CN116741315A - Method for predicting strength of geopolymer concrete - Google Patents

Abstract

The invention discloses a method for predicting the strength of geopolymer concrete. The method includes the following steps: Step 1: Collect geopolymer concrete strength experimental sample data to train a convolutional neural network, and construct an initial model for geopolymer concrete strength prediction. ; Step 2: Use the bat algorithm to optimize the initial meta-parameters in the initial model for geopolymer concrete strength prediction to obtain the optimal value of the meta-parameter; Step 3: Update the initial prediction model according to the optimal value of the meta-parameter to construct an optimization The final geopolymer concrete strength prediction model; step 4: output the prediction value of the compressive strength of concrete according to the polymer concrete strength prediction model; the present invention is conducive to accurately predicting the compressive strength value of geopolymer concrete in industrial production. Evaluate and promote production.

Description

A method for predicting the strength of geopolymer concrete

Technical areas:

The invention relates to the field of judging mechanical properties of concrete, and in particular to a method for predicting the strength of geopolymer concrete.

Background technique:

The compressive strength (CS) of geopolymer concrete is an important indicator of the mechanical properties of concrete and an important parameter that needs to be accurately measured during its production and use. Taking fly ash-slag-based polymer concrete as an example, its compressive strength is affected by various parameters including the type of raw materials, the type and content of alkali activators, and the liquid-to-solid ratio of the mixed materials. The compressive strength of geopolymer concrete is usually measured by conducting pressure tests on concrete specimens after standard curing. This method of obtaining data through tests involves steps such as specimen production, curing, and pressure testing, and the waiting period is long. , and more manpower, material and financial resources need to be invested, and the efficiency is low. Therefore, finding a geopolymer concrete strength prediction method based on deep learning algorithms to replace traditional test methods can achieve the purpose of reducing costs and increasing efficiency, and has very important theoretical value in ensuring the construction progress and quality of geopolymer concrete. and project economic effects.

In recent years, the application of concrete mix design and performance prediction based on artificial intelligence methods has attracted more and more attention. Some scholars use machine learning (ML) methods such as support vector machine (SVM), artificial neural network (ANN), and extreme learning machine (ELM) to predict the 28-day strength of geopolymer concrete. However, for the prediction of the compressive strength of fly ash-slag-based polymer concrete, the accuracy of the prediction results of the above machine learning methods is significantly affected by the selection of model input variables and the setting of model element parameters. Moreover, the traditional machine learning model has a simple and limited architecture, and its application scope and effect are more limited than deep learning.

Deep learning (DL) methods are able to create sophisticated learning models through a combination of feature extraction and pattern recognition to more effectively manage unstructured data by learning at higher levels. This enables DL models to perform complex tasks with high accuracy. . Therefore, they are able to decipher the complex relationship between the mixing parameters and final mechanical properties of concrete raw materials better than traditional ML models. Among many deep learning methods, methods based on convolutional neural networks (CNN) are widely used in the field of concrete material research. The results show that optimizing the meta-parameters of CNN models is crucial.

In order to ensure the accuracy of the CNN model in predicting the compressive strength of geopolymer concrete, the model element parameters need to be optimized. The goal of meta-parameter optimization is to find the global optimal solution by reducing the root mean square error and maximizing the coefficient of determination between the model predicted value and the true value. The Bat Algorithm (BA) has been successfully applied to many optimization problems and has the advantage of rapid convergence. It can optimize the meta-parameters of the network model during the CNN model training process. However, relevant research has proven that BA, like other swarm intelligence algorithms, has insufficient optimization accuracy. , low efficiency, easy to fall into local optimality and other shortcomings.

Contents of the invention:

In view of the technical problems existing in the existing technology, the present invention proposes a geopolymer concrete strength prediction method; this method uses an improved bat algorithm to introduce a new adaptive inertial weight coefficient and random disturbance into the bat speed iterative update formula , improves the local search ability and global search ability; thus, it can more accurately extract and identify the nonlinear characteristics between the various parameters of geopolymer concrete and its strength, and improve the accuracy of the geopolymer concrete compressive strength prediction results; At the same time, the present invention proposes an improved bat algorithm that can solve the problem that the traditional bat algorithm has the risk of failure when faced with highly nonlinear problems.

In order to solve the existing technical problems, the present invention adopts the following technical solutions:

A method for predicting the strength of geopolymer concrete, the method includes the following steps:

Step 1: Collect geopolymer concrete strength experimental training sample data to train the convolutional neural network and build an initial geopolymer concrete strength prediction model;

Step 2: Use the bat algorithm to optimize the initial element parameters in the initial model for geopolymer concrete strength prediction to obtain the optimal value of the element parameters;

Step 3: Update the initial prediction model according to the optimal values of element parameters, and construct an optimized geopolymer concrete strength prediction model;

Step 4: According to the polymer concrete strength prediction model, output the prediction value of the compressive strength of the concrete; where:

The process of optimizing the element parameters in the geopolymer concrete model through the bat algorithm to obtain the optimal values of the element parameters includes the following steps:

Initial values of bat parameters for constructing geopolymer concrete models;

The element parameters in the geopolymer concrete model are positioned using random initialization of individual bats in the population;

Perform individual bat fitness calculations based on the meta-parameters in the geopolymer concrete model after positioning to obtain the optimal meta-parameters;

Update the speed and position of the optimal element parameters in the geopolymer concrete model according to the following formula;

in: Represents the speed of the i-th bat at time t;/> represents the position of the i-th bat at time t minus the optimal solution of all individuals in the current iteration; x ^* represents the optimal solution of all individuals in the current iteration; f _i represents the pulse frequency of the i-th bat;/> represents a random term; x ^r represents the position of a randomly selected bat; lc is the learning factor; τ is the adaptive inertia weight coefficient; that is:

Among them: it represents the current number of iterations; τ _min and τ _max represent the minimum and maximum inertia weights respectively; ω is a constant that adjusts the curve diffusion area, with a value of 10; N _ts is the total number of iterations;

Use the random walk model to perform a local search on the geopolymer concrete model to update the current position of each optimal element parameter;

An iterative search is performed on the geopolymer concrete model to update each optimal element parameter loudness and pulse emissivity;

Re-evaluate the fitness of the optimal element parameters in the geopolymer concrete model to obtain the optimal solution of element parameters;

Determine whether the stopping criterion is met, and output the optimal solution when the number of iterations exceeds the preset maximum number of iterations; otherwise, return to step 3 and recalculate.

Further, in step 1, the process of collecting geopolymer concrete training sample data to train the convolutional neural network and constructing a geopolymer concrete model includes the following steps:

Input the experimental data of the strength influencing factors in the training sample data set into the input layer of the geopolymer concrete strength prediction model;

The concrete strength prediction value is forwardly output through step 4 under the connection weight and bias conditions of the initial geopolymer concrete strength prediction model;

The error between the predicted strength value and the actual measured value is back-propagated from the output layer to the input layer; during the back-propagation process, the geopolymer concrete strength prediction model is continuously updated to make the predicted strength value close to the corresponding true value. ;

Repeat the above steps until the error reaches a certain threshold, ending the geopolymer concrete strength prediction model training. Further, the polymer concrete strength model includes an input layer, a convolution layer, a pooling layer, a fully connected layer and an output layer; where:

The input layer contains 11 neurons, corresponding to 11 influencing factors on the compressive strength of geopolymer concrete;

The input layer is followed by two convolutional layers, which are used for intensity-sensitive feature extraction and deep representation respectively;

The pooling layer downsamples the learned depth features to form a feature map;

The fully connected layer inputs the feature map into the 2 fully connected layers to eliminate redundant features and perform pattern recognition;

The output layer outputs a predicted compressive strength value of geopolymer concrete.

Further, the geopolymer concrete training sample data includes the mass of fly ash and slag in the unit volume mixture, the ratio of fine aggregate and coarse aggregate to binder, the liquid-to-solid ratio between water and solid, Molar ratio of silicon oxide to sodium oxide (Ms), ratio of sodium oxide to binder (nm), plasticizer ratio (SP), curing temperature (CT), oven curing time (CH), concrete age (CA ).

beneficial effects

First: The present invention enhances the local optimization capability when solving multi-dimensional complex problems by introducing an adaptive inertia weight coefficient into the bat speed update formula; and adding a random term increases the diversity of the bat group and improves the search efficiency of the algorithm. Excellent accuracy and efficiency, avoiding the limitations of traditional methods;

Second: The present invention uses the bat algorithm to optimize the element parameters of the deep learning model, proposes an improved bat algorithm to optimize the geopolymer concrete compressive strength prediction model, and establishes the content, proportion and compressive strength of each raw material of the geopolymer mixture. The complex mapping relationship between them realizes the evaluation of the predicted compressive strength of geopolymer concrete.

Third: the present invention is beneficial to accurately predict and evaluate the compressive strength value of geopolymer concrete in industrial production and promote production.

Description of drawings

Figure 1 shows the CNN model structure for strength prediction of fly ash-slag based polymer concrete;

Figure 2 is a flow chart for predicting the compressive strength of geopolymer concrete based on EBA-CNN;

Figure 3 shows the comparison between intensity prediction values and actual measured values based on the proposed method;

Figure 4 is based on the regression analysis of the intensity prediction results of the proposed EBA-CNN;

Figure 5 Regression analysis based on SVM intensity prediction results;

Figure 6 is based on regression analysis of general CNN intensity prediction results.

Detailed ways

The present invention will be described as follows in conjunction with accompanying drawings 1-6:

As shown in Figure 1-2, the present invention provides a method for predicting the strength of geopolymer concrete, which includes the following steps:

Collection and analysis of training sample data: Measurement data on various mix ratios, curing conditions, age and pressure tests of fly ash-slag-based polymer concrete were extensively collected from the published comprehensive experimental database. Analyze the influence of different experimental data on the standard compressive strength of geopolymer concrete from a statistical perspective, including the mass of fly ash and slag in the unit volume mixture, the ratio of coarse/fine aggregate and aggregate to binder, Liquid-solid ratio between water and solid, molar ratio of silica to sodium oxide (Ms), ratio of sodium oxide to binder (nm), plasticizer content (SP), curing temperature (CT), oven curing Time (CH), concrete age (CA). The above data sets will be used for model training and verification of model prediction effects.

Initial configuration of the convolutional neural network; the entire model network consists of an input layer, a convolution layer, a pooling layer, a fully connected layer and an output layer, with a total of 7 layers. The input layer contains 11 neurons, corresponding to the 11 influencing factors of the compressive strength of geopolymer concrete; after the input layer are two convolutional layers, which are used for the extraction of intensity-sensitive features and deep representation respectively; then, the global maximum is used The pooling layer downsamples the learned deep features to form a feature map; then, the feature map is input to the 2-layer fully connected layer for redundant feature elimination and pattern recognition; finally, the geopolymer concrete resistance is output in the output layer. Predicted compressive strength. The process of collecting geopolymer concrete training sample data to train the convolutional neural network and constructing a geopolymer concrete model includes the following steps:

The training process of the initially configured geopolymer concrete strength prediction model;

Input the experimental data of the strength influencing factors in the training sample data set into the input layer of the geopolymer concrete strength prediction model;

The concrete strength prediction value is forwardly output through step 4 under the connection weight and bias conditions of the initial geopolymer concrete strength prediction model;

The error between the predicted strength value and the actual measured value is back-propagated from the output layer to the input layer; during the back-propagation process, the geopolymer concrete strength prediction model is continuously updated to make the predicted strength value close to the corresponding true value. ;

Repeat the above steps until the error reaches a certain threshold, ending the geopolymer concrete strength prediction model training.

Optimization of element parameters in the geopolymer concrete model through the improved bat algorithm: Compared with the traditional method, the bat algorithm in the present invention has two important improvements.

The adaptive inertia weight coefficient τ of the hyperbolic tangent function is introduced into the bat speed update formula, and the expression is as shown in formula (1):

Among them, it represents the current number of iterations; τ _min and τ _max represent the minimum and maximum inertia weights respectively; ω is a constant that adjusts the curve diffusion area, with a value of 10; N _ts is the total number of iterations.

Add a random term to the speed update formula Aiming to increase the diversity of bat colonies. Therefore, after combining the two improvements, the updated expression of bat speed in EBA is as shown in formula (2):

in, Represents the speed of the i-th bat at time t;/> represents the position of the i-th bat at time t minus the optimal solution of all individuals in the current iteration; x ^* represents the optimal solution of all individuals in the current iteration; f _i represents the pulse frequency of the i-th bat;/> represents a random term; x ^r represents the position of a randomly selected bat; lc is the learning factor. The learning factor is shown in formula (3):

lc＝lc _min +(lc _max -lc _min )θ (3)

Among them, lc _min and lc _max represent the minimum and maximum learning factors respectively; θ is a random number between 0 and 1.

Through the above analysis, the calculation process of element parameter optimization in the geopolymer concrete model based on the improved bat algorithm (EBA) includes:

1) Determine the initial parameters according to the geopolymer concrete strength prediction model, including bat population size, solution dimension, initial pulse emission rate, maximum and minimum pulse frequencies, maximum and minimum inertia weights, maximum and minimum learning factors, The maximum number of iterations;

2) According to the initial value of the element parameter of the polymer concrete strength prediction model, the position of each element parameter in the population is randomly initialized;

3) Based on the fitness of each bat, find the single bat with the best meta-parameter fitness, which is regarded as the optimal solution;

4) Update the speed and position of the optimal element parameters in the geopolymer concrete model according to the following formula;

in: represents the speed of the i-th bat at time t; x ^r represents the position of the randomly selected bat; f _i represents the pulse frequency of the i-th bat; lc is the learning factor;

5) In the local search process, a random number between 0 and 1 is generated, and the random walk model is used to perform a local search on the geopolymer concrete model to update the current position of each optimal element parameter;

6) In the initial stage of the iterative search, a random number between 0 and 1 is generated, and the geopolymer concrete model is iteratively searched to update each optimal element parameter loudness and pulse emissivity;

7) Re-evaluate the fitness of the optimal element parameters in the geopolymer concrete model to obtain the optimal solution of element parameters;

8) Determine whether the stopping criterion is met, and output the optimal solution when the number of iterations exceeds the preset maximum number of iterations; otherwise, return to step 3) and recalculate.

Use the data set to retrain the model to obtain an optimized prediction model. Input the 11 factors such as the raw materials, proportions, and curing conditions of geopolymer concrete that need to be predicted into the optimized prediction model, and output the predicted value of the compressive strength of geopolymer concrete.

1. Acquisition and analysis of experimental data

From the published comprehensive experimental database, a total of 850 sets of compressive strength experimental data of fly ash-slag-based polymer concrete with different mix ratios, curing conditions and ages were collected. Analyze the influence of different experimental data on the standard compressive strength of geopolymer concrete from a statistical perspective, including the mass of fly ash and slag in the unit volume mixture, the ratio of fine aggregate and coarse aggregate to binder (FBR&CBR ), liquid-to-solid ratio between water and solid (W/S), molar ratio of silica to sodium oxide (Ms), ratio of sodium oxide to binder (nm), plasticizer ratio (SP), curing Temperature (CT), oven curing time (CH), concrete age (CA). Table 1 shows the statistical analysis results of the geopolymer concrete strength measurement experimental data set, including maximum value, minimum value, median, mode, mean, skewness, kurtosis and standard deviation. Among these statistical indicators, the median, mode, and mean represent the central characteristics of the data set, while the maximum, minimum, skewness, kurtosis, and standard deviation represent the irregular characteristics of the data. Statistical results show that the strength of fly ash-slag-based polymer concrete is widely affected by the above parameters. In order to ensure the accuracy of strength prediction of geopolymer concrete based on the above parameters, the strength prediction model is required to be sufficiently robust.

Table 1 Statistical analysis results of geopolymer concrete strength measurement experimental data:

2. Initial configuration and model training of geopolymer concrete model

Establish a CNN model structure containing a 7-layer network as shown in Figure 1. There are 9 meta-parameters of the model, including initial learning rate, gradient attenuation factor, L2 regularization factor, learning rate reduction period, learning rate reduction factor, the number of cores in the first convolutional layer, the number of cores in the second convolutional layer, The number of neurons in the first FC layer and the number of neurons in the second FC layer. The value range configuration of meta-parameters is as shown in the following table.

Table 2 CNN element parameter value range

595 (70%) of the 850 data sets were randomly selected for model training. The mixture raw materials, proportions and maintenance data of each experimental group are used as input parameters of the prediction model.

3. Optimization of element parameters of geopolymer concrete prediction model

During the model training process, EBA is used to iteratively optimize the model meta-parameters. The optimization objective (fitness) is defined as the root mean square error between the model intensity predicted intensity value and the true result during the training process. The setting parameters of EBA are: bat population size N = 30, solution dimension D = 9, initial pulse emission rate R ₀ = 0, maximum pulse frequency f _max = 2.8, minimum pulse frequency f _min = 1.3, maximum inertial weight τ _max =0.85, minimum inertia weight τ _min =0.15, maximum learning factor lc _max =0.6, minimum learning factor lc _min =0.3, loudness and pulse emission rate enhancement coefficient α =γ =0.9, total iteration number N _ts =200. All meta-parameters reached optimal values within 100 iterations, which fully proves the rapid convergence ability of EBA in model parameter optimization. The optimal values of the CNN model element parameters obtained based on EBA are shown in the table below.

Table 3 Obtains the optimal values of CNN model element parameters based on EBA

Use the optimal values of the meta-parameters in the table to construct the EBA-CNN prediction model, and analyze the training data again.

4. Verification of the effect of geopolymer concrete strength model

Run the EBA-CNN prediction model, input the remaining 255 sets (30%) of corresponding experimental data into the prediction model, obtain the intensity prediction value, and compare the prediction results with the experimental results, as shown in Figure 3. The results show that the 1D-CNN proposed in this article can accurately predict the compressive strength of fly ash-slag-based polymer concrete under various mixing and curing conditions.

In order to prove the superiority of this method, the EBA-CNN prediction method proposed in this invention is used to conduct comparative research with the existing SVM and the 1D-CNN prediction model without meta-parameter optimization. The above three prediction methods were used to analyze 255 sets of experimental data, and the intensity prediction values and actual measured values were subjected to regression analysis.

It is worth noting that for the prediction results of all 3 models, most of the data points in the plot are evenly scattered around the regression line, with only a few data points outside the ±20% boundary line. The coefficient of determination R ² of regression analysis ranges from 0 to 1. The larger the value, the better the performance of the model. Among these geopolymer concrete strength prediction methods, the strength prediction ability based on the proposed EBA-CNN is the best. The determination coefficient R of the test sample is ^0.9577 , which is greater than 0.9042 of the 1D-CNN method without meta-parameter optimization and the SVM method. of 0.9019.

Claims (4)

1. A method for predicting the strength of a geopolymer concrete, the method comprising the steps of:

step 1: training the convolutional neural network by collecting geopolymer concrete strength experiment training sample data, and constructing a geopolymer concrete strength prediction initial model;

step 2: optimizing initial meta-parameters in the initial model of the geopolymer concrete strength prediction by a bat algorithm to obtain an optimal value of the meta-parameters;

step 3: updating the initial prediction model according to the optimum value of the meta-parameter, and constructing an optimized geopolymer concrete strength prediction model;

step 4: outputting and predicting a compressive strength predicted value of the concrete according to the polymer concrete strength predicted model; wherein:

the process for obtaining the optimum value of the initial meta-parameter in the geopolymer concrete strength prediction model by optimizing the initial meta-parameter through a bat algorithm comprises the following steps:

constructing initial values of bat parameters required by the optimization of the parameters of the polymer concrete strength prediction model;

initializing each bat position in the population by adopting a random generation mode to position the meta-parameters in the geopolymer concrete strength prediction model;

performing single bat fitness operation according to the positioned meta-parameters in the geopolymer concrete strength prediction model to obtain optimal meta-parameters;

updating the speed and the position of the optimal element parameters in the geopolymer concrete model according to the following formula;

wherein:indicating the speed of the ith bat at time t; />Representing the position of the ith bat at time t minus the optimal solution for all individuals in the current iteration; x is x ^* The representation is the optimal solution for all individuals in the current iteration; f (f) _i Indicating the ith batPulse frequency of bats; />Representing a random term; x is x ^r Representing the location of the randomly selected bat; lc is a learning factor; τ is an adaptive inertial weight coefficient; namely:

wherein: it represents the current iteration number; τ _min And τ _max Representing minimum and maximum inertial weights, respectively; omega is a constant for adjusting the diffusion area of the curve, and the value is 10; n (N) _ts Is the total iteration number;

local searching is carried out on the parameters of the geopolymer concrete strength prediction model by using a random walk mode, and the current position of each optimal parameter is updated;

iteratively searching the parameters of the geopolymer concrete strength prediction model to update the loudness and pulse emissivity of each optimal parameter;

re-evaluating the adaptability of the optimal meta-parameters in the geopolymer concrete strength prediction model to obtain a meta-parameter optimal solution;

judging whether a stopping criterion is met, and outputting an optimal solution when the iteration times exceed a preset maximum iteration times; otherwise, returning to the step 3 to perform calculation again.

2. The method for predicting the strength of the geopolymer concrete according to claim 1, wherein the step 1 is characterized in that the collected geopolymer concrete strength experimental training sample data is used for training a convolutional neural network to construct a geopolymer concrete strength prediction model process; the method comprises the following steps:

inputting experimental data of the geopolymer concrete strength influence factors in the training sample data set into an input layer of a geopolymer concrete strength prediction model;

outputting the concrete strength predicted value in the forward direction through the step 4 under the connection weight and the bias condition of the initial polymer concrete strength predicted model;

back-propagating the error between the predicted intensity value and the actual measured value from the output layer to the input layer; in the back propagation process, continuously updating the geopolymer concrete strength prediction model to enable the predicted strength value to be close to the corresponding true value;

repeating the steps until the error reaches a certain threshold value, and ending the training of the geopolymer concrete strength prediction model.

3. A method for predicting the strength of a concrete according to claim 1 or 2, wherein the model for predicting the strength of the concrete comprises an input layer, a convolution layer, a pooling layer, a full connection layer and an output layer; wherein:

the input layer comprises 11 neurons, corresponding to 11 influencing factors of the compressive strength of the geopolymer concrete;

the input layer is followed by 2 convolution layers which are respectively used for extracting the intensity sensitive characteristics and deeply characterizing;

the pooling layer performs reduction sampling on the learned depth characteristics to form a characteristic map;

the full-connection layer inputs the characteristic map into the 2-layer full-connection layer to perform redundancy characteristic elimination and pattern recognition;

and the output layer outputs the predicted value of the compression strength of the geopolymer concrete.

4. A method for predicting the strength of a geopolymer concrete according to claim 3, wherein the experimental training sample data of the strength of the geopolymer concrete comprises the mass of fly ash and slag in a unit volume of mixture, the ratio of fine aggregate and coarse aggregate to binder, the liquid-solid ratio between water and solid, the molar ratio of silica to sodium oxide, the ratio of sodium oxide to binder, the ratio of plasticizer, curing temperature, oven curing time, age of concrete and measured strength value.