Chloride ion diffusion coefficient prediction method based on hybrid machine learning model

Abstract

The invention discloses a chloride ion diffusion coefficient prediction method based on a hybrid machine learning model, and the method comprises the steps: widely collecting actual data, carrying out the preprocessing, carrying out the selection of parameters affecting a chloride ion diffusion coefficient through an LASSO regression feature selection method, and preventing the overfitting problem of a regression model on a training set; carrying out regression analysis on a plurality of parameters of the chloride ion diffusion coefficient on the training set by utilizing a distributed gradient boosting algorithm (LightGBM) based on a decision tree algorithm, taking data obtained by carrying out dimension raising processing on an obtained predicted value and an original parameter as input of GPR (Gaussian process regression), taking a true value as output, training a GPR model, analyzing a GPR prediction result, and obtaining a prediction result of the chloride ion diffusion coefficient. And a light GBM-GPR model is constructed by using a Bayesian parameter optimization method. According to the method, the characteristic that the LightGBM prediction error is small is reserved, the uncertainty probability interval is given, and the prediction capacity of the chloride ion diffusion coefficient is improved.

Description

Technical Field

[0001] The present invention relates to the field of prediction of chloride ion diffusion coefficient of concrete, and in particular to a method for predicting chloride ion diffusion coefficient based on a hybrid machine learning model. Background Art

[0002] Concrete is a material with a heterogeneous, multi-scale structure composed of cement, coarse aggregate and fine aggregate. Chloride ions are one of the main causes of steel corrosion in concrete structures. When chloride ions diffuse into the concrete and reach the surface of the steel bars, they will destroy the passivation film of the steel bars and cause steel corrosion. Different concrete mix ratios and material selections will significantly affect the diffusion performance of chloride ions. Studying the chloride ion diffusion coefficient can optimize the concrete mix ratio and material selection and enhance the concrete's resistance to chloride ion erosion. Frequent maintenance and replacement of concrete structures with poor durability not only increases resource consumption, but also has a negative impact on the environment. The prediction of the concrete chloride ion diffusion coefficient is of great significance to the durability design of concrete structures.

[0003] At present, most of the research on the chloride ion diffusion coefficient of concrete adopts traditional experimental methods, analyzing certain special physical structures in concrete, starting from certain concrete containing special materials, and starting from the chloride ion diffusion coefficient of concrete at a certain age. Some existing methods provide specific perspectives for studying the chloride ion diffusion coefficient of concrete, but may face the problem of limited applicability. However, with the rapid development of the computer field, especially the development of artificial intelligence, the method of concrete chloride ion diffusion coefficient based on machine learning algorithms should be explored more deeply. Machine learning regression prediction only needs to collect various characteristic parameters of concrete to make predictions, but generally single machine learning regression algorithms directly act on the field of concrete chloride ion diffusion coefficient prediction. Each has its own advantages, such as some algorithms have higher accuracy and some algorithms are more powerful. Although the overall accuracy of machine learning for the prediction of concrete chloride ion diffusion coefficient is high, it is inevitable that there are errors. The prediction error of the composite machine learning model may not only not converge but diverge, and cannot reveal the possible distribution of the chloride ion diffusion coefficient. Summary of the invention

[0004] The problem to be solved by the present invention is to provide a method for predicting the chloride ion diffusion coefficient based on a hybrid machine learning model. By integrating the LightGBM model and the GPR model, the problem that the error of the composite model not only does not converge but diverges, is solved, thereby improving the prediction accuracy of the chloride ion diffusion coefficient of concrete.

[0005] The present invention adopts the following technical solution: a method for predicting chloride ion diffusion coefficient based on a hybrid machine learning model, comprising the following steps:

[0006] Step 1: Collect measured sample data and clean the sample data;

[0007] Step 2: Perform LASSO regression feature selection on the cleaned sample data: put the potential feature variables of the sample data into LASSO regression for processing. After cross-validation, select the potential feature variable coefficient corresponding to the minimum mean square error, and retain the features that contribute to the target value;

[0008] Step 3: Use random sampling to train sample data and construct training sets and test sets;

[0009] Step 4: Build a distributed gradient boosting algorithm (LightGBM) based on the decision tree algorithm, train the LightGBM model on the training set, and give the predicted values ​​of the trained LightGBM model on the training set and test set;

[0010] Step 5: Use the predicted value of the LightGBM model as an additional column to increase the dimension of the training set and test set data;

[0011] Step 6: Build a Gaussian process regression (GPR) model, use the dimension-upgraded test set data as the input of the GPR model, use the true value of the measured data as the output of the GPR model, and train the GPR model;

[0012] Step 7: Analyze the prediction ability of the GPR model on the training set, automatically optimize the hyperparameters of the GPR model through the Bayesian parameter optimization method, and obtain the optimized GPR model;

[0013] Step 8. Fusion of the LightGBM model and the optimized GPR model to obtain the LightGBM-GPR model: the parameter matrix of the test set after the dimension-upgrading process in step 5 is used as the input of the optimized GPR model, and the prediction result of the LightGBM-GPR model is obtained according to the output of the GPR model;

[0014] Step 9: Perform an overall evaluation of the prediction results of the LightGBM-GPR model to obtain the numerical value of the predicted chloride ion diffusion coefficient and the uncertainty probability interval of the chloride ion diffusion coefficient.

[0015] Specifically, in step 1, a small amount of data in the original data may have missing items, and the missing data can be deleted. There may also be data record errors in the original data due to various reasons. For some abnormal data, some classic data cleaning methods can be used to remove them;

[0016] For example, using the interquartile range (IQR) method, by calculating the interquartile range (IQR) of each parameter column, values ​​that are 1.5 times the IQR below the first quartile or 1.5 times the IQR above the third quartile are considered outliers.

[0017] For example: using the standard score (z-score) method, the data after the missing data deletion operation is standardized according to the parameter column as z = (x-μ) / σ, where μ is the mean of each column and σ is the standard deviation of each column, and outliers with z-score absolute values ​​greater than 3 are deleted.

[0018] Specifically, in step 2, the objective function of LASSO regression feature selection is set as:

[0019]

[0020] Among them, y i is the true value of the chloride ion diffusion coefficient corresponding to each row of data, x i is a row vector with 1 row and j columns, which contains j parameter values, i is the total number of data in the data set, j is the total number of parameters, β is a column vector with j rows and 1 columns, which contains the assignment of j parameters, λ is the penalty coefficient, and ||||1 is the L1 norm;

[0021] LASSO regression uses L1 regularization to compress the sparse coefficients of redundant variables to 0 by adjusting the size of the penalty coefficient to achieve feature screening.

[0022] Specifically, in step 3, in order to balance the samples during training and improve the model performance, random sampling is used for training, so that the training samples account for 80% of the total sample data volume and the test samples account for 20% of the total sample data volume.

[0023] Specifically, in step 4, the LightGBM model is called on the training set for the processed sample data. After the LightGBM model is trained, the corresponding prediction values ​​can be given according to the parameters of the data on the training set and the test set. LightGBM is an optimization and efficient implementation of the classic algorithm gradient boosting decision tree (GBDT). In terms of effect and principle, it has faster training efficiency.

[0024] Build the LightGBM model, based on the Histogram decision tree algorithm. The Histogram of a leaf node can be directly obtained by subtracting the Histogram of the parent node and the Histogram of the sibling node, which greatly improves the training speed and reduces memory usage: The LightGBM model uses the histogram algorithm, which occupies less memory, has lower data segmentation complexity, and higher accuracy: The LightGBM model adopts the Leaf-wise growth strategy. Each time, it finds a leaf with the largest split gain (generally the largest amount of data) from all current leaves, and then splits it, and repeats it in this way. Compared with XGBoost's Level-wise, Leaf-wise can reduce more errors and obtain better accuracy when the number of splits is the same.

[0025] Furthermore, the LightGBM model supports parallel learning: including feature parallelism and data parallelism, and optimizations are made for both parallel methods.

[0026] Specifically, in step 5, it is assumed that the collected original sample data is obtained by deleting missing items and excluding outliers to obtain a data table. After random sampling, taking the training set as an example, the table is a matrix of m rows and n+1 columns, where m is the number of data, n is the number of parameters of the chloride ion diffusion coefficient, the n+1th column is the true value of the chloride ion diffusion coefficient, and each of the first n columns is the value corresponding to the parameter of the column. After LASSO regression feature selection in step 2, it is assumed that the data table becomes a matrix of m rows and q+1 columns, where q≤n, and the q+1th column is still the true value of the chloride ion diffusion coefficient. At this time, if the last column of true values ​​is removed and only the parameter data is considered, an m-row and q-column parameter matrix can be obtained.

[0027] At this time, the predicted value of the LightGBM model is added to the last column of the parameter matrix to perform a data dimension increase operation, that is, the parameter matrix is ​​augmented to obtain a parameter matrix of m rows and q+1 columns. The LightGBM predicted value is added as a parameter, which increases the effective information of the GPR model input.

[0028] Furthermore, in step 6, the dimensionally upgraded data, i.e., the parameter matrix of m rows and q+1 columns, is used as the input of the GPR model, and the true value of the chloride ion diffusion coefficient is used as the output of the GPR model to train the GPR model.

[0029] The GPR model is a nonparametric Bayesian regression method, which is widely used in machine learning and statistics. It is based on Gaussian process and can provide predictions while quantifying the uncertainty of predictions. The GPR principle is to calculate the posterior distribution of the result by the likelihood function of the prior distribution and the data. The core of the process is the kernel function, which is a positive covariance matrix. The autocorrelation of the training set is calculated. Based on the Bayesian framework, the posterior distribution, output value and covariance function of the test set can be obtained. Based on the above results, a 95% confidence interval can be simply calculated.

[0030] Furthermore, in step 7, the prediction ability of GPR on the training set is analyzed, such as comparing the mean square error (MSE), root mean square error (RMSE) and coefficient of determination (R2). If the error is large and the evaluation ability is weak, the Bayesian parameter optimization method is used to automatically optimize the hyperparameters of the GPR model.

[0031] The Bayesian parameter optimization method is suitable for hyperparameter tuning of machine learning models. Bayesian optimization approximates the objective function by using a proxy model (such as a Gaussian process). The proxy model is trained based on known observation data to capture the behavior of the objective function. The acquisition function is used to determine the sampling position of the next parameter. The proxy model and the acquisition function are updated according to the new parameter sampling results, and the above process is repeated until the optimization goal is achieved or the stopping condition is reached.

[0032] Furthermore, in step 8, for the GPR model that has been Bayesian parameter tuned on the training set, the trained LightGBM model is used to give a predicted value on the test set, and the data on the test set is dimensionally upgraded as the input of the tuned GPR. At this time, the GPR model gives the predicted mean and predicted variance of the chloride ion diffusion coefficient corresponding to each data, and calculates the predicted confidence interval corresponding to each data.

[0033] The technical solution of the present invention also provides: an electronic device, comprising:

[0034] one or more processors;

[0035] a storage device having one or more programs stored thereon;

[0036] When the one or more programs are executed by the one or more processors, the one or more processors implement any of the above-mentioned methods for predicting chloride ion diffusion coefficient based on the hybrid machine learning model.

[0037] The technical solution of the present invention also provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the steps in any of the above-mentioned methods for predicting the chloride ion diffusion coefficient based on a hybrid machine learning model are implemented.

[0038] Compared with the prior art, the present invention adopts the above technical solution and has the following technical effects:

[0039] 1. The chloride ion diffusion coefficient prediction method of the present invention integrates the LightGBM model and the GPR model, which not only retains the high prediction accuracy of the LightGBM model, but also retains the function of the GPR model to predict the mean, variance and confidence interval of the data, and has improved the prediction accuracy compared with both the LightGBM model and the GPR model.

[0040] 2. For existing composite GPR models, the first prediction results of other regression models are usually used directly as GPR input, and the true value of the data is used as GPR output to train the model. The method of the present invention retains the original parameter information and performs unique data dimensionality upgrading before using the GPR model, thereby reducing the information loss problem when the composite GPR model is used, making the prediction error smaller and solving the problem that the composite model error not only fails to converge but diverges.

[0041] 3. The chloride ion diffusion coefficient prediction method of the present invention encapsulates the physical laws of concrete characteristics and its chloride ion diffusion coefficient into a machine learning algorithm. The chloride ion diffusion coefficient can be predicted by only obtaining relevant parameters of the concrete, which is simple and accurate.

[0042] 4. The chloride ion diffusion coefficient prediction method of the present invention predicts the chloride ion diffusion coefficient of concrete through a composite model (LightGBM-GPR model). Compared with the direct prediction using the classical machine learning regression algorithm, it has the advantage of high accuracy and provides the function of uncertainty probability interval prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is a flow chart of the method for predicting chloride ion diffusion coefficient based on the hybrid machine learning model of the present invention;

[0044] Figure 2 The result graph of potential feature variable coefficients for parameter selection of LASSO regression feature selection method;

[0045] Figure 3 A regression analysis visualization diagram on a test set according to an embodiment of the present invention;

[0046] Figure 4 Visualization of regression analysis on the LightGBM model test set;

[0047] Figure 5 It is a regression analysis visualization diagram of the BP neural network model on the test set of the instance;

[0048] Figure 6 Visualize the regression analysis of the convolutional neural network (CNN) model on the test set on the instance;

[0049] Figure 7 Visualization of regression analysis of the Temporal Convolutional Network (TCN) model on the test set on the instance;

[0050] Figure 8 Visualization of regression analysis of the least squares support vector machine (LSSVR) model on the test set on the instance;

[0051] Fig. 9 Visualization of regression analysis of the Long Short-Term Memory (LSTM) model on the test set on the instance;

[0052] Fig.10 Visualization of the regression analysis of the GPR model on the test set on the instance;

[0053] Fig.11 The figure is a comparison chart of mean square error (MSE), root mean square error (RMSE) and coefficient of determination (R2) data between the present invention and each single model. DETAILED DESCRIPTION

[0054] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the application is further elaborated in detail below in conjunction with the accompanying drawings. The described embodiments are only a part of the embodiments involved in the present invention. All non-innovative embodiments of other researchers in the field on this embodiment belong to the protection scope of the present invention. At the same time, for the step numbering in the embodiment of the present invention, it is only set for the convenience of explanation, and the order between the steps is not limited in any way. The execution order of each step in the embodiment can be adaptively adjusted according to the understanding of those skilled in the art.

[0055] The purpose of the present invention is to provide a composite machine learning method, that is, to construct a LightGBM-GPR model, and to predict the chloride ion diffusion coefficient of concrete by integrating the LightGBM algorithm and the GPR algorithm. This method has the advantages of high accuracy in predicting the chloride ion diffusion coefficient of concrete using the LightGBM algorithm alone, and the function of providing a confidence interval of the predicted value by the GPR algorithm. Compared with the existing non-machine learning method for predicting the chloride ion diffusion coefficient of concrete, this method has the characteristics of high accuracy, providing a simpler prediction of the distribution of the chloride ion diffusion coefficient and a more convenient prediction process.

[0056] At the same time, compared with the direct use of some regression algorithms in classic machine learning such as BP neural network, convolutional neural network (CNN), temporal convolutional network (TCN), Gaussian process regression (GPR), least squares support vector machine (LSSVR), distributed gradient boosting based on decision tree algorithm (LightGBM) and long short-term memory network (LSTM) algorithm, the LightGBM-GPR model of the present invention still has the function of high accuracy and giving distribution prediction.

[0057] In one embodiment of the present invention, a chloride ion diffusion coefficient prediction method based on a hybrid machine learning model is as follows: Figure 1 As shown, the following steps are included:

[0058] Step 1: Collect and summarize measured data, and perform simple data cleaning on the original data, including deleting missing data and checking for extreme abnormal data;

[0059] Step 2: Perform LASSO regression feature selection on the preliminarily processed data, put the potential feature variables into LASSO regression for processing, select the coefficient corresponding to the minimum mean square error after cross-validation, and select the feature parameters that contribute to the target value;

[0060] Step 3: In order to balance the samples during training and improve the model performance, random sampling is used for training, so that the training samples account for 80% of the total data volume and the test samples account for 20% of the total data volume;

[0061] Step 4: Train the LightGBM model on the training set and give the predicted values ​​of the trained model on the training set and test set;

[0062] Step 5, perform column augmentation operation on the parameter matrix of the training set to increase its parameter column by 1 column, where the added 1 column is the predicted value given by the LightGBM model in step 4. Similarly, perform an augmentation operation on the parameter matrix of the test set to increase the number of columns of the parameter matrix by 1, thus completing the dimensionality increase processing of the data using the predicted value of the LightGBM model in step 4;

[0063] Step 6, using the upgraded test data as the input of the GPR model and the true value of the measured data as the output of the GPR model to train the GPR model;

[0064] Step 7, analyzing the prediction results of the GPR model on the training set, and automatically optimizing the hyperparameters of the GPR model using the Bayesian parameter optimization method to further optimize the secondary prediction value of the GPR model;

[0065] Step 8: Use the parameter matrix of the test set processed in step 5 as the input of the built GPR model. At this time, the prediction result given by GPR is the result of the LightGBM-GPR model, which includes the point estimate and confidence interval estimate of the data on the test set.

[0066] Step 9, organize the prediction results of the LightGBM-GPR model. The LightGBM-GPR model not only predicts the value of the chloride ion diffusion coefficient but also predicts the uncertainty probability interval of the chloride ion diffusion coefficient.

[0067] In this embodiment, step 1 collects a total of 625 pieces of original data. There are no missing items in the original data, and no obvious abnormal values ​​are found, so all the data are retained;

[0068] Furthermore, in step 2, the objective function of LASSO regression feature selection is set as:

[0069]

[0070] Among them, y i is the true value of the chloride ion diffusion coefficient corresponding to each row of data, x i is a row vector with 1 row and j columns, which contains j parameter values, where i is the total number of data in the data set, j is the total number of parameters, β is a column vector with j rows and 1 column, which contains the assignments of j parameters, λ is the penalty coefficient, and ||||1 is the L1 norm.

[0071] It is particularly important to note that when building a prediction model, if too many variables are selected, it is easy to cause the prediction model to overfit, that is, the model performs well in the training set but poorly in the test set. Regularization can effectively solve the problem of model overfitting. Regularization is a form of regression that constrains, adjusts or reduces the coefficient estimate in the direction of zero, thereby reducing the complexity and instability of the model during the learning process and avoiding the risk of overfitting. Therefore, LASSO regression uses L1 regularization to compress the sparse coefficients of redundant variables to 0 by adjusting the size of the penalty coefficient to achieve feature screening.

[0072] In this embodiment, each data contains 12 parameters about the chloride ion diffusion coefficient of concrete, which are: cement (OPC) kg / m 2 、Fly ash (FA) kg / m 3 , blast furnace slag (GGBS) kg / m 3 Silica fume (SF) kg / m 3 、Superplasticiser kg / m 3 , Water (Water)kg / m 3 、Fine aggregate (Fine agg) kg / m3 , coarse aggregate (Coarseagg)kg / m 3 , water-cement ratio (w / b), exposure time (Exposure time) t / a, temperature (Annual meantemperature) T / ℃, chloride ion concentration in seawater ([Cl - ]in seawater)g / L, and the chloride ion diffusion coefficient (Dc) in this state is 10 -12 m 2 / s, therefore, each piece of data contains 13 items of data, and at this time n=625, j=12.

[0073] like Figure 2 As described above, the 12 potential feature variables are put into LASSO regression for processing. After cross-validation, the coefficient corresponding to the minimum mean square error is selected. It can be found that the coefficient of each potential feature variable is not compressed to 0, indicating that all 12 features contribute to the target value, so all features are retained.

[0074] Furthermore, in step 4, the LightGBM model is called on the training set for the processed sample data. After the LightGBM model is trained, the corresponding prediction values ​​can be given according to the parameters of the data on the training set and the test set.

[0075] Specifically, the LightGBM model is built based on the decision tree algorithm of Histogram. The Histogram of a leaf node can be directly obtained by subtracting the Histogram of the parent node and the Histogram of the sibling node, which greatly improves the training speed and reduces memory usage: LightGBM uses the histogram algorithm, which occupies less memory, has lower complexity in data segmentation, and has higher accuracy: LightGBM adopts the Leaf-wise growth strategy, each time finding a leaf with the largest split gain (generally the largest amount of data) from all current leaves, and then splits it, and so on. Compared with XGBoost's Level-wise, Leaf-wise can reduce more errors and obtain better accuracy when the number of splits is the same, and supports parallel learning: LightGBM natively supports parallel learning, and currently supports feature parallelism and data parallelism. LightGBM has optimized both parallel methods;

[0076] The LightGBM model in this embodiment is an optimization and efficient implementation of the classic algorithm gradient boosting decision tree (GBDT). In terms of effect and principle, it has faster training efficiency.

[0077] Further, in step 5, the raw data collected in the present embodiment is obtained by deleting the missing items and excluding the abnormal values ​​to obtain a data table. After random sampling, taking the training set as an example, the table is a matrix of 500 rows and 12+1 columns, wherein 500 is the number of data, 12 is the number of parameters of the chloride ion diffusion coefficient, and the 12+1 column is the true value of the chloride ion diffusion coefficient, and each of the first 12 columns is the value corresponding to the column parameter. After LASSO regression feature selection in step 2, the data table becomes a matrix of 500 rows and 12+1 columns, and the 12+1 column is still the true value of the chloride ion diffusion coefficient. At this time, if the last column of true values ​​is removed and only the parameter data is considered, a parameter matrix of 500 rows and 12 columns can be obtained. At this time, the predicted value of the LightGBM model is added to the last column of the parameter matrix, i.e., the parameter matrix is ​​subjected to column augmentation operation to obtain a parameter matrix of 500 rows and 12+1 columns, which is the so-called data dimensioning operation, which adds the LightGBM predicted value as a parameter, and increases the effective information input by the GPR model.

[0078] Furthermore, in step 6, the dimensionally upgraded data, i.e., the parameter matrix of 500 rows and 12+1 columns, is used as the input of the GPR model, and the true value of the chloride ion diffusion coefficient is used as the output of the GPR model to train the GPR model.

[0079] In particular, in this embodiment, Gaussian process regression (GPR) is a non-parametric Bayesian regression method, which is widely used in machine learning and statistics. It is based on Gaussian process and can provide predictions while quantifying the uncertainty of the predictions. The GPR principle is to calculate the posterior distribution of the result based on the likelihood function of the prior distribution and the data. The core of the process is the kernel function, which is a positive covariance matrix. The autocorrelation of the training set is calculated, and the posterior distribution, output value and covariance function of the test set can be obtained based on the Bayesian framework. Based on the above results, a 95% confidence interval can be simply calculated.

[0080] Furthermore, in step 7, the prediction ability of GPR on the training set is analyzed, such as comparing the mean square error (MSE), root mean square error (RMSE) and coefficient of determination (R2). If the error is large and the evaluation ability is weak, the Bayesian parameter optimization method is used to automatically optimize the hyperparameters of the GPR model.

[0081] The Bayesian parameter optimization method is particularly suitable for hyperparameter tuning of machine learning models. Bayesian optimization approximates the objective function by using a proxy model (such as a Gaussian process). The proxy model is trained based on known observation data to capture the behavior of the objective function. The acquisition function is used to determine the sampling position of the next parameter. The proxy model and acquisition function are updated based on the new parameter sampling results, and the above process is repeated until the optimization goal is achieved or the stopping condition is reached.

[0082] Furthermore, in step 8, for the GPR model that has been Bayesian parameter tuned on the training set, the trained LightGBM model is used to give prediction values ​​on the test set, and the data on the test set is dimensionally upgraded as the input of the tuned GPR. At this time, the GPR model gives the predicted mean and predicted variance of the chloride ion diffusion coefficient corresponding to each data. At this time, it is not difficult to calculate the predicted confidence interval corresponding to each data.

[0083] Further, in step 9, the prediction results of the LightGBM-GPR model of this embodiment are analyzed as a whole, such as Figure 3 At the same time, a single machine learning regression algorithm model is used to predict the chloride ion diffusion coefficient of concrete, and the regression analysis results on the test set are visualized, as shown in Figures 4 to 10 shown. Specifically, Figure 4 This is a regression analysis visualization on the LightGBM model test set. Figure 5 This is a visualization of the regression analysis of the BP neural network model on the test set of the instance. Figure 6 Visualization of the regression analysis of the convolutional neural network (CNN) model on the test set on the instance. Figure 7 Visualization of regression analysis of the temporal convolutional network (TCN) model on the test set on the instance. Figure 8 This is a visualization of the regression analysis of the least squares support vector machine (LSSVR) model on the test set on the instance. Fig. 9 This is a visualization of the regression analysis of the Long Short-Term Memory Network (LSTM) model on the test set on the instance. Fig.10 Visualization of the regression analysis of the GPR model on the test set on the instance.

[0084] Further, Figures 4 to 10 The evaluation parameters for predicting the chloride ion diffusion coefficient of concrete using a single machine learning regression algorithm model and the LightGBM-GPR model of the present invention are compared, such as Fig.11 As shown in the figure, they specifically include: mean square error (MSE), root mean square error (RMSE) and determination coefficient (R2), where MSE and RMSE describe the overall prediction accuracy of the model on the data set, and R2 describes how well the model fits the actual situation.

[0085] from Fig.11It can be seen that the single LightGBM model has the highest overall prediction accuracy and the highest degree of fit compared with other single regression algorithms, and the GPR model provides a prediction of an uncertainty probability interval. The prediction accuracy and degree of fit of the LightGBM-GPR model of the present invention are slightly improved compared with the single LightGBM model, and are more improved than the GPR model. The model of the present invention provides an uncertainty probability interval prediction, and the uncertainty probability interval prediction accuracy is higher than that of the single GPR model.

[0086] It can be seen that, in general, the prediction of chloride ion diffusion coefficient of concrete by the LightGBM-GPR model of the present invention has the highest prediction accuracy and fitting effect, and provides a more reliable uncertainty probability interval.

[0087] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for predicting chloride ion diffusion coefficient based on a hybrid machine learning model, characterized in that: The steps include: Step 1: Collect measured sample data and clean the sample data; Step 2: Perform LASSO regression feature selection on the cleaned sample data: put the potential feature variables of the sample data into LASSO regression for processing. After cross-validation, select the potential feature variable coefficient corresponding to the minimum mean square error, and retain the features that contribute to the target value; Step 3: Use random sampling to train sample data and construct training sets and test sets; Step 4: Build a LightGBM model, train the LightGBM model on the training set, and give the predicted values ​​of the trained LightGBM model on the training set and test set; Step 5: Use the predicted value of the LightGBM model as an additional column to increase the dimension of the training set and test set data; Step 6: Build a GPR model, use the upgraded test set data as the input of the GPR model, use the true value of the measured data as the output of the GPR model, and train the GPR model; Step 7: Analyze the prediction ability of the GPR model on the training set, automatically optimize the hyperparameters of the GPR model through the Bayesian parameter optimization method, and obtain the optimized GPR model; Step 8. Fusion of the LightGBM model and the optimized GPR model to obtain the LightGBM-GPR model: the parameter matrix of the test set after the dimension-upgrading process in step 5 is used as the input of the optimized GPR model, and the prediction result of the LightGBM-GPR model is obtained according to the output of the GPR model; Step 9: Perform an overall evaluation of the prediction results of the LightGBM-GPR model to obtain the numerical value of the predicted chloride ion diffusion coefficient and the uncertainty probability interval of the chloride ion diffusion coefficient.

2. The method for predicting chloride ion diffusion coefficient based on a hybrid machine learning model according to claim 1, characterized in that: In step 1, the data cleaning includes: deleting missing data and checking for abnormal data; the method for checking for abnormal data includes: interquartile range method and standard score method.

3. The method for predicting chloride ion diffusion coefficient based on a hybrid machine learning model according to claim 1, characterized in that: In step 2, the LASSO regression feature selection, the objective function is set as follows: Among them, y i is the true value of the chloride ion diffusion coefficient corresponding to each row of data, x i is a row vector with 1 row and j columns, containing j parameter values, where i is the total number of data in the data set, j is the total number of parameters, β is a column vector with j rows and 1 column, containing the assignments of j parameters, λ is the penalty coefficient, and ||||1 is the L1 norm; LASSO regression uses L1 regularization, and by adjusting the size of the penalty coefficient, the sparse coefficients of redundant feature variables are compressed to 0 for feature screening.

4. The method for predicting chloride ion diffusion coefficient based on a hybrid machine learning model according to claim 1, characterized in that: In step 4, the LightGBM model is built based on the Histogram decision tree algorithm and supports parallel learning. The Histogram of a leaf node is obtained by subtracting the Histogram of the parent node and the Histogram of the sibling node, and the histogram algorithm is used to adopt the Leaf-wise growth strategy. Each time, the leaf with the largest split gain is found from all current leaves, and split and cycle; the parallel learning, Includes feature parallel learning and data parallel learning.

5. The method for predicting chloride ion diffusion coefficient based on a hybrid machine learning model according to claim 1, characterized in that: In step 5, the training set and test set are parameter matrices with m rows and q columns, respectively, where m is the number of data and q is the number of parameters of the chloride ion diffusion coefficient after LASSO regression feature selection in step 2. The parameter matrices of the training set and the test set are augmented by columns, and the predicted values ​​of the LightGBM model on the training set and the test set given in step 4 are added to the last column of the parameter matrices of the training set and the test set, respectively, to obtain a parameter matrix with m rows and q+1 columns.

6. The method for predicting chloride ion diffusion coefficient based on a hybrid machine learning model according to claim 5, characterized in that: In step 6, the GPR model is a Gaussian process regression model, which is based on the Gaussian process and uses the positive definite covariance matrix as the kernel function to calculate the autocorrelation of the training set, obtain the posterior distribution of the test set based on the Bayesian framework, output the value and covariance function, and calculate the 95% confidence interval; The parameter matrix of the test set with m rows and q+1 columns after dimensionality increase is used as the input of the GPR model, and the true value of the chloride ion diffusion coefficient is used as the output of the GPR model to train the GPR model.

7. The method for predicting chloride ion diffusion coefficient based on a hybrid machine learning model according to claim 6, characterized in that: In step 7, the prediction ability of the GPR model on the training set is analyzed, and the method includes: comparing the mean square error, the root mean square error and the determination coefficient, automatically optimizing the hyperparameters of the GPR model through the Bayesian parameter optimization method, and approximating the objective function by using a proxy model, wherein the proxy model includes a Gaussian process, is trained based on known observation data, captures the behavior of the objective function, uses an acquisition function to determine the sampling position of the next parameter, updates the proxy model and the acquisition function based on the new parameter sampling results, and repeats the updating process until the optimization goal is reached or the stopping condition is reached, thereby obtaining the optimized GPR model.

8. The method for predicting chloride ion diffusion coefficient based on a hybrid machine learning model according to claim 1, characterized in that: In step 8, based on the GPR model after Bayesian parameter optimization on the training set, the trained LightGBM model is used to give the predicted value on the test set, and the sample data on the test set is dimensionally upgraded as the input of the optimized GPR model. The output of the GPR model is the predicted mean and predicted variance of the chloride ion diffusion coefficient corresponding to each sample data. The predicted confidence interval corresponding to each sample data is calculated to obtain the prediction result of the LightGBM-GPR model, including the point estimate of the data on the test set and the estimate of the uncertainty probability interval.

9. An electronic device, characterized in that: include: one or more processors; a storage device having one or more programs stored thereon; When the one or more programs are executed by the one or more processors, the one or more processors implement the chloride ion diffusion coefficient prediction method based on the hybrid machine learning model as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and when the program is executed by a processor, the steps in the method for predicting the chloride ion diffusion coefficient based on a hybrid machine learning model described in any one of claims 1 to 8 are implemented.

Images