A RNN concrete adiabatic temperature rise prediction method based on dynamic parameter improvement

Abstract

The present invention discloses a RNN concrete adiabatic temperature rise prediction method for dynamic parameter improvement, which is aimed at temperature control and crack prevention of large-volume concrete structures such as dams. In view of the influence of the diversity of admixtures in modern concrete on the hydration process, an accurate means is provided for predicting adiabatic temperature rise, and a deep learning model of support vector regression, artificial neural network and recurrent neural network is adopted. 15 main variables are selected as input parameters through principal component analysis to predict the adiabatic temperature rise of concrete containing complex mixture components; segmentation technology and bidirectional deep recurrent neural network are used to improve the performance of RNN and improve the accuracy of model prediction; through K-fold cross validation optimization, the effectiveness and reliability of RNN and bidirectional deep recurrent neural network models in complex concrete structures are proved; an advanced solution is provided for predicting adiabatic temperature rise, and the adiabatic temperature rise curve is accurately fitted, which helps to ensure the safety and durability of large-volume concrete structures.

Description

Technical Field

[0001] The present invention relates to the technical field of thermal insulation temperature rise prediction of large-volume concrete structures, and in particular to an RNN concrete thermal insulation temperature rise prediction method for dynamic parameter improvement. Background Art

[0002] Concrete is the most commonly used material in engineering structures everywhere, and has a wide range of construction applications in high-rise buildings, dams, long-span bridges and tunnels. In these projects, mass concrete has become the material of choice for structural components with high strength and durability requirements, such as foundations and abutments. Studies have found that one of the main challenges in using mass concrete is to control the heat generated during the hydration of concrete, which can cause significant temperature gradients, leading to different forms of concrete cracking and damage. Therefore, temperature and thermal crack control is crucial for mass concrete to ensure the safety and durability of the structure. Adiabatic temperature rise is the increase in the temperature inside the concrete during the hydration process, which is a core research area in mass concrete. Direct adiabatic temperature rise experimental methods exist, but they are not always feasible due to their cost, consumables and complexity. Semi-adiabatic temperature rise tests and multiple prediction models have been proposed, but when various admixtures are added to the mix design of modern multi-component concrete, the hydration process of concrete becomes more complicated, which may cause the failure of existing prediction models.

[0003] In recent years, with the rapid development of artificial intelligence and deep learning (DL) technologies, they have shown amazing prediction and classification capabilities in multiple fields. However, DL in concrete science mainly focuses on mechanical properties, workability and durability (carbonation, freeze-thaw, chloride diffusion, alkali-silica reaction, etc.), and the research on cement hydration and thermal behavior of concrete is still limited. There are still challenges in exploring the thermal behavior of concrete using DL. In most cases, the datasets used to train and test DL models are selected and established from individual laboratories or field structures. The applicability and universality of these models have not been verified by datasets from multiple sources. DL models generally require large-capacity data analysis training to improve prediction accuracy, and the available data sources for adiabatic temperature rise of concrete are relatively scarce due to the high cost and time-consuming adiabatic temperature rise experiments. Existing DL thermal analysis method studies have not yet covered the effects of limestone, retarders, accelerators, temperature rise inhibitors and other complex admixtures in modern concrete. It is crucial to analyze and quantify the effects of different mixture components on concrete mixture design. However, existing DL models have not yet revealed the relationship between material composition and adiabatic temperature rise. Summary of the invention

[0004] Purpose of the invention: In view of the problems that the presence of multiple complex admixtures in concrete has an impact on the adiabatic temperature rise, a single source of experimental data, inconsistent experience-based artificial feature extraction, etc., the present invention provides a RNN concrete adiabatic temperature rise prediction method for dynamic parameter improvement, which adopts principal component analysis to select dynamic parameters, and adopts a separate and optimized DL algorithm to predict the relationship between the adiabatic temperature rise of concrete with complex mixture components and the mixture design of concrete. The performance is improved by combining the DL model with segmentation technology and a bidirectional deep recurrent neural network. The effectiveness and reliability of the recurrent neural network (RNN) and the bidirectional deep recurrent neural network model are verified through k-fold cross validation.

[0005] Technical solution: To achieve the above-mentioned purpose, the present invention provides a RNN concrete adiabatic temperature rise prediction method for dynamic parameter improvement, comprising the following steps:

[0006] S1: Obtain data related to the adiabatic temperature rise of concrete, and perform preprocessing and dynamic parameter selection on the data;

[0007] S2: Use Support Vector Regression (SVR), Artificial Neural Network (ANN) and Recurrent Neural Network (RNN) models to predict the adiabatic temperature rise of concrete with various admixtures;

[0008] S3: Analyze the accuracy and performance of DL models based on the prediction results produced by SVR, ANN and RNN models;

[0009] S4: Use segmentation techniques and bidirectional deep recurrent neural networks to further improve the accuracy and performance of DL models;

[0010] S5: Use k-fold cross validation for resampling to reduce the impact of random variations in the training and test data, providing an accurate and unbiased estimate of the model's performance;

[0011] Furthermore, the specific operations of data preprocessing and dynamic parameter selection in step S1 are:

[0012] S11: The collected data set consists of data related to the adiabatic temperature rise of concrete. The principal component analysis method is used to reduce the dimension of the original data features. The specific operations are as follows:

[0013] 1) Input the original data into X in matrix form and perform standardization on the matrix;

[0014] 2) Calculate the covariance matrix of the standardized samples

[0015] 3) Using the eigenvalue decomposition method to calculate the eigenvalues ​​and eigenvectors of the covariance matrix;

[0016] 4) Calculate the variance contribution rate and cumulative contribution rate corresponding to each eigenvalue. According to the cumulative contribution rate, select the first y principal components so that their cumulative contribution rate reaches 85%; the principal component contribution rate is The cumulative contribution rate is

[0017] 5) Project the standardized data matrix onto the selected y principal components to obtain the principal component score matrix; use the principal component score matrix and the transpose of the eigenvector to restore the compressed data; according to the cumulative load value of each characteristic variable, that is, the linear combination coefficient of the principal component, screen out the main influencing factors, and the features with cumulative load values ​​greater than 0.8 are selected as the main features, which are 15 main features affecting the adiabatic temperature rise of concrete, including water (kg / m 3 ), cement (kg / m 3 ) and fly ash (kg / m 3 ), slag (kg / m 3 )、limestone powder (kg / m 3 )、fine powder (kg / m 3 )、fine aggregate (kg / m 3 ), coarse aggregate (kg / m 3 ), condensation reducing agent (kg / m 3 ), gel accelerator (kg / m 3 ), temperature rise inhibitor (kg / m 3 )、Water reducing agent (kg / m 3 ), air entraining agent (kg / m 3 ), initial temperature (°C) and time (h);

[0018] S12: Set the prediction target to adiabatic temperature rise from 0 to 168 hours, where the prediction time interval within 24 hours is 2 hours, the prediction time interval from 24 hours to 48 hours is 3 hours, and the prediction time interval from 48 hours to 168 hours is 24 hours;

[0019] S13: The dataset is randomly divided into two subsets, where the training set includes 70% of the data in the original database, and the remaining data is integrated into the test set;

[0020] S14: Before training and testing, the collected raw data is standardized to reduce the impact of unnecessary feature scaling. The formula is:

[0021]

[0022] Among them, x is the original data, μ is the mean of the original data, σ is the standard deviation of the original data, and N is the data after standardization.

[0023] Furthermore, the specific operation of using SVR, ANN and RNN to predict the adiabatic temperature rise of concrete in step S2 is:

[0024] S21: Construct an SVR model. Through the nonlinear mapping function φ(x), the input vector x is mapped to the high-dimensional feature space H. The optimal decision function in the feature space is constructed using the principle of structural risk minimization. The relationship between the input variables and the output targets is constructed. That is, obtaining the regression function of the adiabatic temperature rise of concrete in the high-dimensional feature space is the main goal of the prediction function to be fitted f(x):

[0025] f(x)=w×φ(x)+b

[0026] Where w represents the weight, φ(x) represents the nonlinear mapping function, and b represents the bias. In order to determine w and b, a non-negative slack variable δ is introduced. i , and penalty factor C, converting the regression problem to be solved into a minimum value problem, where min means solving the minimum value and st means the constraint condition, namely:

[0027]

[0028] Where ε is an insensitive loss parameter used to determine the size of the fitting accuracy, n is the number of samples, and y i is the predicted value of the i-th sample; using the Lagrange multiplier method, the constraints can be transformed into unconstrained equations, where a i , μ i and is the Lagrangian coefficient, and the Lagrangian function is constructed as:

[0029] For the above formula, when w, b, δ i , Under the condition of finding the minimum value, find a i , The maximum value of , and finally the regression prediction function to be fitted according to the optimization conditions is:

[0030]

[0031] Among them, K(x i ,x) represents the kernel function, and e represents the width of the kernel function;

[0032] S22: Construct an ANN model, which contains one hidden layer and eight hidden nodes. The weights connecting the neurons in the input layer are expressed as xi 是 The input neuron is i∈{1, 2, ..., 15}, and the output of a single neuron in the hidden layer is in the following form:

[0033]

[0034] Where n represents the dimension of the input neuron x, represents the bias term, is the weight vector of neuron i, H j Represents the input of a single neuron in the hidden layer; the output layer is expressed as follows:

[0035]

[0036] Where m represents the number of neurons in the hidden layer. represents the weight of the hidden layer, b (2) represents the bias term;

[0037] S23: Build an RNN model, which includes a hidden layer and a linear layer. Each node represents a hidden state. At each time step, the hidden state is updated according to the current input and the previous hidden state. The update method of the hidden layer is as follows:

[0038] h (t) =F(Ux (t) +Wh (t-1) +b)

[0039] Among them, F(·) is the Relu activation function, x (t) is the input, h (t-1) is the previous hidden layer state, U, W represent weights, and t represents time step t.

[0040] Furthermore, the accuracy of the prediction results evaluated in step S3 is quantified using the following verification methods, namely, mean absolute percentage error (MAPE), mean absolute error (MAE), root mean square error (RMSE), Pearson correlation coefficient (R), determination coefficient (R 2 ); MAPE is expressed as a percentage, measuring the difference between the actual temperature rise and the predicted temperature rise. The lower the value, the higher the accuracy. MAE takes the absolute value of the difference between the actual value and the predicted value of each sample and then averages it. RMSE takes the square root of the mean of the squares of the difference between the predicted value and the actual value. R is the covariance of the actual value and the predicted value divided by the product of their standard deviations. 2It reflects the proportion of the total variation of the dependent variable that can be explained by the independent variable through the regression relationship. The value close to 1 indicates that the relationship between the actual temperature rise and the predicted temperature rise is stronger. The calculation formula of the above evaluation method is:

[0041]

[0042] Among them, y i is the true value, y' i is the predicted value, for y i The mean of for y' i The mean of

[0043] S31: Predict the adiabatic temperature rise of concrete with different mix designs in the test set. The loss is the difference between the predicted output and the actual output on the specified data set. The R of the RNN model is 0.9356, and R 2 The R and R of the ANN model are 0.8542, RMSE is 6.6374℃ and MAE is 5.1602℃. 2 , RMSE and MAE are 0.8359, 0.7616, 11.3689℃ and 9.0921℃ respectively, while the experimental results of the SVR model are worse. The results show that the RNN model has a higher prediction accuracy for the adiabatic temperature rise of concrete, better learns the relationship between input and output variables, and exhibits a more stable fitting ability. However, there is an obvious mismatch between the temperature rise predicted by the SVR and ANN models and the actual temperature rise, and it is difficult to provide the same prediction results as the RNN.

[0044] Furthermore, although the experimental results based on the RNN model in step S3 are higher than those of the general deep learning model research, its prediction curve shows that the R 2 is 0.7145, and there are a significant number of test points that deviate greatly from the center line; therefore, the optimization method based on the RNN model in step S4 is:

[0045] S41: Using segmentation technology, the adiabatic temperature rise of concrete is divided into two duration periods according to the temperature rise rate: 0-24 hours and 27-168 hours. Compared with the independent RNN model, R and R 2 The value of the RNN+segment model is slightly higher, while the RMSE and MAE are slightly lower, indicating that the RNN+segment model shows better accuracy. However, the connection between the two segments may distort the adiabatic temperature rise curve predicted by the RNN+segment model, that is, the RNN+segment model may not be able to learn the evolution behavior of the adiabatic temperature rise from the input-output correlation of each segment. Therefore, in order to improve the fitting performance of the concrete adiabatic temperature rise prediction, a bidirectional deep recurrent neural network is proposed to further optimize the RNN model.

[0046] Furthermore, the method for constructing the bidirectional deep recurrent neural network proposed in step S41 is:

[0047] S42: Construct a bidirectional deep recurrent neural network in which two hidden units are connected to the same output in opposite directions, the input is processed in forward and reverse time order, and three linear layers are added to make the gradient from the input layer to the output layer smoother; the bidirectional deep recurrent neural network model can obtain the correlation between cause and effect and provide a reliable trend during the early adiabatic temperature rise period; the calculation formula of the hidden layer is as follows:

[0048]

[0049] in represents the forward hidden state update, Indicates a reverse hidden state update.

[0050] Furthermore, the result of the bidirectional deep recurrent neural network in step S42 is:

[0051] S43: R and R of the prediction results 2 , RMSE, and MAE are 0.9615, 0.9245, 4.3826℃, and 3.3371℃, respectively. All statistical parameters show that among the four deep learning models studied above, the bidirectional deep recurrent neural network has the highest accuracy and is more suitable for predicting the adiabatic temperature rise curve of concrete containing complex mixtures.

[0052] Furthermore, the k-fold cross validation method adopted in step S5 includes:

[0053] S51: The dataset is randomly divided into five subsets {D1, D2, D3, D4, D5}, and the deep learning model is trained and tested five times using different subset combinations;

[0054] Furthermore, the training and testing of the model in step S51 includes:

[0055] S52: In each iteration, a subset D i For testing, the remaining four subsets D\D i For training, the result value is calculated according to the evaluation index provided in step S3, and the root mean square error of each iteration is Where |D i | represents the number of samples in the i-th subset, y i,j is the true value, ^y i,j is the predicted value, and the result is averaged, that is, Evaluate the accuracy of model predictions;

[0056] Furthermore, the evaluation of the model prediction accuracy in step S52 includes:

[0057] S53: Test the test set according to the model trained in step S52, where R 2 The average values ​​of RMSE and MAE are 0.7637, 8.4817℃ and 6.5916℃ respectively, while the R 2 The average values ​​of RMSE and MAE are 0.8197, 7.9991℃ and 6.2943℃ respectively. The results show that the performance of the bidirectional deep recurrent neural network model is better than that of the RNN model. The model fitting curve is drawn with time as the horizontal axis and the predicted adiabatic temperature rise as the vertical axis. Compared with the actual curve, it shows that the RNN concrete adiabatic temperature rise prediction method for dynamic parameter improvement proposed in this paper, that is, the bidirectional deep recurrent neural network model, can accurately fit the adiabatic temperature rise curve.

[0058] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0059] 1. The paper extracts data sets from tests recorded in articles and technical reports related to concrete adiabatic temperature rise, which solves the problem that in most cases, existing data sets come from single laboratories or field structures, thus improving the applicability and universality of the model.

[0060] 2. It can consider the influence of the complex integration of various admixtures such as limestone and retarder on the prediction of concrete adiabatic temperature rise, solve the singleness of the existing variable selection for predicting concrete adiabatic temperature rise, and improve the prediction performance of the model.

[0061] 3. The principal component analysis method is used for feature extraction. It does not require prior domain expertise in the data and automatically learns the most important features from the data. This solves the problem of inconsistency caused by individual differences in manually extracted features based on experience and subjective judgment.

[0062] 4. Compared with the independent model, the RNN+segmentation model has improved statistical parameters. However, segmentation distortion is found on the adiabatic temperature rise profile predicted by the RNN+segmentation model. However, the bidirectional deep recurrent neural network model also obtains good statistical results without sacrificing the fitting of the adiabatic temperature rise curve. In the early stage, the error between the predicted value and the actual value can be reduced.

[0063] 5. Using k-fold cross validation for evaluation, both RNN and bidirectional deep recurrent neural network provide stable prediction accuracy from various training and testing datasets, with small deviations between statistical checks on different data subsets. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0065] In the attached picture:

[0066] Figure 1 It is a schematic diagram of the process of the present invention;

[0067] Figure 2 Schematic diagram of the ANN model in the method of the present invention;

[0068] Figure 3 Schematic diagram of the RNN model in the method of the present invention;

[0069] Figure 4 Schematic diagram of a bidirectional deep recurrent neural network model in the method of the present invention. DETAILED DESCRIPTION

[0070] The present invention is further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, various equivalent forms of modifications to the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0071] The present invention provides a RNN concrete adiabatic temperature rise prediction method for dynamic parameter improvement, such as Figure 1 As shown, it includes the following steps:

[0072] S1: Collect test data recorded in articles or technical reports related to concrete insulation temperature rise, as well as laboratory experimental data in cooperation with insulation temperature rise test equipment, and perform data preprocessing and dynamic parameter selection;

[0073] The steps of data preprocessing and dynamic parameter selection in this embodiment are as follows:

[0074] S11: The collected data set consists of data related to the adiabatic temperature rise of concrete. The principal component analysis method is used to reduce the dimension of the original data features. The specific operations are as follows:

[0075] 1) Input the original data into X in matrix form and perform standardization on the matrix;

[0076] 2) Calculate the covariance matrix of the standardized samples

[0077] 3) Using the eigenvalue decomposition method to calculate the eigenvalues ​​and eigenvectors of the covariance matrix;

[0078] 4) Calculate the variance contribution rate and cumulative contribution rate corresponding to each eigenvalue. According to the cumulative contribution rate, select the first y principal components so that their cumulative contribution rate reaches 85%; the variance contribution rate is The cumulative contribution rate is

[0079] 5) Project the standardized matrix onto the selected y principal components to obtain the principal component score matrix; use the principal component score matrix and the transpose of the eigenvector to restore the compressed data; according to the cumulative load value of each characteristic variable, that is, the linear combination coefficient of the principal component, screen out the main influencing factors, and the features with cumulative load values ​​greater than 0.8 are selected as the main features, which are 15 features that affect the adiabatic temperature rise of concrete, including water (kg / m 3 ), cement (kg / m 3 ) and fly ash (kg / m 3 ), slag (kg / m 3 )、limestone powder (kg / m 3 )、fine powder (kg / m 3 )、fine aggregate (kg / m 3 ), coarse aggregate (kg / m 3 ), condensation reducing agent (kg / m 3 ), gel accelerator (kg / m 3 ), temperature rise inhibitor (kg / m 3 )、Water reducing agent (kg / m 3 ), air entraining agent (kg / m 3 ), initial temperature (°C) and time (h);

[0080] S12: Set the prediction target to adiabatic temperature rise from 0 to 168 hours, where the prediction time interval within 24 hours is 2 hours, the prediction time interval from 24 hours to 48 hours is 3 hours, and the prediction time interval from 48 hours to 168 hours is 24 hours;

[0081] S13: The dataset is randomly divided into two subsets, where the training set includes 70% of the data in the original database, and the remaining data is integrated into the test set;

[0082] S14: Before training and testing, the collected raw data is standardized to reduce the impact of unnecessary feature scaling. The formula is:

[0083]

[0084] Among them, x is the original data, μ is the mean of the original data, σ is the standard deviation of the original data, and N is the data after standardization.

[0085] S2: Use Support Vector Regression (SVR), Artificial Neural Network (ANN) and Recurrent Neural Network (RNN) models to predict the adiabatic temperature rise of concrete with various admixtures;

[0086] Reference Figure 2 , Figure 3 , the specific steps of constructing SVR, ANN and RNN models in this embodiment are:

[0087] S21: Construct an SVR model. Through the nonlinear mapping function φ(x), the input vector x is mapped to the high-dimensional feature space H. The optimal decision function in the feature space is constructed using the principle of structural risk minimization. The relationship between the input variables and the output targets is constructed. That is, the regression function of the adiabatic temperature rise of concrete in the high-dimensional feature space is the main goal of the prediction function to be fitted f(x):

[0088] f(x)=w×φ(x)+b

[0089] Among them, w represents the weight, φ(x) represents the nonlinear mapping function, and b represents the bias. In order to determine w and b, a non-negative slack variable δ is introduced i , and penalty factor C, converting the regression problem to be solved into a minimum value problem, where min means solving the minimum value and st means the constraint condition, namely:

[0090]

[0091]

[0092] Where ε is an insensitive loss parameter used to determine the size of the fitting accuracy, n is the number of samples, and y i is the predicted value of the i-th sample; using the Lagrange multiplier method, the constraints can be transformed into unconstrained equations, where a i , μ i and is the Lagrangian coefficient, and the Lagrangian function is constructed as:

[0093]

[0094] For the above formula, when w, b, δ i ,, Under the condition of finding the minimum value, find a i ,, The maximum value of , and finally the regression prediction function to be fitted according to the optimization conditions is:

[0095]

[0096] Where K(x i ,x) represents the kernel function, and e represents the width of the kernel function;

[0097] S22: Construct an ANN model, which contains one hidden layer and eight hidden nodes. The weights connecting the neurons in the input layer are expressed as x i is the input neuron, and i∈{1, 2, ..., 15}, the output of a single neuron in the hidden layer takes the following form:

[0098]

[0099] Where n represents the dimension of the input neuron x, represents the bias term, is the weight vector of neuron i, H j Represents the input of a single neuron in the hidden layer; the output layer is expressed as follows:

[0100]

[0101] Where m represents the number of neurons in the hidden layer. represents the weight of the hidden layer, b (2) represents the bias term;

[0102] S23: Build an RNN model, which includes a hidden layer and a linear layer. Each node represents a hidden state. At each time step, the hidden state is updated according to the current input and the previous hidden state. The update method of the hidden layer is as follows:

[0103] h (t) =F(Ux (t) +Wh (t-1) +b)

[0104] Among them, F(·) is the Relu activation function, x (t) is the input, h (t-1) is the previous hidden layer state, U, W represent weights, and t represents the tth time step.

[0105] S3: Analyze the accuracy and performance of the DL models given the prediction results produced by the SVR, ANN, and RNN models;

[0106] The specific evaluation criteria in this embodiment are:

[0107] S31: The accuracy of the prediction results is quantified using the following validation methods: mean absolute percentage error (MAPE), mean absolute error (MAE), root mean square error (RMSE), Pearson correlation coefficient (R), determination coefficient (R 2 ); MAPE is expressed as a percentage, measuring the difference between the actual temperature rise and the predicted temperature rise. The lower the value, the higher the accuracy. MAE takes the absolute value of the difference between the actual value and the predicted value of each sample and then averages it. RMSE takes the square root of the mean of the squares of the difference between the predicted value and the actual value. R is the covariance of the actual value and the predicted value divided by the product of their standard deviations. 2 It reflects the proportion of the total variation of the dependent variable that can be explained by the independent variable through the regression relationship. The value close to 1 indicates that the relationship between the actual temperature rise and the predicted temperature rise is stronger. The calculation formula of the above evaluation method is:

[0108]

[0109] Among them, y i is the true value, y' i is the predicted value, for y i The mean of for y' i The mean of

[0110] S32: Predict the adiabatic temperature rise of concrete with different mix designs in the test set, and the loss is the difference between the predicted output and the actual output on the specified data set; the experimental results show that the RNN model has a higher prediction accuracy for the adiabatic temperature rise of concrete and exhibits a more stable fitting ability, however, the SVR and ANN models are difficult to provide the same prediction results as the RNN.

[0111] S4: Use segmentation techniques and bidirectional deep recurrent neural networks to further improve the accuracy and performance of DL models;

[0112] The specific steps of using the segmentation technology in this embodiment are:

[0113] S41: Using segmentation technology, the adiabatic temperature rise of concrete is divided into two segments according to the temperature rise rate: 0-24 hours and 27-168 hours. Compared with the independent RNN model, the RNN+segmentation model shows better accuracy; however, the connection between the two segments may distort the adiabatic temperature rise curve predicted by the RNN+segmentation model, that is, the RNN+segmentation model may not be able to perfectly learn the evolution behavior of the adiabatic temperature rise from the input-output correlation of each segment; therefore, in order to improve the fitting performance of the concrete adiabatic temperature rise prediction, a bidirectional deep recurrent neural network is proposed to further optimize the RNN model;

[0114] Reference Figure 4, the specific steps of constructing a bidirectional deep recurrent neural network model in this embodiment are:

[0115] S42: Construct a bidirectional deep recurrent neural network in which two hidden units are connected to the same output in opposite directions, the input is processed in forward and reverse time order, and three linear layers are added to make the gradient from the input layer to the output layer smoother; the bidirectional deep recurrent neural network model can obtain the correlation between cause and effect and provide a reliable trend during the early adiabatic temperature rise period; the calculation formula of the hidden layer is as follows:

[0116]

[0117] in represents the forward hidden state update, Indicates a reverse hidden state update.

[0118] S5: Use k-fold cross validation for resampling to reduce the impact of random variations in the training and test data, providing an accurate and unbiased estimate of the model's performance;

[0119] In this embodiment, the k-fold cross validation method is used. The specific steps are:

[0120] S51: The dataset is randomly divided into five subsets {D1, D2, D3, D4, D5}, and the deep learning model is trained and tested five times using different subset combinations;

[0121] S52: Train and test the model according to step S51. In each iteration, a subset D i For testing, the remaining four subsets D\D i For training, the result value is calculated according to the evaluation index provided in step S3, and the root mean square error of each iteration is Where |D i | represents the number of samples in the i-th subset, y i,j is the true value, ^y i,j is the predicted value, and the result is averaged, that is, Evaluate the accuracy of model predictions;

[0122] S53: According to the model trained in step S52, the test set is tested, and the model fitting curve is plotted with time as the horizontal axis and the predicted adiabatic temperature rise as the vertical axis to compare with the actual curve, indicating that the bidirectional deep recurrent neural network model can fit the adiabatic temperature rise curve more accurately.

[0123] According to the above examples, it can be seen that in order to solve the problems of the presence of multiple complex admixtures in concrete that affect the adiabatic temperature rise, a single source of experimental data, inconsistent experience-based manual feature extraction, etc., the method of the present invention selects dynamic parameters by using principal component analysis, and uses SVR, ANN and RNN models to predict the adiabatic temperature rise of concrete. At the same time, in order to improve the accuracy, the RNN-based segmentation technology optimization method and the bidirectional deep recurrent neural network are used for prediction. The results can fit the adiabatic temperature rise curve of concrete well.

[0124] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0125] Finally, it should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A RNN concrete adiabatic temperature rise prediction method for dynamic parameter improvement, characterized in that: The steps include: S1: Obtain data related to the adiabatic temperature rise of concrete, and perform preprocessing and dynamic parameter selection on the data; S2: Prediction of adiabatic temperature rise of concrete with various admixtures using SVR, ANN and RNN; S3: Analyze the accuracy and performance of DL models based on the prediction results produced by SVR, ANN, and RNN; S4: Use segmentation techniques and bidirectional deep recurrent neural networks to improve the accuracy and performance of DL models; S5: Use k-fold cross validation for resampling to reduce the impact of random variations in the training and test data, providing an accurate and unbiased estimate of the model's performance; The RNN optimization model in S4, where the optimization method is: S41: Using segmentation technology, the adiabatic temperature rise of concrete is segmented according to the duration period according to the temperature rise rate; S42: Construct a bidirectional deep recurrent neural network in which two hidden units are connected to the same output in opposite directions, the input is processed in forward and reverse time order, and three linear layers are added; the hidden layer is calculated as follows: in, represents the forward hidden state update, Indicates reverse hidden state update; S5 includes: S51: The dataset is randomly divided into five subsets, and the deep learning model is trained and tested five times using different subset combinations; S52: training and testing the model according to step S51. In each iteration, one subset is used for testing and the remaining four subsets are used for training. The result value is calculated according to the evaluation index provided in step S3, and the result is averaged to evaluate the accuracy of the model prediction. S53: According to the model trained in step S52, the test set is tested, and the model fitting curve is plotted with time as the horizontal axis and the predicted adiabatic temperature rise as the vertical axis to compare with the true curve, so as to obtain the fitting effect of the model in this paper.

2. According to the RNN concrete adiabatic temperature rise prediction method for dynamic parameter improvement according to claim 1, it is characterized in that: The specific operations of data preprocessing and dynamic parameter selection in S1 are: S11: The collected data set consists of data related to the adiabatic temperature rise of concrete. The principal component analysis method is used to reduce the dimension of the original data features. The specific operations are as follows: 1) Input the original data in matrix form and standardize the matrix; 2) Calculate the covariance matrix of the standardized samples; 3) Use the eigenvalue decomposition method to calculate the eigenvalues ​​and eigenvectors of the covariance matrix; 4) Calculate the variance contribution rate and cumulative contribution rate corresponding to each eigenvalue, and select the first y principal components according to the cumulative contribution rate so that its cumulative contribution rate reaches 85%; 5) Project the standardized matrix onto the selected y principal components to obtain the principal component score matrix; Use the principal component score matrix and the transpose of the eigenvector to restore the compressed data; According to the cumulative load value of each characteristic variable, that is, the linear combination coefficient of the principal component, screen out the main influencing factors, and the features with cumulative load values ​​greater than 0.8 are selected as the main features, which are 15 features that affect the adiabatic temperature rise of concrete, including water, cement, fly ash, slag, limestone powder, fine powder, fine aggregate, coarse aggregate, setting weakener, gel accelerator, temperature rise inhibitor, water reducer, air entraining agent, initial temperature and time; S12: Set the prediction target to adiabatic temperature rise from 0 to 168 hours, where the prediction time interval within 24 hours is 2 hours, the prediction time interval from 24 hours to 48 hours is 3 hours, and the prediction time interval from 48 hours to 168 hours is 24 hours; S13: The dataset is randomly divided into two subsets, where the training set includes 70% of the data in the original database, and the remaining data is integrated into the test set; S14: Before training and testing, the collected raw data is standardized to reduce the impact of unnecessary feature scaling. The formula is: Among them, x is the original data, μ is the mean of the original data, σ is the standard deviation of the original data, and N is the data after standardization.

3. The RNN concrete adiabatic temperature rise prediction method for dynamic parameter improvement according to claim 1 is characterized in that: The specific operation of using SVR, ANN and RNN to predict the adiabatic temperature rise of concrete in S2 is: S21: Construct an SVR model. Through a nonlinear mapping function, the input vector is mapped to a high-dimensional feature space. The optimal decision function in the feature space is constructed using the principle of structural risk minimization. The relationship between the input variable and the output target is constructed, that is, the regression function of the adiabatic temperature rise of concrete obtained in the high-dimensional feature space is used as the objective function of the prediction function to be fitted f(x): f(x)=w×φ(x)+b Among them, w represents the weight, φ(x) represents the nonlinear mapping function, and b represents the bias; the non-negative slack variable δ is introduced i , And the penalty factor C, determine w and b, and convert the regression problem to be solved into a minimum value problem. Min means solving the minimum value, and st means the constraint condition, that is: Among them, ε is an insensitive loss parameter used to determine the size of the fitting accuracy, n is the number of samples, and y i is the predicted value of the i-th sample; using the Lagrange multiplier method, the constraints can be transformed into unconstrained equations, where a i , μ i and is the Lagrangian coefficient, and the Lagrangian function is constructed as: In the case of w, b, δ i , Under the condition of finding the minimum value, find a i , The maximum value of , and finally the regression prediction function to be fitted according to the optimization conditions is: Among them, K(x i ,x) represents the kernel function, and e represents the width of the kernel function; S22: Construct an ANN model, which includes one hidden layer and eight hidden nodes. The weights connecting the neurons in the input layer are expressed as x i is the input neuron, and i∈{1, 2, ..., 15}, the output of a single neuron in the hidden layer takes the following form: Where n represents the dimension of the input neuron x, represents the bias term, H j Represents the input of a single neuron in the hidden layer; the output layer is expressed as follows: Where m represents the number of neurons in the hidden layer. represents the weight of the hidden layer, b (2) represents the bias term; S23: Build an RNN model, which includes a hidden layer and a linear layer. Each node represents a hidden state. At each time step, the hidden state is updated according to the current input and the previous hidden state. The update method of the hidden layer is as follows: h (t) =F(Ux (t) +Wh (t-1) +b) Among them, F(·) is the Relu activation function, x (t) is the input, h (t-1) is the previous hidden layer state, U, W represent weights, and t represents the time step.

4. The RNN concrete adiabatic temperature rise prediction method for dynamic parameter improvement according to claim 1 is characterized in that: The accuracy of the prediction results evaluated in S3 is quantified using the following verification methods: mean absolute percentage error (MAPE), mean absolute error (MAE), root mean square error (RMSE), Pearson correlation coefficient (R), and determination coefficient (R). 2 ; MAPE is expressed as a percentage, measuring the difference between the actual temperature rise and the predicted temperature rise. The lower the value, the higher the accuracy; MAE takes the absolute value of the difference between the actual value and the predicted value of each sample and then calculates the average; RMSE takes the square root of the average of the squares of the difference between the predicted value and the actual value; R is the covariance of the actual value and the predicted value divided by the product of their standard deviations, R 2 It reflects the proportion of the total variation of the dependent variable that can be explained by the independent variable through the regression relationship. The calculation formula of the above evaluation method is: Among them, y i is the true value, y' i is the predicted value, for y i The mean of for y' i The mean of S31: Predict the adiabatic temperature rise of concrete for different mix designs in the test set, with the loss being the difference between the predicted output and the actual output on the specified dataset.

Images