Interpretation method and interpretation system for predicting oil recovery based on multimodal characteristic disturbance analysis

Abstract

The present invention belongs to the field of computer systems based on specific computing models, and in particular relates to an interpretation method and interpretation system for predicting oil recovery rate based on multimodal feature perturbation analysis. The present invention introduces a variety of feature perturbation types, such as noise, scaling, shifting, random deletion, random interpolation, and nonlinear transformation, to comprehensively explore the potential relationship between features that improve oil recovery rate, breaking through the current limitation of focusing on a single feature in feature importance, evaluating the effect of multiple combination features on the output of the prediction model, and providing solid data support for the optimization and understanding of the model; by introducing asynchronous computing methods and using the MapReduce model for parallel processing, the computing efficiency is significantly improved. The present invention aims to use feature perturbation technology to optimize the oil production prediction model, improve the accuracy, and improve the interpretability of the machine learning model.

Description

Technical Field

[0001] The present invention belongs to the field of computer systems based on specific calculation models, and in particular relates to an interpretation method and an interpretation system for predicting oil recovery rate based on multimodal characteristic disturbance analysis. Background Art

[0002] In recent years, machine learning models have been widely used in industries such as finance, healthcare, and industrial production. While they demonstrate excellent accuracy, they are often "black-box" applications, and some models have been questioned due to their lack of interpretability. Currently, commonly used model interpretation tools include the SHAP method, the LIME method, and the Permutation Importance method.

[0003] SHAP, short for SHapley Additive exPlanations (also known as Shapley Additive Explanations), is a powerful model interpretation tool. However, it still has certain limitations in practical applications. First, its computational complexity is high. For large datasets and complex models, calculating SHAP values ​​is time-consuming. Second, the calculation assumes feature independence, making it difficult to capture complex interactions between features. This makes SHAP interpretation complex and difficult to understand.

[0004] LIME, short for Local Interpretable Model-agnostic Explanations (LIME), stands for Local Interpretable Model-agnostic Explanations. While it excels at local explanations, it still has numerous shortcomings. First, its limitation lies in the limited scope of its explanations. The LIME method's excessive focus on local explanations makes it difficult to fully and accurately present the model's overall behavior, resulting in significant information loss when evaluating the model globally. Second, the sampling nature of the LIME method leads to its limited stability. Due to the randomness of the sampling process, multiple runs of the same prediction task may produce different explanations.

[0005] The permutation importance method, also known as the permutation importance method, is a method used in machine learning to assess feature importance and offers the advantage of a relatively simple computational process. However, this method also has a number of significant limitations. From an analytical perspective, the permutation importance method can only rank features by their relative importance and cannot deeply analyze the synergistic effects between features. In complex real-world datasets, multiple features often interact and jointly influence model output. This method lacks analytical capabilities in this regard, making it difficult to uncover these deep relationships. Furthermore, when analyzing nonlinear relationships, it cannot accurately characterize and explain the complex nonlinear correlations between data features.

[0006] The interpretation methods for machine learning models in oil recovery prediction have evolved from being driven by physical models to being data-driven. Early numerical simulations primarily relied on physical equations such as Darcy's law. While these methods had clear physical meaning, they were computationally complex and relied on precise parameters. With the introduction of machine learning, interpretation tools such as SHAP and LIME have gradually become industry standards, but they still face limitations such as high computational complexity and insufficient feature interaction analysis. The introduction of deep learning has revealed that well log curves are based on time series, and LSTM can capture more continuous relationships, improving model importance assessment and interpretability. However, not all models adhere to time series, which limits this approach. Existing technologies still face challenges such as insufficient depth in feature combination analysis and imperfect modeling of nonlinear relationships.

[0007] In the field of oil production forecasting, accurate production estimates are crucial for the rational development of oil resources, production planning, and economic decision-making. Traditional methods include solving or fitting empirical physics formulas, numerical simulations, and attenuation curve analysis. Empirical physics formulas introduce many idealized assumptions to simplify calculations, which differ significantly from actual reservoir conditions. Numerical simulations, while based on flow mechanics and computer technology, are complex, costly, and heavily reliant on prior geological models. Attenuation curve analysis, however, fails to adequately consider control factors during the production process, limiting the accuracy of predictions.

[0008] With the development of deep learning, several deep neural network models, such as LSTM, TCN, and Transformer, have been applied to oil production forecasting. However, oil production dynamics are uniquely complex, with production curves consisting of trends, local fluctuations, and noise. Actual oil reservoirs are subject to both natural influences, such as changes in reservoir permeability and pressure, which lead to natural production declines, and human intervention, such as engineers adjusting drilling parameters and implementing production-enhancing measures like acidizing or fracturing. These interventions can cause localized production fluctuations, and their characteristics resemble those of noise, increasing the difficulty of forecasting. Furthermore, most deep learning models exhibit strong nonlinearities, resulting in poor interpretability of forecast results. This makes it difficult for engineers to fully understand the underlying basis for model decisions and effectively utilize the model results to optimize production plans and increase oil recovery.

[0009] For example, Chinese patent document CN118917499A describes an interpretable multi-head oil production prediction method and system that predicts oil production in the future M time steps by interpreting the Transformer model. The interpretable Transformer model includes a local amplifier, an interpretable exogenous attention IEA module, and an interpretable temporal attention ITA module. Both the IEA module and the ITA module use aggregated multi-head attention to intuitively present the relevant influence relationship, so that the constructed interpretable Transformer model learns the decoupled local trend representation of production dynamics and achieves interpretability in terms of time patterns and exogenous factors. However, the computational complexity of the aggregated multi-head attention mechanism using the Transformer architecture is O(T 2 ), it is difficult to meet the real-time prediction needs of high-frequency data. The local amplifier relies on convolution filtering, which can easily misjudge high-frequency noise as effective engineering fluctuation signals. Fixed weights associate engineering measures with production changes and cannot adapt to dynamic changes in oil reservoirs. The time series model will limit the widespread application of interpretable algorithms.

[0010] In terms of feature analysis, existing methods are often not comprehensive and in-depth enough when studying the impact of features on oil production prediction models. Some methods only consider a single feature perturbation method and cannot fully capture the impact of features on the model under different change conditions; and there is less research on the combined impact of features, ignoring the complex interactions between features. In addition, in the construction and optimization process of oil production prediction models, there is a lack of systematic feature perturbation analysis methods to assist in determining key features and optimizing model structure, which limits the performance improvement of the model. The feature perturbation analysis method proposed in this invention aims to solve these problems, comprehensively evaluate the impact of features on the oil production prediction model through multiple perturbation methods, and provide strong support for model optimization.

[0011] This invention relates to the application of feature perturbation analysis in model optimization within the field of oil production forecasting. Specifically, it provides a method and related system for analyzing the impact of features on oil production forecasting models using multiple feature perturbation methods. By supplementing the inherent relationship between the impact of combined features and production, this invention aims to utilize feature perturbation technology to optimize oil production forecasting models, improve their accuracy, and enhance the interpretability of machine learning models. Summary of the Invention

[0012] In response to the technical problems existing in the existing technology, the present invention proposes a technical solution based on multimodal feature perturbation analysis to enhance the interpretability of machine learning models. Multimodal feature perturbation analysis can perturb features from multiple angles and methods, comprehensively and deeply understand the impact of features on model output, and discover complex relationships and potential patterns between features that are difficult to capture with traditional methods.

[0013] In order to solve the above problems, the present invention provides an interpretation method for predicting oil recovery rate based on multimodal characteristic disturbance analysis, the method comprising:

[0014] S1. Data preprocessing: cleaning and converting raw data to construct the original feature matrix X , providing high-quality data input for subsequent analysis. The original data are petrophysical and fluid property data of the oil field, including porosity, permeability, depth, crude oil gravity, crude oil viscosity, crude oil temperature, net thickness and initial oil saturation; the original feature matrix X is composed of the original eigenvalues composition, Indicates the sample number, Indicates the feature number;

[0015] S2. Perform perturbation on the data pre-processed in step S1: According to the specific analysis requirements, flexibly select the perturbation type, and set the feature combination method and the initial value of the step length. and threshold , performing disturbance control on the data pre-processed in step S1 to generate characteristic disturbance data with the above-mentioned specific parameter information. The flexible selection of disturbance type means that there are multiple disturbance types, and one of the appropriate disturbance types is selected according to the actual application scenario, which is not limited here;

[0016] S3. The characteristic disturbance data with the above-mentioned specific parameter information generated in step S2 is transmitted to the machine learning model through the model interface layer, triggering the machine learning model to perform prediction calculations and form a prediction result. At the same time, the model interface layer receives the prediction result returned by the machine learning model, i.e., the oil recovery rate;

[0017] S4. The model interface layer transmits the prediction result returned in step S3 to the step-size optimization controller. The step-size optimization controller uses an iterative optimization strategy to determine the optimal feature perturbation step size for each feature, and sends the perturbation impact under the optimal feature perturbation step size to the result analyzer. The perturbation impact is calculated based on the difference between the model prediction result and the original result after each perturbation. The original result is the oil recovery rate directly predicted by the machine learning model based on the original feature values ​​of the current oil well without perturbation processing;

[0018] S5. The result analyzer sorts the disturbance effects under the optimal step size output in step S4, and calculates the correlation between the features and the original results based on the original results and the received disturbance effects;

[0019] S6. Visually present the analysis results generated by the result analyzer in an intuitive and easy-to-understand manner.

[0020] Preferably, the disturbance types in step S2 include: noise disturbance, scaling disturbance, displacement disturbance, random deletion disturbance, interpolation disturbance, and nonlinear transformation.

[0021] Preferably, the feature combination method in step S2 includes a combination of a single feature and a combination of multiple features. Perturbing a single feature combination is called single feature perturbation, and perturbing a combination of multiple features is called combined feature perturbation.

[0022] Preferably, the formula for the disturbance type in step S2 is specifically:

[0023] Single feature perturbation:

[0024] (1) Noise perturbation: For the original feature matrix X The elements in are the original eigenvalues , Indicates the sample number, Represents the characteristic number, from the normal distribution Generate noise And add it to the original eigenvalue to get , the noise generation formula (1) is:

[0025] (1)

[0026] in It is The standard deviation of the features, is the step size parameter, is the noise generated after the disturbance;

[0027] The calculation formula of the eigenvalue after disturbance is:

[0028] (2)

[0029] in, is the perturbed eigenvalue, is the original eigenvalue, is the noise generated after the disturbance;

[0030] This perturbation is suitable for data scenarios with random or uncertain eigenvalues, is more suitable for nonlinear models, and usually performs best.

[0031] (2) Scaling perturbation: For the original eigenvalue The scaling factor is , the eigenvalue after perturbation , the resulting formula is:

[0032] (3)

[0033] in, is the perturbed eigenvalue, is the original eigenvalue, is the step size parameter;

[0034] This perturbation is used to examine the impact of features on the model at different scales and is suitable for features with scale sensitivity.

[0035] (3) Displacement perturbation: give the original eigenvalue Add the step size parameter , the eigenvalue after perturbation for:

[0036] (4)

[0037] is the perturbed eigenvalue, is the original eigenvalue, is the step size parameter;

[0038] This perturbation is used to explore the effects of features at different positions and is suitable for features with positional correlation.

[0039] (4) Random deletion perturbation: Randomly remove some feature values. This perturbation is used to evaluate the impact of feature loss on the model and is suitable for evaluating feature necessity.

[0040] (5)

[0041] is the perturbed eigenvalue, is the original eigenvalue, is the step size parameter; is a uniformly distributed random number in the interval [0,1);

[0042] (5) Interpolation perturbation: For each column of the feature matrix , randomly select two sample indices and , from a uniform distribution Generate interpolation ,Will Copy to a randomly selected sample index , Determine the number of interpolations, that is

[0043] (6)

[0044] in, is the perturbed eigenvalue, From a uniform distribution The interpolation generated in ; This perturbation is used to randomly interpolate between eigenvalues ​​to create new data points, which is suitable for sparse data distribution;

[0045] (6) Nonlinear transformation: Logarithmic transformation of the original eigenvalue To avoid the situation where the independent variable in the logarithmic function is zero, we use Function, step parameter It is used to control the intensity or amplitude of nonlinear transformation. The eigenvalue after perturbation is:

[0046] (7)

[0047] Square root transformation: eigenvalues ​​after perturbation for:

[0048] (8)

[0049] Exponential transformation: eigenvalues ​​after perturbation for:

[0050] (9)

[0051] in, is the perturbed eigenvalue, is the original eigenvalue, is the step size parameter; this perturbation is used to process features with specific distribution shapes, such as skewed distribution data;

[0052] For the combined feature perturbation, the formula is transformed from the original single feature change to the synchronous change of multiple features. The change formula is similar to the above single feature to observe the impact on the final result.

[0053] Preferably, the machine learning model in step S3 is one of random forest, decision tree, and SVM.

[0054] Preferably, the iterative optimization strategy in step S4 is specifically:

[0055] The disturbance effect of the feature is calculated according to the prediction result returned in step S3, and the change amplitude of the feature is calculated according to the disturbance effect of the feature. When the change amplitude does not meet the termination condition, the step length is optimized according to the step length dynamic adjustment formula, and the optimized step length is returned to the machine learning model for iterative calculation until the change amplitude meets the termination condition. The step length at this time is the optimal feature disturbance step length.

[0056] Further preferably, the step size dynamic adjustment formula is as follows:

[0057] (10)

[0058] in, For the n After adjusting the step size, is the multiple of each adjustment step size, For the n -1 adjusted step size;

[0059] The disturbance impact calculation formula is as follows:

[0060] The calculation formula for the disturbance effect of a single characteristic disturbance is:

[0061] (11)

[0062] in, For the j Sample The first feature n The prediction results after the perturbation are: is the original result without perturbation, For the The first feature n The disturbance effect after the disturbance, N is the number of all samples;

[0063] The calculation formula for the disturbance effect of the combined characteristic disturbance is:

[0064] (12)

[0065] in, For the combined features, j Feature combination of samples No. n The prediction results after the joint perturbation are: is the original result without perturbation, For feature combination No. n The disturbance effect after the joint disturbance is Nis the number of all samples;

[0066] The formula for calculating the change range is as follows:

[0067] The calculation formula for the single feature disturbance change amplitude is as follows:

[0068] (13)

[0069] in, For the i The first feature n The disturbance effect after the disturbance, For the i The first feature n -The impact of disturbance after 1 disturbance, For the The first feature n The magnitude of the change after the disturbance.

[0070] The formula for calculating the change amplitude of the combined characteristic disturbance is as follows:

[0071] (14)

[0072] in, For feature combination No. n The disturbance effect after the joint disturbance is For feature combination No. n -1 disturbance effect after joint disturbance, For feature combination The change amplitude after the nth disturbance.

[0073] The termination conditions are:

[0074] After adjusting the step size and performing perturbation operations for multiple times, the perturbation stops when the termination condition is met:

[0075] Single feature perturbation: (15)

[0076] Combined feature perturbation: (16)

[0077] in, is a pre-set minimum threshold. , and is a very small number, for example , m is the number of disturbances and , For the The change amplitude of the feature after the nth disturbance, For feature combination The change amplitude after the nth disturbance.

[0078] Preferably, the result analyzer in step S5 performs sorting according to the disturbance influence under the optimal step length output in step S4 as follows: for a single feature or a combination of features, after being disturbed, the influence on the model prediction result is measured, and the disturbance influence of the optimal step length obtained is sorted;

[0079] Single feature perturbation:

[0080] (17);

[0081] The calculation of combined feature importance is consistent with that of individual features.

[0082] (18).

[0083] Preferably, the correlation between the features in step S5 and the original results is used to measure the linear correlation between a single feature or a combination of features and the predicted results, and is calculated using the Pearson correlation coefficient, which is as follows:

[0084] (19)

[0085] Finally, the positive and negative correlation between a single feature and the combination is judged, regardless of the dimension of the combination, so the same applies to a single feature, the Pearson correlation coefficient The calculation formula of remains unchanged;

[0086] in It is The combined features of samples, is the mean of the combined features, The original feature true value, is the mean of the true values ​​of the original features.

[0087] The coefficients of influence are sorted to obtain positive and negative correlations.

[0088] >0 combination features are positively correlated with the result change, <0 indicates negative correlation.

[0089] The present invention also provides an interpretation system for predicting oil recovery rate based on multimodal characteristic disturbance analysis, the system comprising:

[0090] External data source module: used to obtain the data set used in model training, which serves as the original data source for the entire feature perturbation analysis process;

[0091] Data preprocessing module: performs a series of preprocessing operations and feature engineering related operations on the raw data obtained from external data sources;

[0092] Feature perturbation and controller module: performs various perturbation operations on the processed data features to generate perturbation data with specific parameter information;

[0093] Model interface layer: Accurately transmits the perturbation data with parameters to the machine learning model, which performs prediction calculations and receives the prediction results returned by the machine learning model, and then passes these results to subsequent modules;

[0094] Step size optimization controller module: This module obtains the prediction results from the model interface layer, analyzes the changes in the model prediction results under different perturbation step sizes, uses an iterative optimization strategy to determine the optimal feature perturbation step size for each feature, and sends the perturbation impact under the optimal step size to the result analyzer;

[0095] Result Analyzer Module: The result analyzer uses the perturbation effects under the optimal step size determined by the step size optimizer to sort the results, and calculates the positive and negative correlation between the perturbation effects of individual features and combined features under different perturbation types and the original results;

[0096] Result output module: Based on the analysis results generated by the result analyzer, this module is responsible for visualizing them in an intuitive way.

[0097] Preferably, the preprocessing operations include data cleaning and feature engineering related operations.

[0098] Preferably, the result analyzer module includes a combination feature analysis module.

[0099] Preferably, the combination feature analysis module innovatively introduces parallel computing technology and, based on the MapReduce model, achieves efficient and accurate evaluation of feature combinations, greatly improving analysis efficiency and effectiveness, as follows:

[0100] During the task distribution phase, the module breaks down complex feature combinations into multiple independent subtasks, then evenly and rationally distributes them across multiple computing nodes. This parallel processing significantly reduces overall computation time. Once each computing node completes its evaluation task, the results integration phase begins. This phase involves a comprehensive and detailed summary and analysis of the evaluation results from each node, which are then ranked and calculated to ultimately produce an accurate and reliable assessment of the importance of the feature combination.

[0101] Compared with the prior art, the present invention has the following beneficial effects:

[0102] (1) Feature combination coverage: This paper introduces a variety of feature perturbation types, such as noise, scaling, shifting, random deletion, random interpolation, and nonlinear transformations, to comprehensively explore the potential relationships between features that enhance oil recovery. This method breaks through the current limitation of focusing on a single feature in feature importance analysis and evaluates the effects of multiple combinations of features on the output of the prediction model, providing solid data support for model optimization and understanding.

[0103] (2) Significantly improved computational efficiency: This paper introduces an asynchronous computing method and uses the MapReduce model for parallel processing, which significantly improves computational efficiency. When processing large-scale high-dimensional data, its computational efficiency is significantly higher than that of the traditional SHAP method. For example, in single feature analysis, the computational speed of the feature perturbation method is nearly 61 times higher than that of the SHAP value method. At the same time, the combined feature analysis module also achieves efficient and accurate evaluation through parallel computing technology, making up for the defect that SHAP has no built-in method in the combined feature.

[0104] (3) Automatic optimization of disturbance intensity: Develop an algorithm with adaptive capabilities that can automatically find the appropriate disturbance intensity based on data characteristics and model characteristics, thereby improving the accuracy and stability of disturbance analysis. All step sizes and thresholds can be manually adjusted to meet the needs of small data volumes but high accuracy requirements.

[0105] (4) Significantly enhance model interpretability: The complex model behavior and feature relationships are transformed into an intuitive and easy-to-understand form, including the importance of individual features, the positive and negative correlation of individual features, the importance of combined features, and the positive and negative correlation of combined features. These are displayed in the form of bar charts, which enables people to understand the basis and process of model decision-making, thereby better utilizing the model for decision-making and analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0106] Figure 1 1 is a flow chart of an interpretation method for predicting oil recovery rate based on multimodal characteristic disturbance analysis according to the present invention;

[0107] Figure 2 Schematic diagram of the process of the step size optimization controller module of the present invention;

[0108] Figure 3 It is a schematic diagram of the flow of the combined feature analysis module of the present invention;

[0109] Figure 4 Schematic diagram of the importance of a single feature disturbance in Example 1 of the present invention;

[0110] Figure 5 This is a schematic diagram of the importance of a single feature SHAP in Example 1 of the present invention;

[0111] Figure 6 This is a schematic diagram of the positive and negative correlation of a single feature disturbance in Example 1 of the present invention;

[0112] Figure 7 This is a schematic diagram of the positive and negative correlation of a single feature SHAP in Example 1 of the present invention;

[0113] Figure 8 Schematic diagram of the importance of combined feature disturbances in Example 1 of the present invention;

[0114] Figure 9 This is a schematic diagram of the positive and negative correlation of the combined characteristic disturbances in Example 1 of the present invention. DETAILED DESCRIPTION

[0115] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0116] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present invention belongs.

[0117] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0118] like Figure 1 As shown, the present invention provides an interpretation method for predicting oil recovery rate based on multimodal characteristic disturbance analysis, the method comprising:

[0119] S1. Data preprocessing: cleaning and converting raw data to construct the original feature matrix X , providing high-quality data input for subsequent analysis. The original data are petrophysical and fluid property data of the oil field, including porosity, permeability, depth, crude oil gravity, crude oil viscosity, crude oil temperature, net thickness and initial oil saturation; the original feature matrix X is composed of the original eigenvalues composition, Indicates the sample number, Indicates the feature number;

[0120] S2. Perform perturbation on the data pre-processed in step S1: According to the specific analysis requirements, flexibly select the perturbation type, and set the feature combination method and the initial value of the step length. and threshold , performing disturbance control on the data pre-processed in step S1 to generate characteristic disturbance data with the above-mentioned specific parameter information. The flexible selection of disturbance type means that there are multiple disturbance types, and one of the appropriate disturbance types is selected according to the actual application scenario, which is not limited here;

[0121] Specifically, the disturbance types include: noise disturbance, scaling disturbance, displacement disturbance, random deletion disturbance, interpolation disturbance, and nonlinear transformation; the feature combination methods include the combination of a single feature and the combination of multiple features. The disturbance of a single feature is called single feature disturbance, and the disturbance of a combination of multiple features is called combined feature disturbance.

[0122] Specifically, the formula for the disturbance type in step S2 is:

[0123] Single feature perturbation:

[0124] (1) Noise perturbation: For the original feature matrix X The elements in are the original eigenvalues , Indicates the sample number, Represents the characteristic number, from the normal distribution Generate noise And add it to the original eigenvalue to get , the noise generation formula (1) is:

[0125] (1)

[0126] in It is The standard deviation of the features, is the step size parameter, is the noise generated after the disturbance;

[0127] The calculation formula of the eigenvalue after disturbance is:

[0128] (2)

[0129] is the perturbed eigenvalue, is the original eigenvalue, is the noise generated after the disturbance;

[0130] This perturbation is suitable for data scenarios with random or uncertain eigenvalues, is more suitable for nonlinear models, and usually performs best.

[0131] (2) Scaling perturbation: For the original eigenvalue The scaling factor is , the eigenvalue after perturbation , the resulting formula is:

[0132] (3)

[0133] is the perturbed eigenvalue, is the original eigenvalue, is the step size parameter;

[0134] This perturbation is used to examine the impact of features on the model at different scales and is suitable for features with scale sensitivity.

[0135] (3) Displacement perturbation: give the original eigenvalue Add the step size parameter , the eigenvalue after perturbation for:

[0136] (4)

[0137] is the perturbed eigenvalue, is the original eigenvalue, is the step size parameter;

[0138] This perturbation is used to explore the effects of features at different positions and is suitable for features with positional correlation.

[0139] (4) Random deletion perturbation: Randomly remove some feature values. This perturbation is used to evaluate the impact of feature loss on the model and is suitable for evaluating feature necessity.

[0140] (5)

[0141] in, is the perturbed eigenvalue, is the original eigenvalue, is the step size parameter; is a uniformly distributed random number in the interval [0,1);

[0142] (5) Interpolation perturbation: For each column of the feature matrix , randomly select two sample indices and , from a uniform distribution Generate interpolation ,Will Copy to a randomly selected sample index , Determine the number of interpolations, that is

[0143] (6)

[0144] in, is the perturbed eigenvalue, From a uniform distribution The interpolation generated in ; This perturbation is used to randomly interpolate between eigenvalues ​​to create new data points, which is suitable for sparse data distribution;

[0145] (6) Nonlinear transformation:

[0146] Logarithmic transformation: For the original eigenvalue To avoid the situation where the independent variable in the logarithmic function is zero, we use Function, step parameter It is used to control the intensity or amplitude of nonlinear transformation. The eigenvalue after perturbation is:

[0147] (7)

[0148] Square root transformation: eigenvalues ​​after perturbation for:

[0149] (8)

[0150] Exponential transformation: eigenvalues ​​after perturbation for:

[0151] (9)

[0152] in, is the perturbed eigenvalue, is the original eigenvalue, is the step size parameter;

[0153] This perturbation is used to process features with specific distribution shapes, such as skewed distribution data;

[0154] Each perturbation method has unique applicable scenarios and advantages and disadvantages. For example, nonlinear transformation can process skewed distribution data and improve model accuracy and interpretability; noise perturbation can simulate feature randomness, but excessive noise may mask the influence of real features. It can be considered that the method of optimizing feature perturbation can meet the needs of different data.

[0155] For the combined feature perturbation, the formula is transformed from the original single feature change to the synchronous change of multiple features. The change formula is similar to the above single feature to observe the impact on the final result.

[0156] S3. The disturbance data with the above-mentioned specific parameter information generated in step S2 is transmitted to the machine learning model through the model interface layer, triggering the model to perform prediction calculations. At the same time, the model interface layer receives the prediction results returned by the machine learning model, i.e., the oil recovery rate;

[0157] Specifically, the machine learning model in step S3 is one of random forest, decision tree, and SVM.

[0158] S4. The model interface layer transmits the prediction result returned in step S3 to the step-size optimization controller. The step-size optimization controller uses an iterative optimization strategy to determine the optimal feature perturbation step size for each feature, and sends the perturbation impact under the optimal feature perturbation step size to the result analyzer. The perturbation impact is calculated based on the difference between the model prediction result and the original result after each perturbation. The original result is the oil recovery rate directly predicted by the machine learning model based on the original feature values ​​of the current oil well without perturbation processing;

[0159] Throughout the entire process, the step size optimizer will dynamically update the step size parameters based on the analysis results and promptly feed back to the model interface layer, enabling the model to use the adjusted perturbation data in the next round of predictive analysis, further optimizing the analysis process to obtain more accurate and stable feature importance assessments.

[0160] Specifically, such as Figure 2 As shown, the step-size optimization controller calculates the disturbance effect of the feature according to the prediction result returned in step S3, and calculates the change amplitude of the feature according to the disturbance effect of the feature. When the change amplitude does not meet the termination condition, the step size is optimized according to the step size dynamic adjustment formula, and the optimized step size is returned to the machine learning model for iterative calculation and looped until the change amplitude meets the termination condition. The step size at this time is the optimal feature disturbance step size:

[0161] 1. Initial step length setting: The initial step length is the initial value of the step length described in step S2 , it is a relatively small value, the specific value can be set according to the actual situation, the default is ;

[0162] 2. Step size dynamic adjustment formula: The multiple of each adjustment step size is , No. n The adjusted step size is , the formula is:

[0163] (10)

[0164] After each feature perturbation, the perturbed feature matrix is ​​input into the model to obtain a predicted result. This result is then compared with the original result, and the difference is calculated. The larger the difference, the greater the impact of the perturbation on the model's prediction. In an oil recovery rate prediction model, if the predicted recovery rate after a feature perturbation differs significantly from the original value, it indicates a significant perturbation impact. Based on the difference in the predicted results, the step size output is determined to determine whether it meets the stopping threshold. If not, the iterative calculation continues until the termination condition is met. In particular, a smaller step size can improve the accuracy of the method, but also increases the time complexity.

[0165] 3. Calculation of the change in assessment results:

[0166] The calculation formula for the disturbance effect of a single characteristic disturbance is:

[0167] (11)

[0168] in, For the j Sample The prediction result after the n-th perturbation of the feature, is the original result without perturbation, For the The first feature n The disturbance effect after the disturbance, N is the number of all samples;

[0169] The calculation formula for the disturbance effect of the combined characteristic disturbance is:

[0170] (12)

[0171] in, For the combined features, j Feature combination of samples No. n The prediction results after the joint perturbation are: is the original result without perturbation, For feature combination No. n The disturbance effect after the joint disturbance is N is the number of all samples;

[0172] The calculation formula for the single feature disturbance change amplitude is as follows:

[0173] (13)

[0174] in, For the i The first feature n The disturbance effect after the disturbance, For the i The first feature n -The impact of disturbance after 1 disturbance, For the The first feature n The magnitude of change after the disturbance;

[0175] The formula for calculating the change amplitude of the combined characteristic disturbance is as follows:

[0176] (14)

[0177] in, For feature combination No. nThe disturbance effect after the joint disturbance is For feature combination No. n-1 The disturbance effect after the joint disturbance is For feature combination The change amplitude after the nth disturbance;

[0178] 4. Termination conditions:

[0179] After adjusting the step size and performing perturbation operations for multiple times, the perturbation stops when the termination condition is met:

[0180] Single feature perturbation: (15)

[0181] Combined feature perturbation: (16)

[0182] in, is a pre-set minimum threshold. , and is a very small number, for example , m is the number of disturbances and , For the The first feature n The change amplitude after the disturbance, For feature combination No. n The magnitude of the change after the disturbance.

[0183] The present invention sets different thresholds according to different working scenarios, and setting the thresholds is well known to those skilled in the art.

[0184] If the effect of a perturbation is not significant, it means that the current step size may be too small to fully stimulate the model's response. In this case, the step size will be increased by a pre-set multiple, thereby expanding the exploration range and increasing the strength of the perturbation. After adjusting the step size and performing perturbations repeatedly, if the change in the evaluation result falls below a pre-set minimum threshold, it is determined that there is no significant change, and the perturbation is stopped.

[0185] Figure 2 The dynamic adjustment mechanism of the step size is presented in detail. Starting from the initial step size, the system initiates multiple iterations. In each iteration, a comprehensive and detailed evaluation of the perturbation effect at the current step size is conducted. After the evaluation is complete, the step size is precisely adjusted using the pre-set algorithms and strategies based on the feedback obtained. This iteration-evaluation-adjustment cycle continues, and the system continuously approaches the optimal perturbation effect. The entire step size adjustment process ends when the evaluation results indicate that the pre-set optimal perturbation standard has been met.

[0186] This dynamic adjustment is based on profound mathematical principles. Through continuous trial and error, it seeks the optimal fit between the complex characteristics of the data and the inherent structure of the model, accurately determining the perturbation step size that best suits the current data and model. This process not only ensures the accuracy of evaluation results, avoiding errors and deviations caused by inappropriate step sizes, but also significantly reduces unnecessary waste of computing resources and improves computational efficiency.

[0187] Compared with the existing methods that require manual intervention to set the step size, the adaptive step size optimization mechanism of the present invention has significant advantages. Existing methods often rely on manual experience to determine the step size, which is not only time-consuming and labor-intensive, but also difficult to adapt to the complex changes of different data and models. The mechanism of the present invention can rely on its built-in intelligent algorithm to automatically and in real time adjust the disturbance intensity according to the diversity of data and the complexity of the model, realizing truly intelligent operation. At the same time, taking into account the diversity and flexibility requirements of models in practical applications, the present invention also allows users to independently adjust the initial step size and disturbance threshold according to the specific model characteristics, further improving the applicability, accuracy and flexibility of the entire mechanism in different scenarios, and providing a more powerful and reliable tool for multimodal feature disturbance analysis.

[0188] S5. The result analyzer sorts the disturbance effects under the optimal step size output in step S4, and calculates the correlation between the features and the original results based on the original results and the received disturbance effects.

[0189] In step S5, the result analyzer ranks the features according to the disturbance effects under the optimal step length output in step S4. Specifically, after the disturbance, the influence of the single feature or the combined feature on the model prediction result is measured, and the disturbance effects of the optimal step length obtained are ranked.

[0190] Single feature perturbation:

[0191] (17);

[0192] The calculation of combined feature importance is consistent with that of individual features.

[0193] (18).

[0194] Specifically, the correlation between the features in step S5 and the original results is used to measure the linear correlation between a single feature or a combination of features and the predicted results, and is calculated using the Pearson correlation coefficient, as shown in the following formula:

[0195] (19)

[0196] Finally, the positive and negative correlation between a single feature and the combination is judged. Regardless of the dimension of the combination, the same applies to a single feature. The Pearson correlation coefficient The calculation formula of remains unchanged;

[0197] in It is The combined features of samples, is the mean of the combined features, The original feature true value, is the mean of the true values ​​of the original features.

[0198] The order of the coefficients of influence can obtain positive and negative correlations. >0 combination features are positively correlated with the result change, <0 indicates negative correlation.

[0199] S6. Visually present the analysis results generated by the result analyzer in an intuitive and easy-to-understand manner.

[0200] The present invention also provides an interpretation system for predicting oil recovery rate based on multimodal characteristic disturbance analysis, the system comprising:

[0201] External data source module: This module is used to obtain the dataset used for model training. It serves as the original data source for the entire feature perturbation analysis process. Its key role is to provide basic information for subsequent analysis. The quality and diversity of the data source will directly affect the reliability and comprehensiveness of the final analysis results.

[0202] Data preprocessing module: performs a series of preprocessing operations on the raw data obtained from external data sources, and strictly aligns with the data preprocessing steps in the model training process. Specific preprocessing operations include data cleaning, such as removing duplicate records in the data, handling missing values, such as correcting erroneous data through strategies such as filling and deletion; and feature engineering-related operations, such as standardizing or normalizing numerical features and encoding conversion of categorical features.

[0203] Feature perturbation and controller module: The processed data is transmitted to this module. This module can flexibly select a variety of different perturbation types according to specific analysis needs through the multimodal perturbation engine within the module. In addition to the common noise addition and scaling operations, it also includes random deletion-randomly setting the eigenvalues ​​to missing in a certain proportion, random interpolation-random interpolation between eigenvalues, and nonlinear transformations, such as logarithmic transformation, square root transformation, exponential transformation, etc. At the same time, the module can accurately set the corresponding perturbation parameters, such as the standard deviation of the noise, the scaling factor, the proportion of random deletion, the number of samples for random interpolation, etc. By performing these diverse perturbation operations on the data features, perturbation data with specific parameter information is generated, providing different input scenarios for subsequent exploration of the impact of feature changes on the model;

[0204] The multimodal perturbation engine is a core component for perturbation control of data. It has multiple perturbation modes and can be flexibly combined to generate characteristic perturbation data with specific parameter information.

[0205] Model interface layer: This layer is responsible for loading various trained machine learning models and receiving prediction results from them. It serves as a bridge between feature perturbation data and machine learning models, accurately passing perturbation data with parameters to the machine learning models, triggering the models to perform prediction calculations. It also receives prediction results returned by the machine learning models and passes these results to subsequent modules for in-depth analysis.

[0206] Step size optimization controller module: This module obtains the machine learning model from the model interface layer, systematically analyzes the changes in the model prediction results under different perturbation step sizes, and uses an iterative optimization strategy to determine the optimal feature perturbation step size for each feature. By calculating the difference between the model prediction results and the original results after each perturbation, the perturbation impact is calculated, and by continuously adjusting the step size, a suitable step size sequence is gradually found. At the same time, the perturbation impact based on these optimal step sizes is sent to the result analyzer.

[0207] Result Analyzer Module: The result analyzer uses the perturbation effects of the optimal step size determined by the step size optimizer to sort, and calculates the positive and negative correlation between the perturbation effects of single features and combined features under different perturbation types and the original results. In addition, the result analyzer module also includes a combined feature analysis module.

[0208] Result output module: Based on the analysis results generated by the result analyzer, this module is responsible for visually presenting them in an intuitive and easy-to-understand manner.

[0209] Specifically, the combination feature analysis module innovatively introduces parallel computing technology and, based on the MapReduce model, enables efficient and accurate evaluation of feature combinations, greatly improving analysis efficiency and effectiveness. The details are as follows:

[0210] During the task distribution phase, the module uses a unique task allocation algorithm to break down complex feature combinations into multiple independent subtasks, then evenly and rationally distributes them across multiple computing nodes. These computing nodes can simultaneously and independently perform perturbations and evaluations on their respective feature combinations. This parallel processing approach, like multiple production lines operating simultaneously, significantly reduces overall computation time.

[0211] Once each computing node completes its evaluation task, the results aggregation phase begins. At this point, the system automatically initiates a results integration mechanism, comprehensively and meticulously summarizing and analyzing the evaluation results distributed across various nodes. Through a series of rigorous algorithms and rules, these results are integrated, filtered, and refined, ultimately yielding an accurate and reliable assessment of the importance of feature combinations.

[0212] Figure 3 The paper clearly demonstrates the specific process of achieving efficient computing using the MapReduce model. As a powerful distributed computing paradigm, MapReduce can cleverly decompose and distribute large-scale combined feature analysis tasks to multiple computing nodes, enabling parallel processing and significantly improving computing efficiency.

[0213] In the initial phase, the combined feature data to be processed is input into the system. The Map phase, the starting point of the entire process, plays a crucial role in task decomposition. Based on pre-set rules and algorithms, it divides the large-scale input task into numerous small, independent subtasks. These subtasks are then evenly and rationally distributed to different compute nodes. Upon receiving an assigned subtask, each compute node processes it independently and in parallel, much like multiple craftsmen simultaneously working on a part at their respective workbenches, significantly improving processing speed.

[0214] After each compute node completes its subtasks, it enters the Reduce phase. The core task of this phase is to aggregate and integrate the output from each compute node. Using specific algorithms and rules, the distributed results are unified, redundant information is removed, and valuable data is extracted, ultimately yielding a complete and accurate result.

[0215] Compared to traditional SHAP methods, this module achieves a quantum leap in computational speed. Traditional SHAP methods are often limited by their serial computation mode when processing large-scale feature combinations, resulting in lengthy and inefficient calculations. However, this module's parallel computation mode fully leverages distributed computing resources, significantly accelerating the computational process. Practical testing and application verification have shown that this module significantly improves computational speed compared to traditional SHAP methods when processing feature combinations of similar scale and complexity. This enables faster analysis results for users, effectively supporting time-sensitive applications.

[0216] Example 1

[0217] An interpretation method for predicting oil recovery yield based on multimodal characteristic disturbance analysis, the method comprising:

[0218] S1. Data preprocessing: To make the data more suitable for training machine learning models, the data is normalized;

[0219] The original data comes from the Midway Sunset oil field in California, USA, and is intended to predict hydrocarbon oil recovery.

[0220] The selected input features include porosity, permeability, md, depth, ft, oil gravity, API, oil viscosity, cp, oil temperature, °F, net thickness, ft, and oil saturation at start, %. The output feature is oil recovery efficiency, b / d.

[0221] S2, select the scaling perturbation method to perturb the input features and generate feature perturbation data with specific parameter information to analyze the impact of feature changes on model output. In this method, two features are combined into a group, and the initial step size is The value is 0.001 and the threshold is 0.1.

[0222] S3. Divide the dataset into a training set and a test set at a ratio of 70% to 30%. The 30% test set is used to evaluate the performance of the machine learning model (SVM). During model training, a linear kernel-based support vector regression (SVR) algorithm is selected. The perturbation data with specific parameter information generated in step S2 is passed to the machine learning model through the model interface layer, triggering the model to perform prediction calculations. Simultaneously, the model interface layer receives the prediction results returned by the machine learning model.

[0223] S4. The model interface layer transmits the prediction result returned by step S3 to the step-size optimization controller. The step-size optimization controller uses an iterative optimization strategy to determine the optimal feature perturbation step size for each feature, and sends the perturbation impact under the optimal feature perturbation step size to the result analyzer.

[0224] Specifically, the disturbance effect of the feature is calculated according to the prediction result returned in step S3, and the change amplitude of the feature is calculated according to the disturbance effect of the feature. When the change amplitude does not meet the termination condition, the step length is optimized according to the step length dynamic adjustment formula: the step length will be updated at a ratio of 1.3 times, and the optimized step length will be returned to the machine learning model for iterative calculation, and the cycle will continue until the change amplitude meets the termination condition. The step length at this time is the optimal feature disturbance step length.

[0225] S5. The result analyzer sorts the features according to the disturbance effects under the optimal step size output in step S4, and calculates the correlation between the features and the original results based on the disturbance effects and the original results.

[0226] S6. Output the importance of a single feature, the positive and negative correlation of a single feature, the importance of the combined feature, and a histogram of the positive and negative correlation of the combined feature.

[0227] Figure 4-Figure 9 A comprehensive interpretability analysis was conducted on the trained model. Specifically, for individual features, the importance and positive and negative correlations with the target variable were thoroughly analyzed, clearly demonstrating the contribution and direction of each feature in the model's decision-making process. Furthermore, for combined features, the importance and positive and negative correlations with the target variable were meticulously explored, fully revealing the combined impact of different feature combinations on the model's results. This analysis provides a deep understanding of the model's operating mechanisms from multiple dimensions, providing strong support for further optimizing model performance and rationally utilizing the model.

[0228] In terms of single feature perturbation, by comparison Figure 4 and Figure 5 It can be found that the feature perturbation method and the SHAP value method are consistent in reflecting the importance ranking of features. Figure 6 and Figure 7 It can be found that the feature perturbation method and the SHAP value method show high similarity in reflecting the positive and negative correlation of features.

[0229] Figure 8 is a schematic diagram of the importance of combined feature perturbations, Figure 9 It is a schematic diagram of the positive and negative correlation of the combined feature disturbance.

[0230] The specific calculation time is as follows: the total analysis time for feature perturbation importance and positive and negative correlation is 0.02 seconds, the total calculation time for SHAP value importance and positive and negative correlation is 1.23 seconds, and the total time for combined feature importance and positive and negative correlation analysis is 2.19 seconds.

[0231] Compared with the SHAP calculation method, the feature perturbation method showed relatively better results in single feature analysis, with the calculation rate increased by nearly 61 times.

[0232] The combined feature analysis module of the present invention achieves efficient and accurate evaluation of feature combinations by introducing parallel computing technology and based on the MapReduce model, greatly improving the analysis efficiency and effect.

[0233] In terms of combined feature analysis, the present invention has powerful deep mining capabilities, which can deeply analyze the interactive relationships between complex features and accurately capture the comprehensive impact of different feature combinations on model output.

Claims

1. An interpretation method for predicting oil recovery based on multimodal characteristic disturbance analysis, characterized in that: The method comprises: S1. Data preprocessing: cleaning and converting raw data to construct the original feature matrix X , providing high-quality data input for subsequent analysis, the original data is the petrophysical and fluid property data of the oil field; the original feature matrix X From the original eigenvalue composition, Indicates the sample number, Indicates the feature number; S2. Perform perturbation on the data pre-processed in step S1: According to the specific analysis requirements, flexibly select the perturbation type, and set the feature combination method and the initial value of the step length. and threshold , performing disturbance control on the data pre-processed in step S1 to generate characteristic disturbance data with the above-mentioned specific parameter information; S3. The characteristic disturbance data with the above-mentioned specific parameter information generated in step S2 is transmitted to the machine learning model through the model interface layer, triggering the machine learning model to perform prediction calculations and form a prediction result. At the same time, the model interface layer receives the prediction result returned by the machine learning model, i.e., the oil recovery rate; S4. The model interface layer transmits the prediction result returned in step S3 to the step-size optimization controller. The step-size optimization controller uses an iterative optimization strategy to determine the optimal feature perturbation step size for each feature, and sends the perturbation impact under the optimal feature perturbation step size to the result analyzer. The perturbation impact is calculated based on the difference between the model prediction result and the original result after each perturbation. The original result is the oil recovery rate directly predicted by the machine learning model based on the original feature values ​​of the current oil well without perturbation processing; S5. The result analyzer sorts the disturbance effects under the optimal step size output in step S4, and calculates the correlation between the features and the original results based on the original results and the received disturbance effects; S6. Visually present the analysis results generated by the result analyzer in an intuitive and easy-to-understand manner.

2. The interpretation method for predicting oil recovery rate based on multimodal characteristic disturbance analysis according to claim 1, characterized in that: The disturbance types in step S2 include: noise disturbance, scaling disturbance, displacement disturbance, random deletion disturbance, interpolation disturbance, and nonlinear transformation; the feature combination methods include single feature combination and multiple feature combination. The disturbance of a single feature combination is called single feature disturbance, and the disturbance of a combination of multiple features is called combined feature disturbance.

3. The interpretation method for predicting oil recovery rate based on multimodal characteristic disturbance analysis according to claim 2, characterized in that: The specific formula for the perturbation type is: Single feature perturbation: (1) Noise perturbation: For the original feature matrix X The elements in are the original eigenvalues , Indicates the sample number, Represents the characteristic number, from the normal distribution Generate noise And add it to the original eigenvalue to get , the noise generation formula (1) is: （1） in It is The standard deviation of the features, is the step size parameter, is the noise generated after the disturbance; The calculation formula of the eigenvalue after disturbance is: （2） in, is the perturbed eigenvalue, is the original eigenvalue, is the noise generated after the disturbance; This perturbation is suitable for data scenarios with random or uncertain eigenvalues, is more suitable for nonlinear models, and usually performs best. (2) Scaling perturbation: For the original eigenvalue The scaling factor is , the eigenvalue after perturbation , the resulting formula is: （3） in, is the perturbed eigenvalue, is the original eigenvalue, is the step size parameter; This perturbation is used to examine the impact of features on the model at different scales and is suitable for features with scale sensitivity. (3) Displacement perturbation: give the original eigenvalue Add the step size parameter , the eigenvalue after perturbation for: （4） in, is the perturbed eigenvalue, is the original eigenvalue, is the step size parameter; This perturbation is used to explore the effects of features at different positions and is suitable for features with positional correlation. (4) Random deletion perturbation: Randomly remove some feature values. This perturbation is used to evaluate the impact of feature loss on the model and is suitable for evaluating feature necessity. （5） in, is the perturbed eigenvalue, is the original eigenvalue, is the step size parameter; is a uniformly distributed random number in the interval [0,1); (5) Interpolation perturbation: For each column of the feature matrix , randomly select two sample indices and , from a uniform distribution Generate interpolation ,Will Copy to a randomly selected sample index , Determine the number of interpolations, that is （6） in, is the perturbed eigenvalue, From a uniform distribution The interpolation generated in ; This perturbation is used to randomly interpolate between eigenvalues ​​to create new data points, which is suitable for sparse data distribution; (6) Nonlinear transformation: Logarithmic transformation: For the original eigenvalue To avoid the situation where the independent variable in the logarithmic function is zero, we use Function, step parameter It is used to control the intensity or amplitude of nonlinear transformation. The eigenvalue after perturbation is: （7） Square root transformation: eigenvalues ​​after perturbation for: （8） Exponential transformation: eigenvalues ​​after perturbation for: （9） in, is the perturbed eigenvalue, is the original eigenvalue, is the step size parameter; this perturbation is used to process features with specific distribution shapes, such as skewed distribution data; For the combined feature perturbation, the formula is transformed from the original single feature change to the synchronous change of multiple features. The change formula is similar to the above single feature to observe the impact on the final result.

4. The interpretation method for predicting oil recovery rate based on multimodal characteristic disturbance analysis according to claim 1, characterized in that: The machine learning model in step S3 is one of random forest, decision tree, and SVM.

5. The interpretation method for predicting oil recovery rate based on multimodal characteristic disturbance analysis according to claim 1, characterized in that: The iterative optimization strategy in step S4 is specifically: The disturbance effect of the feature is calculated according to the prediction result returned in step S3, and the change amplitude of the feature is calculated according to the disturbance effect of the feature. When the change amplitude does not meet the termination condition, the step length is optimized according to the step length dynamic adjustment formula, and the optimized step length is returned to the machine learning model for iterative calculation until the change amplitude meets the termination condition. The step length at this time is the optimal feature disturbance step length.

6. The interpretation method for predicting oil recovery rate based on multimodal characteristic disturbance analysis according to claim 5, characterized in that: The step size dynamic adjustment formula is as follows: （10） in, For the n After adjusting the step size, is the multiple of each adjustment step size, For the n -1 adjusted step size; The disturbance impact calculation formula is as follows: The calculation formula for the disturbance effect of a single characteristic disturbance is: （11） in, For the j Sample The first feature n The prediction results after the perturbation are: is the original result without perturbation, For the The first feature n The disturbance effect after the disturbance, N is the number of all samples; The calculation formula for the disturbance effect of the combined characteristic disturbance is: （12） in, For the combined features, j Feature combination of samples No. n The prediction results after the joint perturbation are: is the original result without perturbation, For feature combination No. n The disturbance effect after the joint disturbance is N is the number of all samples; The formula for calculating the change range is as follows: The calculation formula for the single feature disturbance change amplitude is as follows: （13） in, For the i The first feature n The disturbance effect after the disturbance, For the i The first feature n-1 The disturbance effect after the disturbance, For the The first feature n The magnitude of change after the disturbance; The formula for calculating the change amplitude of the combined characteristic disturbance is as follows: （14） in, For feature combination No. n The disturbance effect after the joint disturbance is For feature combination No. n-1 The disturbance effect after the joint disturbance is For feature combination No. n The magnitude of change after the disturbance; The termination conditions are: After adjusting the step size and performing perturbation operations for multiple times, the perturbation stops when the termination condition is met: Single feature perturbation: (15) Combined feature perturbation: (16) in, is a pre-set minimum threshold. , and is a very small number, for example , m is the number of disturbances and , For the The first feature n The change amplitude after the disturbance, For feature combination The change amplitude after the nth disturbance.

7. The interpretation method for predicting oil recovery rate based on multimodal characteristic disturbance analysis according to claim 1, characterized in that: In step S5, the result analyzer ranks the features according to the disturbance effects under the optimal step length output in step S4. Specifically, after the disturbance, the influence of the single feature or the combined feature on the model prediction result is measured, and the disturbance effects of the optimal step length obtained are ranked. Single feature perturbation: （17）； The calculation of combined feature importance is consistent with that of individual features. （18）。 8. The interpretation method for predicting oil recovery rate based on multimodal characteristic disturbance analysis according to claim 1, characterized in that: The correlation between the features described in step S5 and the original results is used to measure the linear correlation between a single feature or a combination of features and the predicted results, and is calculated using the Pearson correlation coefficient, as shown in the following formula: (19) Finally, the positive and negative correlation between a single feature and the combination is judged. Regardless of the dimension of the combination, the same applies to a single feature. The Pearson correlation coefficient The calculation formula of remains unchanged; in It is The combined features of samples, is the mean of the combined features, The original feature true value, is the mean of the true values ​​of the original features; The order of the coefficients of influence can obtain positive and negative correlations. >0 combination features are positively correlated with the result change, <0 indicates negative correlation.

9. The interpretation system for predicting oil recovery rate based on multimodal characteristic disturbance analysis according to any one of claims 1 to 8, characterized in that: The interpretation system includes: External data source module: used to obtain the data set used in model training, which serves as the original data source for the entire feature perturbation analysis process; Data preprocessing module: performs a series of preprocessing operations and feature engineering related operations on the raw data obtained from external data sources; Feature perturbation and controller module: performs various perturbation operations on the processed data features to generate perturbation data with specific parameter information; Model interface layer: Accurately transmits the perturbation data with parameters to the machine learning model, which performs prediction calculations and receives the prediction results returned by the machine learning model, and then passes these results to subsequent modules; Step size optimization controller module: This module obtains the prediction results from the model interface layer, analyzes the changes in the model prediction results under different perturbation step sizes, uses an iterative optimization strategy to determine the optimal feature perturbation step size for each feature, and sends the perturbation impact under the optimal step size to the result analyzer; Result Analyzer Module: The result analyzer uses the perturbation effects under the optimal step size determined by the step size optimizer to sort the results, and calculates the positive and negative correlation between the perturbation effects of individual features and combined features under different perturbation types and the original results; Result output module: Based on the analysis results generated by the result analyzer, this module is responsible for visualizing them in an intuitive way.

10. The interpretation system according to claim 9, characterized in that The result analyzer module includes a combination feature analysis module.

Images