A method for predicting the installation quality of green and environmentally friendly wall materials

Abstract

The present invention proposes a method for predicting the installation quality of green environmentally friendly wall materials, and relates to the technical field of machine learning. The present invention proposes an Eqformer prediction model, which includes a spatiotemporal feature extraction module, a multi-frequency feature fusion module, and a prediction module. Specifically, the spatiotemporal feature extraction module is used to process the correlation between the time dimension and the space dimension, the multi-frequency feature fusion module is used to process the correlation between the feature position dimension and the feature environment dimension, and the prediction module is used to integrate the multi-dimensional features and convert them into available installation quality prediction results. The modules cooperate with each other to achieve accurate prediction of the installation quality of green environmentally friendly wall materials.

Description

Technical Field

[0001] The present invention belongs to the technical field of machine learning, and in particular relates to a method for predicting the installation quality of green environmentally friendly wall materials. Background Art

[0002] With the promotion of green buildings and sustainable development concepts, green and environmentally friendly wall materials have been widely used in the construction industry. Such materials usually have the characteristics of low carbon, environmental protection and high efficiency, which can significantly reduce the negative impact of construction on the environment. However, due to the complexity of environmental conditions, construction technology and material properties during the installation process, the installation quality of wall materials may be affected by many factors. In order to ensure the performance and long-term durability of the materials, it is particularly important to predict and evaluate the effectiveness of installation quality. Through accurate installation quality prediction, the construction process can be optimized and the overall quality and performance of the building can be improved.

[0003] Due to the large differences in environmental conditions at the construction site (such as temperature, humidity, etc.) and operating specifications between different construction sites, these factors may lead to uneven installation quality of materials. Secondly, the characteristics of wall materials during the installation process are highly nonlinear and non-stationary, which makes it difficult for traditional prediction methods to accurately evaluate the long-term performance of materials. In addition, the complex interactions among the multiple variables involved in the installation process (such as material stress, construction time, environmental conditions, etc.) increase the difficulty of effectively predicting the installation quality.

[0004] The data at construction sites are often non-stationary and are greatly affected by fluctuations in the external environment and construction technology. Deep learning, especially neural network-based architectures, can automatically extract complex multi-dimensional features from the data. Compared with traditional manual feature selection, deep learning can effectively capture the nonlinear relationship between various variables in the construction process, thereby greatly improving the prediction accuracy. This is especially important for complex and nonlinear data, such as the complex relationship between the performance of wall materials and construction conditions. Summary of the invention

[0005] The present invention provides a method for predicting the installation quality of green environmentally friendly wall materials. Aiming at the installation quality data of green environmentally friendly wall materials with non-stationary and multivariate factors, an Eqformer model is proposed, which consists of a spatiotemporal feature extraction module, a multi-frequency feature fusion module and a prediction module.

[0006] The technical solution adopted by the present invention to achieve the above-mentioned purpose specifically includes the following steps:

[0007] S1. Collect data related to the installation quality of green environmentally friendly wall materials, including features and target variables, and pre-process the collected data related to the installation quality of green environmentally friendly wall materials;

[0008] S2. Use of installation quality data of pre-treated green environmentally friendly wall materials The method is standardized, and the green environmental protection wall material installation quality data is divided into training set and test set;

[0009] S3. Construct a spatiotemporal feature extraction module to process the correlation between the time dimension and the space dimension. The specific steps are as follows:

[0010] S31. Input the physical parameter data of green environmentally friendly wall materials , decomposed into oscillatory components Apply an exponential function to each IMF component Perform frequency shift and combine all IMF components into a matrix , converted into a feature embedding matrix through a fully connected layer ;

[0011] S32, input position association matrix and the feature matrix , generated by aggregation through the GraphSAGE algorithm and combined into a position embedding matrix , use the fully connected layer to transform and get the final position embedding matrix , then enter the installation environment parameter data , processed using a fully connected layer, to obtain the environment embedding matrix ;

[0012] S33. Embedding features into matrix , the final position embedding matrix Add and fuse to get the feature position matrix , and embed the final position into the matrix , environment embedding matrix Add and fuse to get the environment position matrix ;

[0013] S4. Construct a multi-frequency feature fusion module to process the correlation between the feature position dimension and the feature environment dimension. The specific steps are as follows:

[0014] S41. Propose dynamic multi-head feature position self-attention and calculate activation vector and the weight distribution vector , input feature position matrix Calculate the attention score, perform dynamic head selection and dynamic weight allocation, and obtain the feature position matrix ;

[0015] S42, propose dynamic multi-scale feature environment self-attention, input environment position matrix Through multi-scale convolutional layers, we get the embedded representation of the environment position , introduce the dynamic weight matrix , multiply the environment position embedding representation with the dynamic weight matrix to obtain the fused feature representation And calculate the attention score to obtain the feature environment matrix ;

[0016] S43, feature position matrix and the feature environment matrix Input the fully connected layer for dynamic weighted fusion operation to obtain weighted feature representation ;

[0017] S5. Build a prediction module and represent the weighted features Through dense connection and batch normalization processing, and using two multi-layer perceptrons and linear layers, the predicted value of the installation quality of green wall materials is obtained. .

[0018] Preferably, in step S1, the green environmental protection wall material installation quality related data collected include physical parameter data and installation environment parameter data of the green environmental protection wall material, wherein the physical parameter data include green environmental protection wall material thickness data, green environmental protection wall material density data, green environmental protection wall material strength data, green environmental protection wall material thermal conductivity data, and the installation environment parameter data include installation environment temperature data, installation environment humidity data, and installation environment light condition data. The method deals with outliers in the data. The specific formula is as follows:

[0019] ;

[0020] In the formula, It is the first in the quality data of green and environmentally friendly wall materials installation The value of the data point, It is the mean value of the installation quality data of green and environmentally friendly wall materials. is the standard deviation.

[0021] Preferably, in step S2, the pre-processed green environmentally friendly wall material installation quality data is used The method is standardized, and the specific formula is as follows:

[0022] ;

[0023] In the formula, Installation quality data of green and environmentally friendly wall materials after pretreatment. It is the mean value of the installation quality data of green and environmentally friendly wall materials. is the standard deviation, and then the green and environmentally friendly wall material installation quality data is divided into training set and test set according to the proportion.

[0024] Preferably, in the steps S3 and S31, the physical parameter data of the green environmentally friendly wall material is input. , using variational mode decomposition, the physical parameter data of green environmentally friendly wall materials are gradually decomposed into several oscillation components IMF with specific frequency characteristics. The specific formula of each IMF form is as follows:

[0025] ;

[0026] In the formula, For the The first IMF component is the physical parameter data of the green wall material. The oscillatory components on the scale, is the amplitude varying with time, is the phase, representing the frequency modulated part of the signal, and an exponential function is applied to each IMF component Perform frequency shift so that the center frequency of the data is at the specified frequency The spectrum of the data is concentrated on a specific frequency band. The specific formula is as follows:

[0027] ;

[0028] ;

[0029] In the formula, For the The physical parameter data of green environmental protection wall materials IMF components, For the The physical parameter data of the green wall material corresponds to the The center frequency of the IMF component, For time The partial derivative on , is the unit pulse function, is an imaginary unit, is the convolution operation, is the number of IMF components, The physical parameter data of the original green environmentally friendly wall material. To define the constraints in the optimization problem, we have IMF components After that, all IMF components are combined into a matrix ,in is the number of physical parameters of green wall materials, is the number of time steps, and then the matrix is ​​transformed through the fully connected layer Convert to feature embedding matrix ,in is the dimension of the hidden state, that is, the size of the feature space after transformation.

[0030] Preferably, through variational mode decomposition, the physical parameters are decomposed into several oscillating components with specific frequency characteristics, and frequency shift is achieved through an exponential function to generate an embedded feature matrix. Variational mode decomposition and frequency shift processing enable the model to extract multi-frequency components in the data and enhance the ability to capture implicit features in time series. Feature decomposition and embedding improve the diversity and richness of feature expression, thereby enhancing the predictive ability of the model.

[0031] Preferably, the position association matrix is ​​input in steps S3 and S32 and the feature matrix ,in The number of locations where green wall materials have been installed. For the number of physical parameters at each installation location, the GraphSAGE algorithm is used to aggregate the adjacent node features of each node where green environmentally friendly wall materials are installed to generate an embedding. The aggregation function uses mean aggregation. The embedding formula of the layer is as follows

[0032] ;

[0033] In the formula, is the activation function ReLU, For the The weight matrix of the layer, is the aggregation function of adjacent node features, From the spatial incidence matrix Definition, which is the target location node where green wall materials are installed The adjacent node set, feature matrix As the initial input, the initial embedding is defined at layer 0, i.e. ,go through After the aggregation operation of the layers, the final embedding representation of each green wall material location node is obtained , the embedding combination of all green wall material position nodes forms a position embedding matrix ,in is the dimension of the hidden state, and then the position is embedded in the matrix Further transformation is performed through the fully connected layer to obtain the final position embedding matrix, ensuring that its dimension is consistent with other modules in the model. The specific formula is:

[0034] ;

[0035] In the formula, is the fully connected layer, Embed the matrix for the location, then enter the installation environment parameter data ,in is the number of time steps, is the characteristic number of environmental parameters, is the number of material installation locations. The fully connected layer is used to process the installation environment parameter data to obtain the environment embedding matrix. The specific formula is as follows:

[0036] ;

[0037] In the formula, is the activation function ReLU, For the The weight matrix of the layer, For the The bias vector of the layer, The installation environment parameter data to be entered.

[0038] Preferably, the GraphSAGE algorithm is used to perform feature aggregation on the installation location, and the installation environment parameters are embedded in the fully connected layer to generate a location and environment embedding matrix. The GraphSAGE algorithm captures the correlation between installation locations, and the fully connected layer extracts the impact of environmental factors on installation quality, comprehensively integrating spatial and environmental features. The embedding of location and environment significantly enhances the model's understanding of spatial and environmental factors, and improves the accuracy of predicting installation quality.

[0039] Preferably, in steps S3 and S33, the feature embedding matrix and the final position embedding matrix Add and fuse to get the feature position matrix. The specific formula is as follows:

[0040] ;

[0041] In the formula, is the feature embedding matrix, Embed the matrix for the final position, and then embed the final position into the matrix and the environment embedding matrix Perform addition fusion to obtain the environment position matrix. The specific formula is as follows:

[0042] ;

[0043] In the formula, is the final position embedding matrix, Environment embedding matrix.

[0044] Preferably, the feature embedding matrix is ​​fused with the position and environment embedding matrices to form a comprehensive feature matrix, which is used to characterize the comprehensive influencing factors of installation quality. The feature fusion operation effectively integrates multi-dimensional information to ensure that the model can fully consider the comprehensive effects of physical parameters, position associations and environmental conditions. The fused feature matrix further improves the integrity of feature expression and the prediction effect of the model.

[0045] Preferably, the steps S4 and S41 define For each green wall material installation unit, the hidden state in different quality feature dimensions is For the Layer target installation unit In Dimension The hidden state of is the dimension of the hidden state, For the number of all installed units, a dynamic multi-head feature position self-attention is proposed to calculate the activation vector and weight distribution vector. The specific formula is as follows:

[0046] ;

[0047] ;

[0048] In the formula, is the activation function, , is the learnable weight matrix, For the Layer target installation unit In Dimension The hidden state of , is a learnable bias vector, and then inputs the feature position matrix The attention score is calculated, and the activation vector and weight vector are applied simultaneously for dynamic head selection and dynamic weight allocation. The specific formula is as follows:

[0049] ;

[0050] ;

[0051] ;

[0052] In the formula, is the number of attention heads, For the The activation vector of the attention head, For the The weight distribution vector of the attention heads, For the The attention head Installation units and the target installation unit The attention score indicates the correlation between the two. For the Layer target installation unit In Dimension The hidden state of For the Nonlinear projection function in attention heads , applied to the hidden state , is the collection of all installed units, is the unnormalized attention score, indicating the The attention head Installation units and the target installation unit The size of the correlation between is the natural exponential function, , is a nonlinear projection function, For fusion operation, is the scaling factor, is the feature position matrix, and then outputs the feature position matrix ,in For the Target installation unit in layer In Dimension The hidden state of is the number of all installed units.

[0053] Preferably, the activation vector and weight distribution of the feature position matrix are calculated through a dynamic multi-head self-attention mechanism, and the attention head is dynamically selected and weights are assigned, which can more accurately capture the dependencies between feature positions, thereby improving the feature extraction capability of the model, significantly improving the modeling accuracy of the feature positions, and enhancing the model's ability to capture complex spatial features.

[0054] Preferably, in steps S4 and S42, dynamic multi-scale feature environment self-attention is proposed, and the environment position matrix is ​​input The multi-scale convolutional layer captures the features of different time scales and obtains the embedded representation of the environment position. The specific formula is as follows:

[0055] ;

[0056] In the formula, is the environment position matrix, is the convolution operation, and the convolution kernel size is , and then introduce the dynamic weight matrix , which is generated by the fully connected layer, the specific formula is as follows:

[0057] ;

[0058] In the formula, is the activation function ReLU, is the fully connected layer, is the environment position embedding representation, which is multiplied by the dynamic weight matrix and weighted summed to obtain the fusion feature representation. The specific formula is as follows:

[0059] ;

[0060] In the formula, is the dynamic weight matrix, is the environment position embedding representation, is the number of convolution kernels, and then the attention score is calculated. The specific formula is as follows:

[0061] ;

[0062] ;

[0063] In the formula, is the number of attention heads, For the Layer target installation unit In Dimension The hidden state of is the collection of all installed units, For the The attention head Installation units and the target installation unit The attention score, For the Nonlinear projection function in attention heads , applied to the hidden state , is the natural exponential function, , is a nonlinear projection function, For fusion operation, is the scaling factor, To fuse feature representation, the feature environment matrix is ​​output ,in For the Target installation unit in layer In Dimension The hidden state of is the number of all installed units.

[0064] Preferably, a multi-scale convolutional layer and a dynamic weight matrix are combined to capture the different time scale characteristics of the environmental location, and the feature fusion is optimized by weight adjustment, which can effectively capture the temporal characteristics of the environmental data. The dynamic weight matrix ensures the flexibility and accuracy of feature fusion, which improves the model's ability to understand temporal characteristics and enhances its adaptability to dynamic environmental changes.

[0065] Preferably, in steps S4 and S43, the feature position matrix and the feature environment matrix Input the fully connected layer and perform dynamic weighted fusion operation to obtain weighted feature representation. The specific formula is as follows:

[0066] ;

[0067] ;

[0068] In the formula, is the activation function ReLU, , is the learnable weight matrix, is the learnable bias term, is the feature position matrix, is the characteristic environment matrix, is \left [ {0, 1} \right ] The matrix between is used to control the weighted ratio. It is an element-wise multiplication operation.

[0069] Preferably, the feature position matrix and the environment matrix are weightedly fused through a fully connected layer to generate a weighted feature representation. The dynamic weighted fusion operation can integrate multi-dimensional information to ensure that the output feature representation is more accurate, which significantly improves the prediction effect and generalization ability of the model.

[0070] Preferably, in step S5, for the weighted feature representation Through dense connection and batch normalization, the output is obtained , and use two multi-layer perceptrons for linear transformation to obtain the processed feature representation. The specific formula is as follows:

[0071] ;

[0072] ;

[0073] In the formula, is the activation function, is the output of the first layer of multi-layer perceptron, is the weight matrix of the first layer of multi-layer perceptron, To prevent overfitting, is the weight matrix of the second layer of multilayer perceptron, , is the bias term, and finally a linear transformation is performed through the linear layer to obtain the predicted value of the installation quality of green environmentally friendly wall materials. The specific formula is as follows:

[0074] ;

[0075] In the formula, is a linear layer operation, is the feature representation after processing, Prediction value of installation quality of green and environmentally friendly wall materials.

[0076] Preferably, by processing weighted feature representation through dense connections and batch normalization, a multi-layer perceptron and a linear layer are used to generate the final prediction value, which can extract complex nonlinear relationships and generate prediction results at the same time, ensuring the prediction accuracy of the model, thereby generating high-precision prediction results.

[0077] In summary, due to the adoption of the technical solution, the beneficial effects of the present invention are as follows: the present invention proposes an Eqformer prediction model, which is applied to the green and environmentally friendly wall material installation quality prediction scenario, including a spatiotemporal feature extraction module, a multi-frequency feature fusion module and a prediction module. The spatiotemporal feature extraction module is used to process the correlation between the time dimension and the space dimension, the multi-frequency feature fusion module is used to process the correlation between the feature position dimension and the feature environment dimension, and the prediction module is used to integrate the multi-dimensional features and convert them into usable installation quality prediction results. The modules cooperate with each other to achieve accurate prediction of the installation quality of green and environmentally friendly wall materials. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] Figure 1 This is a step-by-step diagram for a method for predicting the installation quality of green and environmentally friendly wall materials.

[0079] Figure 2 This is the structure diagram of the Eqformer prediction model.

[0080] Figure 3 This is the structural diagram of the spatiotemporal feature extraction module.

[0081] Figure 4 This is the structure diagram of the multi-frequency feature fusion module.

[0082] Figure 5 The Eqformer prediction model realizes the fitting effect diagram of the green and environmentally friendly wall material installation quality prediction. DETAILED DESCRIPTION

[0083] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0084] See also Figure 1-Figure 5 The present invention provides a technical solution: a method for predicting the installation quality of green environmentally friendly wall materials, which can process the correlation between the time dimension and the space dimension by constructing a spatiotemporal feature extraction module, and can process the correlation between the feature position dimension and the feature environment dimension. The prediction module is used to integrate the multi-dimensional features and convert them into usable installation quality prediction results. The specific steps are as follows Figure 1 shown.

[0085] Construct the Eqformer prediction model, whose structure is as follows Figure 2 As shown, the specific steps are as follows:

[0086] S1. Collect data related to the installation quality of green and environmentally friendly wall materials, including features and target variables, and pre-process the collected data related to the installation quality of green and environmentally friendly wall materials.

[0087] Furthermore, the green environmental protection wall material installation quality related data are collected, including the physical parameter data and installation environment parameter data of the green environmental protection wall material, wherein the physical parameter data include the green environmental protection wall material thickness data, the green environmental protection wall material density data, the green environmental protection wall material strength data, and the green environmental protection wall material thermal conductivity data, and the installation environment parameter data include the installation environment temperature data, the installation environment humidity data, and the installation environment light condition data, and use The method deals with outliers in the data. The specific formula is as follows:

[0088] ;

[0089] In the formula, It is the first in the quality data of green and environmentally friendly wall materials installation The value of the data point, It is the mean value of the installation quality data of green and environmentally friendly wall materials. is the standard deviation.

[0090] S2. Use of installation quality data of pre-treated green environmentally friendly wall materials The method is standardized, and the green and environmentally friendly wall material installation quality data is divided into training set and test set.

[0091] Furthermore, the installation quality data of the pre-treated green environmentally friendly wall materials is used The method is standardized, and the specific formula is as follows:

[0092] ;

[0093] In the formula, Installation quality data of green and environmentally friendly wall materials after pretreatment. It is the mean value of the installation quality data of green and environmentally friendly wall materials. is the standard deviation, and then the green and environmentally friendly wall material installation quality data are divided into training set and test set in a ratio of 7:3.

[0094] S31. Input the physical parameter data of green environmentally friendly wall materials , decomposed into oscillatory components Apply an exponential function to each IMF component Perform frequency shift and combine all IMF components into a matrix , converted into a feature embedding matrix through a fully connected layer .

[0095] Furthermore, a spatiotemporal feature extraction module is constructed, and its structure is as follows: Figure 3 As shown, input the physical parameter data of green environmentally friendly wall materials , using variational mode decomposition, the physical parameter data of green environmentally friendly wall materials are gradually decomposed into several oscillation components IMF with specific frequency characteristics. The specific formula of each IMF form is as follows:

[0096] ;

[0097] In the formula, For the The first IMF component is the physical parameter data of the green wall material. The oscillatory components on the scale, is the amplitude varying with time, is the phase, representing the frequency modulated part of the signal, and an exponential function is applied to each IMF component Perform frequency shift so that the center frequency of the data is at the specified frequency The spectrum of the data is concentrated on a specific frequency band. The specific formula is as follows:

[0098] ;

[0099] ;

[0100] In the formula, For the The physical parameter data of green environmental protection wall materials IMF components, For the The physical parameter data of the green wall material corresponds to the The center frequency of the IMF component, For time The partial derivative on , is the unit pulse function, is an imaginary unit, is the convolution operation, is the number of IMF components, The physical parameter data of the original green environmentally friendly wall material. To define the constraints in the optimization problem, we have IMF components After that, all IMF components are combined into a matrix ,in is the number of physical parameters of green wall materials, is the number of time steps, and then the matrix is ​​transformed through the fully connected layer Convert to feature embedding matrix ,in is the dimension of the hidden state, that is, the size of the feature space after transformation.

[0101] S32, input position association matrix and the feature matrix , generated by aggregation through the GraphSAGE algorithm and combined into a position embedding matrix , use the fully connected layer to transform and get the final position embedding matrix , then enter the installation environment parameter data , processed using a fully connected layer, to obtain the environment embedding matrix .

[0102] Furthermore, the input position association matrix and the feature matrix ,in The number of locations where green wall materials have been installed. For the number of physical parameters at each installation location, the GraphSAGE algorithm is used to aggregate the adjacent node features of each node where green environmentally friendly wall materials are installed to generate an embedding. The aggregation function uses mean aggregation. The embedding formula of the layer is as follows

[0103] ;

[0104] In the formula, is the activation function ReLU, For the The weight matrix of the layer, is the aggregation function of adjacent node features, From the spatial incidence matrix Definition, which is the target location node where green wall materials are installed The adjacent node set, feature matrix As the initial input, the initial embedding is defined at layer 0, i.e. ,go through After the aggregation operation of the layers, the final embedding representation of each green wall material location node is obtained , the embedding combination of all green wall material position nodes forms a position embedding matrix ,in is the dimension of the hidden state, and then the position is embedded in the matrix Further transformation is performed through the fully connected layer to obtain the final position embedding matrix, ensuring that its dimension is consistent with other modules in the model. The specific formula is:

[0105] ;

[0106] In the formula, is the fully connected layer, Embed the matrix for the location, then enter the installation environment parameter data ,in is the number of time steps, is the characteristic number of environmental parameters, is the number of material installation locations. The fully connected layer is used to process the installation environment parameter data to obtain the environment embedding matrix. The specific formula is as follows:

[0107] ;

[0108] In the formula, is the activation function ReLU, For the The weight matrix of the layer, For the The bias vector of the layer, The installation environment parameter data to be entered.

[0109] S33. Embedding features into matrix , the final position embedding matrix Add and fuse to get the feature position matrix , and embed the final position into the matrix , environment embedding matrix Add and fuse to get the environment position matrix .

[0110] Furthermore, the feature embedding matrix and the final position embedding matrix Add and fuse to get the feature position matrix. The specific formula is as follows:

[0111] ;

[0112] In the formula, is the feature embedding matrix, Embed the matrix for the final position, and then embed the final position into the matrix and the environment embedding matrix Perform addition fusion to obtain the environment position matrix. The specific formula is as follows:

[0113] ;

[0114] In the formula, is the final position embedding matrix, Environment embedding matrix.

[0115] S41. Propose dynamic multi-head feature position self-attention and calculate activation vector and the weight distribution vector , input feature position matrix Calculate the attention score, perform dynamic head selection and dynamic weight allocation, and obtain the feature position matrix .

[0116] Furthermore, a multi-frequency feature fusion module is constructed, and its structure is as follows: Figure 4 As shown, the definition For each green wall material installation unit, the hidden state in different quality feature dimensions is For the Layer target installation unit In Dimension The hidden state of is the dimension of the hidden state, For the number of all installed units, a dynamic multi-head feature position self-attention is proposed to calculate the activation vector and weight distribution vector. The specific formula is as follows:

[0117] ;

[0118] ;

[0119] In the formula, is the activation function, , is the learnable weight matrix, For the Layer target installation unit In Dimension The hidden state of , is a learnable bias vector, and then inputs the feature position matrix The attention score is calculated, and the activation vector and weight vector are applied simultaneously for dynamic head selection and dynamic weight allocation. The specific formula is as follows:

[0120] ;

[0121] ;

[0122] ;

[0123] In the formula, is the number of attention heads, For the The activation vector of the attention head, For the The weight distribution vector of the attention heads, For the The attention head Installation units and the target installation unit The attention score indicates the correlation between the two. For the Layer target installation unit In Dimension The hidden state of For the Nonlinear projection function in attention heads , applied to the hidden state , is the collection of all installed units, is the unnormalized attention score, indicating the The attention head Installation units and the target installation unit The size of the correlation between is the natural exponential function, , is a nonlinear projection function, For fusion operation, is the scaling factor, is the feature position matrix, and then outputs the feature position matrix ,in For the Target installation unit in layer In Dimension The hidden state of is the number of all installed units.

[0124] S42, propose dynamic multi-scale feature environment self-attention, input environment position matrix Through multi-scale convolutional layers, we get the embedded representation of the environment position , introduce the dynamic weight matrix , multiply the environment position embedding representation with the dynamic weight matrix to obtain the fused feature representation And calculate the attention score to obtain the feature environment matrix .

[0125] Furthermore, a dynamic multi-scale feature environment self-attention is proposed, and the environment position matrix is ​​input The features of different time scales are captured by multi-scale convolutional layers to obtain the embedded representation of the environment position. The specific formula is as follows:

[0126] ;

[0127] In the formula, is the environment position matrix, is the convolution operation, and the convolution kernel size is , and then introduce the dynamic weight matrix , which is generated by the fully connected layer, the specific formula is as follows:

[0128] ;

[0129] In the formula, is the activation function ReLU, is the fully connected layer, is the environment position embedding representation, which is multiplied by the dynamic weight matrix and weighted summed to obtain the fusion feature representation. The specific formula is as follows:

[0130] ;

[0131] In the formula, is the dynamic weight matrix, is the environment position embedding representation, is the number of convolution kernels, and then the attention score is calculated. The specific formula is as follows:

[0132] ;

[0133] ;

[0134] In the formula, is the number of attention heads, For the Layer target installation unit In Dimension The hidden state of is the collection of all installed units, For the The attention head Installation units and the target installation unit The attention score, For the Nonlinear projection function in attention heads , applied to the hidden state , is the natural exponential function, , is a nonlinear projection function, For fusion operation, is the scaling factor, To fuse feature representation, the feature environment matrix is ​​output ,in For the Target installation unit in layer In Dimension The hidden state of is the number of all installed units.

[0135] S43, feature position matrix and the feature environment matrix Input the fully connected layer for dynamic weighted fusion operation to obtain weighted feature representation .

[0136] Furthermore, the feature position matrix and the feature environment matrix Input the fully connected layer and perform dynamic weighted fusion operation to obtain weighted feature representation. The specific formula is as follows:

[0137] ;

[0138] ;

[0139] In the formula, is the activation function ReLU, , is the learnable weight matrix, is the learnable bias term, is the feature position matrix, is the characteristic environment matrix, is \left [ {0, 1} \right ] The matrix between is used to control the weighted ratio. It is an element-wise multiplication operation.

[0140] S5. Build a prediction module and represent the weighted features Through dense connection and batch normalization processing, and using two multi-layer perceptrons and linear layers, the predicted value of the installation quality of green wall materials is obtained. .

[0141] Furthermore, for the weighted feature representation Through dense connection and batch normalization, the output is obtained , and use two multi-layer perceptrons for linear transformation to obtain the processed feature representation. The specific formula is as follows:

[0142] ;

[0143] ;

[0144] In the formula, is the activation function, is the output of the first layer of multi-layer perceptron, is the weight matrix of the first layer of multi-layer perceptron, To prevent overfitting, is the weight matrix of the second layer of multilayer perceptron, , is the bias term, and finally a linear transformation is performed through the linear layer to obtain the predicted value of the installation quality of green environmentally friendly wall materials. The specific formula is as follows:

[0145] ;

[0146] In the formula, is a linear layer operation, is the feature representation after processing, Prediction value of installation quality of green and environmentally friendly wall materials.

[0147] Furthermore, the Eqformer prediction model is written in Python, the experiment is run on Windows operating system, Pytorch is used as the framework in CUDA11.27 environment, and the training is performed on GeForce RTX 3090. The optimizer is , the initial learning rate is set to 0.001, the training batch is set to 64, the training cycle is set to 100, and the data set is 60 days of green and environmentally friendly wall material installation quality related data, which is input into the Eqformer prediction model after preprocessing.

[0148] Furthermore, the Eqformer prediction model realizes the prediction of the installation quality of green wall materials. Figure 5As shown in the figure, the horizontal axis is the installation cycle of green environmentally friendly wall materials (days), the vertical axis is the qualified rate of green environmentally friendly wall materials installation (%), the gray dotted line and dots are real data, and the black solid line and squares are predicted data. It can be seen from the figure that the predicted curve and the real curve have a very close trend, especially in the early stage of the installation cycle, the two curves basically coincide, indicating that the model performs well in dealing with time characteristics within a short period. With the increase of the installation cycle, the qualified rate of green environmentally friendly wall materials installation shows a downward trend. This phenomenon is that the long installation period of green environmentally friendly wall materials may cause high-frequency oscillation components (such as dynamic changes in material properties) to gradually show negative effects on the installation quality. In addition, environmental factors may have more complex nonlinear effects on quality in the later stage of the installation cycle. In summary, the overall prediction effect of the model is good, especially in the short period. It shows a high fitting accuracy. The prediction model can better cope with complex long-term quality change characteristics and has good prediction accuracy.

Claims

1. A method for predicting the installation quality of green and environmentally friendly wall materials, characterized in that: The following steps are involved: S1. Collect data related to the installation quality of green environmentally friendly wall materials, including features and target variables, and pre-process the collected data related to the installation quality of green environmentally friendly wall materials; S2. Standardize the preprocessed green environmental protection wall material installation quality data using the Z-Score method, and divide the green environmental protection wall material installation quality data into a training set and a test set; S3. Construct a spatiotemporal feature extraction module to process the correlation between the time dimension and the space dimension. The specific steps are as follows: S31. Input the physical parameter data of green environmentally friendly wall materials v , decomposed into oscillatory components u i (t) Apply an exponential function to each IMF component Perform frequency shift and combine all IMF components into matrix e IMF , converted into a feature embedding matrix e through a fully connected layer T ; S32, input the position association matrix A and the feature matrix X, aggregate and generate embedding through the GraphSAGE algorithm, and combine them into the position embedding matrix e S , use the fully connected layer to transform and get the final position embedding matrix e Smlp , then input the installation environment parameter data I, use the fully connected layer to process it, and get the environment embedding matrix e W ; S33, embed the features into matrix e T , the final position embedding matrix e Smlp Add and fuse to get the feature position matrix e ST , and embed the final position into the matrix e Smlp , environment embedding matrix e W Add and fuse to get the environment position matrix e WS ; S4. Construct a multi-frequency feature fusion module to process the correlation between the feature position dimension and the feature environment dimension. The specific steps are as follows: S41. Propose dynamic multi-head feature position self-attention, calculate the activation vector g and weight distribution vector w, and input the feature position matrix e ST Calculate the attention score, perform dynamic head selection and dynamic weight allocation, and obtain the feature position matrix S42, propose dynamic multi-scale feature environment self-attention, input environment position matrix e WS Through multi-scale convolutional layers, we get the embedded representation of the environment position Introducing dynamic weight matrix Multiply the environment position embedding representation with the dynamic weight matrix to obtain the fused feature representation And calculate the attention score to obtain the feature environment matrix S43, feature position matrix and the feature environment matrix Input the fully connected layer for dynamic weighted fusion operation to obtain weighted feature representation S5. Build a prediction module and represent the weighted features Through dense connection and batch normalization processing, and using two multi-layer perceptrons and linear layers, the predicted value Y of the installation quality of green and environmentally friendly wall materials is obtained.

2. A method for predicting the installation quality of green and environmentally friendly wall materials according to claim 1, characterized in that: In step S31, the physical parameter data f of the green environmentally friendly wall material is input. v , using variational mode decomposition, the physical parameter data of green environmentally friendly wall materials are gradually decomposed into several oscillation components IMF with specific frequency characteristics. The specific formula of each IMF form is as follows: u i (t)=A i (t)cos(φ i (t)); In the formula, u i (t) is the i-th IMF component, which is the oscillating component at the i-th scale in the physical parameter data of the green environmental protection wall material. i (t) is the amplitude that changes with time, φ i (t) is the phase, which represents the frequency modulation part of the signal. An exponential function is applied to each IMF component. Perform frequency shift so that the center frequency of the data is at the specified frequency w v，i The spectrum of the data is concentrated on a specific frequency band. The specific formula is as follows: In the formula, u v，i is the ith IMF component of the physical parameter data of the vth green environmentally friendly wall material, w v，i is the center frequency of the ith IMF component corresponding to the physical parameter data of the Wth green environmentally friendly wall material, is the partial derivative at time t, δ(t) is the unit impulse function, j is the imaginary unit, * is the convolution operation, M is the number of IMF components, and f v is the physical parameter data of the original green environmentally friendly wall material, st is the constraint condition in defining the optimization problem, and M IMF components {u v，i }={u v，1 ,u v，2 ,…,u v，M }, all IMF components are combined into the matrix e IMF ={{u 1，i },{u 2，i }, …, {u N，i }}∈R P×N×M , where N is the number of physical parameters of green wall materials, P is the number of time steps, and then the matrix e is transformed through the fully connected layer IMF Converted to feature embedding matrix e T ∈R D×N×M , where D is the dimension of the hidden state, that is, the size of the feature space after transformation.

3. A method for predicting the installation quality of green and environmentally friendly wall materials according to claim 2, characterized in that: In step S32, the position correlation matrix is ​​input and the feature matrix Where N * is the number of locations where green wall materials are installed, F is the number of physical parameters of each installation location, and the GraphSAGE algorithm is used to aggregate the adjacent node features of each node where green wall materials are installed to generate embedding. The aggregation function uses mean aggregation. The embedding formula of the lth layer is as follows Where σ is the activation function ReLU, W (l) is the weight matrix of the lth layer, AGGREGATE is the aggregation function of the adjacent node features, It is defined by the spatial association matrix A, which is the set of adjacent nodes of the target location node i where the green wall material is installed. The feature matrix X is used as the initial input to define the initial embedding at the 0th layer, that is, After L layers of aggregation operations, the final embedding representation of each green wall material location node is obtained. The embedding combination of all green wall material position nodes forms a position embedding matrix Where D is the dimension of the hidden state, and then the position is embedded in the matrix e S Further transformation is performed through the fully connected layer to obtain the final position embedding matrix, ensuring that its dimension is consistent with other modules in the model. The specific formula is: e Smlp =MLP(e S ); In the formula, MLP is the fully connected layer, e S is the position embedding matrix, and then inputs the installation environment parameter data I∈R P×n×C , where P is the number of time steps, C is the number of features of environmental parameters, and n is the number of material installation locations. The installation environment parameter data is processed using a fully connected layer to obtain the environment embedding matrix. The specific formula is as follows: e W =s l (s l-1 (…σ1(I·W1+b1)·W2+b2)…W l-1 +b l-1 )·W l +b l ); Where σ is the activation function ReLU, W l is the weight matrix of the lth layer, b l is the bias vector of the lth layer, and I is the input installation environment parameter data.

4. A method for predicting the installation quality of green and environmentally friendly wall materials according to claim 3, characterized in that: In step S33, the feature embedding matrix e T and the final position embedding matrix e Smlp Add and fuse to get the feature position matrix. The specific formula is as follows: And ST =(and Smlp +e T ); In the formula, e T is the feature embedding matrix, e Smlp The final position is embedded in the matrix, and then the final position is embedded in the matrix e Smlp and the environment embedding matrix e W Perform addition fusion to obtain the environment position matrix. The specific formula is as follows: And WS =(and Smlp +e W ); In the formula, e Smlp is the final position embedding matrix, e W Environment embedding matrix.

5. A method for predicting the installation quality of green and environmentally friendly wall materials according to claim 4, characterized in that: The step S41 defines For each green wall material installation unit, the hidden state in different quality feature dimensions is is the hidden state of the target installation unit v in the lth layer in dimension d, where D is the dimension of the hidden state, N * For the number of all installed units, a dynamic multi-head feature position self-attention is proposed to calculate the activation vector and weight distribution vector. The specific formula is as follows: In the formula, Sigmoid is the activation function, W g , W w is the learnable weight matrix, The hidden state of unit v in dimension d for the l-th layer target installation, b g , b w is a learnable bias vector, and then inputs the feature position matrix e ST The attention score is calculated, and the activation vector and weight vector are applied simultaneously for dynamic head selection and dynamic weight allocation. The specific formula is as follows: Where K is the number of attention heads, g k is the activation vector of the kth attention head, W k is the weight assignment vector for the kth attention head, is the i-th installation unit v in the k-th attention head i The attention score between the target installation unit v indicates the correlation between the two. The hidden state of unit v in dimension d is installed for the l-th layer target, is the nonlinear projection function in the kth attention head Applied to hidden state V is the set of all installed units, is the unnormalized attention score, representing the i-th installation unit v in the k-th attention head i The size of the correlation between the target installation unit v, exp is the natural exponential function, is a nonlinear projection function, || is a fusion operation, is the scaling factor, e ST is the feature position matrix, and then outputs the feature position matrix in is the hidden state of the target installation unit v in dimension d in the lth layer, N * is the number of all installed units.

6. A method for predicting the installation quality of green and environmentally friendly wall materials according to claim 5, characterized in that: In step S42, a dynamic multi-scale feature environment self-attention is proposed, and the environment position matrix e is input. WS The multi-scale convolutional layer captures the features of different time scales and obtains the embedded representation of the environment position. The specific formula is as follows: In the formula, e WS is the environment position matrix, is the convolution operation, and the convolution kernel size is κ i , then introduce the dynamic weight matrix W(κ i ), which is generated by the fully connected layer, the specific formula is as follows: In the formula, σ is the activation function ReLU, MLP is the fully connected layer, is the environment position embedding representation, which is multiplied by the dynamic weight matrix and weighted summed to obtain the fusion feature representation. The specific formula is as follows: In the formula, is the dynamic weight matrix, is the embedded representation of the environment position, m is the number of convolution kernels, and then the attention score is calculated. The specific formula is as follows: Where K is the number of attention heads, is the hidden state of the target installation unit v in the lth layer in dimension d, V is the set of all installation units, is the i-th installation unit v in the k-th attention head i The attention score between the target installation unit v, is the nonlinear projection function in the kth attention head Applied to hidden state exp is the natural exponential function, is a nonlinear projection function, || is a fusion operation, is the scaling factor, To fuse feature representation, the feature environment matrix is ​​output in is the hidden state of the target installation unit v in dimension d in the lth layer, N * is the number of all installed units.

7. A method for predicting the installation quality of green and environmentally friendly wall materials according to claim 6, characterized in that: In step S43, the feature position matrix and the feature environment matrix Input the fully connected layer and perform dynamic weighted fusion operation to obtain weighted feature representation. The specific formula is as follows: In the formula, σ is the activation function ReLU, B ρ，1 , B ρ，2 is the learnable weight matrix, b ρ is the learnable bias term, is the feature position matrix, is the characteristic environment matrix, ρ is a matrix between [0, 1], which is used to control the weighted ratio, and ⊙ is the element-by-element multiplication operation.

8. A method for predicting the installation quality of green and environmentally friendly wall materials according to claim 6, characterized in that: In step S5, for the weighted feature representation Through dense connection and batch normalization, the output is obtained Two multi-layer perceptrons are used for linear transformation to obtain the processed feature representation. The specific formula is as follows: In the formula, ReLU is the activation function, is the output of the first-layer multilayer perceptron, W1 is the weight matrix of the first-layer multilayer perceptron, dropout is an operation to prevent overfitting, W2 is the weight matrix of the second-layer multilayer perceptron, b1 and b2 are bias terms, and finally a linear transformation is performed through the linear layer to obtain the predicted value of the installation quality of green environmentally friendly wall materials. The specific formula is as follows: Where Linear is the linear layer operation, is the processed feature representation, and Y is the predicted value of the installation quality of green and environmentally friendly wall materials.

9. A method for predicting the installation quality of green and environmentally friendly wall materials according to claim 1, characterized in that: In order to predict the installation quality of green and environmentally friendly wall materials, the collected data related to the installation quality of green and environmentally friendly wall materials include the physical parameter data and installation environment parameter data of green and environmentally friendly wall materials, among which the physical parameter data include the thickness data of green and environmentally friendly wall materials, the density data of green and environmentally friendly wall materials, the strength data of green and environmentally friendly wall materials, and the thermal conductivity data of green and environmentally friendly wall materials; the installation environment parameter data include the installation environment temperature data, the installation environment humidity data, and the installation environment lighting condition data. The collected relevant data are preprocessed to ensure that there are no outliers in the data, and then the processed data are divided into training set and test set for training and evaluating the performance of the green and environmentally friendly wall material installation quality prediction model.

Images