SHAP-based enhanced neural network prediction model and weather-resistant steel atmospheric corrosion rate prediction method

Abstract

The invention belongs to the technical field of metal material corrosion informatization, and particularly relates to an enhanced neural network prediction model based on SHAP and a weather-resistant steel atmospheric corrosion rate prediction method. Wherein the sample data of the weathering resistant steel to be tested comprises determined key features in the enhanced neural network prediction model based on SHAP; and inputting the preprocessed data into the enhanced neural network prediction model based on the SHAP to obtain the corrosion rate of the weathering resistant steel to be detected. The method not only effectively solves the problems of insufficient prediction precision, high cost, difficult evaluation of long-term performance and the like in a traditional method, but also provides reliable data support and decision basis for the weathering resistant steel material in the field of practical engineering application, and can be suitable for rapid and reliable evaluation of the long-term service performance of the weathering resistant steel.

Description

Technical Field

[0001] This invention belongs to the field of information technology for corrosion of metallic materials, specifically involving an enhanced neural network prediction model based on SHAP and a method for predicting the atmospheric corrosion rate of weathering steel. Background Technology

[0002] Weathering steel possesses excellent resistance to atmospheric corrosion and superior mechanical properties, making it widely used in engineering structures exposed to the atmosphere for extended periods, such as photovoltaic supports, bridge structures, railway tracks, and outdoor facilities. However, its corrosion rate is subject to complex interactions influenced by factors such as alloy composition, environmental conditions, and exposure time. Traditional exposure testing methods are time-consuming, costly, and difficult to predict efficiently. Furthermore, existing prediction models primarily employ simple linear or nonlinear regression, which struggles to capture the complex nonlinear variations in corrosion rates, resulting in limited prediction accuracy.

[0003] Therefore, there is an urgent need to develop an accurate, rapid, and efficient method for predicting the atmospheric corrosion rate of weathering steel in order to assess its long-term performance. Summary of the Invention

[0004] To address the shortcomings of the existing technologies, this invention provides an enhanced neural network prediction model based on SHAP and a method for predicting the atmospheric corrosion rate of weathering steel, thereby solving the problems of low accuracy, high cost, and difficulty in accurately predicting the long-term service performance of traditional weathering steel in complex environments.

[0005] This invention provides, in one aspect, an enhanced neural network prediction model based on SHAP, the method for constructing this prediction model specifically including: S11. Collect long-term exposure test data of weathering steel from different sources and merge them into a dataset, which includes alloy element data, environmental factors, exposure time and corrosion rate. S12. Preprocess the dataset, including data cleaning and feature engineering. S13. Construct a neural network model and train the neural network model based on the preprocessed dataset; S14. Set the hyperparameters and optimize them based on the trained neural network model and optimization strategy to obtain the optimal hyperparameters; S15. Keeping the optimal hyperparameters determined in step S14, perform feature importance analysis on the trained neural network model, determine key features, and reconstruct and train the initial augmented neural network prediction model based on SHAP based on the key features. S16. Use independent data to evaluate the accuracy of the initial SHAP-based augmented neural network prediction model after training. After passing the accuracy evaluation criteria, the SHAP-based augmented neural network prediction model is obtained.

[0006] Preferably, the alloying elements of weathering steel include carbon, sulfur, phosphorus, manganese, silicon, chromium, nickel, molybdenum, copper, niobium, vanadium, titanium, aluminum and rhenium; environmental factors include temperature, relative humidity, wetting time, sulfur dioxide deposition rate and chloride ion deposition rate.

[0007] Preferably, data cleaning includes outlier and missing value detection and processing, wherein outlier processing adopts the interquartile range method and the empirical threshold method, and missing value processing adopts the method of directly deleting missing values; Feature engineering processing includes: performing maximum and minimum normalization on temperature and relative humidity among environmental factors; standardizing the humidification time among environmental factors to [0,1] based on physical meaning; and performing logarithmic transformation on the exposure time, where the logarithmic transformation of the exposure time is shown in the following formula: , In the formula, Represents the eigenvalues ​​after logarithmic transformation. X Indicates the original exposure time.

[0008] Preferably, the model constructed in steps S13 and S15 includes one input layer, three hidden layers, and one output layer; the number of points in the three hidden layers decreases layer by layer, all of which use the Rectified Linear Unit activation function, and a random deactivation regularization layer is set between the second and third hidden layers; the output layer has 1 node and uses a linear activation function; wherein, the input layer of the model constructed in step S13 has 20 points, and the input layer of the model constructed in step S15 has 14 points; In addition, the five-fold cross-validation method is used to train the model in both steps S13 and S15.

[0009] Preferably, the hyperparameters set in step S14 include: number of training rounds, batch size per training session, and learning rate, wherein the combination range of hyperparameters is: 1000 to 2000 training rounds, batch size per training session is 32 to 128, and learning rate is 0.0001 to 0.001. Furthermore, during the hyperparameter optimization process, the ultimate goal is to achieve the best prediction performance of the trained neural network model on the validation set. Based on the optimization strategy adopted, the specific objective function is determined, and the trained neural network model is trained and evaluated in multiple rounds to search for and determine the optimal combination of hyperparameters.

[0010] The preferred and optimal combination of hyperparameters is: 1500 training epochs, batch size of 64 per training epoch, and learning rate of 0.0003. Furthermore, during the hyperparameter optimization process in step S14: When the dimension of the hyperparameter space is low, a grid search strategy is used to traverse all preset hyperparameter combinations. In this process, the optimal hyperparameter combination is determined by minimizing the average value of the loss functions in cross-validation. The objective function is shown below: , In the formula, Indicates the search space of hyperparameters Selecting the values ​​that result in the average loss of the model in K-fold cross-validation. The smallest combination of hyperparameters, For folds, Respectively The validation set input features and the true labels are folded; Indicating the combination of hyperparameters Below is the model's predicted output based on the input data; When the hyperparameter space has a high dimensionality, a Bayesian optimization method is used, specifically by constructing a surrogate function. To simulate the mapping relationship between hyperparameter combinations and model performance, and based on the acquisition function This guides the selection of the next set of hyperparameters, thereby efficiently finding the optimal parameter combination with minimal model evaluations. The objective function is shown below: , In the formula, This represents the set of hyperparameter-performance pairs that have been evaluated so far. To obtain the function used to evaluate candidate hyperparameter combinations Potential optimization value under the proxy model.

[0011] Preferably, in step S15, the SHAP method is used to perform feature importance analysis on the trained neural network model. By calculating the average absolute SHAP value of each feature across all samples, key features that significantly contribute to the trained neural network model are identified. The formula for calculating the SHAP value of each feature in each sample is as follows: , In the formula, Representation of features The SHAP value, where M represents the total number of features and S represents the number of features not included. Feature subset, This represents the model's predicted output based on the input feature subset S. Indicates features The prediction results of the model after adding to subset S. Represents all possible feature subsets Summation; The formula for calculating the mean absolute SHAP value of each feature across all samples is as follows: , In the formula, Representation of features The mean absolute SHAP value, where N is the total number of samples. This represents the SHAP value of the feature in the i-th sample.

[0012] Preferably, in step S15, 3% of the maximum average absolute SHAP value among all features is used as the feature importance threshold. Features with an average absolute SHAP value higher than this threshold are retained as key features, while features with an average absolute SHAP value lower than this threshold are filtered out as non-key features.

[0013] Preferably, in step S16, independent data refers to actual exposure experimental data that did not participate in model training; the accuracy evaluation criterion is that the relative error of the model prediction value is less than the preset relative error threshold.

[0014] Another aspect of this invention provides a method for predicting the atmospheric corrosion rate of weathering steel, which uses an enhanced neural network prediction model based on SHAP to obtain the prediction rate. The specific steps include: S21. Collect sample data of the weathering steel to be tested and preprocess it, wherein the sample data of the weathering steel to be tested includes the key features identified in the SHAP-based enhanced neural network prediction model. S22. Input the preprocessed data into the SHAP-based enhanced neural network prediction model to obtain the corrosion rate of the weathering steel to be tested.

[0015] Compared with the prior art, the present invention has the following beneficial effects: The technical solution of this invention has significant advantages such as high prediction accuracy, strong generalization ability, simple implementation and low cost. It not only effectively solves the problems of insufficient prediction accuracy, high cost and difficulty in long-term performance evaluation in traditional methods, but also provides reliable data support and decision-making basis for weathering steel materials in practical engineering applications. It can be applied to the rapid and reliable evaluation of the long-term service performance and life of weathering steel materials such as photovoltaic brackets and bridge structures. Attached Figure Description

[0016] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0017] Figure 1 A flowchart illustrating the method for constructing an enhanced neural network prediction model based on SHAP provided by this invention; Figure 2 A diagram showing the SHAP values ​​of all the features provided by this invention; Figure 3 A comparison chart of the actual corrosion rate and the predicted corrosion rate of WP720 weathering steel in Taiyuan City after one year of exposure, provided for the purposes of this invention; Figure 4 The graph shows the trend of corrosion rate as a function of exposure years under different ISO 9224 environmental levels, as predicted by the prediction model provided by this invention. Figure 5 The annual average corrosion rate of WP720 weathering steel under C3, C4, and C5 corrosion environments as defined in ISO 9224, with a prediction period of 30 years, is provided by this invention. Figure 6 The flowchart of a method for predicting the atmospheric corrosion rate of weathering steel provided by the present invention is shown. Detailed Implementation

[0018] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0019] like Figure 1 As shown, the first aspect of this invention provides an enhanced neural network prediction model based on SHAP, and the method for constructing this prediction model specifically includes: S11. Collect long-term exposure test data of weathering steel from different sources and merge them into a dataset, which includes alloy element data, environmental factors, exposure time and corrosion rate.

[0020] In this embodiment, long-term exposure test data of different types of weathering steel were collected from Beijing, Wuhan, Guangzhou, Qingdao, Jiangjin, Qionghai, Taiyuan in China, and Cabo Vilano in Spain to form a unified dataset. The alloying elements of weathering steel include carbon, sulfur, phosphorus, manganese, silicon, chromium, nickel, molybdenum, copper, niobium, vanadium, titanium, aluminum, and rhenium; environmental factors include temperature, relative humidity, wetting time, sulfur dioxide deposition rate, and chloride ion deposition rate.

[0021] It should be noted that the "long term" refers to several decades. This application embodiment collected weathering steel exposure test data for up to 16 years, including data from the 1st, 2nd, 4th, 8th, and 16th years. However, in actual experiments, data can be collected monthly, annually, or at intervals of several years, depending on the actual needs of the personnel.

[0022] This application collects long-term exposure experimental data of various types of weathering steel materials from multiple regions. The data collection scope covers the alloy element composition data of weathering steel, environmental factor data, exposure time and corresponding measured corrosion rate data. Through data integration, a unified dataset with sufficient scale and rich data types is formed to ensure the training and generalization performance of subsequent models.

[0023] S12. Preprocess the dataset, including data cleaning and feature engineering.

[0024] In this application, data cleaning includes outlier and missing value detection and processing. Outlier processing uses the interquartile range method and the empirical threshold method, while missing value processing uses the method of directly deleting missing values.

[0025] In this embodiment of the application, the upper and lower limits based on the interquartile range are used to identify outliers. Data below the lower limit or above the upper limit will be considered outliers and removed. The calculation formula for the interquartile range method is as follows: , , , In the formula, Interquartile range, It is the first quartile. It is the third quartile.

[0026] In this application, the feature engineering processing includes: performing maximum and minimum normalization processing on temperature and relative humidity in environmental factors; standardizing the humidification time in environmental factors to [0,1] based on physical meaning; and performing logarithmic transformation on the exposure time.

[0027] In this embodiment of the application, the calculation formula for the maximum and minimum normalization processing is as follows: , In the formula, Represents the original data. This represents the normalized value.

[0028] In this embodiment of the application, the logarithmic change of the exposure time is shown in the following formula: , In the formula, Represents the eigenvalues ​​after logarithmic transformation. X Indicates the original exposure time.

[0029] It should be noted that since the corrosion mechanism of weathering steel is to first form a dense oxide film to prevent further corrosion, this application uses a logarithmic change in the exposure time, which closely matches the corrosion rate curve of weathering steel over time, thus improving the accuracy compared to directly using the exposure time.

[0030] S13. Construct a neural network model and train the neural network model based on the preprocessed dataset.

[0031] In this embodiment of the application, the neural network model constructed in step S13 is a fully connected feedforward neural network model.

[0032] In this application, the neural network model constructed in step S13 includes an input layer, three hidden layers, and an output layer. The input layer has 20 points, including alloy element data, environmental factors, and exposure time. The number of points in the three hidden layers decreases layer by layer, and all three use the Rectified Linear Unit (ReLU) activation function. A dropout regularization layer is set between the second and third hidden layers. The output layer has one node, which is the corrosion rate, and uses a linear activation function.

[0033] In this embodiment, the first hidden layer has 256 nodes, the second hidden layer has 128 nodes, and the third hidden layer has 64 nodes; a random deactivation regularization layer is set between the second and third hidden layers, with a deactivation rate of 0.1.

[0034] In this application, by setting a random deactivation regularization layer, the activation outputs of some neurons are randomly discarded during the model training phase, thereby effectively preventing overfitting and improving the model's generalization ability.

[0035] In this application, step S13 uses a five-fold cross-validation method to train the neural network model. Specifically, the preprocessed dataset is first shuffled and then randomly divided into five parts. Four of these parts are used to train the neural network model, and the remaining part is used as a validation set to test the prediction accuracy of the neural network model.

[0036] In this application, a five-fold cross-validation method is used to systematically evaluate the performance of the neural network model in order to improve its robustness and generalization ability.

[0037] S14. Set the hyperparameters and optimize them based on the trained neural network model and optimization strategy to obtain the optimal hyperparameters.

[0038] In this application, the hyperparameters set include: number of training rounds, batch size for each training round, and learning rate.

[0039] In this application, the hyperparameter optimization process aims to achieve the best prediction performance of the trained neural network model on the validation set. After determining the specific objective function based on the adopted optimization strategy, the trained neural network model is trained and evaluated multiple times to search for and determine the optimal hyperparameter combination.

[0040] Preferably, when the dimension of the hyperparameter space is low, a grid search strategy is used to traverse all preset hyperparameter combinations. In this process, the optimal hyperparameter combination is determined by minimizing the average value of each loss function in cross-validation. The objective function is as follows: , In the formula, Indicates the search space of hyperparameters Selecting the values ​​that result in the average loss of the model in K-fold cross-validation. The smallest combination of hyperparameters, For folds, Respectively The validation set input features and the true labels are folded; Indicating the combination of hyperparameters Below, the model's predicted output based on the input data.

[0041] Preferably, when the hyperparameter space has a high dimension, a Bayesian optimization method is used, specifically by constructing a surrogate function. To simulate the mapping relationship between hyperparameter combinations and model performance, and based on the acquisition function This guides the selection of the next set of hyperparameters, thereby efficiently finding the optimal parameter combination with minimal model evaluations. The objective function is shown below: , In the formula, This represents the set of hyperparameter-performance pairs that have been evaluated so far. To obtain the function used to evaluate candidate hyperparameter combinations Potential optimization value under the proxy model.

[0042] In addition, Bayesian optimization methods can be used when the search space is complex or the training cost is high.

[0043] In this embodiment of the application, the desired improved EI criterion is used as the acquisition function, and its expression is as follows: , In the formula, Representative proxy function The mathematical expectation of the predicted distribution under the current observation conditions. This indicates that the best validation performance has been achieved among the hyperparameter combinations. The loss is used to validate the predictions of the current candidate hyperparameter combinations under the surrogate model.

[0044] It should be noted that the overall objective in this application is to achieve optimal prediction performance on the validation set, such as minimizing MSE or maximizing R². The objective in grid search is to minimize the average loss during cross-validation; in Bayesian optimization, the objective is to maximize the desired function, such as improving the EI criterion. This is an indirect optimization of performance. Although the EI function in Bayesian optimization has a different form, its essence is to select the parameter combination that "most likely improves validation performance," and it will eventually converge to the optimal performance. Therefore, the objective functions of both grid search and Bayesian optimization are strategic objective functions that serve to achieve the overall goal.

[0045] In this application, the hyperparameter space has a high dimension, the search space is complex, and the training cost is large. Therefore, the Bayesian optimization method is used to optimize the hyperparameters in this application.

[0046] It should be noted that by employing the Bayesian optimization method, the generalization performance and prediction accuracy of the model can be ensured to reach their optimal levels.

[0047] In this embodiment of the application, the preferred range of hyperparameter combinations is: 1,000 to 2,000 training rounds, batch size of each training round of 32 to 128, and learning rate of 0.0001 to 0.001.

[0048] In this embodiment of the application, the optimal combination of hyperparameters is: 1500 training rounds, batch size of 64 per training round, and learning rate of 0.0003.

[0049] S15. Keeping the optimal hyperparameters determined in step S14, perform feature importance analysis on the trained neural network model, determine key features, and reconstruct and train the initial augmented neural network prediction model based on SHAP based on the key features.

[0050] In this application, the SHAP method is used to perform feature importance analysis on the trained neural network model. By calculating the average absolute SHAP value of each feature in all samples, the key features that make a significant contribution to the trained neural network model are identified.

[0051] In this embodiment of the application, the formula for calculating the SHAP value of each feature in each sample is as follows: , In the formula, Representation of features The SHAP value, where M represents the total number of features and S represents the number of features not included. Feature subset, This represents the model's predicted output based on the input feature subset S. Indicates features The prediction results of the model after adding to subset S. Represents all possible feature subsets Sum.

[0052] To quantify the importance of features, this embodiment further calculates the average absolute SHAP value of each feature across all samples as a measure of its global importance. The calculation formula is as follows: , In the formula, The average absolute SHAP value represents the feature, where N is the total number of samples. This represents the SHAP value of the feature in the i-th sample.

[0053] The SHAP method, based on Shapley value theory in game theory, assesses the importance of each input feature by quantifying its marginal contribution to the model's prediction results under different feature combinations. This application utilizes the SHAP value formula to enable the SHAP method to accurately evaluate the average marginal contribution of each input feature in different contexts, thereby reflecting its influence on the predictions of the trained neural network model.

[0054] In this application, in step S15, 3% of the maximum average absolute SHAP value among all features is used as the feature importance threshold. Features with an average absolute SHAP value higher than this threshold are retained as key features, while features with an average absolute SHAP value lower than this threshold are filtered out as non-key features.

[0055] In this embodiment of the application, feature importance analysis is performed on the trained neural network model, and the average absolute SHAP value of each feature is as follows: Figure 2 As shown. From Figure 2 It can be clearly seen that among all features, the average absolute SHAP value of exposure time is the largest, at 11.36. Therefore, 3% of the average absolute SHAP value of exposure time is 0.34. Subsequently, features with an average absolute SHAP value below 0.34 are screened out as non-critical features, while features with an average absolute SHAP value above 0.34 are retained as critical features. Therefore, in this embodiment, the non-critical features screened out include phosphorus, titanium, sulfur, molybdenum, aluminum, and rhenium, while the critical features retained include exposure time, sulfur dioxide deposition rate, chloride ion deposition rate, chromium, nickel, silicon, wetting time, relative humidity, temperature, copper, carbon, vanadium, manganese, and niobium.

[0056] It should be noted that the feature of importance analysis in step S15 is the input layer function of the neural network model constructed in step S13.

[0057] In this application, the initial enhanced neural network prediction model based on SHAP constructed in step S15 includes an input layer, three hidden layers, and an output layer. The input layer has 14 points, including exposure time, sulfur dioxide deposition rate, chloride ion deposition rate, chromium, nickel, silicon, wetting time, relative humidity, temperature, copper, carbon, vanadium, manganese, and niobium. The number of points in the three hidden layers decreases layer by layer, and all three use the Rectified Linear Unit (ReLU) activation function. A dropout regularization layer is set between the second and third hidden layers. The output layer has one node, which is the corrosion rate, and uses a linear activation function.

[0058] In this embodiment, the first hidden layer has 256 nodes, the second hidden layer has 128 nodes, and the third hidden layer has 64 nodes; a random deactivation regularization layer is set between the second and third hidden layers, with a deactivation rate of 0.1.

[0059] In this application, step S15 uses a five-fold cross-validation method to train the initial augmented neural network prediction model based on SHAP.

[0060] In this application, an initial enhanced neural network prediction model based on SHAP is reconstructed and trained based on key features to further improve the model's prediction performance and computational efficiency.

[0061] S16. Use independent data to evaluate the accuracy of the initial SHAP-based augmented neural network prediction model after training. After passing the accuracy evaluation criteria, the SHAP-based augmented neural network prediction model is obtained.

[0062] In this application, independent data refers to actual exposure experimental data that did not participate in model training; the accuracy evaluation criterion is that the relative error of the model prediction value is less than a preset relative error threshold.

[0063] In this embodiment, the preset relative error threshold is 5%, but it can be set according to the actual engineering requirements in practical applications.

[0064] It is important to note that when evaluating the accuracy of the initial augmented neural network prediction model based on SHAP after training, it is preferable to use actual exposure experimental data that was not used in the model training, in order to objectively evaluate the actual predictive ability and generalization performance of the prediction model.

[0065] like Figure 3As shown, independent verification was performed using corrosion data of WP720 weathering steel exposed for one year in Taiyuan City. The actual corrosion rate was 16.86 µm / a, while the corrosion rate predicted by the prediction model constructed in this method was 17.25 µm / a, with a prediction error of only 2.31%. This meets the accuracy evaluation standard and is significantly better than the accuracy requirement of 5% for traditional models, verifying that the model has high accuracy and generalization ability.

[0066] like Figure 6 As shown, the second aspect of this invention provides a method for predicting the atmospheric corrosion rate of weathering steel, which uses an enhanced neural network prediction model based on SHAP to obtain the prediction rate. The specific steps include: S21. Collect sample data of the weathering steel to be tested and preprocess it, wherein the sample data of the weathering steel to be tested includes the key features identified in the SHAP-based enhanced neural network prediction model. S22. Input the preprocessed data into the SHAP-based enhanced neural network prediction model to obtain the corrosion rate of the weathering steel to be tested.

[0067] It is important to emphasize that the SHAP-based enhanced neural network prediction model constructed in this application can predict the corrosion rate at any point in time, such as a certain month or year. In practical applications, the sample data of the weathering steel to be tested can be directly input into the prediction model to obtain the corresponding corrosion rate. To calculate the average annual corrosion rate over a certain period, such as the average annual corrosion rate over 30 years, the average annual corrosion rate over 30 years can be obtained by using the prediction model to obtain the corrosion rate for each year within those 30 years.

[0068] In this application, the key features obtained from the SHAP-based enhanced neural network prediction model include exposure time, sulfur dioxide deposition rate, chloride ion deposition rate, chromium, nickel, silicon, wetting time, relative humidity, temperature, copper, carbon, vanadium, manganese, and niobium. Therefore, in the process of predicting the atmospheric corrosion rate of weathering steel, the sample data collected also include exposure time, sulfur dioxide deposition rate, chloride ion deposition rate, chromium, nickel, silicon, wetting time, relative humidity, temperature, copper, carbon, vanadium, manganese, and niobium.

[0069] It should be noted that preprocessing in prediction methods also includes data cleaning and feature engineering.

[0070] It is important to emphasize that a time span of several decades implies an extremely long period. Attempting to obtain actual data within this long timeframe is clearly out of touch with the actual needs of engineering projects. After all, in real-world engineering, waiting decades to collect complete data is neither realistic nor would it significantly delay project progress. Furthermore, existing models have significant shortcomings; they cannot accurately determine the simulated time frame, which introduces considerable uncertainty into long-term predictions in engineering applications.

[0071] The proposed solution features a unique and scientific design. First, regarding input variables, it explicitly incorporates the key time dimension of "exposure time" and models it using a logarithmic transformation, enabling the model to consider time factors from its fundamental construction. Second, it uses data with an exposure time of up to 16 years for model training. This large dataset comprehensively covers corrosion over many years, providing rich information support for the model's long-term predictions. Finally, by integrating the SHAP+Bayesian optimization mechanism, it enhances model performance while ensuring the interpretability and reliability of long-term prediction results, avoiding the confusion caused by "black box" models. Combining these three advantages, the model constructed in this application not only possesses the technical foundation and theoretical feasibility for long-term prediction but also accurately defines the simulated time period, providing a more precise and reliable solution for corrosion prediction in the engineering field.

[0072] Figure 4 This paper demonstrates the trend of corrosion rate variation with exposure years under three typical atmospheric corrosion environment levels (C3, C4, and C5) according to ISO 9224, based on the SHAP-based enhanced neural network prediction model provided by this invention. Figure 4 It can be clearly seen that the corrosion rate under each environmental level gradually decreases and tends to stabilize over time, which is consistent with the evolution law of weathering steel forming a dense protective layer through initial rusting in the atmospheric environment, thereby slowing down the later corrosion rate. This result verifies the rationality and engineering applicability of the prediction model constructed in this application in long-term corrosion prediction.

[0073] like Figure 5 As shown, this application also uses an enhanced neural network prediction model based on SHAP to predict the average annual corrosion rate of WP720 weathering steel over 30 years in the C3, C4, and C5 corrosion environments defined by ISO 9224. The model prediction results are 4.33 µm / a (C3), 6.18 µm / a (C4), and 18.11 µm / a (C5), respectively, all within the range recommended by the ISO standard. Furthermore, the cumulative corrosion depths are 129.86 µm, 185.31 µm, and 543.19 µm, respectively, all significantly lower than the standard allowable limits. This also fully demonstrates that the prediction model constructed in this application has practical application value and engineering applicability, and can guide the actual engineering application of weathering steel.

[0074] In summary, the SHAP-based enhanced neural network prediction model constructed in this invention has high prediction accuracy, high generalization ability, and potential for practical engineering applications in predicting the corrosion rate of weathering steel, and can effectively meet the needs of long-term performance evaluation and life prediction of weathering steel.

[0075] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An enhanced neural network prediction model based on SHAP, characterized in that, The specific methods for constructing this prediction model include: S11. Collect long-term exposure test data of weathering steel from different sources and merge them into a dataset, which includes alloy element data, environmental factors, exposure time and corrosion rate. S12. Preprocess the dataset, including data cleaning and feature engineering. S13. Construct a neural network model and train the neural network model based on the preprocessed dataset; S14. Set the hyperparameters and optimize them based on the trained neural network model and optimization strategy to obtain the optimal hyperparameters; S15. Keeping the optimal hyperparameters determined in step S14, perform feature importance analysis on the trained neural network model, determine key features, and reconstruct and train the initial augmented neural network prediction model based on SHAP based on the key features. S16. Use independent data to evaluate the accuracy of the initial SHAP-based augmented neural network prediction model after training. After passing the accuracy evaluation criteria, the SHAP-based augmented neural network prediction model is obtained.

2. The SHAP-based enhanced neural network prediction model according to claim 1, characterized in that, Weathering steel alloy elements include carbon, sulfur, phosphorus, manganese, silicon, chromium, nickel, molybdenum, copper, niobium, vanadium, titanium, aluminum, and rhenium; environmental factors include temperature, relative humidity, wetting time, sulfur dioxide deposition rate, and chloride ion deposition rate.

3. The SHAP-based enhanced neural network prediction model according to claim 1, characterized in that, Data cleaning includes outlier and missing value detection and handling. Outlier handling uses the interquartile range method and empirical threshold method, while missing value handling uses the method of directly deleting missing values. Feature engineering processing includes: performing maximum and minimum normalization on temperature and relative humidity among environmental factors; standardizing the humidification time among environmental factors to [0,1] based on physical meaning; and performing logarithmic transformation on the exposure time, where the logarithmic transformation of the exposure time is shown in the following formula: ， In the formula, Represents the eigenvalues ​​after logarithmic transformation. X Indicates the original exposure time.

4. The SHAP-based enhanced neural network prediction model according to claim 1, characterized in that, The models constructed in steps S13 and S15 consist of one input layer, three hidden layers, and one output layer. The number of points in the three hidden layers decreases progressively, and all three use the Rectified Linear Unit activation function. A random deactivation regularization layer is set between the second and third hidden layers. The output layer has one node and uses a linear activation function. Specifically, the input layer of the model constructed in step S13 has 20 points, and the input layer of the model constructed in step S15 has 14 points. In addition, the five-fold cross-validation method is used to train the model in both steps S13 and S15.

5. The SHAP-based enhanced neural network prediction model according to claim 1, characterized in that, The hyperparameters set in step S14 include: number of training epochs, batch size per training session, and learning rate. The combination of hyperparameters ranges from 1,000 to 2,000 training epochs, batch size per training session is 32 to 128, and learning rate is 0.0001 to 0.

001. Furthermore, during the hyperparameter optimization process, the ultimate goal is to achieve the best prediction performance of the trained neural network model on the validation set. Based on the optimization strategy adopted, the specific objective function is determined, and the trained neural network model is trained and evaluated in multiple rounds to search for and determine the optimal combination of hyperparameters.

6. The SHAP-based enhanced neural network prediction model according to claim 5, characterized in that, The optimal combination of hyperparameters is: 1500 training epochs, batch size of 64 per training epoch, and learning rate of 0.0003. Furthermore, during the hyperparameter optimization process in step S14: When the dimension of the hyperparameter space is low, a grid search strategy is used to traverse all preset hyperparameter combinations. In this process, the optimal hyperparameter combination is determined by minimizing the average value of the loss functions in cross-validation. The objective function is shown below: ， In the formula, Indicates the search space of hyperparameters Selecting the values ​​that result in the average loss of the model in K-fold cross-validation. The smallest combination of hyperparameters, For folds, Respectively The validation set input features and the true labels are folded; Indicating the combination of hyperparameters Below is the model's predicted output based on the input data; When the hyperparameter space has a high dimensionality, a Bayesian optimization method is used, specifically by constructing a surrogate function. To simulate the mapping relationship between hyperparameter combinations and model performance, and based on the acquisition function This guides the selection of the next set of hyperparameters, thereby efficiently finding the optimal parameter combination with minimal model evaluations. The objective function is shown below: ， In the formula, This represents the set of hyperparameter-performance pairs that have been evaluated so far. To obtain the function used to evaluate candidate hyperparameter combinations Potential optimization value under the proxy model.

7. The SHAP-based enhanced neural network prediction model according to claim 1, characterized in that, In step S15, the SHAP method is used to perform feature importance analysis on the trained neural network model. By calculating the average absolute SHAP value of each feature across all samples, key features that significantly contribute to the trained neural network model are identified. The formula for calculating the SHAP value of each feature in each sample is shown below: ， In the formula, Representation of features The SHAP value, where M represents the total number of features and S represents the number of features not included. Feature subset, This represents the model's predicted output based on the input feature subset S. Indicates features The prediction results of the model after adding to subset S. Represents all possible feature subsets Summation; The formula for calculating the mean absolute SHAP value of each feature across all samples is as follows: ， In the formula, Representation of features The mean absolute SHAP value, where N is the total number of samples. This represents the SHAP value of the feature in the i-th sample.

8. The SHAP-based enhanced neural network prediction model according to claim 7, characterized in that, In step S15, 3% of the maximum average absolute SHAP value among all features is used as the feature importance threshold. Features with an average absolute SHAP value higher than this threshold are retained as key features, while features with an average absolute SHAP value lower than this threshold are filtered out as non-key features.

9. The SHAP-based enhanced neural network prediction model according to claim 1, characterized in that, In step S16, independent data refers to actual exposure experimental data that did not participate in model training; the accuracy evaluation criterion is that the relative error of the model prediction value is less than the preset relative error threshold.

10. A method for predicting the atmospheric corrosion rate of weathering steel, characterized in that, The prediction is obtained using an enhanced neural network prediction model based on SHAP. The specific steps include: S21. Collect sample data of the weathering steel to be tested and preprocess it, wherein the sample data of the weathering steel to be tested includes the key features identified in the SHAP-based enhanced neural network prediction model. S22. Input the preprocessed data into the SHAP-based enhanced neural network prediction model to obtain the corrosion rate of the weathering steel to be tested.

Images