid neural network model construction, dynamic optimization switching, and perturbation enhancement multi-strategy integrated adaptive training, providing technical support for complex lithology classification.

[0144] Example 2.

[0145] The lithology identification model application method based on three-network collaboration provided in this application embodiment can be applied to, for example... Figure 1 The application environment shown depicts a terminal communicating with a server via a network. A data storage system stores the data the server needs to process. This system can be set up independently, integrated into the server, or located in the cloud or on another server. The terminal can send application requests to the server. Upon receiving the request, the server retrieves logging parameters from multiple consecutive depth points belonging to the same well. Using these parameters as input, it employs a lithology identification model to determine the lithology of the last depth point within that set. The server can then feed back the obtained lithology result (related to the application request) to the terminal.

[0146] Furthermore, in some embodiments, the application method of the lithology identification model based on three-network collaboration can also be implemented by the server or the terminal alone. For example, the terminal can directly process the application request to be processed, or the server can obtain the application request to be processed from the data storage system and process it.

[0147] In one exemplary embodiment, such as Figure 9As shown, a lithology identification model application method based on three-network collaboration is provided. This method is executed by a computer device, specifically by a terminal or server alone, or by both a terminal and a server. In this embodiment, the method is applied to... Figure 1 The following steps are used as an example of a server in the example.

[0148] Step T1: Obtain logging parameters for multiple consecutive depth points belonging to the same well.

[0149] Step T2: Using logging parameters from multiple consecutive depth points belonging to the same well as input, the lithology of the last depth point among the multiple consecutive depth points belonging to the same well is determined using a lithology identification model; the lithology identification model is a model trained using the dynamic optimization training method for lithology identification model based on three-network collaboration described in Example 1.

[0150] This application also provides an application scenario in which the above-described lithology identification model application method based on three-network collaboration is applied. Specifically, the lithology identification model application method based on three-network collaboration provided in this embodiment can be applied in a lithology identification scenario. A lithology identification scenario includes an identification stage and a display stage. The identification stage is used to identify lithology, and the display stage is used to display lithology. The lithology identification model application method based on three-network collaboration provided in this embodiment belongs to the identification stage.

[0151] Example 3.

[0152] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 10 As shown, the computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores data. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network. When the computer program is executed by the processor, it implements either a dynamic optimization training method for a lithology identification model based on three-network collaboration or an application method for a lithology identification model based on three-network collaboration.

[0153] Those skilled in the art will understand that Figure 10 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0154] In one exemplary embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the dynamic optimization training method for the lithology identification model based on three-network collaboration in Embodiment 1 or the application method for the lithology identification model based on three-network collaboration in Embodiment 2.

[0155] Example 4.

[0156] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the dynamic optimization training method for the lithology identification model based on three-network collaboration in Embodiment 1 or the application method for the lithology identification model based on three-network collaboration in Embodiment 2.

[0157] Example 5.

[0158] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the dynamic optimization training method for the lithology identification model based on three-network collaboration in Embodiment 1 or the application method for the lithology identification model based on three-network collaboration in Embodiment 2.

[0159] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Moreover, the collection, use and processing of the relevant data are carried out in compliance with the relevant data protection laws and policies of the country where the location is located, and with the authorization granted by the owner of the corresponding device.

[0160] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0161] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A dynamic optimization training method for a lithology identification model based on three-network collaboration, characterized in that, The dynamic optimization training method for the lithology identification model based on three-network collaboration includes: Obtain a dataset; the dataset includes multiple sample data and label data corresponding to each sample data, the sample data includes logging parameters of multiple consecutive depth points belonging to the same well, and the label data includes the lithology of the last depth point among the multiple consecutive depth points belonging to the same well; A dynamic optimization model is constructed; the dynamic optimization model includes a three-network model and an optimizer. The three-network model includes a multi-scale ResCNN module, a BiLSTM module, a global optimization attention mechanism module, and a multi-layer fully connected module connected in sequence. The optimizer is an Adam-SGD dynamic switching optimizer. Using the dataset as input, the optimizer is used to train the three-network model to obtain the lithology identification model.

2. The dynamic optimization training method for the lithology identification model based on three-network collaboration according to claim 1, characterized in that, The multi-scale ResCNN module includes a first branch, a second branch, and a third branch connected in sequence; The first branch includes a first input layer, a first convolutional layer, a first batch normalization layer, a first ReLU activation function layer, a first pooling layer, a first addition layer, and a first dropout layer connected in sequence. The input of the first addition layer is also connected to the output of the first input layer. The input of the first input layer is the input of the first branch, and the output of the first dropout layer is the output of the first branch. The second branch includes a second input layer, a second convolutional layer, a second batch normalization layer, a second ReLU activation function layer, a second pooling layer, a second addition layer, and a second Dropout layer connected in sequence. The input of the second addition layer is also connected to the output of the second input layer. The input of the second input layer is the input of the second branch, and the output of the second Dropout layer is the output of the second branch. The third branch includes a third input layer, a third convolutional layer, a third batch normalization layer, a third ReLU activation function layer, a third addition layer, and a third Dropout layer connected in sequence. The input of the third addition layer is also connected to the output of the third input layer. The input of the third input layer is the input of the third branch, and the output of the third Dropout layer is the output of the third branch. The kernel sizes of the first convolutional layer, the second convolutional layer, and the third convolutional layer are different.

3. The dynamic optimization training method for the lithology identification model based on three-network collaboration according to claim 1, characterized in that, The BiLSTM module includes a BiLSTM network, a fourth batch normalization layer, and a fourth Dropout layer connected in sequence.

4. The dynamic optimization training method for the lithology identification model based on three-network collaboration according to claim 1, characterized in that, The global optimization attention mechanism module is used to calculate the relevance score of each time step based on the output vector of each time step output by the BiLSTM module and the learnable global query vector. The relevance score of each time step is normalized to obtain the attention weight of each time step. Based on the attention weight of each time step, the output vector of each time step is weighted and summed to obtain the context vector. The formula for calculating the relevance score is as follows: ； in, For time steps The relevance score; The projection vector; This is the first learnable weight matrix; A learnable global query vector; This is the learnable second weight matrix; For time steps The output vector.

5. The dynamic optimization training method for the lithology identification model based on three-network collaboration according to claim 1, characterized in that, The multi-layer fully connected module includes a first fully connected layer, a fifth batch normalization layer, a fourth ReLU activation function layer, a fifth Dropout layer, a second fully connected layer, a sixth batch normalization layer, a fifth ReLU activation function layer, a third fully connected layer, a sixth Dropout layer, and an output layer connected in sequence.

6. The dynamic optimization training method for the lithology identification model based on three-network collaboration according to claim 1, characterized in that, Obtaining the dataset specifically includes: Obtain logging curves for each well in multiple wells; the logging curves include sonic transit time curves, density curves, compensated neutron curves, natural gamma curves, resistivity curves of the invaded zone, true formation resistivity curves, and spontaneous potential curves; The logging curves are characterized to obtain constructed logging curves, and the constructed logging curves are then filtered to obtain filtered logging curves. For each well, the data points of the target logging curve are sorted in order of depth from smallest to largest or from largest to smallest to obtain multiple depth points, and the logging parameters and lithology of each depth point are determined; the target logging curve includes the logging curve and the filtered logging curve, and the logging parameters are the values ​​of the target logging curve; A dataset is constructed based on the logging parameters and lithology at each depth point of each well.

7. The dynamic optimization training method for lithology identification model based on three-network collaboration according to claim 6, characterized in that, The well logging curves are characterized to obtain constructed well logging curves, and the constructed well logging curves are then filtered to obtain filtered well logging curves, specifically including: The compensated neutron curve is corrected using the caliper curve to obtain the corrected compensated neutron curve. The sonic transit time curve, the density curve, the corrected compensated neutron curve, the natural gamma curve, the resistivity curve of the invaded zone, the true formation resistivity curve, and the spontaneous potential curve are combined to form the corrected logging curve. The caliper curve and the compensated neutron curve belong to the same well. For each of the corrected logging curves, the corrected logging curves are independently transformed to obtain a first transformed logging curve; the independent transformation includes at least one of square transformation, logarithmic transformation, square root transformation, difference transformation, gradient transformation and sliding window transformation, and the sliding window transformation is a statistic within a sliding calculation window, and the statistic includes at least one of mean and standard deviation. For any two corrected logging curves, an interactive transformation is performed on the two corrected logging curves to obtain a second transformed logging curve; the interactive transformation includes at least one of addition transformation, subtraction transformation, multiplication transformation and division transformation. The first transformed logging curve and the second transformed logging curve are combined to form the constructed logging curve; The post-construction logging curves were subjected to feature screening using the F-test method, mutual information method, and XGBoost feature importance method, respectively, to obtain a first screening set, a second screening set, and a third screening set. The post-construction logging curves that simultaneously belong to the first screening set, the second screening set, and the third screening set were selected as the screened post-logging curves.

8. A method for applying a lithology identification model based on three-network collaboration, characterized in that, The application method of the lithology identification model based on three-network collaboration includes: Obtain logging parameters for multiple consecutive depth points belonging to the same well; Using logging parameters from multiple consecutive depth points belonging to the same well as input, the lithology of the last depth point among the multiple consecutive depth points belonging to the same well is determined using a lithology identification model; the lithology identification model is a model trained using the dynamic optimization training method for lithology identification model based on three-network collaboration as described in any one of claims 1-7.

9. A computer device, comprising: The memory, the processor, and the computer program stored in the memory and capable of running on the processor are characterized in that the processor executes the computer program to implement the dynamic optimization training method for the lithology identification model based on three-network collaboration as described in any one of claims 1-7 or the application method for the lithology identification model based on three-network collaboration as described in claim 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the dynamic optimization training method for the lithology identification model based on three-network collaboration as described in any one of claims 1-7, or the application method for the lithology identification model based on three-network collaboration as described in claim 8.

Images