Sucker rod corrosion fatigue influence factor analysis method based on double-model fusion

Abstract

The invention discloses a sucker rod corrosion fatigue influence factor analysis method based on double-model fusion, and belongs to the field of safety assessment of oil exploitation equipment. In order to solve the problems that in the prior art, analysis dimensions are limited, and prediction precision and model interpretability are difficult to consider at the same time, the method comprises the steps that firstly, a high-dimensional parameterized data set is generated through finite element simulation; secondly, analyzing and visualizing the independent nonlinear influence trend of each factor by adopting a generalized additive model (GAM); meanwhile, a high-precision XGBoost prediction model is constructed, and the global importance, local dependence and complex interaction of each factor are quantified in combination with an SHAP method; and finally, fusing double-model results to carry out cross validation and comprehensive evaluation. According to the method, high-precision prediction and high interpretability are combined, comprehensive insight of a multi-factor coupling effect is realized, and the robustness and credibility of an evaluation conclusion are remarkably improved through a built-in verification mechanism.

Description

Technical Field

[0001] This invention belongs to the field of safety assessment technology for oil extraction equipment, and in particular relates to a quantitative analysis method for the influencing factors of sucker rod corrosion fatigue that combines finite element analysis, statistics and machine learning. Background Technology

[0002] In the oil extraction industry, sucker rods, as key load-bearing components of the oil production system, are particularly vulnerable to fatigue fracture under corrosive well fluids and high-frequency alternating loads, which is one of the main causes of production interruptions and economic losses. Corrosion pits on the surface of sucker rods act as micro-crack initiations, causing severe stress concentration and significantly reducing their fatigue life.

[0003] Currently, research on corrosion fatigue of sucker rods mainly relies on physical experiments and finite element numerical simulations (FEM). In recent years, researchers have begun to use data-driven methods, such as regression analysis or basic machine learning algorithms, to construct predictive models of corrosion parameters and fatigue life based on simulation or experimental data.

[0004] Despite some progress in existing technologies, systemic bottlenecks remain in addressing complex fatigue problems such as sucker rods. First, at the data acquisition level, the computational resource consumption of numerical simulations increases exponentially when dealing with multidimensional variables such as corrosion pit depth, morphology, location, and load conditions. This often leads to simplification of research into low-dimensional parameter analysis, failing to comprehensively capture the coupling effects of multiple factors. Second, at the data analysis level, traditional statistical models struggle to fit strongly nonlinear relationships, while emerging machine learning methods, although offering high predictive accuracy, suffer from opaque decision-making processes due to their "black box" nature, failing to provide clear physical insights and causal explanations. Finally, because existing technological approaches generally rely on single analytical models and lack a comprehensive verification framework that combines statistical interpretability with the predictive power of machine learning, the analytical conclusions are often one-sided, and their robustness and credibility urgently need improvement. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the existing technology, namely, the limited analysis dimensions, the difficulty in balancing model accuracy and interpretability, and the lack of a systematic verification framework when analyzing the factors affecting the corrosion fatigue of sucker rods. Therefore, this invention provides a quantitative evaluation method for the factors affecting the fatigue of sucker rods based on finite element simulation and dual-model fusion.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A method for quantitatively evaluating the influencing factors of sucker rod corrosion fatigue, characterized by comprising the following steps:

[0008] Step 1: Construct a high-dimensional parameterized fatigue performance dataset.

[0009] Using finite element analysis software, a digital model of a sucker rod with corrosion pits was established. By systematically changing the values ​​of multiple corrosion and load parameters, batch fatigue life simulation calculations were performed to generate a structured dataset containing multi-dimensional input features and fatigue life outputs.

[0010] Step 2: Nonlinear relation analysis based on the generalized additive model (GAM).

[0011] The dataset is fitted with a generalized additive model, and a nonlinear smooth function is fitted for each influencing parameter, thereby visually revealing the independent and interpretable nonlinear dependencies between each factor and fatigue life.

[0012] Step 3: Global and local importance analysis based on XGBoost and SHAP.

[0013] A high-precision fatigue life prediction model is trained using the Gradient Boosting Decision Tree (XGBoost) algorithm, and the trained model is interpreted using the SHAP method to obtain the global importance ranking, local dependencies, and coupling interactions of each factor.

[0014] Step 4: Fusion and comprehensive evaluation of the results from the two models.

[0015] By comparing and integrating the analysis results of steps two and three, the rationality of the SHAP analysis is verified by using the intuitive physical trends provided by GAM, and the parameter interactions revealed by XGBoost-SHAP are used to supplement the independence assumption of GAM, ultimately forming a comprehensive assessment conclusion of influencing factors.

[0016] Compared with the prior art, the present invention has the following beneficial effects:

[0017] 1. Achieves a balance between high-precision prediction and strong interpretability: This method innovatively combines a high-precision prediction model (XGBoost) with a highly interpretable model (GAM, SHAP), breaking the dilemma in existing technologies where prediction accuracy and model interpretability are difficult to balance. It can guarantee the accuracy of predicting the complex nonlinear fatigue life of sucker rods (as shown in R in this embodiment). 2 >0.98), and can completely open the "black box" of machine learning through the trend curve of GAM and the quantitative attribution of SHAP, so that the analysis results are both accurate and easy for engineers and technicians to understand, trust and apply to actual decision-making.

[0018] 2. Provides comprehensive and multi-dimensional insights into the influence of factors: Through the generation of limited metadata and dual-model analysis, this method not only provides a global importance ranking of each factor but also reveals their independent nonlinear influence patterns. More importantly, it can quantitatively reveal complex interactions between parameters that cannot be discovered by existing low-dimensional analysis methods. This comprehensive insight into main effects, nonlinear effects, and interaction effects overcomes the analytical limitations caused by computational constraints or model limitations in existing studies.

[0019] 3. A robust analysis framework with mutual verification was constructed: This method provides a systematic and reliable analytical paradigm, rather than relying on a single analytical model. By analyzing the same dataset using two models with different mechanisms, GAM and XGBoost-SHAP, the results can corroborate each other (e.g., the trend of GAM is highly consistent with the ranking of SHAP) and complement each other (the independence assumption of GAM is improved by the interaction analysis of SHAP). This built-in cross-validation mechanism greatly improves the robustness and credibility of the final evaluation conclusions, providing strong technical support for solving complex industrial safety assessment problems. Attached Figure Description

[0020] Figure 1 This is an overall method flowchart according to an embodiment of the present invention.

[0021] Figure 2 a is a partial dependence diagram of the influence of "corrosion pit location" on fatigue life generated by GAM analysis according to an embodiment of the present invention.

[0022] Figure 2 b is a partial dependence diagram of the influence of "corrosion pit diameter" on fatigue life generated by GAM analysis according to an embodiment of the present invention.

[0023] Figure 2 c is a partial dependence diagram of the influence of "sucker rod stress" on fatigue life generated by GAM analysis according to an embodiment of the present invention.

[0024] Figure 2 d is a partial dependence diagram of the influence of "corrosion pit depth" on fatigue life generated by GAM analysis according to an embodiment of the present invention.

[0025] Figure 3 This is a SHAP summary diagram generated by XGBoost-SHAP analysis according to an embodiment of the present invention, used to demonstrate the global importance of each influencing factor.

[0026] Figure 4 According to an embodiment of the present invention, a SHAP dependency graph for "sucker rod stress" is generated, and the interaction with "corrosion pit depth" is shown in color.

[0027] Figure 5 According to an embodiment of the present invention, a SHAP force map is generated for a single high-risk sample and used to predict attribution diagnosis. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this invention clearer, the following is combined with... Figures 1 to 5 The method of the present invention will be further described in detail with reference to a preferred embodiment using an H-class sucker rod.

[0029] Step S100: Construct a high-dimensional parameterized fatigue performance dataset.

[0030] This step uses finite element numerical simulation to construct a dataset of H-class sucker rod fatigue performance for data-driven modeling.

[0031] In the finite element software ANSYS Workbench, a three-dimensional model of an H-class sucker rod conforming to the API Spec 11B standard was established, and it was given the material properties of AISI 4330M high-strength steel. According to the API Spec 11B standard, the modified Goodman fatigue limit diagram corresponding to its allowable stress amplitude and mean stress was defined.

[0032] Secondly, parametric modeling is performed on the key factors affecting the fatigue life of the sucker rod. In the downhole environment, corrosion pits on the surface of the sucker rod are the main cause of fatigue fracture. In this embodiment, corrosion pits are simplified as semi-ellipsoidal geometric defects, and four key influencing factors are selected for parameterization, with five discrete levels set for each factor:

[0033] 1. Corrosion pit depth (D): Defined as the maximum radial dimension of the corrosion pit on the rod.

[0034] 2. Corrosion pit opening major axis (W): The opening size on the surface of the sucker rod.

[0035] 3. Corrosion pit location (Pos): Defined as the axial distance from the center of the corrosion pit to the center of the rod.

[0036] 4. Maximum cyclic load (Pmax): Simulates the maximum tensile load that the sucker rod experiences at the peak of the upstroke.

[0037] To comprehensively study the four parameters and their interactions, this embodiment employs a full factorial experimental design, requiring 5^4 = 625 different combinations of operating conditions. By writing Python scripts to call ANSYS's Parametric Design Language (APDL), automatic modeling, mesh generation, load application, fatigue analysis, and life result extraction for all operating conditions were achieved.

[0038] Finally, all simulation results were compiled into a structured dataset. This dataset contains 625 samples, each with 4 input features (D, W, Pos, Pmax) and 1 target output (fatigue life Nf, in terms of cycle number), providing a high-quality data foundation for the subsequent analysis steps S200 to S400.

[0039] Step S200: Nonlinear relation analysis based on the Generalized Additive Model (GAM)

[0040] This step aims to utilize the statistical tool of Generalized Additive Models (GAMs) to analyze and visualize the independent, non-linear influence trends between each influencing factor and the fatigue life of the sucker rod. The core advantage of choosing GAMs lies in their excellent interpretability; they can decompose complex black-box problems into a series of intuitive, additive functional relationships.

[0041] In this embodiment, this step is implemented in the following way:

[0042] 1. Data Preprocessing

[0043] Before model fitting, the dataset obtained in step S100 is preprocessed. Considering that fatigue life (Nf) typically spans multiple orders of magnitude and its distribution is skewed, a logarithmic transformation is performed on fatigue life to stabilize the variance and make it more consistent with the model's assumptions. Specifically, the logarithmic value log_life is calculated using the log1p(log(1+x)) function and used as the model's response variable. Furthermore, for discrete or categorical variables, such as corrosion pit locations (Pos), they are converted to categorical data types so that the model can treat them as factor terms.

[0044] 2. Model Building and Training

[0045] This embodiment uses the Python programming language and its scientific computing libraries pandas and numpy for data manipulation, and uses the pygam library to build and train GAM.

[0046] Specifically, a linear generalized additive model is constructed. The structure of the model is defined as the sum of multiple terms, including:

[0047] Factor: Corresponds to the categorical variable "location of erosion pits", which is represented as f(0) in pygam.

[0048] Spline smoothing term: Corresponds to all continuous variables, including "crater depth", "crater opening major axis", and "maximum cyclic load", and is represented as s(i) in pygam. To give the model sufficient flexibility to capture complex nonlinear modes, the number of basis functions (n_splines) for each spline term is set to 12.

[0049] The model was trained using the grid search function built into the pygam library. This function can automatically search within the range of the penalty term coefficients to find the optimal smoothness, thereby achieving the best balance between goodness of fit and overfitting.

[0050] 3. Results Analysis and Visualization

[0051] After the model is trained, a partial dependency graph is generated for each of the four influencing factors (i.e., each item in the model) to visualize their marginal effect on the response variable log_life.

[0052] Please see Figure 2 a to Figure 2 d shows the partial dependency plots of each influencing factor obtained from the GAM analysis in this embodiment. The generation process of these plots is as follows:

[0053] For continuous variables (such as pit depth), the horizontal axis of the graph represents the value of the variable, and the vertical axis represents the variable's contribution to log_life (i.e., partial dependence). The graph also includes 95% confidence intervals, represented by gray bands, to assess the uncertainty of the prediction.

[0054] For categorical variables (such as the location of corrosion pits), the graph is displayed as a bar chart, where each bar represents a location category, its height represents the average contribution of that location to log_life, and the error bars represent its 95% confidence interval.

[0055] like Figure 2 As shown in Figure d, the fatigue life of the sucker rod exhibits a significant nonlinear decreasing trend with increasing corrosion pit depth. These visualized and interpretable curves provide a physically clear benchmark for subsequent cross-validation with the XGBoost-SHAP model and offer engineers an intuitive and quantitative understanding of the impact patterns.

[0056] Step S300: Global and Local Importance Analysis Based on XGBoost and SHAP

[0057] This step aims to leverage the high prediction accuracy of the Limiting Gradient Boosting Model and combine it with the powerful attribution capabilities of the SHAP method to quantitatively assess the overall importance of each influencing factor and reveal the complex coupling interactions between them.

[0058] 1. Construction and Training of XGBoost High-Precision Prediction Model

[0059] In this embodiment, a regression prediction model is constructed using the Python programming language and its xgboost library. XGBRegressor is selected as the main model, with its objective function set to reg: squarederror. The logarithm of the fatigue life generated in step S100 (log1p(Nf)) is used as the prediction target (y), and the other four influencing parameters (corrosion pit depth, corrosion pit opening major axis, corrosion pit location, and maximum cyclic load) are used as input features (X). The dataset is randomly divided into training and test sets in an 8:2 ratio. To ensure the model's generalization ability and prediction performance, key hyperparameters are set and optimized. For example, the number of decision trees in the ensemble (n_estimators) is set to 100, the learning rate (learning_rate) to 0.1, and the maximum tree depth (max_depth) to 5. By fitting the model to the training set and validating it on the test set, the final trained prediction model achieves a determination coefficient R² of over 0.98, demonstrating that the model has extremely high prediction accuracy for sucker rod fatigue life.

[0060] 2. SHAP-based global and local interpretability analysis

[0061] After obtaining a high-precision prediction model, the shap library in Python is used to provide an in-depth interpretation of this "black box" model in order to reveal its internal decision-making logic.

[0062] First, a SHAP interpreter is initialized using the trained XGBoost model. Then, this interpreter computes the SHAP value for each sample and each feature in the dataset. The SHAP value fairly distributes the model's individual predictions across the input features, thus quantifying the contribution of each feature to the prediction.

[0063] Based on the calculated SHAP value, this embodiment performed the following three types of visualization analysis:

[0064] a) Global Feature Importance Analysis (SHAP Summary Diagram)

[0065] Please see Figure 3 The graph displays a SHAP summary plot. It aggregates the SHAP values ​​for each feature across all samples as scatter plots. The vertical axis is sorted by the average absolute SHAP value of the features, visually illustrating the global feature importance. The horizontal axis represents the SHAP value, indicating the direction and magnitude of the feature's influence on the model output (log(fatigue life)). Figure 3 As shown, the "maximum cyclic load" is the most important factor affecting fatigue life, far exceeding the importance of other factors; followed by the "corrosion pit depth" and "corrosion pit diameter"; while the "corrosion pit location" has the least impact.

[0066] b) Feature interaction analysis (SHAP dependency graph)

[0067] Please see Figure 4 This graph displays a SHAP dependency plot for "maximum cyclic load" and uses color to show the interaction with "crater depth." The horizontal axis represents the actual value of "maximum cyclic load," and the vertical axis represents its corresponding SHAP value. The graph not only reveals the negative impact of maximum cyclic load itself on lifetime (SHAP values ​​decrease with increasing stress) but also reveals the interactions between features through the color of the points (determined by the value of "crater depth"). Figure 4 As shown, in areas of high stress, the sample points representing high corrosion depth (reddish color) generally have lower SHAP values. This quantitatively demonstrates the synergistic damage effect between high load and deep corrosion pits, i.e., high load significantly amplifies the harmfulness of deep corrosion pits.

[0068] c) Single-sample predictive attribution analysis (SHAP force map)

[0069] Please see Figure 5 To attribute risks to specific operating conditions, this embodiment also generates a SHAP force map. This map explains the model's prediction process for a single high-risk sample (sample 149). The force map clearly shows the model's baseline prediction and how each feature value of this sample "pulls down" the final prediction from the baseline (blue band). This analysis has extremely high engineering application value for failure diagnosis and risk tracing in specific operating conditions.

[0070] Through the above systematic analysis based on XGBoost-SHAP, this step not only constructed a high-precision prediction model, but more importantly, it completely opened the "black box" of the model, providing quantitative global importance ranking, evidence of interaction, and in-depth insights into single-sample attribution for the subsequent S400 fusion evaluation.

[0071] Step S400: Fusion and Comprehensive Evaluation of Dual Model Results

[0072] First, this step enhances the robustness of the evaluation conclusions through consistency verification. Specifically, the GAM partial dependency graph (e.g., Figure 2 The key influencing factors and their physical trends revealed in (c) and (2d) are consistent with the SHAP abstract diagram ( Figure 3 The importance ranking and direction of influence of the two models were compared. In this embodiment, the two models reached highly consistent conclusions in identifying key factors and determining physical laws, which greatly enhanced the reliability of the evaluation results.

[0073] Building upon this foundation, complementary fusion is employed to gain deeper insights. This process utilizes the intuitive and interpretable independent influence curves provided by GAM as fundamental physical understanding, and then supplements it with higher-order interactions revealed by XGBoost-SHAP that GAM cannot represent. For example, GAM ( Figure 2 d) clearly reveals an independent nonlinear relationship between "corrosion pit depth" and fatigue life, while XGBoost-SHAP ( Figure 4 This further reveals that this relationship is strongly modulated by the "maximum cyclic load," meaning that high loads significantly amplify the harmfulness of deep corrosion pits.

[0074] Finally, based on the above cross-validation and fusion analysis, a comprehensive evaluation conclusion was formed. This conclusion clearly integrates the global importance ranking derived from the SHAP analysis (e.g., Figure 3 ), and independent nonlinear influence curves derived from GAM analysis (such as Figure 2 (a to 2d), as well as quantitative descriptions and engineering recommendations for key interactions.

[0075] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for analyzing the influencing factors of corrosion fatigue in sucker rods based on dual-model fusion, characterized in that, Includes the following steps: Step 1: Construct a high-dimensional parameterized fatigue performance dataset; establish a finite element model of a sucker rod with corrosion pits; and based on the model, perform batch fatigue life simulations by systematically changing multiple preset corrosion and load parameters, thereby generating a structured dataset containing the parameters as input features and fatigue life as output. Step 2: Nonlinear relationship analysis based on the Generalized Additive Model (GAM); The structured dataset is fitted using the Generalized Additive Model to fit a nonlinear smooth function between each input feature and fatigue life, so as to determine the independent and interpretable influence trend of each input feature; Step 3: Global and local importance analysis based on XGBoost and SHAP; a fatigue life prediction model is trained using the structured dataset with the gradient boosting decision tree algorithm; and the SHAP method is used to interpret the trained prediction model to quantitatively analyze the global importance, local dependence, and interaction between features of each input feature. Step 4: Fusion and comprehensive evaluation of the results from the two models; The global importance obtained in step three is compared with the influence trend determined in step two for cross-validation. Furthermore, by utilizing the interaction between features analyzed in step three, the independent influence trends of each input feature in step two are supplemented and corrected, ultimately forming a comprehensive evaluation conclusion on the influencing factors of the sucker rod corrosion fatigue.

2. The method according to claim 1, characterized in that, The corrosion and load parameters include: corrosion pit depth, corrosion pit opening major axis, corrosion pit location, and maximum cyclic load.

3. The method according to claim 1, characterized in that, In step one, before fitting the generalized additive model, a preprocessing step of performing a logarithmic transformation on the fatigue life output in the dataset is also included.

4. The method according to claim 1 or 2, characterized in that, In step two, the generalized additive model processes the input features as continuous variables into spline smoothing terms and the input features as categorical variables into factor terms.

5. The method according to claim 1, characterized in that, In step three, the global importance is visualized using a SHAP summary graph.

6. The method according to claim 1, characterized in that, In step three, the interaction between the features is visualized through a SHAP dependency graph, where the color of the data points in the graph represents the value of the third feature.

7. The method according to claim 1, characterized in that, In step four, the cross-validation specifically involves comparing the significance of the feature importance ranking obtained from SHAP analysis with the influence trend fitted by the generalized additive model to verify the consistency of the identification of key influencing factors.

8. The method according to claim 1, characterized in that, In step four, the supplementation and correction specifically involve: using the synergistic damage effect between the maximum cyclic load and the pit depth revealed by SHAP analysis to provide a risk supplementary explanation for the independent influence curve of pit depth given by the generalized additive model.

Images