Ship and ocean engineering structure corrosion fatigue crack growth rate prediction method based on machine learning

Abstract

The invention provides a ship and ocean engineering structure corrosion fatigue crack propagation rate prediction method based on machine learning. The method comprises the following steps: S1, obtaining all characteristic variables; s2, screening core influence factors from all feature variables through a feature selection algorithm to obtain feature data; s3, establishing a machine learning prediction model which adopts one of a plurality of machine learning algorithms, and inputting the feature data into the machine learning prediction model to learn a nonlinear mapping relationship between corrosion fatigue crack growth rate data and the selected variables, optimizing the machine learning prediction model by using a hyper-parameter optimization method to obtain a trained model; and S4, inputting feature data of a to-be-predicted target into the trained model for prediction to obtain the corrosion fatigue crack growth rate of the to-be-predicted target. The method is suitable for service safety analysis and life evaluation of key facilities such as ships, ocean platforms and submarine pipelines.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of safety assessment, in particular, especially relates to a ship and marine engineering structure corrosion fatigue crack propagation rate prediction method based on machine learning. BACKGROUND

[0002] Corrosion fatigue crack propagation is the process of crack initiation and propagation under the combined action of corrosion environment and cyclic loading, which eventually leads to the failure of the material. In the field of marine engineering and shipbuilding, metal structures are subjected to complex marine corrosion environment and frequent mechanical loading for a long time. Corrosion fatigue crack propagation is an important factor affecting the safety and reliability of the structures. Corrosion environment can accelerate crack initiation and propagation, reduce the fatigue strength of the material, and cyclic loading can cause crack propagation, eventually leading to material failure. This process involves various physical and chemical mechanisms, including corrosion of the metal surface by the corrosion medium, stress concentration at the crack tip, and microstructure evolution. Accurate prediction and evaluation of corrosion fatigue crack propagation rate is of great significance for ensuring the safe operation of marine engineering structures and ships, prolonging their service life, and reducing maintenance costs.

[0003] Corrosion fatigue crack propagation is one of the core causes of failure of ship and marine engineering structures. The traditional prediction methods based on fracture mechanics theory (such as Paris formula) and BS 7910 standard recommended curve have significant limitations: 1. Low prediction accuracy, unable to effectively handle the nonlinear relationship of mechanical, material, and environmental multi-factor coupling; 2. Poor applicability, dependent on specific materials and environmental conditions, requiring complex classification (such as stress ratio R and cathodic protection potential P segmentation); 3. Insufficient data utilization, limited processing capacity for high-dimensional feature space. With the development of machine learning technology, its application in crack propagation prediction has gradually attracted attention, but existing methods still need to be optimized in feature selection, data preprocessing, and model generalization ability. Although machine learning methods provide a new way to solve such problems, existing technologies still face challenges such as rough data preprocessing, single feature selection, and insufficient model generalization ability. Therefore, it is of great engineering significance to develop a corrosion fatigue crack propagation prediction method that integrates multi-source data, optimizes feature engineering, and improves model adaptability. SUMMARY

[0004] Therefore, the purpose of the present application is to provide a ship and marine engineering structure corrosion fatigue crack propagation rate prediction method based on machine learning to solve the technical problem of rough data preprocessing of existing prediction methods.

[0005] The technical means adopted by the present application are as follows: A ship and marine engineering structure corrosion fatigue crack propagation rate prediction method based on machine learning, comprising the following steps: S1, collect multi-dimensional data, pre-process the multi-dimensional data to obtain all characteristic variables; S2, screen core influencing factors from all characteristic variables through a feature selection algorithm to obtain feature data; S3, establish a machine learning prediction model, the machine learning prediction model adopts one of a plurality of machine learning algorithms, inputs the feature data into the machine learning prediction model to learn a nonlinear mapping relationship between the corrosion fatigue crack propagation rate data and each variable selected, and uses a hyperparameter optimization method to optimize the machine learning prediction model to obtain a trained model; S4, input the feature data of a target to be predicted into the trained model for prediction to obtain the corrosion fatigue crack propagation rate of the target to be predicted.

[0006] Further, it further comprises: S5, analysis and verification of the machine learning prediction model method and the BS 7910 method; selecting an optimal machine learning algorithm from a plurality of machine learning algorithms, predicting variables with the optimal machine learning algorithm, and predicting variables with the optimal machine learning algorithm under the condition and range of BS 7910, classifying data under different conditions, and comparing and analyzing with the BS 7910 specification condition; regression of the predicted data, and comparison and analysis of the regression curve with the BS 7910 specification recommended curve.

[0007] Further, in S1, the multi-dimensional task includes environmental parameters, mechanical conditions and material properties; the environmental parameters include temperature, medium concentration and pH value, the mechanical conditions include stress ratio, loading frequency and stress intensity factor range, and the material properties include yield strength, tensile strength and elongation; The pre-processing step is as follows: using the Z-Score standardization method, converting the multi-dimensional data into a distribution with a mean of 0 and a standard deviation of 1; using the Min-Max method to scale the data to a pre-set range; Using one-hot encoding, a new binary feature is created for each category of each discrete attribute, and for each sample, only one binary feature is 1, indicating that the sample belongs to the corresponding category, and the other features are 0; using serial number encoding to map each category of the discrete feature to an integer serial number; The missing data is filled, and the K-nearest neighbor algorithm is used to fill the missing data, and the neighboring individuals of the sample in the feature space have similar attributes.

[0008] Further, S2 specifically comprises the following steps: S21, perform Pearson correlation analysis on all characteristic variables to select a first characteristic subset; S22, Lasso importance analysis is performed on all feature variables to screen a second feature subset; S23, an intersection feature subset of the first feature subset and the second feature subset is determined; S24, SHAP importance analysis is performed on the intersection feature subset, if a preset condition is met, the intersection feature subset is taken as the feature data, if the preset condition is not met, feature variables are sequentially selected from the feature variables that have not been screened out to combine with the intersection feature subset, and SHAP importance analysis is performed again until the preset condition is met, and the finally determined feature variable set is taken as the feature data.

[0009] Further, in S3, the plurality of machine learning algorithms include a support vector regression algorithm, an artificial neural network algorithm, and an extreme gradient boosting tree algorithm; When the machine learning algorithm is the extreme gradient boosting tree algorithm, the training steps are as follows: S311, the original data set is divided into a training set and a test set; S312, a multi-dimensional hyperparameter search space is constructed, covering learning rate, tree depth, subsample rate, feature sampling rate, and regularization term; S313, based on a joint optimization framework of grid search and cross-validation, taking mean square error as the core objective function, the optimal parameter combination is screened; S314, the optimal hyperparameter configuration is used to reconstruct the model, and the generalization performance indicators are calculated on the cross-validation set and the independent test set, respectively; When the machine learning algorithm is the support vector regression algorithm, the training steps are as follows: S321, define the parameter search space, including the regularization parameter, the tolerance interval, the kernel function and the related parameters thereof; S322, use the Bayesian optimizer to optimize the model parameters of the support vector regression algorithm, set the optimization iteration number and the cross-validation strategy; S323, use the optimal parameter combination to train the SVR model, and evaluate it on the cross-validation set and the independent test set; When the machine learning algorithm is the artificial neural network algorithm, the training steps are as follows: S331, define the parameter search space of the artificial neural network algorithm model, including the number of hidden neurons, the activation function, the L2 regularization coefficient and the learning rate; S332, use the Bayesian optimizer to optimize the model parameters of the artificial neural network algorithm, set the optimization iteration coefficient and the cross-validation strategy; S333, use the optimal parameter combination to train the artificial neural network algorithm model, and evaluate it on the cross-validation set and the independent test set.

[0010] Further, S5 specifically comprises the following steps: S51, determine the applicable variable conditions and ranges according to the BS 7910 method, and select the characteristic variable combination suitable therefor; S52, input the selected characteristic variable combination into the machine learning prediction model corresponding to the optimal machine learning algorithm for prediction; S53, classify the predicted data to form multiple sets of data under different conditions; S54, perform regression analysis on each set of data to construct a regression curve; S55, compare and analyze the regression curve with the recommended curve of the BS 7910 method to verify the accuracy and reliability of the prediction model.

[0011] Further, S51 specifically comprises: S511, determine the applicable conditions of the corrosion fatigue crack propagation curve in the BS 7910 specification, including temperature range, frequency range and corrosion environment conditions; S512, set the characteristic variables based on the applicable conditions: fix the temperature variable, assign values to the frequency variable within the specified range, set the stress intensity factor in increasing order, uniformly set the material ductility, and construct the stress ratio and cathodic protection potential according to the recommended curve classification.

[0012] Further, S53 specifically comprises: S531, divide the predicted data into multiple subsets according to the stage division standard of the BS 7910 recommended curve; S532, perform nonlinear regression fitting on each subset respectively to obtain the regression curve equation of each stage.

[0013] Further, S55 specifically comprises: S551, plot the regression curve and the BS 7910 recommended curve in the same coordinate system; S552, analyze the degree of agreement between the predicted curve and the recommended curve through visual comparison to evaluate the accuracy of the prediction model; S553, calculate the error index between the predicted curve and the recommended curve to further verify the reliability of the prediction model.

[0014] Compared with the prior art, the present application has the following advantages: The application establishes a database by the corrosion fatigue crack propagation rate data of 10 kinds of material types of steel 2015 group in the field of ships, offshore platforms and submarine pipelines, and the main influencing factors of corrosion fatigue crack propagation are not comprehensive, so the mechanical factors (stress intensity factor range, loading frequency, loading stress ratio, loading waveform, pre-crack length) that affect the corrosion fatigue crack propagation rate, material factors (yield strength, tensile strength, elongation) and environmental factors (cathode potential, temperature, PH) are considered.

[0015] The application selects the characteristic variables that play a leading role in corrosion fatigue crack propagation by Pearson correlation analysis, cable importance analysis and SHAP importance analysis, as the input variables of the machine learning model, which can greatly improve the accuracy and robustness of the model.

[0016] The machine learning model proposed in the application is compared and analyzed with the recommended method of the traditional BS7910 specification, and the applicability and accuracy of the machine learning model of the application are verified. In addition, in exploring the corrosion fatigue crack propagation law of high-strength steel materials, the machine learning model of the application avoids the classification of stress ratio and the specialization of cathode protection potential, and compared with the BS7910 specification, it shows less limitation and can provide technical support for practical engineering application. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the application or the prior art, the drawings needed in the embodiment or prior art description will be briefly introduced below. Obviously, the drawings in the following description are some embodiments of the application, and those skilled in the art can also obtain other drawings according to these drawings without creative labor.

[0018] Figure 1 The method flowchart of the application.

[0019] Figure 2 The SHAP importance analysis importance percentage chart of the application.

[0020] Figure 3 The evaluation index chart of three algorithms of the application. DETAILED DESCRIPTION

[0021] In the following, the technical solutions in the embodiments of the present application will be described clearly and completely with reference to the drawings in the embodiments of the present application, so that those skilled in the art can better understand the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work should fall within the scope of protection of the present application.

[0022] It should be noted that the terms "first", "second", and the like in the specification and claims of the present application and the above-described drawings are used to distinguish similar objects, and do not necessarily have to be used to describe a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not have to be limited to only those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0023] As shown in Figure 1 The present application provides a ship and marine engineering structure corrosion fatigue crack propagation rate prediction method based on machine learning, comprising the following steps: S1, collecting multi-dimensional data, pre-processing the multi-dimensional data to obtain all characteristic variables; The multi-dimensional task includes environmental parameters, mechanical conditions and material properties; the environmental parameters include temperature, medium concentration and PH value, the mechanical conditions include stress ratio, loading frequency and ΔK range, and the material properties include yield strength, tensile strength and elongation; The present application considers 11 characteristic variables including stress intensity factor range (ΔK), loading frequency (F), loading stress ratio (R), cathode potential (P), temperature (T), pH, loading waveform (W), pre-crack length (a), yield strength (YS), tensile strength (TS) and elongation (E) as variables affecting corrosion crack propagation rate. The names, parameter ranges and parameter units of input parameters and output parameters are shown in Table 1.

[0024] Table 1: Comparison of input and output parameters

[0025] The pre-processing step is as follows: Data cleaning: In the data cleaning process, based on the influence of the physical level of corrosion fatigue crack propagation, due to the corrosion resistance of stainless steel material and the characteristics of not being commonly used in ship and ocean engineering structure, the corrosion crack propagation rate data of stainless steel is removed. Compressive stress (R<0) will inhibit the corrosion fatigue crack propagation. Therefore, the corrosion fatigue crack propagation data of steel under the condition of stress ratio less than 0 is removed, and the remaining corrosion fatigue crack propagation rate data is normally retained for model training and testing.

[0026] The Z-Score standardization method is used to convert multi-dimensional data into a distribution with a mean of 0 and a standard deviation of 1; the Min-Max method is used to scale the data to a predetermined range; Before normalization, the loaded waveform is converted to numerical data, and the waveform features (sine wave, triangle wave, square wave) processed by the application have sequentiality, so ordinal encoding is used. Z-Score standardization: , Where μ is the sample mean of the feature column, and σ is the sample standard deviation of the feature column. After standardization, the mean of the data feature is 0 and the standard deviation is 1. The influence of the dimension difference of different variables on the machine learning model is eliminated through standardization.

[0027] One-hot encoding is used to create a new binary feature for each category of each discrete attribute. For each sample, only one binary feature is 1, indicating that the sample belongs to the corresponding category, and the other features are 0; ordinal encoding is used to map each category of the discrete feature to an integer sequence number; The missing data is filled by using K-nearest neighbor algorithm (KNN) to fill the missing data. This method is based on the distance measurement between samples, and uses the feature values of the nearest samples to interpolate the missing values. Three nearest neighbors are selected and weighted in inverse proportion to the distance. The filled data set eliminates the limitation of data missing on model training, enhances the universality and compatibility of the model, and provides guarantee for the effective operation of subsequent algorithms.

[0028] S2, filtering core influencing factors from all feature variables by a feature selection algorithm to obtain feature data; S21, performing Pearson correlation analysis on all feature variables to select a first feature subset; In machine learning regression problems, Pearson correlation coefficient is used to measure the linear relationship between variables. When |r|≥0.6 between two variables, it indicates that there is a strong linear correlation. At this time, if both are introduced into the regression model, it may cause multicollinearity problem, affecting the stability and statistical significance of regression coefficient estimation.

[0029] The present application considers various factors affecting the corrosion crack propagation rate, and there may be a high linear relationship between variables, which needs to be screened by correlation analysis. Therefore, the heat map is used to visualize the Pearson correlation coefficient between variables. The tensile strength (TS) is strongly positively correlated with the yield strength (YS) (r=0.87); the correlation coefficient of temperature (T) with TS and YS is 0.66; the pre-crack length (a) is negatively correlated with the elongation (E) (r=-0.69); and the pH value is negatively correlated with the TS (r=-0.68). Therefore, the highly correlated variables need to be screened before modeling to improve the robustness of the model.

[0030] S22, Lasso importance analysis is performed on all feature variables to screen a second feature subset; Lasso regression realizes feature selection and model simplification by introducing L1 regularization, can effectively handle high-dimensional data and multicollinearity problems. It retains key variables by compressing or even setting to zero the coefficients of unimportant features, thereby improving the prediction performance and interpretability of the model. The present application uses Lasso regression to analyze the importance of each variable on the corrosion crack propagation rate, and visually displays the variable importance through a variable importance chart.

[0031] S23, an intersection feature subset of the first feature subset and the second feature subset is determined; S24, SHAP importance analysis is performed on the intersection feature subset, if a preset condition is met, the intersection feature subset is taken as the feature data; if the preset condition is not met, feature variables are selected from the unfiltered feature variables in turn, combined with the intersection feature subset, and SHAP importance analysis is performed again until the preset condition is met, and the finally determined feature variable set is taken as the feature data.

[0032] SHAP can intuitively show the contribution of each feature to the prediction result of the machine learning model. Compared with the traditional gain importance, SHAP effectively avoids misjudgment caused by data noise or multicollinearity by fairly allocating feature contributions, and provides consistent and unbiased feature importance measurement.

[0033] The importance of the feature is obtained by calculating the absolute mean of the SHAP value of each feature in all samples. The vertical axis is the feature, arranged from high to low according to the importance, and the horizontal axis is the SHAP value, representing the positive and negative contribution of the feature to the prediction of a single sample. The importance is based on the sum of the absolute values of the SHAP values of all samples, and the feature importance is determined by the position on the vertical axis, the higher the feature, the greater the influence on the overall prediction of the model.

[0034] The importance percentage of the SHAP importance analysis is Figure 2The importance score of the first six characteristic variables (ΔK, F, R, P, T, and E) can be 98.7%, which can represent the influence of 11 characteristic variables on the corrosion fatigue crack propagation rate. The method effectively reveals the key variables and their influence, and provides a reliable basis for model optimization and theoretical analysis. Therefore, the final selected characteristic inputs are shown in Table 2.

[0035] Table 2 Parameter information table after feature selection

[0036] S3, a machine learning prediction model is established, which uses one of several machine learning algorithms, inputs the characteristic data into the machine learning prediction model to learn the nonlinear mapping relationship between the corrosion fatigue crack propagation rate data and the selected variables, and uses a hyperparameter optimization method to optimize the machine learning prediction model to obtain a trained model; In S3, the several machine learning algorithms include a support vector regression algorithm, an artificial neural network algorithm, and an extreme gradient boosting tree algorithm. When the machine learning algorithm is an extreme gradient boosting tree algorithm, the training steps are as follows: S311, dividing the original data set into a training set and a test set; The original data set is divided into a training set (80%) and an independent test set (20%) according to a ratio of 4:1, and the reproducibility of the experiment is ensured by controlling random_state=42. When initializing the XGB regressor, set the baseline parameters (n_estimators=100) to constrain the model complexity and avoid initial overfitting.

[0037] S312, constructing a multi-dimensional hyperparameter search space covering learning rate, tree depth, subsample rate, feature sampling rate, and regularization term; A multi-dimensional hyperparameter search space (param_grid) is constructed, covering learning rate (learning_rate), tree depth (max_depth), subsample rate (subsample), feature sampling rate (colsample_bytree), and regularization term (gamma, reg_alpha, reg_lambda), and the value range is determined based on the Bayesian optimization prior distribution. Based on the joint optimization framework of grid search (GridSearch) and five-fold cross-validation (5-FoldCV), the mean square error (MSE) is used as the core objective function (refit=’MSE’), and the multi-dimensional indicators such as explained variance (EVS), mean absolute error (MAE), and coefficient of determination , are simultaneously evaluated to screen the optimal parameter combination, and the corrosion fatigue crack propagation rate is predicted by the model with the optimal parameter combination.

[0038] S313, based on the joint optimization framework of grid search and cross validation, taking mean square error as the core objective function, screening the optimal parameter combination; The XGB model is reconstructed by using the optimal hyperparameter configuration, and the generalization performance indicators are calculated on the cross-validation set and the independent test set respectively. Through comparative analysis, the robustness of the model in extrapolating the training data is revealed. The prediction value-real value scatter plot and residual distribution plot are generated to quantitatively evaluate the bias-variance balance characteristics of the model in the nonlinear regression task, wherein >0.9 indicates that the model has strong representation ability for the multi-physical field coupling mechanism of the corrosion fatigue crack propagation rate.

[0039] S314, the model is reconstructed by using the optimal hyperparameter configuration, and the generalization performance indicators are calculated on the cross-validation set and the independent test set respectively; When the machine learning algorithm is a support vector regression algorithm, the training steps are as follows: S321, define the parameter search space, including the regularization parameter, the tolerance interval, the kernel function and its related parameters; The parameter search space of the SVR model involves multiple key parameters, and the regularization parameter C, the tolerance interval epsilon, the kernel function kernel and its related radial basis function (RBF) parameter gamma are optimized. By defining the parameter search space (svr_param_space), the range is defined for subsequent parameter search. The regularization parameter C is used to balance the fitting ability and complexity of the model, to avoid overfitting or underfitting; the tolerance interval epsilon determines the error range allowed by the model; the kernel function kernel and its related parameter gamma affect the mapping and fitting method of the model to the data.

[0040] S322, the model parameters of the support vector regression algorithm are optimized by using the Bayesian optimizer, and the optimization iteration number and the cross-validation strategy are set; The Bayesian optimizer is used to optimize the parameters of the SVR model. The SVR model is instantiated (estimator=SVR()), and the defined parameter search space (search_space=svr_param_space) is passed in. The number of optimization iterations n_iter is set to 50, which means that the Bayesian optimizer will perform 50 iterations of search within the parameter search space to find the optimal parameter combination. At the same time, 5-fold cross-validation (cv=5) is used to evaluate the performance of the model under different parameter combinations. By repeatedly dividing the training set for training and validation, the generalization performance of the model can be more robustly estimated. The evaluation metric is the negative mean squared error (Scoring=’neg_mean_squared_error’), which aims to minimize the mean squared error between the predicted value and the true value, and the random seed (random_state=42) is set to ensure the reproducibility of the experiment.

[0041] S323, train the SVR model using the optimal parameter combination and evaluate it on the cross-validation set and the independent test set; Run the Bayesian optimizer to get the optimal parameter combination through iterative search. Use the optimal parameter combination to train the SVR model to make full use of the information in the data to build the best-performing model. After training, the model is evaluated on the 5-fold cross-validation set and the independent test set, respectively. The coefficient of determination (R2), explained variance score (EVS), mean absolute error (MAE), and mean squared error (MSE) are used as evaluation metrics. The optimal parameter combination is used to predict the corrosion fatigue crack growth rate.

[0042] When the machine learning algorithm is an artificial neural network algorithm, the training steps are as follows: S331, define the parameter search space of the model of the artificial neural network algorithm, including the number of hidden neurons, activation function, L2 regularization coefficient, and learning rate; The parameter search space defines that the performance of the MLPRegressor model is highly sensitive to multiple parameters. This example focuses on the optimization exploration of the number of hidden neurons (hidden_layer_sizes), the activation function (activation), the L2 regularization coefficient (alpha), and the learning rate (learning_rate_init). By defining the parameter search space (mlp_param_space), a clear range is set for subsequent parameter optimization. The number of hidden neurons is directly related to the complexity of the neural network hidden layer, and has a significant impact on the learning potential of the model. The activation function injects nonlinearity into the model, enabling it to learn complex function mapping relationships. The L2 regularization coefficient serves as a control factor for model complexity, effectively suppressing overfitting. The learning rate determines the size of the parameter update step during model training, affecting the stability and convergence speed of the training.

[0043] S332, the model parameters of the artificial neural network algorithm are optimized by using the Bayesian optimizer, and the optimization iteration coefficient and the cross-validation strategy are set; The Bayesian optimizer is used to optimize the parameters of the MLPRegressor model. When instantiating the MLPRegressor model, the early stopping mechanism (early_stopping=True) is enabled to avoid the risk of overfitting during training. The maximum number of iterations (max_iter) is set to 1000, and the random seed (random_state=42) is specified to ensure the reproducibility of the experimental results. Then, the pre-defined parameter search space (mlp_param_space) is passed to the optimizer. The optimization iteration coefficient n_iter is set to 30, i.e., the Bayesian optimizer will perform 30 iterations of search within the parameter search space to find the optimal parameter combination. The 5-fold cross-validation (cv=5) strategy is used to evaluate the performance of the model under different parameter settings. By repeatedly dividing the training set for training and validation, the robustness of the model's generalization performance is achieved. The negative mean squared error (scoring='neg_mean_squared_error') is selected as the evaluation metric, aiming to minimize the mean squared error between the predicted values and the true values. At the same time, the random seed (random_state=42) is fixed to ensure the consistency and reproducibility of the optimization process.

[0044] S333, the model is trained using the optimal parameter combination, and the performance is evaluated on the cross-validation set and the independent test set.

[0045] The Bayesian optimizer is started to determine the optimal parameter combination through multiple rounds of iterative search. Then, the MLPRegressor model is trained using the optimal parameter combination to fully extract the effective information in the data and build a high-performance prediction model. After training, the model is comprehensively evaluated on the 5-fold cross-validation set and the independent test set. The coefficient of determination (R2), the explained variance score (EVS), the mean absolute error (MAE), and the mean squared error (MSE) are selected as evaluation indicators. Among them, R2 and EVS are used to measure the goodness of fit of the model to the data, and the closer the value is to 1, the better the model fitting effect; MAE and MSE are used to measure the error degree between the predicted value and the true value, and the smaller the value, the higher the model prediction accuracy. The optimal parameter combination model is used to predict the corrosion fatigue crack propagation rate.

[0046] S4, the feature data of the target to be predicted is input into the trained model for prediction to obtain the corrosion fatigue crack propagation rate of the target to be predicted.

[0047] S5, analysis and verification of the machine learning prediction model method and the BS 7910 method; select the optimal machine learning algorithm from several machine learning algorithms, use the machine learning prediction model with the optimal machine learning algorithm to predict the variables that adapt to the variable conditions and range of BS 7910, classify the data under different conditions, and compare and analyze with the BS 7910 specification conditions; regression of the predicted data, and comparison and analysis of the regression curve with the BS 7910 specification recommended curve.

[0048] S51, determine the applicable variable conditions and range according to the BS 7910 method, and select the feature variable combination that adapts to it; S511, determine the applicable conditions of the corrosion fatigue crack propagation curve in the BS 7910 specification, including temperature range, frequency range, and corrosion environment conditions; S512, set the feature variables based on the applicable conditions: fix the value of the temperature variable, assign values to the frequency variable within the specified range, set the stress intensity factor in an increasing sequence, uniformly set the material ductility, and construct the stress ratio and cathodic protection potential according to the recommended curve classification.

[0049] S52, input the selected feature variable combination into the machine learning prediction model corresponding to the optimal machine learning algorithm for prediction; The selection method of the optimal machine learning algorithm is as follows: explained_variance_score: the explained variance score of the regression model, the value range is [0, 1], the closer to 1, the more the independent variable can explain the variance change of the dependent variable, the smaller the value, the worse the effect. Mean absolute error (MAE): evaluate the closeness of the predicted results and the real data set, the smaller the value, the better the fitting effect. Mean absolute error (MAE) is the most understandable regression error indicator. We will calculate the residual for each data point, only take the absolute value of each residual, so that negative residuals and positive residuals will not be offset. Mean squared error (MSE): this indicator calculates the mean of the sum of squares of the error of the fitted data and the original data corresponding sample points, the smaller the value, the better the fitting effect. R 2 : the explained variance score of the regression model, the value range is [0, 1], the closer to 1, the more the independent variable can explain the variance change of the dependent variable, the smaller the value, the worse the effect.

[0050] Figure 3 The performance of XGB, SVR and MLP algorithms under different evaluation indicators is shown, and the visualization results of the three evaluation indicators are combined. When evaluating the performance of the model for the data set of the application, the performance of the three algorithms is ranked as XGB>SVR>MLP. Therefore, the XGB model is selected for subsequent analysis and verification.

[0051] S53, classify the predicted data to form multiple data sets under different conditions; S531, according to the stage division standard of the BS 7910 recommended curve, the predicted data is divided into multiple subsets; S532, nonlinear regression fitting is performed on each subset respectively to obtain the regression curve equation of each stage.

[0052] S54, regression analysis is performed on each group of data to construct a regression curve; S55, compare and analyze the regression curve with the recommended curve of the BS 7910 method to evaluate the accuracy and reliability of the prediction model.

[0053] S551, draw the regression curve and the BS 7910 specification recommended curve in the same coordinate system; S552, by visual comparison, analyze the degree of agreement between the predicted curve and the recommended curve to evaluate the accuracy of the prediction model; S553, calculate the error indicators between the predicted curve and the recommended curve to further verify the reliability of the prediction model.

[0054] S5 specific steps as follows: 1, the BS790 crack propagation model conditions selection; Research to six characteristic variables (stress intensity factor range ΔK, loading frequency F, loading stress ratio R, cathode potential P, temperature T, elongation E) to build input characteristic system. In view of the BS7910 curve applicable conditions for 5-10ff temperature interval, 0.17-0.5Hz frequency range and 3% NaCl solution corrosion environment, the application of variable is set as follows: temperature variable T fixed value is 10; the frequency F is assigned in the range of 0.17-0.5Hz; the stress intensity factor ΔK is set in the increasing sequence of 9-63, which reflects the stress intensity change gradient; the material ductility E is uniformly set to 24, which represents the inherent characteristics of the material. The stress ratio R and the cathode protection potential P are constructed according to the BS7910 recommended curve classification: the stress ratio R is divided into two categories of R<0.5 and R≥0.5; the cathode protection potential P is divided into three grades of free corrosion potential-680mV, -850mV and-1100mV. Based on the above classification combination, six groups of characteristic data are generated, which are stored in six independent Excel tables as the standardized input data set of XGB model.

[0055] 2, corrosion fatigue crack propagation mean curve construction; Data driven method is adopted to construct the prediction model: first, read the characteristic data in the Excel table, input it into the XGB model after parameter optimization for prediction. According to the BS7910 recommended curve A, B stage turning point threshold (threshold), the prediction data is divided into two subsets: A stage data is stored in group_below_threshold list, B stage data is stored in group_above_threshold list. Then, the least square method is used to carry out nonlinear regression fitting for A, B stage data respectively, and the regression curve equation and curve of each stage are obtained. The regression curve analytical expression is output, which is used for subsequent quantitative comparison analysis with the slope of BS7910 recommended curve. The intersection coordinates of the two regression curves are determined by numerical calculation, which is used as the boundary to draw the segmented prediction curve: the A stage regression curve is used before the intersection point, and the B stage regression curve is switched to after the intersection point, forming a complete prediction mean curve.

[0056] 3, corrosion fatigue crack propagation two standard deviation curve construction; Based on the statistical characteristics of the mean regression curve, combined with the mean and standard deviation of the predicted value, the two standard deviation expressions of the corrosion fatigue curve are derived. By solving the equation, the intersection of the two standard deviation curves of A and B stages is determined, and the curve segmentation is completed. Using the scientific computing and visualization library of Python, the mean curve and the two standard deviation curves are plotted in the double logarithmic coordinate system, and the corresponding environmental parameters (such as XGB prediction P=850mV, R<0.5) and curve types (Mean curve, Mean+2SD) are marked. Further, the XGB prediction data points under six working conditions, the BS7910 recommended curve and the prediction curve are integrated in the same coordinate system, and the prediction accuracy and reliability of the XGB model are verified by visualization comparison.

[0057] 4. Comparison of machine learning model prediction results and BS7910 recommended curve; The BS7910 recommended steel fatigue crack propagation curve in marine environment is divided into R<0.5 and R>=0.5 according to different stress ratios R, and corrosion potential, respectively, free corrosion potential, the average potential of free corrosion-680mv, cathodic protection potential-850mv and cathodic protection potential-1100mv, combined to form corrosion crack propagation curves under six corrosion environments. The recommended curve divides the curve into two stages, stage A and stage B, and gives the stress intensity factor range at the turning point of stage A and stage B under each condition. In addition, the recommended curve not only gives the mean curve under six conditions, but also gives the curve of two standard deviations. According to the above characteristics, the data to be predicted is also divided into six conditions. For each condition, the stress intensity factor will increase in a certain interval, and the extension, temperature and retention will remain unchanged, the stress ratio and corrosion potential will change with the situation, and the corrosion frequency will also be selected from multiple values within the given frequency range of BS7910 recommended curve. The corrosion fatigue crack propagation data of BS7910 given environment predicted by XGB optimal model is fitted into a curve by least square method, and the corrosion fatigue crack propagation curve predicted by XGB algorithm is compared with the BS7910 recommended steel fatigue crack propagation curve in marine environment to verify the prediction accuracy and reliability.

[0058] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A machine learning based method for predicting corrosion fatigue crack growth rate of a marine vessel and offshore engineering structure, characterized in that, The method comprises the following steps: S1, collecting multi-dimensional data, preprocessing the multi-dimensional data to obtain all characteristic variables; S2, screening core influencing factors from all characteristic variables through a feature selection algorithm to obtain characteristic data; S3, establishing a machine learning prediction model, the machine learning prediction model adopts one of a plurality of machine learning algorithms, inputs the characteristic data into the machine learning prediction model to learn the nonlinear mapping relationship between the corrosion fatigue crack propagation rate data and the selected variables, and uses a hyperparameter optimization method to optimize the machine learning prediction model to obtain a trained model; S4, inputting the characteristic data of a target to be predicted into the trained model for prediction to obtain the corrosion fatigue crack propagation rate of the target to be predicted.

2. The machine learning based ship and offshore structure corrosion fatigue crack growth rate prediction method according to claim 1, characterized in that, Further comprising: S5, analysis and verification of the machine learning prediction model method and the BS 7910 method; selecting an optimal machine learning algorithm from a plurality of machine learning algorithms, predicting variables within the range and condition of BS 7910 with the machine learning prediction model with the optimal machine learning algorithm, classifying data under different conditions, and comparing and analyzing with the BS 7910 specification condition; regression of the predicted data, and comparison and analysis verification of the regression curve and the BS 7910 specification recommended curve.

3. The machine learning based ship and offshore structure corrosion fatigue crack growth rate prediction method according to claim 1, characterized in that, In S1, the multi-dimensional task includes environmental parameters, mechanical conditions and material properties; the environmental parameters include temperature, medium concentration and PH value, the mechanical conditions include stress ratio, loading frequency and stress intensity factor range, and the material properties include yield strength, tensile strength and elongation; The preprocessing step is as follows: using the Z-Score standardization method to convert the multi-dimensional data into a distribution with a mean of 0 and a standard deviation of 1; using the Min-Max method to scale the data to a preset range; Using one-hot encoding, a new binary feature is created for each category of each discrete attribute, and for each sample, only one binary feature is 1, indicating that the sample belongs to the corresponding category, and the other features are 0; using serial number encoding to map each category of the discrete feature to an integer serial number; The missing data is filled, and the K-nearest neighbor algorithm is used for filling, and the neighboring individuals of the sample in the feature space have similar attributes.

4. The machine learning based ship and offshore structure corrosion fatigue crack growth rate prediction method of claim 1, wherein, S2 specifically comprises the following steps: S21, performing Pearson correlation analysis on all characteristic variables to select a first characteristic subset; S22, performing Lasso importance analysis on all characteristic variables to select a second characteristic subset; S23, determining the intersection characteristic subset of the first characteristic subset and the second characteristic subset; S24, performing SHAP importance analysis on the intersection characteristic subset, if the preset condition is met, the intersection characteristic subset is taken as the characteristic data; if the preset condition is not met, the characteristic variables are selected from the characteristic variables that have not been screened out in turn and combined with the intersection characteristic subset, and SHAP importance analysis is performed again until the preset condition is met, and the finally determined characteristic variable set is taken as the characteristic data.

5. The machine learning based ship and offshore structure corrosion fatigue crack growth rate prediction method according to claim 1, characterized in that, In S3, the machine learning algorithms include a support vector regression algorithm, an artificial neural network algorithm, and an extremely gradient boosting tree algorithm. When the machine learning algorithm is the extremely gradient boosting tree algorithm, the training steps are as follows: S311, divide the original data set into a training set and a test set; S312, construct a multi-dimensional hyperparameter search space covering learning rate, tree depth, subsample rate, feature sampling rate, and regularization term; S313, based on a joint optimization framework of grid search and cross-validation, take mean square error as the core objective function, and screen the optimal parameter combination; S314, reconstruct the model using the optimal hyperparameter configuration, and calculate the generalization performance indicators on the cross-validation set and the independent test set, respectively; When the machine learning algorithm is the support vector regression algorithm, the training steps are as follows: S321, define the parameter search space, including the regularization parameter, the tolerance interval, the kernel function, and the related parameters thereof; S322, use a Bayesian optimizer to optimize the model parameters of the support vector regression algorithm, set the optimization iteration number and the cross-validation strategy; S323, use the optimal parameter combination to train the SVR model, and evaluate it on the cross-validation set and the independent test set; When the machine learning algorithm is the artificial neural network algorithm, the training steps are as follows: S331, define the parameter search space of the artificial neural network algorithm, including the number of hidden neurons, the activation function, the L2 regularization coefficient, and the learning rate; S332, use a Bayesian optimizer to optimize the model parameters of the artificial neural network algorithm, set the optimization iteration coefficient and the cross-validation strategy; S333, use the optimal parameter combination to train the artificial neural network algorithm, and evaluate it on the cross-validation set and the independent test set.

6. The machine learning based ship and offshore structure corrosion fatigue crack growth rate prediction method according to claim 2, characterized in that, S5 specifically includes the following steps: S51, determine the applicable variable conditions and ranges according to the BS 7910 method, and select the feature variable combination suitable therefor; S52, input the selected feature variable combination into the machine learning prediction model corresponding to the optimal machine learning algorithm for prediction; S53, classify the predicted data to form multiple data sets under different conditions; S54, perform regression analysis on each data set to construct a regression curve; S55, compare and analyze the regression curve with the recommended curve of the BS 7910 method to verify the accuracy and reliability of the prediction model.

7. The machine learning based ship and offshore structure corrosion fatigue crack growth rate prediction method according to claim 6, characterized in that, S51 specifically includes: S511, determine the applicable conditions of the corrosion fatigue crack propagation curve in the BS 7910 specification, including the temperature range, the frequency range, and the corrosion environment conditions; S512, set the feature variables based on the applicable conditions: fix the temperature variable, assign values to the frequency variable within the specified range, set the stress intensity factor in an increasing sequence, uniformly set the material ductility, and construct the stress ratio and cathodic protection potential according to the recommended curve classification.

8. The machine learning based ship and offshore structure corrosion fatigue crack growth rate prediction method according to claim 6, characterized in that, S53 specifically includes: S531, divide the predicted data into multiple subsets according to the stage division standard of the BS 7910 recommended curve; S532, perform nonlinear regression fitting on each subset to obtain the regression curve equation of each stage.

9. The machine learning based ship and offshore structure corrosion fatigue crack growth rate prediction method of claim 6, wherein, S55 specifically includes: S551、Draw the regression curve and the BS 7910 recommended curve in the same coordinate system; S552, by visual comparison, analyze the degree of agreement between the predicted curve and the recommended curve, and evaluate the accuracy of the prediction model; S553, calculate the error index between the predicted curve and the recommended curve, and further verify the reliability of the prediction model.

Images