Concrete chloride ion diffusion prediction method based on physical enhancement characteristics and XGBoost model

Abstract

The invention discloses a concrete chloride ion diffusion prediction method based on physical enhancement characteristics and an XGBoost model, and the method comprises the steps: forming enhancement input data through constructing physical characteristics, material composition characteristics and durability related indexes which reflect a diffusion process mechanism, and inputting the enhancement input data into the XGBoost model for training and prediction. By reasonably setting a model structure and a training strategy, high-precision modeling and prediction of a diffusion coefficient are realized, and key influence factors are analyzed and identified in combination with feature importance. According to the invention, high-precision prediction of the concrete chloride ion diffusion behavior is realized. On the basis of fully considering concrete material composition and environmental factors, the model generates expansion features with clear physical significance, and improves the fitting ability to the diffusion process. Meanwhile, the performance of the model is optimized through evaluation indexes, it is ensured that the model has good stability and generalization ability under different data conditions, and therefore reliable prediction support is provided for concrete formula design and construction scheme adjustment.

Description

Technical Field

[0001] The present invention belongs to the field of chloride ion diffusion prediction of concrete materials, and in particular relates to a concrete chloride ion diffusion prediction method based on physical enhancement characteristics and an XGBoost model. Background Art

[0002] As one of the most commonly used materials in construction, concrete's durability is crucial to the lifespan and safety of buildings. Chloride ion diffusion in concrete is a major factor in steel corrosion. Chloride ions enter the concrete surface, causing corrosion and subsequently compromising the strength and stability of the structure. Therefore, accurate prediction of chloride ion diffusion is a core issue in concrete durability research, particularly in complex environmental conditions. Accurately predicting chloride ion diffusion behavior has become a key challenge in concrete engineering.

[0003] Traditional methods for predicting chloride ion diffusion rely primarily on experimental data and empirical formulas (such as Fick's law), using simplified mathematical models to describe the chloride ion diffusion process. While these methods provided a reference for early research, they have significant limitations in practical application. Traditional methods fail to fully consider environmental conditions, the heterogeneity of concrete materials, and the interactive effects of multiple factors. Experimental testing methods are not only costly and time-consuming, but also lack sufficient prediction accuracy and adaptability when faced with complex concrete mixes and environmental changes.

[0004] With the development of machine learning technology, data-driven prediction models have gradually become an effective tool to replace traditional methods. However, the application of existing machine learning methods in the prediction of chloride ion diffusion in concrete still faces many challenges. First, most existing models fail to effectively combine physical laws, especially under complex environmental conditions, and cannot fully reflect the physical properties of concrete. Secondly, although machine learning can handle nonlinear relationships, existing models often ignore the multi-level, complex physical and material properties of concrete during feature selection and modeling. Finally, existing machine learning methods mostly rely on the data itself for training and lack deep integration of physical properties, which limits the accuracy and generalization ability of the model. Therefore, how to fully combine machine learning algorithms with actual physical characteristics has become the key to research in this field. Summary of the Invention

[0005] Purpose of the Invention: This invention aims to provide a method for predicting chloride ion diffusion in concrete based on physical enhancement features and the XGBoost model. By generating features based on physical laws and combining them with machine learning algorithms, the accuracy of chloride ion diffusion prediction is significantly improved. This method effectively addresses the lack of accuracy of traditional prediction methods under complex environmental conditions and can help engineers optimize concrete formulations and construction plans, improve concrete durability, reduce maintenance costs, and extend the service life of structures.

[0006] Technical solution: A method for predicting chloride ion diffusion in concrete based on physical enhancement features and the XGBoost model of the present invention comprises the following steps:

[0007] Step 1: Obtain the original experimental data of concrete and perform data preprocessing;

[0008] Step 2: Using the data preprocessed in step 1, Fick's second law is used to construct diffusion time and time decay characteristics. The temperature influence term is calculated using the Arrhenius equation, and the chloride ion binding capacity and concentration decay effects are introduced to form physical mechanism characteristics. Material composition characteristics that reflect the influence of concrete structure on diffusion path are extracted. Durability characteristics are constructed, including strength factor, durability index, diffusion boundary constraint, and comprehensive diffusion index.

[0009] Step 3: Build an XGBoost regression model based on the integrated regression tree structure. The physical mechanism characteristics, material composition characteristics, and durability characteristics of step 2 are integrated to form an XGBoost regression model with enhanced feature input. The model outputs the chloride ion diffusion coefficient of concrete and sets the model's hyperparameters.

[0010] Step 4: Complete the training of the XGBoost model built in step 3 on the training set, input the enhanced features and target variables into the model, optimize the regression performance through multiple rounds of iterative learning, and enable the early stopping mechanism on the validation set;

[0011] Step 5: Calculate the evaluation indicators of the model, including calculating the mean square error (MSE), root mean square error (RMSE), and determination coefficient (R). 2 Indicators, and optimize model parameters. If the following three conditions are met at the same time, the fitting effect is judged to be up to standard and go to step 6;

[0012] Condition 1: The mean square error (MSE) value is between 0 and 1.

[0013] Condition 2: The root mean square error (RMSE) value is between 0 and 1;

[0014] Condition 3: Determination coefficient R 2 The value of is between 0.82-0.88;

[0015] If the above three conditions cannot be met at the same time, the model is optimized by adjusting the model's hyperparameters;

[0016] Step 6: Use the final model to perform actual predictions and feature importance analysis. At the same time, based on the feature importance output function of XGBoost, the relative contribution of each input variable is obtained, the key factors that have the greatest impact on diffusion performance are identified, and a quantitative basis is provided for auxiliary material design and durability control.

[0017] Furthermore, step 1 includes the following steps:

[0018] Step 11: Collect concrete mix data, including the proportions of cement, fly ash, mineral admixtures, water, and aggregate raw materials;

[0019] Step 12: Obtain data on environmental factors affecting concrete, including external environmental parameters such as temperature, humidity, and chloride ion concentration;

[0020] Step 13: Divide the data into training set and test set according to the proportion for model training and evaluation.

[0021] Step 14: Perform data preprocessing, including removing outliers, filling missing values, and standardizing the data. The standardization method is as follows:

[0022]

[0023] Where: μ is the mean of the feature, σ is the standard deviation of the feature.

[0024] In concrete chloride diffusion prediction, the quality of raw data directly determines model performance. Steps 11 through 14, through complete data access, reasonable data partitioning, and standard preprocessing, provide a clearly structured and numerically stable input foundation for subsequent feature construction and model training. This significantly improves the model's trainability and the credibility of the results, laying a solid foundation for subsequent training of the physical feature-enhanced neural network.

[0025] Furthermore, step 2 includes the following steps:

[0026] Step 21: Construct diffusion behavior-related features based on time parameters; calculate the square root of the exposure time t in the original data in years. As a diffusion time characteristic, it is used to approximately characterize the trend of diffusion distance over time, which conforms to the characteristic solution form of Fick's second law in a homogeneous medium;

[0027] Step 22: Based on the time characteristic, a temperature correction term is introduced to construct a temperature-time coupling characteristic. Considering the accelerating effect of temperature on the diffusion process, the temperature influence factor is calculated based on the Arrhenius equation:

[0028]

[0029] Among them, E a =40000 J / mol is the activation energy of chloride ion migration, R = 8.314 J / (mol·K) is the universal gas constant, T is the actual ambient temperature, T ref =293.15K is the reference temperature;

[0030] Step 23: Construct chloride ion concentration-related features to reflect the driving and coupling effects of the external medium on the diffusion process, and calculate the chloride ion concentration effect features:

[0031]

[0032] Among them, C CI is the external chloride ion concentration, in g / L. This characteristic reflects the enhancement effect at high concentrations and the inhibition effect at low concentrations, and the chloride ion binding capacity characteristic is calculated from this:

[0033] B CI =0.5·(1-exp(-0.1·C CI ))

[0034] Step 24. Based on the physical diffusion mechanism, extract the relevant characteristics of concrete material composition to express the indirect effect of the mixing ratio on the diffusion path and calculate the total cementitious material mass:

[0035] M binder =FA+GGBS+SF+OPC

[0036] The pozzolan admixture ratio is calculated from this:

[0037]

[0038] Among them, FA is fly ash, GGBS is slag, SF is silica fume, and OPC is ordinary Portland cement, all in kg / m 3 , further calculate the effective water-binder ratio:

[0039]

[0040] Among them, r wb The original water-binder ratio (WBR) is the ratio of the original water content to the cementitious material. This expression takes into account the dilution effect of the admixture reactivity and reflects the formation trend of the capillary pore structure inside the concrete.

[0041] Step 25: Construct porosity and hydration-related features to supplement the microscopic channel effect of the material structure on ion diffusion; the volcanic ash reaction effect index is calculated as follows:

[0042]

[0043] f porosity =exp(-3·r pz )

[0044] Based on the ratio of water-binder ratio to cementitious material mass, the hydration potential index is constructed:

[0045]

[0046] f hydration =exp(-2·r wb_ratio )

[0047] Where W is the mass of mixing water kg / m 3 , which reflects the changing trend of the degree of hydration reaction and porosity inside concrete and is the key parameter controlling the formation of ion permeation channels;

[0048] Step 26: Construct durability-related indicators, constrain the physical range of the prediction results, and calculate the strength factor and durability index:

[0049]

[0050] f durability =f strength ·f porosity

[0051] in The critical water-binder ratio is set empirically, and the boundary value of the diffusion coefficient is set as a constraint:

[0052] D max =10 -11 ·exp(2·r wb )

[0053] D min =10 -13

[0054] Steps 21 to 26 construct a comprehensive set of physically enhanced feature systems that integrate diffusion dynamics, environmental response, material structure, and hydration evolution. By introducing rigorous physical equations and material scaling models, the model not only improves its ability to model chloride ion diffusion processes but also ensures good physical consistency and interpretability. This feature-enhanced structure provides a high-dimensional, physics-oriented data foundation for subsequent deep learning and ensemble learning models.

[0055] Furthermore, step 3 includes the following steps:

[0056] Step 31: Combine the physical mechanism features, material composition features, and durability features from step 2 to construct a unified enhanced input feature as the data basis for machine learning modeling.

[0057] Step 32: Based on the enhanced input features and their corresponding chloride ion diffusion coefficient labels, define the regression task objectives and clarify the supervised learning modeling paradigm centered on predicting continuous values;

[0058] Step 33: Select the gradient boosting decision tree XGBoost as the regression model framework;

[0059] Step 34. Configure the hyperparameters of the XGBoost model, including the maximum tree depth max_depth, learning rate eta, minimum child weight min_child_weight, column sampling ratio colsample_bytree, subsample ratio subsample, regularization term lambda, alpha, to enhance model fitting ability, control complexity and prevent overfitting;

[0060] Step 35. Define the training objective of the XGBoost model as minimizing the following loss function:

[0061]

[0062] Among them, y i is the true chloride ion diffusion coefficient of the i-th sample, is the predicted value, Ω(f k ) represents the regularization term of the k-th regression tree, including the number of leaves and the penalty of leaf weight; the objective function is automatically optimized by the XGBoost framework, supplemented by RMSE and R 2 As an evaluation indicator, it provides performance guarantee for the subsequent training process.

[0063] Furthermore, step 4 includes the following steps:

[0064] Step 41: Based on the XGBoost model structure, the enhanced features and target values ​​are input into the model to form the data input of the supervised learning task.

[0065] Step 42: Set the maximum number of iterations for model training and enable early stopping. Evaluate the loss on the validation set after each training round. Terminate training early if the loss stops decreasing after several consecutive rounds to prevent overfitting of the model to the training set.

[0066] Step 43: In each iteration, the XGBoost framework is used to construct a new regression tree based on the current residual distribution to fit the prediction error of the previous round and gradually optimize the overall prediction performance.

[0067] Step 44: Dynamically adjust the splitting condition, node weight, and regularization coefficient of each subtree according to the configured hyperparameters, so that the model can control the structural complexity while learning the complex feature structure;

[0068] Step 45: After the training is completed, the set of model parameters with the best performance on the validation set is automatically selected as the final model output for subsequent evaluation and prediction.

[0069] Furthermore, in step 5, the mean square error MSE, root mean square error RMSE and determination coefficient R 2 , which is calculated as follows:

[0070] Mean square error (MSE): used to calculate the difference between the predicted value and the true value.

[0071]

[0072] Among them, among them, Represents the model prediction value, y i Represents the true value of the sample, n is the total number of samples

[0073] Root mean square error RMSE: The square root of MSE is used to provide a measure of the error.

[0074]

[0075] Coefficient of determination R 2 Value: measures how well the Bayesian neural network model fits the data.

[0076]

[0077] The present invention further discloses a computer device, comprising a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method of the present invention.

[0078] The present invention further discloses a computer-readable storage medium having a computer program / instruction stored thereon, which implements the steps of the method of the present invention when the computer program / instruction is executed by a processor.

[0079] The present invention further discloses a computer program product, comprising a computer program / instruction, which implements the steps of the method of the present invention when executed by a processor.

[0080] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0081] (1) The present invention achieves high-precision prediction of chloride ion diffusion behavior in concrete by introducing physical enhancement features and fully combining them with the XGBoost model. The model generates extended features with clear physical meanings based on full consideration of concrete material composition and environmental factors, effectively improving the ability to fit the diffusion process. At the same time, the model optimizes performance through evaluation indicators, ensuring good stability and generalization capabilities under different data conditions, thereby providing reliable prediction support for concrete formula design and construction plan adjustment.

[0082] (2) Through complete data reading, reasonable data division and standard preprocessing operations, a clear structure and numerically stable input foundation is provided for subsequent feature construction and model training, which significantly improves the trainability of the model and the credibility of the results, and provides a solid foundation for the subsequent training of physical feature enhanced neural networks.

[0083] (3) By combining the enhanced features with the XGBoost regression modeling framework and combining it with scientific and reasonable hyperparameter configuration and loss function definition, this set of steps established a prediction model structure with high expressiveness, strong robustness and interpretability, laying a key foundation for subsequent training and generalization performance improvement.

[0084] (4) By introducing an alternating evaluation mechanism of training sets and validation sets, combined with an early stopping strategy and a residual iterative update process, the XGBoost model is ensured to have good generalization capabilities while maintaining a high degree of fit, thereby improving the stability and reliability of the model in actual predictions. BRIEF DESCRIPTION OF THE DRAWINGS

[0085] Figure 1 It is a diagram of the overall idea of ​​the present invention;

[0086] Figure 2 Comparison chart of the actual value and predicted value of the test set: shows the comparison between the actual diffusion coefficient in the test set and the predicted value of the physical enhancement model, and marks the ideal reference line for evaluating the fitting accuracy and generalization ability of the model;

[0087] Figure 3 Forecast error distribution diagram: displays the histogram distribution and kernel density estimation curve of the forecast error in the test set, which is used to analyze the distribution characteristics and skewness of the model error;

[0088] Figure 4 Residual analysis chart: shows the relationship between the predicted value and its corresponding residual, and combines the trend line to determine whether the model has systematic bias and heteroscedasticity;

[0089] Figure 5QQ plot: shows the degree of distribution fit between the standard normal distribution and the model residuals, which is used to evaluate the normality of the error and judge the reliability and stability of the model prediction. DETAILED DESCRIPTION

[0090] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0091] The first part of this embodiment involves building a physically meaningful feature enhancement mechanism by incorporating diffusion physics theory, thermodynamic relationships, and the structural laws of concrete materials. This expands traditional raw input data into a high-dimensional input vector encompassing physical, material, and durability characteristics, significantly improving the model's ability to perceive and interpret complex diffusion processes. This part is one of the core innovations of this invention.

[0092] First, based on the durability prediction problem of concrete materials, the present invention selects multiple features obtained from Excel spreadsheets as input and uses them as the basis for designing a physical enhancement feature system. Specifically, the present invention uses the following feature data:

[0093] OPC (kg / m 2 ):the amount of ordinary Portland cement; FA (kg / m 3 ): amount of fly ash; GGBS (kg / m 3 ):Amount of blast furnace slag powder; SF (kg / m 3 ):Amount of silica fume; Superplasticiser (kg / m 3 ): water reducing agent dosage; Water (kg / m 3 ): water dosage; Fine agg (kg / m 3 ): amount of fine aggregate; Coarse agg (kg / m 3 ):amount of coarse aggregate; w / b (water-binder ratio): water-binder ratio; Exposure time (t / a): exposure time; Annual mean temperature (T / ℃): annual mean temperature; [Cl-]in seawater (g / L): chloride ion concentration in seawater.

[0094] Based on the above variables, the present invention constructs three types of enhancement features: physical features, material features and durability features.

[0095] In terms of physical characteristics, the diffusion time feature is first introduced To reflect the square root relationship between diffusion rate and time in Fick's diffusion law; at the same time, the time decay effect (t+1) is introduced -0.2, to reflect the trend of the diffusion process gradually slowing down over time. For the influence of ambient temperature, the Arrhenius equation is used to model and construct the temperature response term:

[0096]

[0097] Among them, E a =40000 J / mol is the activation energy of the diffusion process, R = 8.314 J / (mol·K) is the gas constant, T is the ambient temperature, T ref = 293.15K is the reference temperature. This temperature factor is used to characterize the effect of ambient temperature on the diffusion rate. Furthermore, a composite term combining temperature and diffusion time is constructed to enhance the model's ability to detect the synergistic effects of multiple variables.

[0098] In terms of material characteristics, the present invention constructs multiple structure-related indicators. First, the pozzolan ratio is calculated by the total proportion of active admixtures such as fly ash, slag, and silica fume to characterize the optimization effect of cement replacement materials on pore structure. The expression is as follows:

[0099]

[0100] Among them, FA, GGBS, SF, and OPC represent the amount of fly ash, slag, silica fume, and cement, respectively (in kg / m 3 ).

[0101] Subsequently, the effective water-binder ratio (Effective w / b) is introduced to reflect the actual water-binder ratio participating in the reaction. Its calculation method is:

[0102] Effective w / b =w / b×(1-0.3·Pozzolanic ratio )

[0103] Among them, w / b is the original water-binder ratio, 0.3 is the empirical penalty coefficient, and Pozzolanic ratio The aforementioned volcanic ash ratio.

[0104] In addition, in order to further reflect the influence of admixture reactivity on porosity, the pozzolanic effect index is introduced:

[0105]

[0106] Among them, the coefficients 0.3, 0.4, and 0.8 represent the weighted effects of different admixtures in reducing porosity.

[0107] And construct the porosity factor accordingly:

[0108] Porosity factor=exp(-3·Pozzolanic effect )

[0109] Finally, to reflect the relationship between water-binder ratio and reactivity, the hydration potential characteristics were also calculated:

[0110]

[0111] Where Water represents the amount of mixing water, Total binder =FA+GGBS+SF+OPC is the total amount of cementitious materials. This characteristic is used to reflect the regulatory effect of the water-binder ratio on the hydration rate.

[0112] In terms of durability characteristics, the present invention designs an intensity factor (based on the exponential decay of the water-binder ratio and the critical value), a durability index (combining the porosity factor and the intensity factor), and theoretically acceptable maximum and minimum diffusion boundaries. Specifically, the intensity factor is expressed as:

[0113]

[0114] Where w / b is the water-binder ratio of the current sample, and 0.45 is the critical water-binder ratio set empirically.

[0115] The durability index is defined as:

[0116] Durability index =Strength factor Porosity factor

[0117] Used to overall reflect the coupling relationship between structural density and strength attenuation.

[0118] In order to avoid the model from producing physically unreasonable prediction results, theoretical diffusion boundary constraints are also set:

[0119] D max =1×10 -11 exp(2·w / b)

[0120] D min =1×10 -13

[0121] Among them D max is the maximum diffusion coefficient prediction threshold, which increases exponentially with the water-binder ratio. min It is the physical lower limit.

[0122] All of the aforementioned physical, material, and durability features are ultimately concatenated with the original input features to form a unified high-dimensional input feature vector. Before entering the neural network, all features undergo missing value completion, outlier trimming, and normalization to ensure feature data integrity and numerical stability.

[0123] The second part of the present invention carries out the structural design and parameter configuration of the XGBoost regression model. In the early stage, the present invention constructs a highly expressive enhanced input matrix by introducing three types of features: physical mechanism, material composition, and durability index, providing rich feature semantics for subsequent modeling. In order to effectively fit the complex mapping relationship between these nonlinear enhancement features and the chloride ion diffusion coefficient, the present invention proposes to use the XGBoost model based on the gradient boosting tree as the main regression prediction framework.

[0124] The basic idea of ​​the XGBoost model used in this paper is to train a series of weak regression trees in a round-by-round stacking manner, and in each round, minimize the loss function of the residual error of the previous round to improve the model prediction performance. Assume that the final model consists of K trees, and the input enhanced feature vector is x i , then the output of the model can be expressed as:

[0125]

[0126] Among them, f k represents the kth regression tree, is a function space containing all learnable tree structures and their leaf node weights. This structure enables the model to gradually learn complex nonlinear relationships and is suitable for the multi-source coupling mechanism in the diffusion behavior of concrete.

[0127] To achieve stable and efficient model training, the present invention systematically designs XGBoost's hyperparameters. The model structure sets a maximum depth (max_depth = 5) to control the split depth of a single tree and avoid overfitting. The learning rate (eta = 0.005) is set to a small value to ensure stable convergence of each learning step. The minimum child weight (min_child_weight = 5) is used to limit the minimum sample weight of leaf nodes and avoid redundant splits. The subsample ratio (subsample = 0.8) and the feature column sampling ratio (colsample_bytree = 0.8) introduce sampling perturbations to improve generalization ability.

[0128] In addition, to further suppress model complexity and control tree structure redundancy, gamma = 0.1 is set as the minimum loss reduction threshold. The tree structure is split only when the loss after division is reduced by more than this value; the regularization parameters include L2 regularization (lambda = 2.0) and L1 regularization (alpha = 0.1), which are used to balance model complexity and parameter robustness.

[0129] In terms of feature processing strategy, this paper adopts a histogram-based training method (tree_method = 'hist') combined with a feature binning mechanism (max_bin = 256), which significantly accelerates model construction and reduces memory usage, making it suitable for large-scale feature input. All input features are normalized to eliminate scale differences between features and ensure that each type of physical enhancement feature receives a reasonable weight in the regression model.

[0130] In the second part, this implementation method reasonably designs the regression model structure and constraint parameters so that the XGBoost model can efficiently and stably learn the nonlinear mapping relationship between physical enhancement features and diffusion performance, ensuring that the prediction results have good accuracy, robustness and interpretability.

[0131] The third part of this example involves the training process and verification mechanism of the XGBoost model. Its core goal is to optimize the model parameters through supervised learning so that the mapping relationship between the physical enhancement features and the chloride ion diffusion coefficient can be stably established with minimal prediction error. After the initial feature generation and model construction are completed, the available dataset is first divided into a training set and a validation set to ensure that the model can obtain sufficient sample information during the learning process and that the training behavior can be supervised in real time through an independent validation set, thereby achieving a balance between model accuracy and generalization ability.

[0132] The input data is normalized by StandardScaler before training to eliminate scale differences between features and improve numerical stability during model training. The normalized training data is encapsulated as a DMatrix object required by XGBoost, which contains the feature matrix and the corresponding target labels, providing a unified interface for subsequent iterative training.

[0133] During model training, RMSE (root mean square error) is used as the main evaluation metric. The goal is to gradually fit the true value on the training set while monitoring the trend of the prediction error on the validation set. In each iteration, XGBoost builds a new regression tree to learn the current residual:

[0134]

[0135] in is the residual of the i-th sample in the t-th round, y i is the true value, is the predicted value of the previous iteration. The residual is used as the learning target of the current round tree to continuously narrow the gap between the model prediction and the actual value, and finally achieve error convergence.

[0136] To prevent overfitting of the model on the training set, this paper introduces a validation set monitoring mechanism during training and sets early_stopping_rounds = 100. If the RMSE on the validation set fails to improve after 100 consecutive iterations, training will automatically terminate early, and the model parameters corresponding to the round with the lowest error will be retained as the final model. This process is implemented by calling the xgb.train() function, which internally sets the number of iterations to 2000 and uses the early stopping mechanism to achieve adaptive training termination.

[0137] After training is completed, the model will have a complete structure and parameter configuration, and can be directly used for prediction tasks on the test set, while also providing a reusable model state for subsequent evaluation stages.

[0138] The final section focuses on performance evaluation and results analysis of the constructed model based on physical enhancement features and XGBoost, aiming to verify the effectiveness and reliability of the proposed method in predicting the chloride ion diffusion coefficient. The evaluation process includes model performance on the test set, robustness testing of cross-validation, visualization of residual distribution and error structure, and ranking of feature importance.

[0139] After completing network training, the next step is to evaluate the performance of the model and perform necessary optimizations. This includes using the test set data to make predictions and using appropriate evaluation metrics to measure the performance of the model. For this patent application, common evaluation metrics include mean square error (MSE), root mean square error (RMSE), and R 2 wait.

[0140] After the training is completed, the present invention uses the test set to evaluate the model. By comparing the prediction results of the model with the real data, the mean square error (MSE), root mean square error (RMSE) and R 2 and other indicators to evaluate the accuracy of the model.

[0141] Mean Squared Error (MSE): used to calculate the difference between the predicted value and the true value;

[0142]

[0143] Root mean square error (RMSE): The square root of the MSE, used to provide a measure of the error;

[0144]

[0145] R 2 Value: measures how well the model fits the data, the closer it is to 1, the better the fit;

[0146]

[0147] After obtaining these evaluation indicators, the present invention optimizes and adjusts the model. If the following three conditions are met at the same time, the fitting effect is judged to be up to standard,

[0148] Condition 1: The root mean square error (RMSE) is between 0 and 1.

[0149] Condition 2: The root mean square error (RMSE) value is between 0 and 1;

[0150] Condition 3: Determination coefficient R 2 The value of is between 0.82-0.88:

[0151] If the above three conditions cannot be met at the same time, the parameters are adjusted by increasing the learning rate, increasing the number of hidden layers, and adding regularization terms to improve the accuracy and reliability of concrete chloride ion diffusion prediction. The specific evaluation indicators of the training set and test set are shown in Table 1:

[0152] MSE RMSE <![CDATA[R 2 ]]> Test set 0.621 0.788 0.828

[0153] Table 1

[0154] In addition, in terms of evaluating the explanatory power of the model structure, through feature importance analysis, it can be found that the top ten most representative input variables are: temperature (Temperature), mineral admixture ratio (Pozzolanic_ratio), chloride ion concentration (Cl_concentration), time factor (Time_effect), porosity factor (Porosity_factor), coarse aggregate content (Coarse_agg), exposure time (Exposure_time), chloride ion effect (Cl_effect), temperature effect (Temperature_effect) and total amount of cementitious materials (Total_binder). These factors can affect the internal microstructure evolution and ion migration path of concrete in terms of physical mechanism, and have clear engineering significance. Feature importance is obtained through the split gain method based on ensemble learning. Its absolute importance value is converted into relative importance and cumulative contribution rate through normalization, which is specifically defined as:

[0155]

[0156] Among them, I iIt represents the contribution value of the i-th feature in the model training process, n is the total number of features, and k represents the cumulative impact of the top k features in the current ranking. The specific feature importance analysis is shown in Table 2:

[0157] Feature_Name Importance Relative_Importance Cumulative_Importance Temperature 4.123638 11.352999 11.352999 Pozzolanic_ratio 2.610181 7.186223 18.539222 Cl_concentration 2.463466 6.782295 25.321517 Time_effect 2.081445 5.730531 31.052048 Porosity_factor 1.946601 5.359286 36.411334

[0158] Table 2

[0159] In summary, this model achieves high-precision prediction on the basis of maintaining physical consistency, with a reasonable residual structure and stable error distribution. It can well reflect the changing trend of chloride ion diffusion behavior of concrete materials under different working conditions and has good prospects for engineering application.

[0160] System implementation and application

[0161] System Architecture

[0162] The system architecture of the present invention includes a data acquisition module, a data preprocessing module, a physical enhancement feature generation module, an XGBoost modeling module, a model training and prediction module, and a result output module. The system first collects concrete mix information and environmental parameters, including basic information such as cement content, water-cement ratio, temperature, and chloride ion concentration, through the data acquisition module; then, the data preprocessing module standardizes the original data, corrects outliers, and fills in missing values ​​to ensure data quality. The processed data is calculated by the physical enhancement feature generation module to form multidimensional input features such as diffusion time, temperature influence factor, chloride ion binding capacity, material composition characteristics, and durability index, which are used to characterize the diffusion behavior of concrete under different working conditions. All enhanced input features will be uniformly sent to the XGBoost modeling module to construct a prediction model based on an integrated regression tree, and the training module will perform iterative optimization on the training set and validation set to complete the automatic learning of model parameters. Finally, the model prediction module outputs the predicted value of the chloride ion diffusion coefficient, and can display the model performance index and feature importance ranking through the result output module, providing a basis for concrete material design and engineering durability evaluation.

[0163] System Application

[0164] This system is suitable for predicting and assessing the durability of concrete. It can assist engineers in optimizing material mixes during the design phase and can also be used for health monitoring and assessment of existing structures. Leveraging the system's intelligent prediction and analysis capabilities, it can effectively improve the safety and economic efficiency of concrete structures and promote the development of intelligent design and management in civil engineering.

Claims

1. A method for predicting chloride ion diffusion in concrete based on physical enhancement features and XGBoost model, characterized in that: The steps include: Step 1: Obtain the original experimental data of concrete and perform data preprocessing; Step 2: Using the data preprocessed in step 1, Fick's second law is used to construct diffusion time and time decay characteristics, the temperature effect term is calculated using the Arrhenius equation, and the chloride ion binding capacity and concentration decay effects are introduced to form physical mechanism characteristics; Extract material composition features that reflect the influence of concrete structural composition on diffusion paths; Construct durability characteristics, including intensity factor, durability index, diffusion boundary constraint and comprehensive diffusion index; Step 3: Build an XGBoost regression model based on the integrated regression tree structure. The physical mechanism characteristics, material composition characteristics, and durability characteristics of step 2 are integrated to form an XGBoost regression model with enhanced feature input. The model outputs the chloride ion diffusion coefficient of concrete and sets the model's hyperparameters. Step 4: Complete the training of the XGBoost model built in step 3 on the training set, input the enhanced features and target variables into the model, optimize the regression performance through multiple rounds of iterative learning, and enable the early stopping mechanism on the validation set; Step 5: Calculate the evaluation indicators of the model, including calculating the mean square error (MSE), root mean square error (RMSE), and determination coefficient (R). 2 Indicators, and optimize model parameters. If the following three conditions are met at the same time, the fitting effect is judged to be up to standard and go to step 6; Condition 1: The mean square error (MSE) value is between 0 and 1. Condition 2: The root mean square error (RMSE) value is between 0 and 1; Condition 3: Determination coefficient R 2 The value of is between 0.82-0.88; If the above three conditions cannot be met at the same time, the model is optimized by adjusting the model's hyperparameters; Step 6: Use the final model to perform actual predictions and feature importance analysis. At the same time, based on the feature importance output function of XGBoost, the relative contribution of each input variable is obtained, the key factors that have the greatest impact on diffusion performance are identified, and a quantitative basis is provided for auxiliary material design and durability control.

2. The method for predicting chloride ion diffusion in concrete based on physical enhancement features and XGBoost model according to claim 1, characterized in that: Step 1 includes the following steps: Step 11: Collect concrete mix data, including the proportions of cement, fly ash, mineral admixtures, water, and aggregate raw materials; Step 12: Obtain data on environmental factors affecting concrete, including external environmental parameters such as temperature, humidity, and chloride ion concentration; Step 13: Divide the data into training set and test set according to the proportion for model training and evaluation. Step 14: Perform data preprocessing, including removing outliers, filling missing values, and standardizing the data. The standardization method is as follows: Where: μ is the mean of the feature, σ is the standard deviation of the feature.

3. The method for predicting chloride ion diffusion in concrete based on physical enhancement features and XGBoost model according to claim 1, characterized in that: Step 2 includes the following steps: Step 21: Construct diffusion behavior-related features based on time parameters; calculate the square root of the exposure time t in the original data in years. As a diffusion time characteristic, it is used to approximately characterize the trend of diffusion distance over time, which conforms to the characteristic solution form of Fick's second law in a homogeneous medium; Step 22: Introduce a temperature correction term based on the time characteristic to construct a temperature-time coupling characteristic; Considering the accelerating effect of temperature on the diffusion process, the temperature influence factor is calculated based on the Arrhenius equation: Among them, E a =40000 J / mol is the activation energy of chloride ion migration, R = 8.314 J / (mol·K) is the universal gas constant, T is the actual ambient temperature, T ref =293.15K is the reference temperature; Step 23: Construct chloride ion concentration-related features to reflect the driving and coupling effects of the external medium on the diffusion process and calculate the chloride ion concentration effect features: Among them, C CI is the external chloride ion concentration, in g / L, reflecting the enhancement effect at high concentrations and the inhibition effect at low concentrations, and the chloride ion binding capacity characteristics are calculated from this: B CI =0.5·(1-exp(-0.1·C CI )) Step 24. Based on the physical diffusion mechanism, extract the relevant characteristics of concrete material composition to express the indirect effect of the mixing ratio on the diffusion path and calculate the total cementitious material mass: M binder =FA+GGBS+SF+OPC The pozzolan admixture ratio is calculated from this: Among them, FA is fly ash, GGBS is slag, SF is silica fume, and OPC is ordinary Portland cement, all in kg / m 3 , further calculate the effective water-binder ratio: Among them, r wb The original water-binder ratio is the ratio of the original water to the cementitious material. This expression takes into account the dilution effect of the admixture's reactivity and reflects the formation trend of the capillary pore structure inside the concrete. Step 25: Construct porosity and hydration-related features to supplement the microscopic channel effect of the material structure on ion diffusion; the volcanic ash reaction effect index is calculated as follows: f porosity =exp(-3·r pz ) Based on the ratio of water-binder ratio to cementitious material mass, the hydration potential index is constructed: f hydration =exp(-2·r wb_ratio ) Where W is the mass of mixing water kg / m 3 , which reflects the changing trend of the degree of hydration reaction and porosity inside concrete and is the key parameter controlling the formation of ion permeation channels; Step 26: Construct durability-related indicators, constrain the physical range of the prediction results, and calculate the strength factor and durability index: f durability =f strength ·f porosity in The critical water-binder ratio is set empirically, and the boundary value of the diffusion coefficient is set as a constraint: D max =10 -11 ·exp(2·r wb ) D min ＝10 -13 。 4. The method for predicting chloride ion diffusion in concrete based on physical enhancement features and XGBoost model according to claim 1, characterized in that: Step 3 includes the following steps: Step 31: Combine the physical mechanism features, material composition features, and durability features from step 2 to construct a unified enhanced input feature as the data basis for machine learning modeling. Step 32: Based on the enhanced input features and their corresponding chloride ion diffusion coefficient labels, define the regression task objectives and clarify the supervised learning modeling paradigm centered on predicting continuous values; Step 33: Select the gradient boosting decision tree XGBoost as the regression model framework; Step 34. Configure the hyperparameters of the XGBoost model, including the maximum tree depth max_depth, learning rate eta, minimum child weight min_child_weight, column sampling ratio colsample_bytree, subsample ratio subsample, regularization term lambda, alpha, to enhance model fitting ability, control complexity and prevent overfitting; Step 35. Define the training objective of the XGBoost model as minimizing the following loss function: Among them, y i is the true chloride ion diffusion coefficient of the i-th sample, is the predicted value, Ω(f k ) represents the regularization term of the k-th regression tree, including the number of leaves and the penalty of leaf weight; the objective function is automatically optimized by the XGBoost framework, supplemented by RMSE and R 2 As an evaluation indicator, it provides performance guarantee for the subsequent training process.

5. The method for predicting chloride ion diffusion in concrete based on physical enhancement features and XGBoost model according to claim 1, characterized in that: Step 4 includes the following steps: Step 41: Based on the XGBoost model structure, the enhanced features and target values ​​are input into the model to form the data input of the supervised learning task. Step 42: Set the maximum number of iterations for model training and enable early stopping. Evaluate the loss on the validation set after each training round. Terminate training early if the loss stops decreasing after several consecutive rounds to prevent overfitting of the model to the training set. Step 43: In each iteration, the XGBoost framework is used to construct a new regression tree based on the current residual distribution to fit the prediction error of the previous round and gradually optimize the overall prediction performance. Step 44: Dynamically adjust the splitting condition, node weight, and regularization coefficient of each subtree according to the configured hyperparameters, so that the model can control the structural complexity while learning the complex feature structure; Step 45: After the training is completed, the set of model parameters with the best performance on the validation set is automatically selected as the final model output for subsequent evaluation and prediction.

6. The method for predicting chloride ion diffusion in concrete based on physical enhancement features and XGBoost model according to claim 1, characterized in that: In step 5, the mean square error MSE, root mean square error RMSE and determination coefficient R 2 , which is calculated as follows: Mean square error (MSE): used to calculate the difference between the predicted value and the true value. in, Represents the model prediction value, y i Represents the true value of the sample, and n is the total number of samples; Root mean square error RMSE: The square root of MSE is used to provide a measure of the error. Coefficient of determination R 2 Value: measures how well the Bayesian neural network model fits the data.

7. A computer device comprising a memory, a processor, and a computer program stored in the memory, wherein: The processor executes the computer program to implement the steps of the method according to claim 1.

8. A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that: When the computer program / instructions are executed by a processor, the steps of the method according to claim 1 are implemented.

9. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of the method according to claim 1 are implemented.

Images