Concrete strength prediction method

Abstract

The invention provides a concrete strength prediction method, and relates to the field of concrete strength prediction. The concrete strength prediction process provided by the invention comprises the following steps: constructing a concrete strength prediction data set, converting time sequence data into frequency domain data through discrete Fourier transform, converting a real part, an imaginary part and an amplitude of each frequency component in the frequency domain data into a high-dimensional feature vector through linear mapping, a frequency domain information fusion module is used for combining high-dimensional feature vectors of a real part, an imaginary part and an amplitude to obtain frequency domain feature information, a multi-layer perceptron obtains time domain feature information of time sequence data, then an improved self-attention mechanism is used for better combining the time domain feature information and the frequency domain feature information to obtain a fusion feature vector, and the fusion feature vector is used for fusion of the real part, the imaginary part and the amplitude. And finally, all fusion feature vectors are spliced, and concrete strength prediction is carried out through fully connected nerves, so that time domain and frequency domain information can be effectively combined, and the prediction accuracy of the method is improved.

Description

Technical Field

[0001] The present invention belongs to the field of concrete strength prediction, and particularly relates to a concrete strength prediction method. Background Art

[0002] Concrete strength refers to the compressive capacity of concrete when subjected to external forces, representing the structural bearing capacity and stability of concrete. Insufficient concrete strength will make the structural safety of buildings unable to be guaranteed, leading to cracks, deformations or even collapses, causing serious economic losses and safety accidents. Therefore, concrete strength prediction is crucial. Predicting the strength changes of concrete under different conditions in advance can help engineers make accurate construction decisions, ensure project quality and safety, and thus effectively reduce the risks and potential hazards caused by insufficient strength.

[0003] In the field of concrete strength prediction, statistical models based on experimental data and machine learning methods are mainly adopted. Although statistical models can provide effective prediction results in some cases, they usually have difficulty in dealing with non-linear and complex multi-variable relationships. Machine learning methods establish prediction models by analyzing various influencing factors such as concrete mix ratio and curing conditions. However, they often fail to fully consider the frequency characteristics of the time series data composed of influencing factors to more comprehensively understand the internal structure of the data, thereby improving the reliability and accuracy of the prediction results.

[0004] Time series data often consists of components with different frequencies. Low-frequency components usually represent long-term trends or periodic fluctuations, while high-frequency components usually represent short-term changes or sudden fluctuations. Therefore, in machine learning methods, by mapping time series data to the frequency domain and combining frequency domain analysis with time domain analysis, it can more comprehensively help the model understand the internal structure of the data and improve the reliability of the prediction results. Summary of the Invention

[0005] The present invention provides a concrete strength prediction method, which improves the prediction accuracy by combining time domain and frequency domain information. First, the discrete Fourier transform is used to convert the time series data of concrete strength factor data into different frequency components, and the expression ability of the frequency components is enhanced through linear mapping. Then, a frequency domain information fusion module is proposed to effectively extract frequency domain feature information. A multi-layer perceptron is used to extract the time domain feature information of the time series data of concrete strength factor data. The self-attention mechanism is improved to fuse the time domain and frequency domain feature information and capture the correlation between the two. Finally, the fully connected neural network integrates the fused information for accurate prediction of concrete strength.

[0006] The technical methods adopted by the present invention to achieve the above object specifically include the following steps: S1. Collect the factor data affecting concrete strength and construct a concrete strength prediction data set; S2. The factor data constitutes time series data. The time series data is converted into different frequency components through discrete Fourier transform to form frequency domain data, and a frequency domain mapping module is constructed. S3. The amplitude is calculated based on the real part and the imaginary part of the frequency component. The real part, the imaginary part, and the amplitude are converted into high-dimensional feature vectors by using linear mapping. S4. A frequency domain information fusion module is proposed. An adaptive feature adjustment factor is introduced to weight and aggregate the three high-dimensional feature vectors of the real part, the imaginary part, and the amplitude of each frequency component in the frequency domain data, and the frequency domain feature information of each frequency component is obtained. S5. By using a multi-layer perceptron, deep learning is performed on the high-dimensional feature relationships in the time series data of the concrete strength factor data, and the time domain feature information containing the global information of the time series is extracted. S6. The self-attention mechanism is improved. An improved self-attention mechanism is used to construct a time domain and frequency domain fusion module. The time domain feature information is used as the query vector, the frequency domain feature information is used as the key and value vectors, and an asymmetric adjustment factor is introduced into the attention score. Based on the time domain feature information, the correlation of the frequency domain feature information is adaptively learned, and a new fusion feature vector is generated by combining the time domain and frequency domain feature information. S7. A fully connected neural network is constructed. All the fusion feature vectors are concatenated to obtain a prediction vector, and the fully connected neural network is used to calculate the prediction vector to obtain the concrete strength prediction value.

[0007] Preferably, in S1, the factor data affecting the concrete strength is collected, including cement content, cement strength, water-cement ratio, pouring height, curing temperature, curing humidity, and curing time. For the missing values that are not collected, the average values of the data three days before and after are used to fill them, and a concrete strength prediction data set is constructed.

[0008] Preferably, in S2, a frequency domain mapping module is constructed. The time series data of the c-th concrete strength factor data in the concrete strength prediction data set is input c ∈ [1, N], where N is the number of types of the collected concrete strength factor data, X c is the time domain data, T is the length of the time series of the input data, is the data value of the c-th concrete strength factor data collected at the t-th time step, t ∈ [1, T]. Discrete Fourier transform is applied to X c to map X c into different frequency components. The specific mathematical model is: In the formula, is the k-th frequency component obtained by performing a discrete Fourier transform on the time series data of the c-th concrete strength factor data, where k is the frequency component number. is in the exponential form of a complex number, where e is the base of the natural logarithm, i is the imaginary unit, and satisfies i 2 = -1, and π is the circumference ratio. is the real part of A c，k , and cos(·) is the cosine function. is the imaginary part of A c，k , and sin(·) is the sine function. A c is the frequency domain data obtained by arranging the frequency components obtained from the c-th concrete strength factor data in ascending order of the frequency component number. A c = [A 。，0 , A c，1 ,... A c，T-1 .

[0009] Preferably, a discrete Fourier transform is performed on the concrete strength prediction data. By converting the time domain data into frequency domain data, different frequency components in the data are extracted. These frequency components can help the model capture various characteristics such as periodicity, long-term trends, and short-term fluctuations in the data, balance the model's learning of low-frequency and high-frequency components, and thus improve the model's prediction ability and robustness.

[0010] Preferably, in S3, according to the real part Re(A c，k ) and the imaginary part Im(A c，k ) of the k-th frequency component A c，k of the c-th concrete strength factor data, the amplitude |A c，k | is calculated. The amplitude can help understand the strength of the frequency component. The specific mathematical model is: In the formula, is the amplitude of the frequency component A c，k . Furthermore, the real part, the imaginary part, and the amplitude are linearly mapped into a richer space for representation to obtain high-dimensional vectors. The specific mathematical model is: and are the high-dimensional feature vectors obtained by linearly mapping the real part, the imaginary part, and the amplitude respectively. d is the dimension of the vector after linear mapping. W Re , W Im and are trainable weight matrices, and b Re , b Im and are trainable bias parameters.

[0011] Preferably, by linearly mapping the real part, imaginary part, and amplitude of the frequency components to generate a high-dimensional feature vector, the characteristic differences between different frequency components can be effectively captured, and the low-dimensional data can be transformed into a high-dimensional feature space, which can improve the expression ability and learning ability of the model. By introducing trainable weight and bias parameters, the model can automatically optimize these mapping processes according to the data, further improving the prediction accuracy.

[0012] Preferably, in S4, a frequency-domain information fusion module is proposed, which specifically includes the following steps: S41. Introduce an adaptive feature adjustment factor, and calculate the adaptive feature adjustment factor according to the intensity of the real part, imaginary part, and amplitude high-dimensional feature vector matrices. The specific mathematical model is: In the formula, are the adaptive feature adjustment factors of the real part, imaginary part, and amplitude respectively, and ||·|| F is the Frobenius norm; Further use this factor to perform weighted summation on the real part, imaginary part, and amplitude high-dimensional feature vectors, dynamically adjust the contribution of the three high-dimensional feature vectors to the final output, and obtain a feature vector that fuses the overall frequency-domain information. The specific mathematical model is: In the formula, and are high-dimensional feature vector matrices formed by arranging the real part, imaginary part, and amplitude of the frequency components in the frequency-domain data of the c-th concrete strength factor according to the frequency component numbers respectively, is the feature vector matrix that fuses the real part, imaginary part, and amplitude high-dimensional feature vectors; S42. Perform a non-linear mapping on Y1 to obtain a new feature vector, perform a residual connection between Y1 and the new feature vector, and then obtain the final frequency-domain feature information through non-linear mapping again. The specific mathematical model is: Y2 = ReLU(Y1W1 + b1); Y c = ReLU((Y2 + Y1)W2 + b2); In the formula, is the new feature vector obtained through non-linear mapping, W1, are trainable weight matrices, b1, are bias parameters, ReLU(·) is the ReLU activation function, is the frequency-domain feature information matrix obtained through non-linear mapping combined with residual connection.

[0013] Preferably, by introducing an adaptive feature adjustment factor, non-linear mapping, and residual connection, the proposed frequency-domain information fusion module can efficiently fuse the real part, imaginary part, and amplitude of three high-dimensional feature vectors, dynamically adjust the contribution of each feature to the final output, and further improve the model's representation ability through a deep network structure. This method can effectively capture complex patterns in the frequency-domain information and improve the robustness of the model in practical applications.

[0014] Preferably, in S5, a multi-layer perceptron is used to learn the time-series data of the c-th concrete strength factor data of the overall time-domain feature information. The multi-layer perceptron includes an input layer, a hidden layer, and an output layer, a total of 3 layers. Among them, the input layer captures the basic features in X through linear transformation and activation function c . The specific mathematical model is: h1 = ReLU(X c W3 + b3); In the formula, is the basic feature vector of the input layer, d h is the dimension of the basic feature of the input layer, are the learnable weight matrix and bias parameter of the input layer respectively, and ReLU(·) is the ReLU activation function; The output layer further extracts the time-domain feature information from the output features of the hidden layer. The specific mathematical model is: h2 = ReLU(h1W4 + b4); In the formula, is the time-series feature of the hidden layer, are the learnable weight matrix and bias parameter of the hidden layer respectively; The hidden layer combines these basic features, extracts deeper time-series relationships to obtain time-series features. The specific mathematical model is: h c = h2W out + b out ; In the formula, is the time-domain feature information containing the c-th concrete strength factor data output by the output layer, are the learnable weight matrix and bias parameter of the output layer respectively.

[0015] Preferably, by using a three-layer multi-layer perceptron to learn the high-dimensional feature relationship of time-series data, deeper time-series information modeling can be achieved. The input layer performs linear transformation on the original data and uses the activation function to enhance the representation of non-linear features. Further, the expression ability of the model is optimized through the hidden layer and the output layer. This process can extract the comprehensive information of time-series data and provide a more accurate time-series feature representation for subsequent analysis and prediction.

[0016] Preferably, in S6, the self-attention mechanism is improved to construct a time-frequency domain fusion module, which specifically includes the following steps: S61. Take the time-domain feature information h c as the query vector, and take the time-domain feature information of each frequency component in the frequency-domain feature vector matrix Y c as the key and value vectors. Y c = [y c，0 , y c，1 ,..., y c，T-1 , where is the frequency-domain feature information of the k-th frequency component of the c-th concrete strength factor data. The similarity between the query vector and the key vector is calculated by dot product to obtain the attention score. The specific mathematical model is: In the formula, is the attention score of the k-th frequency component of the c-th concrete strength factor data, is the transpose of y c，k ; S62. The interaction between the time-domain feature and the frequency-domain feature is asymmetric. An asymmetric adjustment factor is introduced to asymmetrically adjust the attention score, calculate the asymmetric adjustment factor and add it to the attention score s′ c，k . The specific mathematical model is: s c，k = s′ c，k + σ c，k ; In the formula, is the asymmetric adjustment factor of the k-th frequency component of the c-th concrete strength factor data, β, are learnable parameters, is the attention score of the k-th frequency component of the c-th concrete strength factor data after fusing the asymmetric adjustment factor; S63. Normalize s c，k to obtain the attention weight. The specific mathematical model is: In the formula, is the normalized attention weight of s c，k , is the attention score of the i-th frequency component of the c-th concrete strength factor data after fusing the asymmetric adjustment factor; According to the weight, perform weighted summation on the corresponding value vectors to obtain the feature vector after fusing the time-frequency domain feature information. The specific mathematical model is: In the formula, The weighted sum is obtained by weighting all the frequency-domain feature information of the c-th concrete strength factor data according to the corresponding attention weights, and it is added to the time-domain feature information h c to obtain the fused feature vector.

[0017] Preferably, the time-domain feature information and the frequency-domain feature information are effectively fused through an improved self-attention mechanism, and an asymmetric adjustment factor is introduced to adjust the interaction mode between the time-domain and frequency-domain features, thereby enhancing the flexibility and expression ability of the model. The improved self-attention mechanism uses a query vector to perform a weighting operation on each frequency-domain feature information to combine the time-domain and frequency-domain information, so as to more comprehensively understand the data pattern.

[0018] Preferably, in S7, the fused feature vectors obtained from all the concrete strength factor data are concatenated to obtain the final feature vector for prediction. The specific mathematical model is:[[]] In the formula, is the feature vector obtained by concatenating the fused feature vectors of all the concrete strength factor data, is the concatenation operation; The feature vector obtained by concatenation is calculated through a fully connected neural network to obtain the concrete strength prediction value. The specific mathematical model is: P--FCNN(L); In the formula, is the final concrete strength prediction value, and FCNN(·) is the fully connected neural network.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention converts the time-domain data of the concrete strength factor into frequency-domain data through discrete Fourier transform, extracts the real part, imaginary part, and amplitude of each frequency component in the frequency-domain data, maps the real part, imaginary part, and amplitude into high-dimensional feature vectors by linear mapping, and then dynamically fuses the three high-dimensional feature vectors through a frequency-domain information fusion module to obtain the global frequency-domain feature information of the frequency components. At the same time, a multi-layer perceptron is applied to extract the time-domain feature information of the time series data. Subsequently, an improved self-attention mechanism is used to fuse the time-domain and frequency-domain feature information. Finally, all the fused features are concatenated and input into a fully connected neural network for the final concrete strength prediction, improving the prediction accuracy. Description of the Drawings

[0020] Figure 1 It is a flowchart of the steps of a concrete strength prediction method.

[0021] Figure 2 It is a diagram of the frequency-domain information fusion module.

[0022] Figure 3 It is a structure diagram of the multi-layer perceptron.

[0023] Figure 4 It is a diagram of the time-domain and frequency-domain fusion module.

[0024] Figure 5 It is the effect diagram of concrete strength prediction. Specific implementation manners

[0025] The present invention provides a method for predicting concrete strength. This method uses the discrete Fourier transform to convert the time-domain data of concrete strength factors into frequency-domain data, enhances the feature expression of frequency components in the frequency-domain data through linear mapping, proposes a frequency-domain information fusion module to combine the real part, imaginary part, and amplitude information of frequency components to obtain frequency-domain feature information, and a multi-layer perceptron obtains the overall information of the time-series data in the time domain. On this basis, the self-attention mechanism is improved to achieve the deep fusion of time-domain and frequency-domain information, capture more complex feature relationships, and finally splice all the fusion information and calculate the concrete strength prediction result through a fully connected neural network. The technical solutions in the embodiments of the present invention will be described in detail and completely below, including the following steps, as Figure 1 shown.

[0026] S1. Collect the data of factors affecting concrete strength and construct a concrete strength prediction data set.

[0027] Further, in the step S1, collect the data of concrete strength factors for 102 days, including cement content, cement strength, water-cement ratio, pouring height, curing temperature, curing humidity, and curing time. Preprocess the 7 types of collected data. For missing values, fill them with the average values of the data three days before and after the data to ensure the quality and consistency of the data, and construct a high-quality concrete strength prediction data set. Subsequently, divide the data into a training set and a validation set in chronological order, and the division ratio is 8:2. Use the first 7 days of data of each day as the historical data input for model training.

[0028] S2. The factor data forms time-series data, and the time-series data is converted into different frequency components through the discrete Fourier transform to form frequency-domain data, and a frequency-domain mapping module is constructed.

[0029] Further, in the step S2, construct a frequency-domain mapping module to obtain a concrete strength prediction data set, which contains 7 types of data of factors affecting concrete strength. The time series of the c-th type of concrete strength factor data is denoted as c ∈ [1, 7], where T is the length of the time series of the input data, T = 7, is the observed value of the c-th type of concrete strength factor data at the t-th time step, t ∈ [1, 7]. Apply the discrete Fourier transform to convert the time series X of the c-th type of concrete strength factor data cMapped to the frequency domain to obtain the different frequency components of the time series. The specific mathematical model is as follows: In the formula, is the k-th frequency component obtained by the discrete Fourier transform of the time series of the c-th concrete strength factor data. k is the frequency component number. is the exponential form of a complex number, where e is the base of the natural logarithm, i is the imaginary unit, and it satisfies i 2 = -1, and π is the circumference ratio. is the real part of the k-th frequency component, and cos(·) is the cosine function. is the imaginary part of the k-th frequency component, and sin(·) is the sine function. A c is the frequency domain data obtained by arranging the frequency components obtained from the c-th concrete strength factor data in order. A c = [A c，0 , A c，1 , …, A c，6 .

[0030] S3. Calculate the amplitude based on the real part and the imaginary part of the frequency component, and use linear mapping to convert the real part, the imaginary part, and the amplitude into a high-dimensional feature vector.

[0031] Furthermore, in the above S3, the k-th frequency component A c，k in the frequency domain data of the c-th concrete strength factor is composed of the real part Re(A c，k ) and the imaginary part Im(A c，k ). Calculate the amplitude |A c，k | of this frequency component through the real part and the imaginary part to help the model understand the significance of the frequency component. The specific mathematical model is as follows: In the formula, is the amplitude of the frequency component A c，k ; Convert the real part, the imaginary part, and the amplitude of A c，k into a richer high-dimensional feature vector through linear mapping to improve the feature expression ability of the model. The specific mathematical model is as follows: In the formula, and are respectively the 32-dimensional high-dimensional vectors obtained by linearly mapping the real part, the imaginary part, and the amplitude. W Re , W Im and are trainable weight matrices, and b Re , b Im and are trainable bias parameters.

[0032] S4. Propose a frequency-domain information fusion module, introduce an adaptive feature adjustment factor, and weighted aggregate the real part, imaginary part, and amplitude of each frequency component in the frequency-domain data, which are three high-dimensional feature vectors, to obtain the frequency-domain feature information of each frequency component.

[0033] Further, in S4, a frequency-domain information fusion module is proposed, as Figure 2 shown, which specifically includes the following steps: S41. For the real part, imaginary part, and amplitude of all frequency components in the frequency-domain data of the c-th concrete strength factor, form matrices of high-dimensional feature vectors according to the frequency component numbers, and obtain the real part matrix imaginary part matrix and amplitude matrix Calculate the adaptive feature adjustment factor based on the three matrices. The specific mathematical model is: In the formula, are the adaptive feature adjustment factors of the real part, imaginary part, and amplitude respectively, and ||·|| F is the Frobenius norm; Use the adaptive feature adjustment factor to perform weighted summation on the real part, imaginary part, and amplitude high-dimensional feature vectors to obtain a feature vector containing the overall frequency-domain information. The specific mathematical model is: In the formula, is the feature vector matrix combining the real part, imaginary part, and amplitude high-dimensional feature vectors; S42. Perform a non-linear mapping on the feature vector matrix Y1 in sequence and a non-linear mapping combining residual connection to obtain a vector matrix that fully fuses different-dimensional information in the frequency-domain features. The specific mathematical model is: Y2 = ReLU(Y1W1 + b1); Y c = ReLU((Y2 + Y1)W2 + b2); In the formula, is the new feature vector obtained through non-linear mapping, is the trainable weight matrix, is the bias parameter, ReLU(·) is the ReLU activation function, is the frequency-domain feature information matrix obtained through the non-linear mapping combining residual connection, is the trainable weight matrix, is the bias parameter.

[0034] S5. By using a multi-layer perceptron, perform deep learning on the high-dimensional feature relationships in the time-series data of concrete strength factor data, and extract the time-domain feature information containing the global information of the time series.

[0035] Further, in S5, as Figure 3 shown, use a multi-layer perceptron to process the time-series data X of the c-th type of concrete strength factor data c for processing. The multi-layer perceptron includes three layers: an input layer, a hidden layer, and an output layer. The input layer includes a linear transformation and an activation function to capture the basic features in the input data. The specific mathematical model is: h1 = ReLU(X c W3 + b3); In the formula, is the basic feature vector of the input layer, are the learnable weight matrix and bias parameter of the input layer respectively, and ReLU(·) is the ReLU activation function. The hidden layer uses a linear transformation and an activation function again to combine these features and extract deeper time-series feature vectors. The specific mathematical model is: h2 = ReLU(h1W4 + b4); In the formula, is the time-series feature vector of the hidden layer, are the learnable weight matrix and bias parameter of the hidden layer respectively; Finally, the output layer maps the learned features to the final output space to obtain a feature vector containing the overall time-series information. The specific mathematical model is: h c = h2W out + b out ; In the formula, is the time-domain feature information finally output by the output layer, representing the overall time-series information of the c-th type of time-series data, are the learnable weight matrix and bias parameter of the output layer respectively.

[0036] S6. Improve the self-attention mechanism, use the improved self-attention mechanism to construct a time-domain and frequency-domain fusion module, use the time-domain feature information as the query vector, the frequency-domain feature information as the key and value vectors, and introduce an asymmetric adjustment factor in the attention score. Based on the time-domain feature information, adaptively learn the correlation of the frequency-domain feature information, and combine the time-domain and frequency-domain feature information to generate a new fusion feature vector.

[0037] Further, in S6, improve the self-attention mechanism to construct a time-domain and frequency-domain fusion module, asFigure 4 As shown in the figure, it specifically includes the following steps: S61. Use the time-domain feature information h c as the query vector, and use the time-domain feature information of each frequency component in the frequency-domain feature vector matrix Y c as the key and value vectors. Y c = [y c，0 , y c，1 ,..., y c，T-1 , where is the frequency-domain feature information of the k-th frequency component of the c-th concrete strength factor data. Calculate the similarity weight between the query vector and the key vector through the dot product method to obtain the attention score. The specific mathematical model is: In the formula, is the attention score of the k-th frequency component of the c-th concrete strength factor data, is the transpose of y c，k ; S62. Introduce an asymmetric adjustment factor, which reflects the asymmetric relationship between the time-domain feature information and the frequency-domain feature information, and add it to the preliminarily calculated attention score s′ c，k . The specific mathematical model is: s c，k = s′ c，k + σ c，k ; In the formula, is the asymmetric adjustment factor of the k-th frequency component of the c-th concrete strength factor data, β, are learnable parameters, is the attention score of the k-th frequency component of the c-th concrete strength factor data after fusing the asymmetric adjustment factor; S63. Normalize s c，k to obtain the attention weight. The specific mathematical model is: In the formula, is the attention weight after normalizing s c，k , is the attention score of the i-th frequency component of the c-th concrete strength factor data after fusing the asymmetric adjustment factor; Based on the above-normalized attention weight, perform weighted summation on the corresponding value vectors to obtain the feature vector that fuses the frequency-domain and time-domain feature information. The specific mathematical model is: In the formula, For the data of the c-th concrete strength factor, the frequency-domain characteristic information of all frequency components is weighted according to the corresponding attention weights, and after obtaining the weighted sum, it is added to the time-domain characteristic information h c to obtain the fused feature vector.

[0038] S7. Construct a fully connected neural network, splice all the fused feature vectors to obtain a prediction vector, and use the fully connected neural network to calculate the prediction vector to obtain the predicted value of the concrete strength.

[0039] Furthermore, in S7, the fused feature vectors obtained from all the concrete strength factor data are spliced according to the last dimension to obtain the feature vector for prediction. The specific mathematical model is as follows: In the formula, is the feature vector obtained by splicing the fused feature vectors of all the concrete strength factor data, is the splicing operation; Subsequently, the predicted value of the concrete strength is calculated using the fully connected neural network. The specific mathematical model is: P = FCNN(L); In the formula, is the final predicted value of the concrete strength, and FCNN(·) is the fully connected neural network.

[0040] Furthermore, this method is implemented based on the Python 3.8 programming language and the PyTorch framework. In order to accelerate the training process, this method runs in the CUDA 12 environment and uses the NVIDIA RTX 3090 GPU for calculation. During the training process, the learning rate is 1×10 -3 , the batch size is 12, the optimizer is Stochastic Gradient Descent (SGD), and the loss function is Mean Squared Error (MSE).

[0041] Furthermore, the prediction effect of this method is as Figure 5 shown, where the ordinate is the concrete strength (MPa), the abscissa is the time (days), the gray dashed line in the figure represents the actual strength value of the concrete, and the black solid line is the concrete strength value predicted by this method. It can be seen from Figure 5 that there is a high degree of fitting between the predicted value and the actual value, and as time goes by, the change trends of the two are basically the same, indicating that this method can effectively capture the change law of the concrete strength and verify the effectiveness and feasibility of this method in the concrete strength prediction task.

[0042] The above is only the preferred embodiment of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the inventive concept of the present invention, several deformations and improvements can still be made, and these all belong to the protection scope of the present invention.

Claims

1. A method for predicting the strength of concrete, characterized in that, It includes the following steps: S1. Collect data on factors affecting concrete strength and construct a concrete strength prediction dataset; S2. The factor data constitutes time series data. The time series data is converted into different frequency components through discrete Fourier transform to form frequency domain data, and a frequency domain mapping module is constructed; S3. Calculate the amplitude based on the real and imaginary parts of the frequency components, and use linear mapping to convert the real part, imaginary part, and amplitude into high-dimensional feature vectors; S4. Propose a frequency domain information fusion module, introduce an adaptive feature adjustment factor, and weighted aggregate the real part, imaginary part, and amplitude of the three high-dimensional feature vectors of each frequency component in the frequency domain data to obtain the frequency domain feature information of each frequency component; S5. By using a multi-layer perceptron, perform deep learning on the high-dimensional feature relationships in the time series data of the concrete strength factor data, and extract the time domain feature information containing the global information of the time series; S6. Improve the self-attention mechanism, use the improved self-attention mechanism to construct a time domain and frequency domain fusion module, use the time domain feature information as the query vector, the frequency domain feature information as the key and value vectors, and introduce an asymmetric adjustment factor into the attention score. Based on the time domain feature information, adaptively learn the correlation of the frequency domain feature information, and combine the time domain and frequency domain feature information to generate a new fused feature vector; S7. Construct a fully connected neural network, splice all the fused feature vectors to obtain a prediction vector, and use the fully connected neural network to calculate the prediction vector to obtain the concrete strength prediction value.

2. The concrete strength prediction method according to claim 1, characterized in that, In S1, collect data on factors affecting concrete strength, including cement content, cement strength, water-cement ratio, pouring height, curing temperature, curing humidity, and curing time. For the missing values in the data, fill them with the average values of the data three days before and after, and construct a high-quality concrete strength prediction dataset.

3. A method for predicting the strength of concrete according to claim 2, characterized in that, In the above S2, the time series data of the c-th concrete strength factor data in the concrete strength prediction dataset is denoted as where T is the length of the time series, is the data value of the c-th concrete strength factor data collected at the t-th time step, t ∈ [1, T], c ∈ [1, N], N is the number of types of concrete strength factor data collected, and the discrete Fourier transform is used to map X c to different frequency components. The specific mathematical model is as follows: In the formula, is the k-th frequency component obtained by discrete Fourier transform of the time series data of the c-th concrete strength factor data, where k is the frequency component number, is the exponential form of a complex number, where e is the base of the natural logarithm, i is the imaginary unit, and satisfies i 2 = -1, and π is the circumference ratio, is the real part of A c，k , is the imaginary part of A c,k , cos(·) is the cosine function, sin(·) is the sine function, and A c is the frequency domain data obtained by arranging the frequency components of the c-th concrete strength factor data in order, and A c = [A c,0 , A c,1 ,..., A c,T-1 .

4. A method for predicting the strength of concrete according to claim 3, characterized in that, In S3, the k-th frequency component A in the frequency-domain data of the c-th concrete strength factor data c，k is composed of the real part Re(A c，k ) and the imaginary part Im(A c，k ). The amplitude |A c，k | of this frequency component is calculated based on the real part and the imaginary part. The specific mathematical model is as follows: In the formula, is the frequency component A c，k amplitude; Use linear mapping to map the real part, imaginary part, and amplitude into the high-dimensional space respectively to obtain high-dimensional feature vectors. The specific mathematical model is: In the formula, and are high-dimensional feature vectors obtained by linearly mapping the real part, imaginary part, and amplitude of the frequency component A c，k respectively. d is the dimension of the vector after linear mapping. W Re , W Im and are trainable weight matrices, and b Re , b Im and are trainable bias parameters.

5. A method for predicting the strength of concrete according to claim 4, characterized in that, In S4, the proposed frequency domain information fusion module includes the following steps: S41. For the high-dimensional eigenvectors of the real part, imaginary part, and amplitude of all frequency components in the frequency-domain data of the c-th concrete strength factor data, construct a matrix by numbering the frequency components, and obtain the real part matrix Imaginary part matrix And amplitude matrix Introduce an adaptive feature adjustment factor. The specific mathematical model is as follows: wherein, are respectively the adaptive feature adjustment factors of the real part, imaginary part, and amplitude, and ||·|| F is the Frobenius norm; Use the adaptive feature adjustment factor to perform weighted summation on the real part, imaginary part, and amplitude high-dimensional feature vectors to obtain a feature vector containing the overall frequency domain information. The specific mathematical model is: In the formula, is the eigenvector matrix that combines the real part, imaginary part, and amplitude high-dimensional eigenvector; S42. Perform a non-linear mapping on Y1 to obtain a new feature vector, and pass Y1 and the new feature vector through a non-linear mapping combined with a residual connection again to obtain the frequency domain feature information. The specific mathematical model is: Y2 = ReLU(Y1W1 + b1); Y c = ReLU((Y2 + Y1)W2 + b2); wherein, is the new feature vector obtained through non-linear mapping, is the trainable weight matrix, is the bias parameter, and ReLU(·) is the ReLU activation function, is the frequency domain feature information matrix obtained through non-linear mapping combined with residual connection, is the trainable weight matrix, is the bias parameter.

6. A method for predicting the strength of concrete according to claim 5, characterized in that, In S5, for the time series data of the c-th concrete strength factor data Use a multi-layer perceptron to obtain X that contains c The time-domain feature information of the overall time series information. The multi-layer perceptron consists of 3 layers: an input layer, a hidden layer, and an output layer. The input layer captures the basic features in X through linear transformation and activation functions. The specific mathematical model is as follows: c in h1 = ReLU(X c W3 + b3); Wherein, is the basic feature vector obtained by the input layer, and d h is the dimension of the hidden vector of the input layer, are respectively the learnable weight matrix and bias parameter of the input layer, and ReLU(·) is the ReLU activation function; The hidden layer further extracts the deeper time series features in the basic features. The specific mathematical model is: h2 = ReLU(h1W4 + b4); In the formula, is the temporal feature vector obtained by the hidden layer, are the learnable weight matrix and bias parameter of the hidden layer, respectively; The output layer maps the time series features to the final output space to obtain the time domain feature information. The specific mathematical model is: h c = h2W out + b out ; wherein, is the time-domain characteristic information of the overall information of the time series data containing the data of the c-th concrete strength factor output by the output layer, are respectively the learnable weight matrix and bias parameter of the output layer.

7. A method for predicting concrete strength according to claim 6, characterized in that, In S6, improve the self-attention mechanism and construct a time domain and frequency domain fusion model, including the following steps: S61. Take the time-domain feature information h of the c-th concrete strength factor data c as the query vector, and the frequency-domain feature information of each frequency component in the frequency domain as the key and value vectors, Y c = [y c，0 , y c，1 ,..., y c，T-1 , where y c，k ∈ Y c is the frequency-domain feature information of the k-th frequency component of the c-th data. Calculate the attention score between the query vector and the key vector by the dot product method. The specific mathematical model Type: In the formula, is the attention score of the k-th frequency component of the c-th concrete strength factor data, is the transpose of y c，k ; S62. Introduce an asymmetric adjustment factor and add it to the attention score s′ c，k The specific mathematical model is as follows: s c，k = s' c，k + σ c，k ; In the formula, is the asymmetric adjustment factor of the k-th frequency component of the c-th concrete strength factor data, is a learnable parameter, is the attention score of the k-th frequency component fusion asymmetric adjustment factor of the c-th concrete strength factor data; S63. Normalize s c，k to obtain attention weights. The specific mathematical model is as follows: In the formula, is s c，k normalized attention weight, is the attention score of the fusion asymmetric adjustment factor of the i-th frequency component of the c-th concrete strength factor data; According to the weights, perform weighted summation on the corresponding value vectors to obtain a feature vector after fusing the frequency domain and time domain feature information. The specific mathematical model is: In the formula, For the data of the c-th concrete strength factor, the frequency domain feature information of all frequency components is weighted according to the corresponding attention weights, and after obtaining the weighted sum, it is added to the time domain feature information h c to obtain the fused feature vector.

Images