A method for predicting the ultimate compressive strength of concrete based on CTGAN

Abstract

The application discloses a method for predicting the ultimate compressive strength of concrete based on CTGAN. The method comprises the following steps: constructing an index system of factors influencing the compressive strength of concrete, establishing and analyzing an original data set, analyzing the relationship between each ingredient, limiting the proportion between some ingredients, using a CTGAN model for data enhancement, and using a one-dimensional residual convolution model for regression prediction. The application can more accurately predict the ultimate compressive strength of concrete, to a certain extent, solves the problem of lack of data in the engineering field, and has strong portability, improves the efficiency and accuracy of engineering design.

Description

Technical Field

[0001] This invention belongs to the field of concrete ultimate compressive strength prediction technology, and more specifically, relates to a method for predicting concrete ultimate compressive strength based on CTGAN. Background Technology

[0002] In recent years, the effective utilization of industrial solid waste has become a hot topic. Adding industrial waste such as blast furnace slag and fly ash to concrete materials can not only improve workability but also significantly reduce carbon emissions, promoting green development in the civil engineering field. For example, high-performance concrete (HPC) incorporates auxiliary cementitious materials such as fly ash, slag, and chemical admixtures.

[0003] Nowadays, structural failures due to insufficient concrete strength occur frequently in actual engineering projects. As a widely used building material, the structural strength of concrete is receiving increasing attention both domestically and internationally. Concrete plays a crucial role in the construction of rigid pavements. High strength, good stability, and good durability have always been advantages of cement pavements; therefore, the corresponding material properties are critical to the service life and quality of the pavement.

[0004] Currently, many experts at home and abroad have conducted relevant research on the material design of concrete. However, traditional material design is mainly based on indoor laboratories. Due to the randomness of measurement data and the influence of systematic errors, the permeability law of concrete has many uncertainties. Using general statistical methods, the dispersion of experimental observation data is large, which often leads to the distortion of analysis results. Moreover, traditional experimental methods are often a long and complex process with a long experimental cycle, huge workload, and relatively low research efficiency.

[0005] With the continuous development of artificial intelligence, some experts have begun to turn their attention to this field, and research on the performance evaluation of concrete structures based on deep learning has been widely applied. While simple machine learning and backpropagation (BP) neural networks have effectively solved some complex problems in research, reduced errors, and improved research efficiency, the inherent limitations of limited data in engineering fields, such as overfitting, can lead to inappropriate and unreliable predictions. Furthermore, research into deeper relationships requires substantial data support. However, in concrete material testing, collecting sufficient data is difficult due to the significant manpower and time required, and the accuracy of current technologies for expanding original small-scale experimental datasets needs further improvement.

[0006] Based on the aforementioned defects and shortcomings, there is an urgent need in this field to propose a new method for predicting the ultimate compressive strength of concrete, which can effectively expand small experimental datasets and thereby accurately predict the ultimate compressive strength of concrete. Summary of the Invention

[0007] To address the aforementioned deficiencies or improvement needs of existing technologies, this invention provides a method for predicting the ultimate compressive strength of concrete based on CTGAN. Based on limited experimental data, the CTGAN model is used to augment the features of the original dataset after filtering, thereby improving prediction accuracy and obtaining prediction results that are closer to reality. This addresses the problem of predicting the compressive strength of concrete containing industrial waste in practical engineering projects.

[0008] To achieve the above objectives, the technical solution of the present invention is as follows:

[0009] A method for predicting the ultimate compressive strength of concrete based on CTGAN, characterized by comprising the following steps:

[0010] S1: Obtain the original dataset of industrial waste concrete, calculate the ratio of water usage to the sum of cement usage, blast furnace slag usage and fly ash usage in the original dataset, and at the same time calculate the ratio of the sum of coarse aggregate usage and fine aggregate usage to the sum of cement usage, blast furnace slag usage and fly ash usage, and find the maximum and minimum values ​​of the above two ratios in the original dataset.

[0011] The two ratios are restricted to their respective maximum and minimum values, thus forming two restrictive conditions;

[0012] S2 performs correlation analysis on each feature in the original dataset, retains features with a correlation coefficient of no more than 0.5, and recombines them to obtain the screening dataset; the screening dataset is divided into a training set and a test set, the training set of the screening dataset is used as the input of the CTGAN model, the CTGAN model is trained to learn the data distribution of the screening dataset and the correlation between each ingredient and the ultimate compressive strength of concrete, and generates virtual concrete mix proportion data. The generated virtual concrete mix proportion data meets the two constraints in step S1, and the virtual dataset is output.

[0013] S3 constructs a one-dimensional residual convolution model and trains it using a virtual dataset to obtain the final one-dimensional residual convolution model for predicting the ultimate compressive strength of concrete.

[0014] The one-dimensional residual convolutional model comprises four convolutional layers, each followed by a batch normalization layer and a ReLU activation function. The input to the one-dimensional residual convolutional model is a one-dimensional tensor, which, after four convolutions, batch normalization, and ReLU activation, outputs a three-dimensional tensor. The output of the ReLU activation function following the first convolutional layer is connected to a residual module, which comprises a residual layer, a batch normalization layer, and a ReLU function connected in series. The output of the residual module and the three-dimensional tensor are input together into three fully connected layers (FC) connected in series. The first two fully connected layers (FC) are each followed by a batch normalization layer and a ReLU activation function. The output of the last fully connected layer (FC) is the output of the one-dimensional residual convolutional model.

[0015] The CTGAN model includes a generator and a discriminator. The generator consists of three XOR layers connected in series. Each XOR layer is connected to a convolutional layer (Conv), a batch normalization (BN) layer, and a ReLU activation function before being connected to the input of the next XOR layer. The last XOR layer is processed by a convolutional layer (Conv), a BN layer, and a ReLU activation function before being connected to a fully connected layer (FC). The output of the FC layer is processed by the tanh activation function and the gumbel activation function to output a virtual sample. The input of the first XOR layer is random noise.

[0016] The discriminator includes two convolutional layers (Conv). The output of the first convolutional layer (Conv) is processed by a leaky activation function and a drop layer before being connected to the second convolutional layer (Conv). The output of the second convolutional layer (Conv) is processed by a leaky activation function and a drop layer before being connected to a fully connected layer (FC). The fully connected layer (FC) outputs the score of the virtual sample, which is then fed into the generator. The input of the first convolutional layer (Conv) is the virtual sample output by the generator and the sample from the filtered dataset.

[0017] Step S2 specifically includes the following steps:

[0018] S21 Based on the characteristics of the selected dataset, add the two constraints in step S1 between ingredients to the CTGAN model and adjust the hyperparameters of the CTGAN model. The convolutional layers Conv in the CTGAN model are all one-dimensional convolutional layers.

[0019] In the S22CTGAN model, the generator adds random noise to the samples in the selected dataset to generate virtual samples, which are then fed into the discriminator in the CTGAN model. The discriminator then provides the score of the virtual samples back to the generator.

[0020] S23 restricts the features of each virtual sample generated by CTGAN to be positive. At the same time, the create_cus tom_constraint module in the sdv library is called to create custom constraint parameters. After the CTGAN model generates virtual data, it is filtered to exclude virtual samples that do not meet the two constraints and retain virtual samples that meet the requirements, so that the generated virtual samples are more in line with the data range of the filtered dataset.

[0021] S24 Repeat steps S22-S23 to train the generator and discriminator until the discriminator can no longer distinguish between real and fake samples, thus obtaining a CTGAN model that has learned the data distribution of the selected dataset and the correlation between each ingredient and the ultimate compressive strength of concrete.

[0022] S25 saves the CTGAN model obtained in step S24 and generates 40,000-60,000 virtual samples to form a virtual dataset.

[0023] The expressions for the two restrictive conditions are:

[0024]

[0025]

[0026] Wherein, water represents the amount of water used, cement represents the amount of cement used, slag represents the amount of blast furnace slag used, flyash represents the amount of fly ash used, coarse aggregate represents the amount of coarse aggregate used, and fine aggregate represents the amount of fine aggregate used.

[0027] Step S3 specifically includes the following steps:

[0028] S31 uses the raw materials and mix proportion factors that affect the ultimate compressive strength of concrete as input to a one-dimensional residual convolution model, and the ultimate compressive strength as output.

[0029] S32 uses a one-dimensional residual convolution model to predict the ultimate compressive strength. The loss function used during model training is the mean squared error regression loss function, the optimizer is Adam, the learning rate is set to lr = 0.000001, and the number of iterations is epochs = 5000.

[0030] S33 uses two evaluation metrics, mean absolute error and correlation coefficient, to evaluate the prediction results of the one-dimensional residual convolution model when testing it.

[0031] The expression for the Mean Absolute Error (MAE) is:

[0032]

[0033] The correlation coefficient R 2 The expression is:

[0034]

[0035] Where n is the number of samples, and y is the true value of the sample data. The average of the true values ​​of all sample data. These are the predicted values ​​for the sample data.

[0036] As a further optimization, before training the CTGAN model, data analysis is performed on each factor in the concrete ultimate compressive strength index system in the original dataset, including Pearson correlation analysis between variables. The weak correlation between variables helps the prediction model capture more information and features. Therefore, features with a correlation coefficient of no more than 0.5 are selected for retention to form the screening dataset.

[0037] In summary, compared with the prior art, the above-described technical solutions conceived by this invention mainly possess the following technical advantages:

[0038] 1. The CTGAN model of this invention improves the model's ability to capture deeper intrinsic features in the original dataset to a certain extent, making the virtual samples more reliable. Compared with traditional GAN ​​and CTGAN models, the CTGAN model based on a one-dimensional convolutional network can learn the distribution of the original data and the correlation between various features more accurately, thereby generating more reliable virtual sample data.

[0039] 2. In the training process of the CTGAN model based on a one-dimensional convolutional network, the create_custo m_constraint module in the sdv library is introduced to create custom constraints, so that the generated virtual data is more in line with the data range of the original dataset.

[0040] 3. The one-dimensional residual convolution model used in this invention improves the expressive power and accuracy of the model by passing information between different convolutional layers through residual modules. Compared with traditional machine learning models, it can learn deeper relationships between various features of the data, effectively accelerate the training and convergence speed of the model, and has a very obvious effect on suppressing overfitting during the training process.

[0041] 4. The method of the present invention has strong scalability and can be applied to the performance prediction and optimization design of other materials.

[0042] 5. The method of the present invention can quickly and accurately predict the ultimate compressive strength of concrete, thereby improving the efficiency and accuracy of engineering design. Attached Figure Description

[0043] Figure 1 This is a schematic diagram of the CTGAN model and the one-dimensional residual convolution model in a method for predicting the ultimate compressive strength of concrete based on CTGAN, which is an embodiment of the present invention.

[0044] Figure 2 This is a graph showing the correlation between the original dataset and the virtual dataset in a method for predicting the ultimate compressive strength of concrete based on CTGAN, which is an embodiment of the present invention.

[0045] Figure 3 This is a t-SNE visualization comparison chart of virtual datasets obtained by different GAN models in a method for predicting the ultimate compressive strength of concrete based on CTGAN, which is an embodiment of the present invention.

[0046] Figure 4 This is a loss trend diagram of a one-dimensional residual convolution model during training in a method for predicting the ultimate compressive strength of concrete based on CTGAN, which is an embodiment of the present invention. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0048] like Figure 1 As shown in the figure, an embodiment of the present invention provides a method for predicting the ultimate compressive strength of concrete based on CTGAN, comprising the following steps:

[0049] Step 1: Construct a concrete ultimate compressive strength index system based on the raw material and mix proportion factors affecting the ultimate compressive strength of concrete. Collect sample data of this structural strength index system, denoted as the original dataset, and divide it into a training set and a test set at a 9:1 ratio. The original dataset consists of concrete that has solidified under normal conditions for a period of time, with the compressive strength obtained through typical compressive strength tests on 150mm high cylindrical standard specimens. The original dataset contains 9 parameters, each of which has a certain influence on the ultimate compressive strength.

[0050] The concrete ultimate compressive strength index system includes factors affecting ultimate compressive strength and evaluation indexes for ultimate compressive strength. Factors affecting ultimate compressive strength include cement content, blast furnace slag content, fly ash content, water content, high-efficiency water-reducing agent content, coarse aggregate content, fine aggregate content, and curing cycle. The evaluation index for compressive strength is the ultimate compressive strength of concrete.

[0051] In this embodiment, a publicly available dataset is used as the original dataset to ensure data accuracy and better compare with existing models. A concrete ultimate compressive strength index system is constructed based on the raw materials and mix proportion factors affecting the ultimate compressive strength of concrete, and sample data for this index system are collected.

[0052] Calculate the ratio of water usage to the sum of cement usage, blast furnace slag usage, and fly ash usage in the original dataset. At the same time, calculate the ratio of the sum of coarse aggregate usage and fine aggregate usage to the sum of cement usage, blast furnace slag usage, and fly ash usage. Find the maximum and minimum values ​​of the above two ratios in the original dataset.

[0053] The two ratios are restricted to their respective maximum and minimum values, thus forming two restrictive conditions;

[0054] S2 performs correlation analysis on each feature in the original dataset, retains features with a correlation coefficient of no more than 0.5, and recombines them to obtain the screening dataset; the screening dataset is divided into a training set and a test set in a 9:1 ratio. The training set of the screening dataset is used as the input of the CTGAN model to train the CTGAN model to learn the data distribution of the screening dataset and the correlation between each ingredient and the ultimate compressive strength of concrete, generating virtual concrete mix proportion data. The generated virtual concrete mix proportion data meets the two constraints in step S1, and the virtual dataset is output.

[0055] Pearson correlation analysis was performed on the variables in the concrete ultimate compressive strength index system. If there is a strong linear correlation between the input variables, it can lead to data redundancy and multicollinearity, reducing the model's generalization ability and stability. This invention uses the Pearson correlation coefficient to measure the correlation between features, such as... Figure 2 As shown, the Pearson correlation coefficient ranges from -1 to 1. The darker the color of the square, the larger the Pearson correlation coefficient between the two variables, and the stronger the correlation. Conversely, the darker the color of the square, the weaker the correlation.

[0056] Statistical analysis was performed on the variables in the ultimate compressive strength index system of concrete, that is, correlation analysis was conducted among the variables. Figure 2 The data in this embodiment shows that the correlation between the input variables is weak, that is, there is a certain non-linear correlation, which reflects the inherent structure and pattern of the data and helps the neural network capture more information and features.

[0057] S3 constructs a one-dimensional residual convolution model and trains it using a virtual dataset to obtain the final one-dimensional residual convolution model for predicting the ultimate compressive strength of concrete.

[0058] The one-dimensional residual convolutional model comprises four convolutional layers, each followed by a batch normalization layer and a ReLU activation function. The input to the one-dimensional residual convolutional model is a one-dimensional tensor, which, after four convolutions, batch normalization, and ReLU activation, outputs a three-dimensional tensor. The output of the ReLU activation function following the first convolutional layer is connected to a residual module, which comprises a residual layer, a batch normalization layer, and a ReLU function connected in series. The output of the residual module and the three-dimensional tensor are input together into three fully connected layers (FC) connected in series. The first two fully connected layers (FC) are each followed by a batch normalization layer and a ReLU activation function. The output of the last fully connected layer (FC) is the output of the one-dimensional residual convolutional model.

[0059] This invention uses a one-dimensional residual convolutional model to predict the ultimate compressive strength of concrete. The one-dimensional residual convolutional model uses convolutional layers to extract features from the input data and employs skip connections in the residual module to capture low-level and high-level features, improving the model's accuracy. The residual connection layers in the residual module connect the input of preceding convolutional layers to the output of subsequent convolutional layers, effectively mitigating the gradient vanishing problem, helping the model better retain original information, reducing overfitting during training, and increasing network depth without increasing parameters, thereby improving the model's expressive power, predictive performance, and generalization ability.

[0060] The loss function used in training the one-dimensional residual convolutional model is the mean squared error regression loss function, and the Adam optimizer is employed. When testing the one-dimensional residual convolutional model, two evaluation metrics are used to assess the prediction results: mean absolute error and correlation coefficient. The expression for the mean absolute error is as follows:

[0061]

[0062] The expression for the correlation coefficient is as follows:

[0063]

[0064] Where n is the number of samples, and y is the true value of the sample data. The average of the true values ​​of all sample data. These are the predicted values ​​for the sample data. The virtual dataset is divided into training and test sets in a 9:1 ratio.

[0065] Based on the above evaluation metrics, the structure of the one-dimensional residual convolution model with the best prediction performance is saved.

[0066] The expressions for the two restrictive conditions are:

[0067]

[0068]

[0069] Wherein, water represents the amount of water used, cement represents the amount of cement used, slag represents the amount of blast furnace slag used, flyash represents the amount of fly ash used, coarse aggregate represents the amount of coarse aggregate used, and fine aggregate represents the amount of fine aggregate used.

[0070] The CTGAN model includes a generator and a discriminator. The generator consists of three XOR layers connected in series. Each XOR layer is connected to a convolutional layer (Conv), a batch normalization (BN) layer, and a ReLU activation function before being connected to the input of the next XOR layer. The last XOR layer is processed by a convolutional layer (Conv), a BN layer, and a ReLU activation function before being connected to a fully connected layer (FC). The output of the FC layer is processed by the tanh activation function and the gumbel activation function to output a virtual sample. The input of the first XOR layer is random noise.

[0071] The discriminator includes two convolutional layers (Conv). The output of the first convolutional layer (Conv) is processed by a leaky activation function and a drop layer before being connected to the second convolutional layer (Conv). The output of the second convolutional layer (Conv) is processed by a leaky activation function and a drop layer before being connected to a fully connected layer (FC). The fully connected layer (FC) outputs the score of the virtual sample, which is then fed into the generator. The input of the first convolutional layer (Conv) is the sample from the filtered dataset and the virtual sample output by the generator.

[0072] The generator takes vectors drawn from a noisy distribution as input and transforms them into data similar to the selected dataset. The generator consists of four convolutional layers, each transformed by a non-linear activation function to produce more complex non-linear mappings.

[0073] The discriminator is another feedforward neural network used to evaluate whether a given data sample comes from either the filtered dataset or the generator. The discriminator receives virtual samples generated by the generator and real samples from the filtered dataset and attempts to distinguish between them. The discriminator consists of two convolutional layers and one fully connected layer (FC), where each layer is also transformed by a non-linear activation function to produce a more complex non-linear mapping. Finally, the discriminator outputs a score to evaluate how realistic the input data sample is.

[0074] The hyperparameters of the CTGAN model are appropriately adjusted based on the characteristics of the selected dataset, and prior constraints are added to construct the CTGAN model. In this embodiment, the CTGAN model's batch_size is set to 60, epochs to 10000, and discriminator_steps to 3, meaning that during training, the discriminator is updated 3 times each time the generator is updated. The generator in the CTGAN model adds random noise to the samples in the selected dataset to generate virtual samples, which are then fed back to the discriminator in the CTGAN model. The discriminator provides the score of the virtual samples to the generator.

[0075] The generator and discriminator neural networks are continuously trained until the discriminator can no longer distinguish between real and fake samples, resulting in a CTGAN model that has learned the data distribution of the selected dataset and the correlation between each ingredient and the ultimate compressive strength of concrete. This CTGAN model is then saved and fitted to generate 50,000 virtual samples, forming a virtual dataset.

[0076] Similarly, to demonstrate the reliability of the virtual dataset obtained in this application, data analysis was performed on the virtual dataset, including Pearson correlation analysis and t-SNE visualization analysis between the variables.

[0077] like Figure 2 As shown, the left figure reflects the correlation between variables in the selected dataset, and the right figure reflects the correlation between variables in the virtual dataset. It can be seen that the correlation between features in the original dataset and the virtual dataset is similar and the correlation is low, which is beneficial for the subsequent training of the one-dimensional residual convolution model.

[0078] A t-SNE visualization analysis was performed on the selected dataset and the virtual dataset, visualizing the nine-dimensional concrete dataset as a two-dimensional graph, such as... Figure 3 As shown, the distribution of virtual samples generated by CTGAN is more similar to that of real samples, indicating that CTGAN performs better in generating virtual samples. t-SNE uses the t-distribution to establish the probability distribution between the high-dimensional and low-dimensional spaces, and then optimizes the low-dimensional coordinates by minimizing the KL divergence between these two distributions. The formula for the KL divergence is as follows:

[0079]

[0080] Where p(j|i) represents a given data point x i At that time, another data point x j The conditional probability; q(j|i) represents the conditional probability of a given mapping point y. i At that time, another mapping point y j The conditional probability, P i Let Q represent the probability of the i-th event in the true probability distribution.i represents the probability of the i-th event in the fitted probability distribution, and C represents the KL divergence, which is an indicator that measures the difference between two probability distributions. It measures the information loss when the probability distribution Q is used to fit the true probability distribution P.

[0081] We replaced the CTGAN model with Wasserstein GAN, and performed t-SNE visualization analysis on the virtual data obtained by Wasserstein GAN processing, comparing it with the CTGAN model. Compared with Wasserstein GAN, CTGAN uses a conditional generator that can generate synthetic data according to user-specified conditions, thereby improving the diversity and controllability of the synthetic data. Furthermore, CTGAN uses a pattern regularization method to enhance the generator's ability to capture multimodal distributions in real data, thus improving the quality and accuracy of the synthetic data. Therefore, the CTGAN model is more stable and reliable, less susceptible to problems such as gradient vanishing, has a smoother training process, and generates more consistent data quality. Figure 3 As shown, the left figure is a t-SNE visualization analysis of the virtual data generated by Wasserstein GAN, and the right figure is a t-SNE visualization analysis of the virtual data generated by CTGAN.

[0082] A comparative analysis of the prediction results of a one-dimensional residual convolutional model with the same structure before and after data augmentation showed a significant improvement in prediction performance, confirming the effectiveness of using the CTGAN model for data augmentation. The changes in mean absolute error and correlation coefficient during training are shown below. Figure 4 As shown, both prediction evaluation metrics converged after 3000 iterations. To verify the effectiveness of training the one-dimensional residual convolutional model using a virtual dataset, 100 data samples were randomly selected from both the original and virtual datasets to form new validation sets. The trained one-dimensional residual convolutional model was then used to predict on this validation set, further confirming the effectiveness of using the CTGAN model for data augmentation and the validity of the one-dimensional residual convolutional model.

[0083] Example 1

[0084] The method for predicting the ultimate compressive strength of concrete based on CTGAN proposed in this invention mainly includes the following steps:

[0085] (1) Obtain the original dataset and perform data analysis.

[0086] The original dataset consists of 1030 experimental data points on industrial waste concrete collected by I.-C. Yeh from GitHub. It includes nine variables: cement dosage, blast furnace slag dosage, fly ash dosage, water dosage, high-efficiency water-reducing agent dosage, coarse aggregate dosage, fine aggregate dosage, curing age, and ultimate compressive strength. The data are shown in Table 1.

[0087] Table 1 Sample Data

[0088]

[0089]

[0090] By analyzing the original dataset, two constraints needed when generating data were identified and added to the CTGAN model.

[0091] The expressions for the two constraints are:

[0092]

[0093]

[0094] Wherein, water represents the amount of water used, cement represents the amount of cement used, slag represents the amount of blast furnace slag used, flyash represents the amount of fly ash used, coarse aggregate represents the amount of coarse aggregate used, and fine aggregate represents the amount of fine aggregate used.

[0095] (2) Perform data augmentation on the original dataset

[0096] At this point, the features in the original dataset are weakly correlated, which is why the dataset was filtered. Nine variables—cement usage, blast furnace slag usage, fly ash usage, water usage, high-efficiency water-reducing agent usage, coarse aggregate usage, fine aggregate usage, curing age, and ultimate compressive strength—were directly used as input to the CTGAN model. The two constraints mentioned above were added to the CTGAN model, and the hyperparameters were set to epoch 10000, batch size 60, and discriminator steps 3. After training, the CTGAN model was saved, and 50,000 virtual samples were generated to form a virtual dataset.

[0097] (3) Regression prediction of ultimate compressive strength

[0098] The input of the one-dimensional residual convolution model consists of eight variables: cement usage, blast furnace slag usage, fly ash usage, water usage, high-efficiency water-reducing agent usage, coarse aggregate usage, fine aggregate usage, and curing age. The ultimate compressive strength is used as the output.

[0099] The original dataset was divided into training and test sets in a 9:1 ratio. During training, the mean squared error regression loss function was used, and the Adam optimizer was employed. The learning rate was set to lr = 0.00001, and the number of epochs was 5000. The mean absolute error and correlation coefficients of predictions under different ultimate compressive strength prediction model structures and hyperparameters were recorded, as shown in Table 2.

[0100] Table 2

[0101]

[0102] The one-dimensional residual convolutional model established in this invention has the smallest sum of mean absolute error and correlation coefficient. Similarly, the virtual dataset is divided into training and test sets at a 9:1 ratio. Using the same hyperparameters, optimizer, and losser, the saved one-dimensional residual convolutional model is retrained. Simultaneously, 100 samples are randomly selected from both the original and virtual datasets to form a validation set, which is used as model input. The mean absolute error and correlation coefficient are calculated similarly. The mean absolute error and correlation coefficient obtained by using the same one-dimensional residual convolutional model for regression prediction on the original dataset, virtual dataset, and validation set are compared. The results are shown in Table 3.

[0103] Table 3

[0104]

[0105] The results in Table 3 show that after data augmentation of the original dataset using the CTGAN model, the performance of regression prediction using the one-dimensional residual convolution model is significantly improved. The mean absolute error and correlation coefficient of the one-dimensional residual convolution model in this application are 3.84 and 0.98, respectively, which significantly improves the accuracy. It has good application prospects in the field of prediction of ultimate compressive strength of concrete based on materials and mix proportions.

[0106] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0107] Any aspects not covered in this invention are applicable to existing technologies.

Claims

1. A method for predicting the ultimate compressive strength of concrete based on CTGAN, characterized in that, The method comprises the following steps: S1: Obtain an original data set of industrial waste concrete, calculate the proportion of water consumption to the sum of cement consumption, blast furnace slag consumption and fly ash consumption in the original data set, and calculate the proportion of the sum of coarse aggregate consumption and fine aggregate consumption to the sum of cement consumption, blast furnace slag consumption and fly ash consumption, and find the maximum and minimum values of the two proportions in the original data set; The two proportions are limited to the maximum and minimum values respectively to obtain values, forming two constraint conditions; S2: Perform correlation analysis on each feature in the original data set, retain features with a correlation coefficient not greater than 0.5, and recombine to obtain a screening data set; divide the screening data set into a training set and a test set, use the training set of the screening data set as the input of the CTGAN model, train the CTGAN model to learn the data distribution of the screening data set and the correlation between each ingredient and the ultimate compressive strength of concrete, generate virtual concrete mix proportion data, the generated virtual concrete mix proportion data meets the two constraint conditions in step S1, and output a virtual data set; S3: Construct a one-dimensional residual convolution model and train it using the virtual data set to obtain a final one-dimensional residual convolution model for predicting the ultimate compressive strength of concrete; The raw materials and mix proportion factors affecting the ultimate compressive strength of concrete are used as the input of the one-dimensional residual convolution model, and the ultimate compressive strength is used as the output; The one-dimensional residual convolution model comprises four convolution layers, each of which is provided with a batch normalization layer and a ReLU activation function; the input of the one-dimensional residual convolution model is a one-dimensional tensor, which is processed by four times of convolution, batch normalization and ReLU activation function to output a three-dimensional tensor; the output of the ReLU activation function after the first convolution layer is connected to a residual module, the residual module comprises a residual layer, a BN layer and a ReLU function connected in series, and the output of the residual module and the three-dimensional tensor are input into three fully connected layers FC connected in series, a batch normalization layer and a ReLU activation function are connected after the first two fully connected layers FC, and the output of the last fully connected layer FC is the output of the one-dimensional residual convolution model.

2. The method for predicting the ultimate compressive strength of concrete based on CTGAN according to claim 1, wherein The CTGAN model comprises a generator and a discriminator, the generator comprises three XOR layers connected in series, each XOR layer is connected to a convolution layer Conv, a BN layer and a ReLU activation function, and then connected to the input of the next XOR layer, the last XOR layer is connected to a fully connected layer FC after being processed by a convolution layer Conv, a BN layer and a ReLU activation function, the output of the fully connected layer FC is processed by a tanh activation function and a gumbel activation function respectively to output a virtual sample; the input of the first XOR layer is random noise. The discriminator includes two convolutional layers Conv, the output of the first convolutional layer Conv is connected to the second convolutional layer Conv after being processed by a leaky activation function and a drop layer, and the output of the second convolutional layer Conv is connected to a fully connected layer FC after being processed by a leaky activation function and a drop layer, and the fully connected layer FC outputs the score Score of the virtual sample to the generator; the input of the first convolutional layer Conv is the sample in the screening data set and the virtual sample output by the generator.

3. The method of predicting the ultimate compressive strength of concrete based on CTGAN according to claim 1, wherein, Step S2 specifically includes the following steps: S21, according to the characteristics of the screening data set, adding two limiting conditions between ingredients in step S1 in the CTGAN model, and adjusting the hyperparameters of the CTGAN model, wherein the convolutional layers Conv in the CTGAN model are all one-dimensional convolutional layers; S22, the generator in the CTGAN model adds random noise to the sample in the screening data set to generate a virtual sample for the discriminator in the CTGAN model, and the discriminator gives the score of the virtual sample to the generator for feedback; S23, limiting the features of each virtual sample generated by the CTGAN to be positive, and simultaneously, calling the create_custom_constraint module in the sdv library to create a custom constraint parameter constraints, after the CTGAN model generates virtual data, screening is performed to exclude virtual samples that do not meet the two limiting conditions, and virtual samples that meet the requirements are retained, so that the generated virtual samples are more consistent with the data range of the screening data set; S24, repeating steps S22-S23 to train the generator and the discriminator until the discriminator cannot distinguish the true and false of the sample, and obtaining the CTGAN model that learns the data distribution of the screening data set and the correlation between each ingredient and the ultimate compressive strength of concrete; S25, saving the CTGAN model obtained in step S24, generating 40000-60000 virtual samples to form a virtual data set.

4. The method of predicting the ultimate compressive strength of concrete based on CTGAN according to claim 1, wherein, The expressions of the two limiting conditions are: Wherein, water is the water amount, cement is the cement amount, slag is the blast furnace slag amount, flyash is the fly ash amount, coarseaggregate is the coarse aggregate amount, and fineaggregate is the fine aggregate amount.

5. The method of predicting the ultimate compressive strength of concrete based on CTGAN according to claim 1, wherein, Step S3 specifically includes the following steps: S31, taking the raw materials and mix proportion factors affecting the ultimate compressive strength of concrete as the input of the one-dimensional residual convolutional model, and taking the ultimate compressive strength as the output; S32, using the one-dimensional residual convolutional model to predict the ultimate compressive strength, using the mean square error regression loss function as the loss function during model training, using Adam as the optimizer, setting the learning rate lr=0.000001, and setting the number of iterations epochs=5000; S33, using the mean absolute error MAE and the correlation coefficient as two evaluation indexes to evaluate the prediction results of the one-dimensional residual convolutional model when testing the one-dimensional residual convolutional model; The expression of the mean absolute error MAE is: The correlation coefficient R 2 The expression is: where n is the number of samples, y is the real value of sample data, is the average value of real values of all sample data, is the predicted value of sample data.

Images