A Well Logging Fluid Identification Method Based on Hybrid Expert Model Architecture Integration

Abstract

This application belongs to the field of oil and gas exploration and development technology, and relates to a logging fluid identification method based on a hybrid expert model architecture integration. It acquires raw logging data from the target exploration block, combines it with dynamic data from oil testing and production to complete fluid calibration, feature screening, and standardized preprocessing, and constructs a high-quality dataset suitable for model input. Then, it selects four types of heterogeneous expert models—kernel logistic regression, support vector machine, Bagging decision tree, and neural network—for independent pre-training and parameter fixing. Subsequently, it builds a trainable gated model to deeply analyze the logging feature attributes of the samples, dynamically activates the suitable expert models, assigns contribution weights to the outputs of each heterogeneous expert model, and performs collaborative weighting. This allows for adaptive discrimination and result fusion for samples with different feature complexities. Finally, it relies on the fully trained integrated model to conduct reservoir fluid identification, achieving efficient dynamic collaboration of heterogeneous expert models, significantly improving the accuracy and robustness of logging fluid identification in complex reservoir scenarios, and making the identification results more aligned with actual exploration needs.

Description

Technical Field

[0001] This application belongs to the field of oil and gas exploration and development technology, and more specifically, relates to a logging fluid identification method based on a hybrid expert model architecture integration. Background Technology

[0002] Well logging fluid identification is a core and fundamental task for oil and gas reservoir evaluation, well location deployment, and reserve upgrading, directly determining the efficiency and effectiveness of oil and gas exploration and development. As oil and gas exploration gradually advances into complex blocks such as tight sandstone and deep unconventional oil and gas reservoirs, exploration blocks represented by the Jingbian area of ​​the Ordos Basin generally face challenges such as complex geological conditions, extremely strong reservoir heterogeneity, outdated and incomplete well logging data, the intermingling of low-resistivity oil layers and high-resistivity water layers, and disordered oil-water relationships. Traditional well logging fluid identification methods based on cross plots and linear templates are difficult to adapt to the well logging response characteristics of complex reservoirs, generally resulting in low interpretation accuracy and high misjudgment rates, which seriously restricts the progress of oil and gas exploration and development in complex blocks.

[0003] To overcome the technical bottlenecks of traditional linear identification methods, machine learning algorithms have been gradually introduced into the field of well logging fluid identification. Relying on single models such as kernel logistic regression (KLR), support vector machines (SVM), Bagging Trees, and neural networks (NN), multi-dimensional well logging curve features can be learned and fluid classification modeled, improving the intelligence level of fluid identification to some extent. However, in practical applications, single machine learning models are limited by factors such as limited training sample size, high hyperparameter sensitivity, and weak model generalization ability. They exhibit poor identification stability for complex and difficult samples, making it difficult to adapt to the fluid identification needs of complex geological blocks.

[0004] To compensate for the shortcomings of single models, existing technologies are gradually adopting ensemble learning strategies such as Boosting and Stacking for well logging fluid identification, improving identification stability by integrating the output results of multiple models. However, these ensemble strategies all adopt a static ensemble mode. On the one hand, they cannot achieve adaptive discrimination and weight allocation for complex and difficult samples, making it difficult to leverage the unique advantages of different heterogeneous models. On the other hand, they have low ensemble support and poor collaborative efficiency for heterogeneous expert models such as kernel logistic regression and support vector machines, and lack a trainable dynamic gating mechanism to achieve intelligent routing and weighted fusion of expert models. As a result, in complex blocks with tight sandstone and low-resistivity oil layers, the accuracy, generalization and robustness of fluid identification still cannot meet the actual production needs of oilfields.

[0005] In summary, existing well logging fluid identification technologies generally suffer from core technical problems such as the inability to achieve dynamic collaborative weighting of heterogeneous expert models, lack of adaptive discrimination capability for difficult samples, and insufficient accuracy and robustness in identifying complex reservoirs. Summary of the Invention

[0006] This invention provides a well logging fluid identification method based on a hybrid expert model architecture integration, which aims to solve the core technical problems that existing well logging fluid identification technologies generally suffer from: inability to achieve dynamic collaborative weighting of heterogeneous expert models, lack of adaptive discrimination capability for difficult samples, and insufficient accuracy and robustness in identifying complex reservoirs.

[0007] This invention provides a well logging fluid identification method based on a hybrid expert model architecture integration, comprising the following steps: S1: Obtain the raw logging data of the target exploration block, combine the oil testing results and production dynamic data to complete the reservoir fluid type calibration, screen effective logging features and carry out standardized data preprocessing, and construct a standardized logging dataset that is adapted to the model input. S2: Construct a hybrid expert model architecture, select kernel logistic regression, support vector machine, Bagging decision tree, and neural network as four types of heterogeneous expert models, perform independent pre-training and parameter optimization for each expert model, and fix the parameters of each type of expert model after pre-training is completed; S3: Construct a trainable gating model, train the gating model using the same input data as the expert model, analyze the logging feature attributes of the input samples through the gating model, dynamically allocate the output contribution weights of each expert model, and perform weighted fusion of the fluid recognition results of each expert model based on the weights to obtain the trained integrated fluid recognition model; S4: Input the standardized well logging dataset into the trained integrated fluid identification model to complete the identification and determination of fluid type in the target block reservoir.

[0008] This invention acquires raw well logging data from the target exploration block and combines it with oil testing and production dynamic data to complete fluid calibration, feature screening, and standardized preprocessing, constructing a high-quality dataset adapted to the model input, providing a data foundation for subsequent model training and recognition. Then, it selects four types of differentiated heterogeneous expert models—kernel logistic regression, support vector machine, Bagging decision tree, and neural network—for independent pre-training and parameter fixing, fully leveraging the unique advantages of each model in well logging fluid recognition scenarios and compensating for the limitations of single models. Subsequently, a trainable gated model is built, utilizing it to deeply analyze the well logging feature attributes of the input samples, achieving dynamic allocation and collaborative weighting of the output contribution weights of each heterogeneous expert model. Adaptive discrimination and weighted fusion of results are performed for samples with different feature complexities. Finally, reservoir fluid recognition is carried out based on the fully trained integrated model, fundamentally overcoming the core defects of existing technologies. This not only achieves efficient dynamic collaboration of heterogeneous expert models but also possesses adaptive discrimination capabilities for difficult samples, significantly improving the accuracy and robustness of well logging fluid recognition in complex reservoir scenarios, making the fluid recognition results more aligned with actual exploration needs.

[0009] Preferably, the effective logging features are obtained by principal component analysis and include sonic transit time, natural gamma, spontaneous potential, resistivity and porosity. The selected effective logging features are then standardized and preprocessed as the core input features of the hybrid expert model.

[0010] Preferably, the data preprocessing includes: Box plots are used to locate outlier data. The outlier data exceeding the threshold is constrained by the tail reduction method while preserving the relative relationship of extreme values. After extracting the incomplete data segment, min-max normalization is used to compress the data value range to [0,1].

[0011] Preferably, the kernel logistic regression acts as a nonlinear probability expert, responsible for processing linearly inseparable samples and outputting calibration probability discrimination results; The support vector machine, acting as a boundary expert, is responsible for constructing the optimal classification interval boundary for high-dimensional logging data. The Bagging decision tree, acting as a robust consensus expert, is responsible for generating a stable prediction result with minimal bias. The neural network, acting as a deep feature expert, is constructed with a single hidden layer and 25 neurons, and employs the ReLU activation function to extract deep nonlinear coupling features from well logging data.

[0012] Preferably, the gating model adopts a logistic regression model.

[0013] The training objective of the gating model is to minimize the overall classification error of the integrated fluid recognition model. During the training process, the gating model and the four types of expert models share the same input samples. By learning the mapping relationship between sample features and expert model expertise, intelligent routing matching of samples is achieved.

[0014] Preferably, the gating model normalizes the original weights using the softmax function to generate a standard weight distribution with a sum of 1. At the same time, a sparse activation mechanism is adopted, which activates only 2 to 3 expert models with a matching degree higher than the threshold of the current sample's logging features for a single input sample to participate in the calculation. The output contribution weights of the unactivated expert models are directly set to zero to shield invalid interference.

[0015] Preferably, the weighted fusion logic of the integrated fluid recognition model is as follows: The weights of each expert model output by the gating model are multiplied with the recognition results of the corresponding expert models, and the fluid recognition results are summed. The gating model introduces L2 regularization to suppress overfitting during the training phase and uses five-fold cross-validation to optimize the hyperparameters.

[0016] Preferably, the reservoir fluid type is divided into four categories according to the oil test calibration results, including: dry layer and water layer, oil-bearing water layer, oil-water co-layer, and oil layer, wherein the dry layer and water layer are of the same category.

[0017] Preferably, after the integrated fluid recognition model is trained, generalization verification is carried out using blind logging data that was not used in the training. The recognition performance is evaluated by accuracy, precision, recall, and F1 score. The effectiveness is verified by two sets of ablation experiments.

[0018] Preferably, the two sets of ablation experiments are as follows: The first is a static integration comparison experiment with the removal of gating models, and the second is a leave-one-out experiment with the sequential proposal of single-class expert models.

[0019] The beneficial effects of this invention include: This invention acquires raw well logging data from the target exploration block and combines it with oil testing and production dynamic data to complete fluid calibration, feature screening, and standardized preprocessing, constructing a high-quality dataset adapted to the model input, providing a data foundation for subsequent model training and recognition. Then, it selects four types of differentiated heterogeneous expert models—kernel logistic regression, support vector machine, Bagging decision tree, and neural network—for independent pre-training and parameter fixing, fully leveraging the unique advantages of each model in well logging fluid recognition scenarios and compensating for the limitations of single models. Subsequently, a trainable gated model is built, utilizing it to deeply analyze the well logging feature attributes of the input samples, achieving dynamic allocation and collaborative weighting of the output contribution weights of each heterogeneous expert model. Adaptive discrimination and weighted fusion of results are performed for samples with different feature complexities. Finally, reservoir fluid recognition is carried out based on the fully trained integrated model, fundamentally overcoming the core defects of existing technologies. This not only achieves efficient dynamic collaboration of heterogeneous expert models but also possesses adaptive discrimination capabilities for difficult samples, significantly improving the accuracy and robustness of well logging fluid recognition in complex reservoir scenarios, making the fluid recognition results more aligned with actual exploration needs. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art 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.

[0021] Figure 1 The logging curves of the low-resistivity oil layer in the Jingbian area are provided for embodiments of the present invention.

[0022] Figure 2 This is a structural diagram of an expert model provided for an embodiment of the present invention.

[0023] Figure 3 This is a structural diagram of an integrated model based on a hybrid expert architecture provided for an embodiment of the present invention.

[0024] Figure 4 This is a schematic diagram of the confusion matrix of the LR-gated hybrid expert integration model provided in an embodiment of the present invention.

[0025] Figure 5 The fluid identification result diagram of well Y41-23 provided in the embodiment of the present invention.

[0026] Figure 6 The image shows the effect after hole filling in well Y41-23 provided in an embodiment of the present invention. Detailed Implementation

[0027] To make the technical problems, solutions, and beneficial effects of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0028] This embodiment uses the Yan 8 to Chang 9 reservoirs in the Jingbian area of ​​the Ordos Basin as the target exploration block to provide a detailed explanation of the logging fluid identification method based on a hybrid expert model architecture integration described in this invention. The target exploration block in this embodiment is the Jingbian area in the central-western part of the Ordos Basin, spanning Hengshan District and Jingbian County of Yulin City, Shaanxi Province, and Etuoke Banner of Inner Mongolia Autonomous Region. It belongs to a key oil and gas enrichment area within the basin, with the main oil-bearing strata covering the Chang 6, Chang 7, and Chang 2 oil layers of the Yanchang Formation and the Yan 9 oil layer of the Yan'an Formation. Among them, the Chang 6 and Chang 7 sections are typical tight sandstone reservoirs with ultra-low porosity and ultra-low permeability characteristics; see [link to previous document]. Figure 1 As shown, the block is generally characterized by low-resistivity oil layers and high-resistivity water layers, with complex oil-water relationships. Traditional linear logging identification methods have extremely low interpretation accuracy, and existing logging data is outdated with some incomplete sequences, which seriously restricts the progress of oil and gas exploration and development. To address the above-mentioned practical engineering problems, the logging fluid identification method based on the integrated hybrid expert model architecture of this invention is adopted to carry out operations, which can effectively overcome the problem of identifying complex reservoir fluids and improve identification accuracy and engineering practicality.

[0029] The specific steps of this embodiment are as follows: S1: In this embodiment, the original logging data of the target exploration block is obtained. According to the oil testing results and production dynamics, the reservoir fluid types in the exploration block are divided into four categories: ① dry and water layers, ② oil-bearing water layers, ③ oil-water co-layers, and ④ oil layers; and labeled in the order of 0-3.

[0030] In this embodiment, the sample selection is limited to the intervals where the well logging interpretation conclusion is a reservoir, and there must be clear oil testing or production data as the basis for fluid property calibration to ensure the reliability of the training labels and provide reliable geological evidence for the model.

[0031] Principal component analysis (PCA) was used to evaluate the sensitivity of the logging data (see Table 1) to quantify the response intensity of each logging curve to oil and gas, as shown in Table 1 below: Table 1. Rotation factor loading matrix of logging curves based on principal component analysis

[0032] In this embodiment, principal component analysis is used to conduct sensitivity assessment on the basic logging data. The data is then processed by rotation using the maximum variance method, combined with the rotation factor loading matrix, KMO test value (KMO=0.754), and cumulative variance contribution rate (73.24%), to quantify the response intensity of each logging curve to the oil and gas reservoir. Redundant features with strong lithological responses and weak fluid responses are eliminated. Finally, sonic transit time, natural gamma, spontaneous potential, resistivity, and porosity are selected as core input features. Based on this selection method, data dimensionality is reduced, feature redundancy is decreased, and the model focuses on logging responses strongly correlated with fluid properties, improving training efficiency and recognition accuracy. In this embodiment, a total of 639.5 m well sections and 5116 data points are extracted to form a dataset.

[0033] However, due to the diverse sources of logging sequences and the mixed instrument models in the Jingbian area, the curve types, dimensions, and physical meanings are not uniform, and missing values ​​and outliers coexist. If the raw data is directly input, the weight bias will be amplified and the model will be overfitted. Therefore, data preprocessing is required before the model is established. First, the data is analyzed by box plot to locate potential outliers. However, considering the special nature of logging curves, directly deleting outliers would be considered as truncating the abnormal information of reservoir response. In particular, it is easy to misjudge extreme signals of low-resistivity oil layers, high-resistivity water layers, or thin interlayers as outliers and lose geological significance. Therefore, in this embodiment, the tail reduction method is used to reduce the outlier data of the upper and lower thresholds of the box plot. For example, when the normal value range of a certain index is [10, 100], the outlier value 120 will be replaced with the maximum value 100, thereby eliminating the influence of extreme noise while preserving the relative relationship of extreme values. Simultaneously, missing values ​​are deleted; then, the data after the above processing is normalized to eliminate dimensional differences. The normalized data ranges from [0,1], and is then divided into training and test sets in an 8:2 ratio to provide high-quality data support for subsequent model building.

[0034] S2: To ensure the diversity and generalization ability of the expert pool in the hybrid expert architecture, this embodiment follows the principles of complementary advantages and anti-overfitting, and selects candidate models that perform well on the training set. In the initial attempt, wide neural networks and KNN were included in the expert pool due to their excellent performance on the training and test sets. However, both showed significant overfitting during blind testing, which seriously dragged down the generalization ability of the ensemble model. The reason for this phenomenon is that the memory rather than learning characteristic of wide neural networks impairs the generalization performance under limited data. Secondly, as an lazy learning algorithm, the performance of KNN is heavily dependent on the local sample density and quality. When facing complex patterns such as low-resistivity oil layers with high data noise levels, its voting mechanism based on simple distance metrics is prone to overfitting, resulting in unstable discrimination on unseen data. In view of this, this embodiment abandons the wide neural network and KNN models that are prone to overfitting, and selects kernel logistic regression, support vector set, Bagging decision tree, and neural network (NN) as four types of heterogeneous expert models. Each model has a clear division of labor and complementary functions, and their specific positioning is as follows: KLR (Kernel Logistic Regression) is a nonlinear probabilistic classification model based on kernel methods. Its core feature lies in implicitly mapping the original input space to a high-dimensional feature space through a kernel function, and then performing linear logistic logic within that space. KLR utilizes the inner product operation of the kernel function to avoid explicit computation of high-risk mapping functions, thus efficiently constructing the nonlinear decision boundary while retaining the natural advantage of logistic regression's output probability interpretation. Its objective function inherits and extends the maximum likelihood estimation framework of logistic regression, and can be expressed as: ; In the formula: Represent the objective function; This represents the weight vector of the sample to be determined; Indicates the first j The weight coefficients corresponding to each training sample; Indicates a bias scalar; This represents the regularization coefficient; N is the total number of training samples. This is a kernel function used to measure the similarity between two samples; Indicates the first i One sample label; Indicates the first i Feature variables of each sample; It is the transpose operator, represented in the formula as transpose; log It is the natural logarithm function; n The total number of training samples; exp It is an exponential function; It is the norm squared of the model weights (or decision functions) in the high-dimensional feature space; K represents the kernel matrix, which is a... A symmetric positive semi-definite matrix.

[0035] In this embodiment, in the integrated fluid identification model based on the hybrid expert model architecture, KLR plays the role of a nonlinear probability expert, specifically responsible for handling complex data patterns that are linearly inseparable in the original logging feature space but can be clearly interpreted probabilistically after nonlinear mapping. For example, in the Jingbian area, there is a common problem of overlapping and confusion between low-resistivity oil layers and high-resistivity water layers in conventional cross-plots. This model constructs a more complex classification interface through its nonlinear mapping capability and outputs the discrimination results in the form of calibrated probabilities, providing an important uncertainty quantification reference for integrated decision-making. It provides robust probabilistic discrimination based on statistical learning theory and nonlinear transformation, making up for the inherent shortcomings of SVM in probability output, thereby enhancing the integrated model's ability to judge complex and fuzzy samples.

[0036] The Support Vector Machine (SVM) is a supervised learning model based on statistical learning theory, primarily used for classification and regression analysis. It finds an optimal hyperplane in the feature space that separates samples of different classes, maximizing the margin between the two classes on this hyperplane. For linearly inseparable problems, SVM maps the original data to a high-dimensional space using a kernel function, making it linearly separable in the new space. Its objective function is: In the formula: This represents the penalty coefficient / regularization parameter; This represents the slack variable vector, which allows the i-th sample to deviate from the hard margin to a certain extent, and is used to handle inseparable data. Represents the normal vector of the hyperplane; Indicates a bias scalar; This means finding an optimal set of conditions that satisfy specific rules. This minimizes the value of the objective function. n This represents the total number of training samples.

[0037] In integrated fluid recognition models based on a hybrid expert model architecture, SVM can play the role of a boundary expert, adept at handling data subsets with clear geometric separation boundaries in the feature space, and providing robust discrimination opinions based on the maximum margin principle.

[0038] Bagging decision trees are a typical implementation of Bootstrap Aggregating. They generate multiple training subsets through bootstrapping, train several decision trees in parallel, and then integrate the output using voting or averaging. In the ensemble fluid recognition model based on a hybrid expert model architecture in this embodiment, the Bagging decision tree acts as a "robust consensus expert." By integrating multiple decision trees, it achieves robust consensus, providing a stable and low-biased prediction for samples with locally ambiguous or perturbed data (such as small-amplitude noise in well logging curves), effectively smoothing out the uncertainties that a single model might produce. Bagging Trees achieved the highest accuracy among all expert models on the test set, serving as the best-performing single model and providing a robust performance starting point for subsequent ensemble models.

[0039] The neural network (NN) is a machine learning algorithm that can automatically learn deep patterns and complex features of data through multi-layer nonlinear transformations. It constructs a hierarchical abstraction of feature interactions through a hidden layer structure that balances depth and width. Through cascaded transformations of multiple hidden layers, the original input features are progressively mapped to a higher-dimensional, more abstract representation space, thereby automatically mining the complex nonlinear relationships and hierarchical patterns inherent in the data. In this embodiment, during the actual model construction process, a connection layer with 25 neurons was built, using ReLU as the activation function. Its objective function is as follows: ; In the formula: Represents the objective function of the neural network; It is the neural network model itself; It is the first The input feature vector of each training sample; Let be the true label or target value of the i-th training sample; R is the set of all trainable parameters of a neural network; The regularization term () is used to penalize the model complexity and prevent overfitting. Represents the regularization coefficient; n The total number of training samples; It is the loss function.

[0040] In the integrated fluid identification model based on the hybrid expert model architecture, the neural network (NN) acts as a deep feature expert, responsible for learning the complex, nonlinear, and hierarchical feature interaction patterns in well logging data. For example, for the complex coupling relationship and deep response law between multiple curves such as resistivity, sonic transit time, and natural gamma in the tight sandstone reservoir of Jingbian area, the NN can effectively extract and model them through multi-layer nonlinear mapping capabilities. It provides a discrimination criterion based on deep feature abstraction, making up for the lack of deep pattern learning ability of other models, thereby enhancing the integrated model's overall understanding and discrimination ability of complex geological responses.

[0041] In this embodiment, a phased training method is adopted to construct the integrated fluid recognition model. First, four expert models—KLR, SVM, Bagging Trees, and NN—are pre-trained independently, optimizing their respective parameters until convergence; for example... Figure 2 As shown, the four expert models each have different characteristics. For example, KLR is good at providing calibrated probability outputs through nonlinear mapping, SVM is good at constructing optimal classification boundaries, NN has powerful deep nonlinear feature learning and representation capabilities, and Bagging Trees can provide a stable prediction with small bias. After training each expert model, its model parameters are fixed and it is used as a fixed expert member in the hybrid expert architecture.

[0042] S3: Construct a trainable gating model, train the gating model using the same input data as the expert model, analyze the logging feature attributes of the input samples through the gating model, dynamically allocate the output contribution weights of each expert model, and perform weighted fusion of the fluid recognition results of each expert model based on the weights to obtain the trained integrated fluid recognition model; In this embodiment, the gating model is the intelligent routing core in the hybrid expert architecture. It is responsible for dynamically allocating the weights of each expert model based on the features of the input samples. Building upon step S2, the gating model is trained using the same training data, enabling it to learn to allocate expert weights based on sample features and minimize the overall classification error of the ensemble model. In this embodiment, three different gating models were constructed and compared to explore their adaptability in complex well logging data environments. The first type is the Adaptive Boosting (AdaBoost) gated model. This model dynamically increases the weights of misclassified samples through iterative training, thereby adaptively focusing on difficult samples. It possesses powerful nonlinear fitting and pattern recognition capabilities, making it suitable for capturing complex feature interactions in well logging responses. The second type is the Random Forest (RF) gated model. This model constructs multiple decision trees through bootstrapping and integrates their outputs. It has natural robustness, noise resistance, and tolerance for outliers; it is suitable for handling missing and inconsistent issues in well logging data in the Jingbian area. The third type is the Logistic Regression (LR) gated model. This model achieves classification decisions through linear weighting and mapping with a logistic function. Its advantages are simple model structure, high training efficiency and strong interpretability, which can provide a stable and easy-to-understand linear benchmark for the decision-making of the gated model. See Figure 3 As shown, in this embodiment, corresponding hybrid expert ensemble models are constructed based on the above-mentioned gating models.

[0043] During the process, for a new input sample, the gating model first calculates the weights of each expert and allocates the data; then, all activated experts process the sample in parallel and produce outputs. The final prediction result is a weighted fusion of all expert outputs, with the weights being the values ​​calculated by the gating model. This mechanism ensures that the model can adaptively integrate the judgments of the most relevant experts for each sample, and the final output of the integrated model is: ; In the formula: This represents the gating weight of the i-th expert (calculated by the Softmax function); This represents the output of the i-th expert network; Indicates the total number of experts; This represents the input vector of the model.

[0044] Furthermore, in this embodiment, in response to the challenges of limited logging data scale, significant curve noise, and imbalanced samples in the Jingbian area, this study implemented multiple anti-overfitting and optimization strategies. An L2 regularization term was introduced in the gating model training to constrain the model weights and prevent overfitting caused by excessively large parameters. Five-fold cross-validation was used to fine-tune the model hyperparameters to ensure that the model selection is based on robust generalization performance estimation.

[0045] In this embodiment, the model is analyzed in the example. The fluid recognition accuracy of the single model KLR, SVM, Bagging Trees and NN on the test set is 71.55%, 64.51%, 82.28% and 74.97% respectively. The accuracy of the conventional ensemble models Stacking and Boosting on the test set reaches 83.09% and 82.80% respectively.

[0046] The hybrid expert ensemble model, gated by AdaBoost, RF, and LR, achieved accuracy of 78.89%, 80.45%, and 84.20%, precision of 80.95%, 83.18%, and 86.18%, recall of 79.38%, 81.68%, and 85.30%, and F1 score of 79.80%, 81.90%, and 85.50% on the test set.

[0047] Among them, the hybrid expert ensemble model with LR as the gate performed best in all aspects, with an average accuracy improvement of 11.37% compared to the single model, and also showed significant improvement over conventional ensemble models such as Stacking and Boosting (see [link to relevant documentation]). Figure 4 (As shown in the figure); it is demonstrated that the LR-gated model has better adaptability and optimal performance in the Jingbian area compared to the AdaBoost and RF-gated models.

[0048] The reason why the hybrid expert ensemble model with LR gating performs best is that its gating mechanism is deeply compatible with the data characteristics of the Jingbian area. The well logging response patterns in the Jingbian area are complex, with numerous nonlinear feature interactions (such as the complex response relationship between low-resistivity oil layers and high-water-resistivity layers). The LR-gated model, with its simple linear weighting structure and strong interpretability, provides a stable and easily understood linear benchmark for sample routing, avoiding the overfitting risk that complex gating models might introduce, and ensuring that the outputs of each expert model are robustly integrated. In contrast, while the RF-gated model has excellent robustness, its averaging mechanism is slightly insufficient when characterizing extremely complex decision boundaries; although AdaBoost gating can adaptively focus on difficult samples, its iterative weighting mechanism is prone to overfitting when faced with well logging data noise, leading to instability in the gating model's decisions.

[0049] In this embodiment, to further verify the model's effectiveness and test its generalization ability, a hybrid expert ensemble model gated by LR was applied to four wells—T25-12, T16-26, L41-21, and L50-28—for blind testing. These four wells had not participated in the model's training and were completely unfamiliar to the model. The final results showed an accuracy of 94% compared to the actual oil testing results, indicating good model application performance (Table 2), demonstrating the model's reliability and generalization ability.

[0050] Table 2. Fluid prediction results of various models in blind well testing.

[0051] Furthermore, in this embodiment, to verify the necessity and effectiveness of dynamic intelligent gating and heterogeneous expert pools in the proposed hybrid expert integration model, an ablation experiment is constructed, which is conducted from two dimensions: The first dimension: verify whether the gated network is better than the static strategy; The second dimension: verifying whether each expert makes a unique contribution to the system; Architectural validity dissolution: To evaluate the value of the dynamic gating model, the gating model is compared with two simplified integration strategies: Group A: LR-gated ensemble model; Group B: LR-gated model removed, final prediction made by simple averaging of the output probabilities of the four experts; Group C: Simulating an invalid gating system, weighted fusion of expert weights randomly generated for each input sample. The comparison results are shown in Table 3. Table 3. Results of ablation experiments on architectural effectiveness

[0052] The comparative results show that the model performance significantly decreased after removing the intelligent gating. Specifically, the accuracy of the model using the static averaging strategy (Group B) decreased by 3.75%, indicating that the static and undifferentiated ensemble method cannot fully utilize the strengths of each expert. The random weight ensemble (Group C) had the lowest performance, further proving that the effectiveness of the gating model stems from its learned intelligent routing strategy, rather than a simple combination of weights. The random weights are equivalent to a noisy gating, and their performance is even lower than that of the static averaging. The significant advantage of the complete model (Group A) confirms that the trainable gating model, through a dynamic and adaptive routing mechanism, is the key to achieving efficient collaboration among expert models.

[0053] Expert Contribution Ablation: To explore the unique contribution of each member in the expert pool, a leave-one-out ablation experiment was performed. That is, in the complete LR-gated ensemble model architecture, one expert model was removed sequentially, and the system was reconstructed based on the remaining expert models to evaluate their contributions. The results are shown in Table 4. Table 4. Ablation Experiment Results of Expert Model Contribution

[0054] As shown in Table 4, all ablation experiments resulted in performance degradation, verifying that each member of the expert pool has an irreplaceable contribution. The magnitude of the degradation reveals the differences in the roles of different experts: ① Core structural expert: Bagging Trees. Removing Bagging Trees caused a sharp performance drop (8.93%), far exceeding the drop of other experts. This indicates that Bagging Trees, as a "robust consensus expert," is the structural cornerstone of this ensemble model. It provides a reliable performance baseline for the entire ensemble model, dominates decision-making on the vast majority of samples, and ensures the basic discriminative ability of the model. ② Specialized functional experts: SVM, NN, KLR. Removing any one of SVM, NN, or KLR caused an overall decrease in accuracy, but the decrease was relatively small (0.89%-1.05%). This result shows that they leverage their specialized advantages in their respective local regions (boundary samples, nonlinear patterns, complex feature interactions), and through the dynamic allocation of gating models, they jointly achieved a performance breakthrough in the ensemble model. At the same time, there is some functional overlap among the three in handling complex boundaries. However, this overlap is not simply redundancy; it forms a functional complementarity and synergy when dealing with complex nonlinear responses. This ensures the reliability of the model's decisions under complex conditions and jointly enhances the model's robustness.

[0055] The ablation experiments above demonstrate that dynamic intelligent gating, a hybrid expert architecture-based ensemble model, surpasses traditional static ensembles in its core strength. The learned routing strategies are crucial for achieving personalized ensemble training. Furthermore, the currently constructed expert pool is an optimized collaborative system, comprising both foundational experts contributing to global performance (Bagging Trees) and functional experts enhancing specific task performance (SVM, KLR, NN), forming an expert team with both strong discriminative power and high robustness. The hybrid expert ensemble model using LR as the gating mechanism achieves further improvements in generalization and accuracy compared to the traditional ensemble model Stacking, further proving that the performance differences revealed by the ablation experiments are reliable and not accidental.

[0056] S4: Input the standardized well logging dataset into the trained integrated fluid identification model to complete the identification and determination of fluid type in the target block reservoir.

[0057] See Figure 5 and Figure 6 In this embodiment, the integrated fluid identification model identified multiple oil layers. Well Y41-23 underwent perforation repair based on the fluid identification results, achieving a good production increase effect. This further verified the reliability and application value of the model in actual production.

[0058] This invention addresses the challenge of fluid identification in the Jingbian region by treating four sub-models—KLR, SVM, Bagging Trees, and NN—as expert models, constructing an integrated model based on a hybrid expert model architecture. It utilizes a gating model to select and activate the most suitable expert model and assign weights, leveraging the expertise of each model to achieve specialized division of labor. The hybrid expert ensemble model exhibits strong adaptability and demonstrates good accuracy, stability, and generalization ability in fluid identification in the Jingbian region.

[0059] In this invention, the integrated model based on the hybrid expert model architecture performed excellently in fluid identification in the Jingbian area, significantly improving the identification accuracy. In the test set and blind test, the accuracy of the integrated model was significantly improved compared with other models, and the effect was significantly better than other models. It also showed high reliability in fluid identification at actual well locations, helped to discover multiple new oil layers, and successfully guided the hole filling operation, verifying its engineering practicality and economic value.

[0060] Dynamic intelligent gating is the core of the hybrid expert architecture ensemble model. The dynamic gating model can dynamically allocate and weight the input data, enabling the ensemble model to effectively adapt to the complex situation in Jingbian area. Ablation experiments show that the model performance drops significantly after removing the gating mechanism (accuracy decreases by 3.75%~5.22%), confirming the effectiveness of the dynamic routing strategy.

[0061] The ensemble model with LR as the gate performed best among all ensemble models. With its simple linear weighting structure and strong interpretability, the LR-gated model is better suited to the complex geological conditions in Jingbian area. In contrast, the hybrid expert ensemble model with AdaBoost and RF as gates performed poorly, even worse than the best single model. This shows that the appropriate selection of the gated model plays a key role in fully realizing the effectiveness of the ensemble model.

[0062] Heterogeneous expert models exhibit strong complementarity, enhancing system robustness. The four types of expert models—KLR, SVM, Bagging Trees, and NN—each possess distinct mathematical modeling advantages, forming functional complementarity within the ensemble framework. Ablation experiments effectively demonstrate that Bagging Trees, as a "core structural expert," significantly contributes to the performance of the ensemble model; while NN, SVM, and KLR, as "functional experts," leverage their strengths in their respective domains, enhancing the model's ability to identify challenging samples through functional complementarity and synergy.

[0063] The above are merely preferred embodiments of this application and are not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A well logging fluid identification method based on a hybrid expert model architecture integration, characterized in that, Includes the following steps: S1: Obtain the raw logging data of the target exploration block, combine the oil testing results and production dynamic data to complete the reservoir fluid type calibration, screen effective logging features and carry out standardized data preprocessing, and construct a standardized logging dataset that is adapted to the model input. S2: Construct a hybrid expert model architecture, select kernel logistic regression, support vector machine, Bagging decision tree, and neural network as four types of heterogeneous expert models, perform independent pre-training and parameter optimization for each expert model, and fix the parameters of each type of expert model after pre-training is completed; S3: Construct a trainable gating model, train the gating model using the same input data as the expert model, analyze the well logging feature attributes of the input sample through the gating model, dynamically activate the adapted expert model and assign the output contribution weight of each expert model, and perform weighted fusion of the fluid recognition results of each expert model based on the weight to obtain the trained integrated fluid recognition model; S4: Input the standardized well logging dataset into the trained integrated fluid identification model to complete the identification and determination of fluid type in the target block reservoir.

2. The well logging fluid identification method based on a hybrid expert model architecture integration as described in claim 1, characterized in that, The effective logging features were selected by principal component analysis and included sonic transit time, natural gamma, natural point, resistivity and porosity. The selected effective logging features were standardized and preprocessed as the core input features of the hybrid expert ensemble model.

3. The well logging fluid identification method based on a hybrid expert model architecture integration as described in claim 1, characterized in that, The data preprocessing includes: Box plots are used to locate outlier data. The outlier data exceeding the threshold is constrained by the tail reduction method while preserving the relative relationship of extreme values. After extracting the incomplete data segment, min-max normalization is used to compress the data value range to [0,1].

4. The well logging fluid identification method based on a hybrid expert model architecture integration as described in claim 1, characterized in that, The kernel logistic regression, as a nonlinear probability expert, is responsible for processing linearly inseparable samples and outputting calibration probability discrimination results; The support vector machine, acting as a boundary expert, is responsible for constructing the optimal classification interval boundary for high-dimensional logging data. The Bagging decision tree, acting as a robust consensus expert, is responsible for generating a stable prediction result with minimal bias. The neural network, acting as a deep feature expert, is constructed with a single hidden layer and 25 neurons, and employs the ReLU activation function to extract deep nonlinear coupling features from well logging data.

5. The well logging fluid identification method based on a hybrid expert model architecture integration as described in claim 1, characterized in that, The gated model adopts a logistic regression model. The training objective of the gated model is to minimize the overall classification error of the integrated fluid recognition model. During the training process, the gated model and four types of expert models share the same input samples. By learning the mapping relationship between sample features and expert model expertise, intelligent routing matching of samples is achieved.

6. The well logging fluid identification method based on a hybrid expert model architecture integration according to claim 1, characterized in that, The gated model normalizes the original weights using the softmax function to generate a standard weight distribution with a sum of 1. At the same time, it adopts a sparse activation mechanism, which activates only 2 to 3 expert models with a matching degree higher than the threshold with the current sample's logging features for a single input sample. The output contribution weights of the unactivated expert models are directly set to zero to shield invalid interference.

7. The well logging fluid identification method based on a hybrid expert model architecture integration as described in claim 1, characterized in that, The weighted fusion logic of the integrated fluid recognition model is as follows: The weights of each expert model output by the gating model are multiplied with the recognition results of the corresponding expert models, and the fluid recognition results are summed. The gating model introduces L2 regularization to suppress overfitting during the training phase and uses five-fold cross-validation to optimize the model hyperparameters.

8. The well logging fluid identification method based on a hybrid expert model architecture integration according to claim 1, characterized in that, The reservoir fluid types are classified into four categories according to the oil test calibration results: dry and water layers, oil-bearing water layers, oil-water co-layers, and oil layers, with dry and water layers belonging to the same category.

9. The well logging fluid identification method based on a hybrid expert model architecture integration according to claim 1, characterized in that, After the integrated fluid recognition model is trained, generalization verification is carried out using blind logging data that was not used in the training. The recognition performance is evaluated by accuracy, precision, recall, and F1 score. The effectiveness is verified by two sets of ablation experiments.

10. The well logging fluid identification method based on a hybrid expert model architecture integration as described in claim 9, characterized in that, The two sets of ablation experiments verified the following: The first is a static integration comparison experiment with the removal of gating models, and the second is a leave-one-out experiment with the sequential proposal of single-class expert models.

Images