Construction process carbon emission prediction method for intelligent manufacturing

Abstract

The application discloses a construction process carbon emission prediction method for intelligent manufacturing, relates to the technical field of carbon emission data processing, and comprises the following steps: collecting carbon emission data, adopting a self-encoder based on sparse features to perform feature dimension reduction, and adopting a neural network based on a quantum superposition state as a feature extraction model, wherein the self-encoder based on sparse features is used to reduce and compress high-dimensional carbon emission features, the construction data is subjected to nonlinear dimension reduction through sparse regularization, feature redundancy is reduced, and the ability to capture key features is enhanced, the neural network model based on the quantum superposition state is adopted to model complex nonlinear carbon emission data, and the global optimization ability of parameter initialization is realized through quantum state superposition, so that the adaptability of the model to a complex nonlinear data environment is enhanced.

Description

Technical Field

[0001] This invention relates to the field of carbon emission data processing technology, and in particular to a method for predicting carbon emissions during construction processes for intelligent manufacturing. Background Technology

[0002] Currently, the construction industry faces challenges such as high carbon emissions and enormous energy consumption in achieving sustainable development goals. Carbon emissions during construction originate from diverse sources, including equipment energy consumption, material consumption, and environmental factors, and these factors exhibit complex nonlinear relationships. To achieve low-carbon construction, accurate prediction and analysis of carbon emissions at each stage of construction are essential.

[0003] In the prior art, Chinese invention patent with publication number CN119250300A proposes a method, device, and electronic device for predicting carbon emissions from buildings in a university area. This method includes: acquiring historical carbon emission data for each building within the university area; initially classifying each building according to its function; filtering buildings in each type based on historical carbon emission data, eliminating buildings with poor carbon emission correlation; establishing a regional carbon emission model corresponding to each type, and predicting the total regional carbon emissions of buildings in each type to obtain predicted carbon emission values ​​for various types of buildings in the university area. Chinese invention patent with publication number CN119250569A proposes a method for predicting carbon emissions from highway bridges based on random forests, including the following steps: constructing a carbon emission dataset; preprocessing the dataset; performing carbon emission prediction feature variable analysis on the collected data through correlation analysis and grey relational analysis, selecting feature variables as input to the bridge carbon emission prediction model; and processing the dataset. The method and system for predicting carbon emissions in the construction industry based on the random forest algorithm are proposed in Chinese invention patent CN119250845A. The method involves: acquiring historical carbon emission data from the construction industry; ranking the importance of influencing factors in the historical carbon emission data based on the random forest regression model, extracting the feature dataset related to carbon emissions, and dividing it into training and testing sets; inputting the training set into a pre-set LSTM model for computation, evaluating the LSTM prediction results using the testing set, and outputting the fully trained LSTM model; combining the random forest regression model with the fully trained LSTM model to predict the carbon emissions of existing buildings.

[0004] The existing technologies have the following drawbacks: 1. Regarding the dimensionality reduction and compression of high-dimensional carbon emission features, existing methods cannot effectively distinguish between key features and redundant features when processing multimodal high-dimensional data during construction, which easily leads to noise interference and reduces the efficiency and accuracy of the prediction model; 2. Regarding the modeling of complex nonlinear carbon emission data, traditional modeling methods lack the ability to adapt to complex nonlinear relationships, are prone to getting trapped in local optima, and the models are difficult to show good generalization ability in nonlinear feature extraction and classification tasks; 3. Regarding the adaptation and updating of the prediction model, the parameters of existing models are fixed and cannot adapt to the dynamic changes of equipment, processes and environment during construction, resulting in a decrease in prediction accuracy over time, which is difficult to meet actual needs. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide a method for predicting carbon emissions during construction processes for intelligent manufacturing.

[0006] The technical solution adopted to solve the above-mentioned technical problems is: a carbon emission prediction method for construction processes in intelligent manufacturing, including the following steps: S1, collecting carbon emission data and manually labeling the collected carbon emission data, with labeling types including low carbon emission, medium carbon emission, and high carbon emission; S2, using an autoencoder based on sparse features for feature dimensionality reduction to improve the effective representation and compression capabilities of features. By applying sparse constraints to the feature vectors of the autoencoder, the autoencoder model can mine key features and effectively suppress redundant information; S3, using a neural network based on quantum superposition as a feature extraction model to extract features from the carbon emission data. During the training process, a gradient search mechanism under quantum superposition is used to improve optimization efficiency and generalization ability, enhancing the neural network's ability to process complex nonlinear carbon emission data; S4, automatically training, validating, and updating the prediction model to ensure that the model can adapt to the dynamic changes in the construction process; S5, predicting the carbon emission situation of each stage of the construction process in real time based on the trained model, providing a basis for optimizing and adjusting the construction process.

[0007] Furthermore, the carbon emission data in S1 includes equipment operating time Da, material usage Ca, ambient temperature Ma, ambient humidity Ka, dust concentration Pa, noise level Qa, vibration index Sa, overtime operation Ua, and energy-saving mode duration Ha.

[0008] Furthermore, the training process of the sparse feature-based autoencoder in S2 includes the following steps:

[0009] S201, initialize the parameters of the autoencoder, including the weights and biases. The weight initialization method for the autoencoder is expressed as follows:

[0010]

[0011] In the formula, Here, represents the weights of the l-th layer of the autoencoder, and randn() is a function that generates random numbers following a normal distribution. This represents the number of nodes in the (l-1)th layer of the autoencoder. Let l be the number of nodes in the l-th layer of the autoencoder, where l is the layer index of the autoencoder.

[0012] The bias initialization method for an autoencoder is expressed as follows:

[0013]

[0014] In the formula, Here, represents the bias of the l-th layer of the autoencoder, and zeros() is a function that generates a vector containing all zeros.

[0015] S202, based on the preliminary analysis results of carbon emission data, the learning rate parameter of the autoencoder is dynamically configured, as follows:

[0016]

[0017] In the formula, η r (t) represents the learning rate of the autoencoder in the t-th iteration, η r,init δ represents the initial learning rate of the autoencoder. r The decay rate of the autoencoder is denoted by t, and int(t) is the current iteration number.

[0018] S203 involves sparse encoding learning of carbon emission data. L1 regularization is used to achieve sparse feature representation, improving the model's ability to capture key features. When processing dimensionality reduction of carbon emission data, the sparse feature autoencoder captures key features through L1 regularization constraints, as shown below:

[0019]

[0020] In the formula, J r (W r ,b r ) represents the total loss function of the autoencoder, m r The number of training samples input to the autoencoder is given by i, where i is the index of the input sample to the autoencoder, and x is the index of the input sample. r,i For the i-th input sample of the autoencoder, The reconstructed output of the autoencoder, ∥∥ is the L2 norm, ∥∥1 is the L1 norm, λ r h is the sparse regularization coefficient. r () is the encoding function of the autoencoder, h r (x r,i ) represents the encoded output of the autoencoder layer for the i-th sample;

[0021] To enhance the sparsity of feature encoding in the autoencoder, the L1 regularization term is calculated as follows:

[0022]

[0023] In the formula, n r α is the number of nodes in the encoding layer, j is the index of the first autoencoder neuron, and α is the number of nodes in the encoding layer. r β is the first dynamic adjustment factor of the autoencoder. r h is the second dynamic adjustment factor of the autoencoder. r,j (x r,i ) represents the encoded output of the j-th neuron of the autoencoder for the i-th sample;

[0024] S204 uses dynamic tensor coupling decomposition to reduce the dimensionality and reconstruct the hidden layer features, eliminating noise and redundant information. The calculation method is expressed as follows:

[0025] H r =fold(tdc(unfold(h) r (X r )),[R1,R2,R3]))

[0026] In the formula, H r The feature representation is after dimensionality reduction. fold() represents the tensor folding operation, h r (X r ) represents the encoded output of the autoencoder layer on the sample, tdc() represents the Tucker decomposition operation, unfold() represents the tensor expansion operation, and X r The input data for the autoencoder layer is [R1,R2,R3], where R1 is the first kernel tensor dimension, R2 is the second kernel tensor dimension, and R3 is the third kernel tensor dimension.

[0027] The calculation method of Tucker decomposition is expressed as follows:

[0028]

[0029] In the formula, G r For the core tensor, The first mode matrix, This is the second mode matrix. The third mode matrix is ​​obtained by singular value decomposition, where the first, second, and third mode matrices are obtained through singular value decomposition. ×1 represents the first step of matrix multiplication, ×2 represents the second step of matrix multiplication, and ×3 represents the third step of matrix multiplication.

[0030] S205 employs a dynamic adjustment layer in the autoencoder, dynamically selecting and adjusting the activation function and model parameters based on the characteristics of carbon emission data. This addresses carbon emission data distributions of varying complexity and prevents model overfitting, as shown below:

[0031] σ r (x rdc ) = c r tanh(α r x rdc )+(1-c r ReLU(α) r x rdc )

[0032] In the formula, σ r () is the activation function of the dynamic adjustment layer, x rdc c is the input to the dynamic adjustment layer. r α is the adjustment coefficient for the activation function ratio under the current data distribution, tanh() is the hyperbolic tangent function, and α r ReLU() is the scaling factor for the dynamically adjusted layer, and ReLU() is the ReLU activation function.

[0033] This coefficient is automatically adjusted based on the average activity value of carbon emission data. The adjustment coefficient is set as follows:

[0034] c r =Sig(γ) r mean(abs(h r (X r ))))

[0035]

[0036] In the formula, Sig() is the Sigmoid activation function, and γ r This is a dynamic adjustment factor used to ensure the stability of dynamic adjustment when the ratio of data variance to standard deviation changes significantly. mean() is the function to calculate the mean, abs() is the function to take the absolute value, std() is the function to calculate the standard deviation, and var() is the function to calculate the variance.

[0037] S206 eliminates redundant features, simplifies the model, improves inference speed, and avoids the situation where there are still interrelated features after dimensionality reduction. Correlation analysis is performed on the dimensionality-reduced features to remove redundant features. The calculation method is expressed as follows:

[0038]

[0039] In the formula, For the i-th a The dimensionality reduction feature and the j-th feature a The correlation coefficients between the reduced-dimensional features, and cov() performs covariance calculations. Features standard deviation Features standard deviation For the i-th a A dimensionality reduction feature, For the j-th a A dimensionality reduction feature, i a j is the index of the first dimensionality reduction feature. a The index for the second dimensionality reduction feature;

[0040] S207 employs an adaptive feature weight adjustment method, which combines local and global statistical information during dimensionality reduction to dynamically weight the correlation between features and the target variable. This increases the focus on key features and suppresses irrelevant or noisy features. The calculation method is expressed as follows:

[0041]

[0042] In the formula, For the i-th a The mutual information between each feature and the target variable. is the target variable, representing the optimal value of the eigenvalues ​​after dimensionality reduction, and p() is the probability density function;

[0043] The weight of each feature in the model is dynamically adjusted to improve the model's sensitivity to key features. The feature weights are calculated as follows:

[0044]

[0045] In the formula, For the i-th dimension after dimensionality reduction a The weights of each feature, where exp() is an exponential function. For the i-th a The mutual information between each feature and the target variable. For the j-th a The mutual information between each feature and the target variable, λ r Hyperparameters used to adjust the degree of influence of mutual information;

[0046] Feature weights are embedded into the autoencoder update process. By calculating the parameter increments of the autoencoder, adaptive adjustment of feature weights is achieved during autoencoder training. The entire network is fine-tuned to optimize reconstruction error and feature retention. The calculation method is expressed as follows:

[0047]

[0048] In the formula, η represents the update amount of the weights in the l-th layer of the autoencoder. r(t) represents the learning rate of the autoencoder in the t-th iteration, where t is the index of the iteration number. J is the sign of the partial derivative. r The loss function of the autoencoder;

[0049] The update amount of the autoencoder bias is calculated using the same update amount calculation method as the autoencoder weight. The autoencoder weight and bias are updated based on the update amount of the autoencoder weight and bias. The update method is to add the autoencoder weight and bias of the current iteration to the update amount, which is used as the autoencoder weight and bias of the next iteration.

[0050] S208 employs a dynamic energy dissipation control method, enabling each neuron in the autoencoder to flexibly adjust its energy dissipation rate based on the complexity of the input carbon emission data during feature extraction and dimensionality reduction. The energy dissipation rate is calculated as follows:

[0051]

[0052] In the formula, The energy dissipation rate of the l-th layer of the self-encoder. This represents the number of nodes in the (l-1)th layer of the autoencoder. For the weights of the l-th layer of the autoencoder, and the weights of the i-th layer... b The element corresponding to each input, i b Here is the index of the second autoencoder neuron, and Sig() is the sigmoid activation function. For the i-th b One dimensionality reduction feature;

[0053] To dynamically control the information propagation speed during autoencoder training and to measure the activation threshold within network layers, the threshold constant is defined as follows:

[0054]

[0055] In the formula, γ is the information propagation threshold constant of the l-th layer of the autoencoder. rec var() is the threshold adjustment factor for the autoencoder, and var() is the variance calculation function;

[0056] The feature output of the autoencoder is updated based on a threshold constant, and the calculation method is expressed as follows:

[0057]

[0058] In the formula, This is the output of the l-th layer of the autoencoder. This is the output of the (l-1)th layer of the self-encoder;

[0059] S209. Repeat the above steps until the preset stopping iteration condition is met, which means that the model training is complete.

[0060] Furthermore, the training process of the neural network based on quantum superposition states in S3 includes the following steps:

[0061] S301 uses the quantum state superposition method to randomly initialize the parameters of the neural network model. The nondeterminism of quantum state superposition enhances the network's ability to explore unknown spaces, improving the model's adaptability and convergence speed in complex nonlinear carbon emission data environments. This is expressed as:

[0062]

[0063] In the formula, These are the initial weights of the neural network. For the i-th c The quantum probability amplitude of a quantum ground state. For the i-th c A quantum ground state, i c np is the index of the quantum ground state, and np is the total number of quantum ground states;

[0064] S302 employs dynamic information bottleneck regularization in each iteration to constrain the complexity of the information flow. This allows the model to focus more on features that are decisive for the final output during the learning process, reducing interference from redundant information. This results in more discriminative features in complex nonlinear carbon emission data analysis, improving the model's robustness and generalization ability. A dynamic regularization method based on differential privacy is used to balance carbon emission data privacy protection and model performance when handling sensitive data. Dynamic information bottleneck regularization limits the mutual information of the hidden layers, making the model focus more on the decisive role of key features in the emission results while reducing dependence on redundant features. This can be expressed as:

[0065] L p =L0(W p ,X p )+β p ·I(Y p Z p |X p )+λ pe ·Lap(μ pe ,b pe )

[0066] In the formula, L p Let W be the total loss function of the neural network, L0 be the basic loss function of the neural network, and W be the total loss function of the neural network. p X represents the weights of the neural network. p β is the carbon emission data input to the neural network. p Y is the regularization parameter.p Z is the label for the carbon emission data input to the neural network. p For the feature representation of the intermediate layer of the neural network, I(Y) p Z p |X p ) represents the state of a given input X p Output Y under the condition p With hidden representation Z p Mutual information between them, I() is the mutual information function, λ pe Lap(μ) is the privacy protection strength coefficient. pe ,b pe Characterized by μ pe Centered on, b pe The Laplace noise distribution at scale μ pe Let b be the mean of the current neural network input feature vector. pe The Laplace noise scale;

[0067] The regularization coefficient is dynamically adjusted based on the uncertainty of the hidden layer information, thereby achieving a balance between information compression and preservation. The regularization parameter is dynamically adjusted, and the calculation method is expressed as follows:

[0068]

[0069] In the formula, σ p kc is the regularization intensity adjustment factor. p To adjust the steepness of the response curve, H p (Y p |Z p Given a hidden representation Z p Time tag Y p The conditional entropy, τ p The target threshold for entropy, ∈ pe For privacy cost parameters, d p (X p ,X′ p The input carbon emission data X to the neural network is... p Carbon emission data X′ with added Laplace noise p Distance metric;

[0070] S303 employs a quantum superposition gradient search strategy to update parameters. By exploring multiple gradient directions simultaneously, it improves the efficiency and accuracy of finding the global optimum, avoids the network getting trapped in local optima, and enhances the model's generalization performance. Carbon emission data features, due to their diversity and nonlinearity, may easily lead to local optima during training. The quantum superposition gradient search strategy, by exploring multiple gradient directions simultaneously, is expressed as:

[0071]

[0072] In the formula, Let be the weights of the neural network in the (t+1)th iteration. Let η be the weight of the neural network in the t-th iteration. p The learning rate used to update the weight parameters of the neural network. For the i-th c The probability amplitude of each quantum state Let be the gradient of the total loss function of the neural network with respect to the weights;

[0073] S304, to improve the network's structural adaptability during training, employs nonlinear dynamic topology adjustment. This treats the neural network's topology as a system state influenced by nonlinear dynamic equations, and combines this with quantum superposition feedback to achieve real-time adjustment of inter-layer connections and node activation functions. The calculation method is expressed as follows:

[0074]

[0075] In the formula, Let G be the topological change in the t-th iteration. p () is a nonlinear adjustment function. Let be the network state in the t-th iteration. Let be the loss value in the t-th iteration. Characterize the quantum state information in the t-th iteration;

[0076] The nonlinear adjustment function incorporates a feedback integration mechanism for the quantum state, and its calculation method is expressed as follows:

[0077]

[0078] In the formula, ν pe To adjust the amplitude coefficient, K pc To adjust the linear factor in the function, Sig() is the Sigmoid activation function and tanh() is the hyperbolic tangent function;

[0079] S305, repeat the above steps until the preset stopping iteration condition is met, which means that the model training is complete.

[0080] The beneficial effects of the present invention are as follows: (1) In the carbon emission prediction task, the present invention reduces and compresses the high-dimensional carbon emission features by using a sparse feature autoencoder and performs nonlinear dimensionality reduction on the construction data through sparse regularization, thereby reducing feature redundancy and enhancing the ability to capture key features.

[0081] (2) In the carbon emission prediction task, this invention combines the dynamic tensor coupling decomposition method to compress the multimodal features in high-dimensional carbon emission data into a core low-dimensional representation, while retaining key information directly related to carbon emissions.

[0082] (3) In the carbon emission prediction task, the present invention has a dynamic energy dissipation control mechanism that dynamically adjusts the energy dissipation rate of neurons according to the data complexity and optimizes the feature extraction capability.

[0083] (4) In the carbon emission prediction task, this invention adopts a neural network model based on quantum superposition state for modeling complex nonlinear carbon emission data. The model achieves global optimization capability of parameter initialization through quantum state superposition, thereby enhancing the adaptability of the model in complex nonlinear data environment.

[0084] (5) In the carbon emission prediction task, the present invention adopts the dynamic information bottleneck regularization method to suppress the interference of redundant information on modeling, and at the same time adopts differential privacy protection technology to protect sensitive information in carbon emission data.

[0085] (6) In the carbon emission prediction task, the present invention adopts a gradient search strategy of quantum superposition state to avoid the model getting trapped in local optima and improve the nonlinear feature extraction capability.

[0086] (7) In the carbon emission prediction task, the present invention adopts a dynamic adjustment layer and a nonlinear dynamic topology adjustment method in the process of adaptive and updated prediction model. The model structure and parameters are adjusted in real time according to the complex distribution of construction data to improve the ability to capture key features. Attached Figure Description

[0087] Figure 1 This is a heatmap of the noise characteristics before dimensionality reduction in this invention.

[0088] Figure 2 This is a heatmap of the noise characteristics after dimensionality reduction according to the present invention.

[0089] Figure 3 This is an experimental data graph showing the ability of the dynamic adjustment layer to adjust complex data distributions.

[0090] Figure 4 This is a bar chart showing the improvement in classification accuracy due to feature sparsity.

[0091] Figure 5 This is a flowchart of the training process for a neural network based on quantum superposition states. Detailed Implementation

[0092] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0093] The carbon emission prediction method for construction processes in this embodiment for intelligent manufacturing includes the following steps:

[0094] S1 collects carbon emission data and manually labels the collected carbon emission data. The labeling types include low carbon emissions, medium carbon emissions, and high carbon emissions.

[0095] The carbon emission data acquisition module obtains data from the construction site, including equipment operation data, energy consumption data, environmental monitoring data, and construction material data. Carbon emission data acquisition relies on embedded sensor networks and Internet of Things (IoT) technology to ensure the real-time performance and accuracy of the data.

[0096] The collected carbon emission data undergoes preliminary preprocessing, including cleaning and formatting, before being stored in a distributed database system in a structured format, such as JSON or XML, to support subsequent data processing and analysis needs.

[0097] Carbon emission data includes: Equipment operating time Da: total operating time of equipment during the monitoring period; Material usage Ca: total amount of materials used during construction; Ambient temperature Ma: average temperature of the construction site; Ambient humidity Ka: average humidity of the construction site; Dust concentration Pa: concentration of dust in the air at the construction site; Noise level Qa: average noise level of the construction site; Vibration index Sa: vibration intensity generated by the construction equipment; Overtime operation Ua: records of equipment exceeding normal operating time; Energy-saving mode duration Ha: duration of equipment using energy-saving mode.

[0098] S2 employs an autoencoder based on sparse features for feature dimensionality reduction, improving the effective representation and compression capabilities of features. By applying sparse constraints to the feature vectors reduced by the autoencoder, the autoencoder model can mine key features and effectively suppress redundant information.

[0099] The training process for a sparse feature-based autoencoder includes the following steps:

[0100] S201, initialize the parameters of the autoencoder, including the weights and biases. The weight initialization method for the autoencoder is expressed as follows:

[0101]

[0102] In the formula, Here, represents the weights of the l-th layer of the autoencoder, and randn() is a function that generates random numbers following a normal distribution. This represents the number of nodes in the (l-1)th layer of the autoencoder. Let l be the number of nodes in the l-th layer of the autoencoder, where l is the layer index of the autoencoder.

[0103] The bias initialization method for an autoencoder is expressed as follows:

[0104]

[0105] In the formula, is the bias of the l-th layer of the autoencoder, and zeros() is a function that generates a vector with all elements being 0.

[0106] S202, based on the preliminary analysis results of carbon emission data, dynamically configures the learning rate parameter of the autoencoder to address the issue of inconsistent distribution of different batches of data in the feature space. This ensures rapid convergence in the early stages of training and maintains stability in the later stages. Given the characteristics of carbon emission data, the distribution of different data batches varies significantly, such as the significant changes in energy consumption at different construction stages. Dynamic learning rate adjustment enables the model to quickly converge to the overall data distribution in the early stages, and gradually optimizes local features of specific stages in the later stages, avoiding model instability caused by inconsistent feature distributions. For example, when processing material consumption features, the model can dynamically adjust the weights, focusing on learning emission-related features, as shown below:

[0107]

[0108] In the formula, η r (t) represents the learning rate of the autoencoder in the t-th iteration, η r,init Let η be the initial learning rate of the autoencoder. r,init Set to 0.01, δ r δ represents the attenuation rate of the self-encoder. r Set to 0.95, where int(t) is the current iteration number.

[0109] S203 employs sparse coding learning on carbon emission data to address the problem of excessive redundant features in high-dimensional feature spaces. L1 regularization is used to achieve sparse feature representation, improving the model's ability to capture key features. When reducing the dimensionality of carbon emission data, the sparse feature autoencoder captures key features through L1 regularization constraints. Carbon emission data often contains many redundant features or noise; for example, ambient temperature and humidity may have a relatively small impact on emissions. Through sparse coding, the autoencoder can focus on key features such as energy consumption and material consumption, suppressing the influence of irrelevant features. For instance, by assigning higher sparsity weights to ambient humidity features, which have a low correlation with emissions, the model automatically ignores secondary features, improving the data representation after dimensionality reduction. This is represented as follows:

[0110]

[0111] In the formula, J r (W r ,b r ) represents the total loss function of the autoencoder, m r The number of training samples input to the autoencoder is given by i, where i is the index of the input sample to the autoencoder, and x is the index of the input sample. r,iFor the i-th input sample of the autoencoder, The reconstructed output of the autoencoder, ∥∥ is the L2 norm, ∥∥1 is the L1 norm, λ r λ is the sparse regularization coefficient. r Set to 0.5 to maintain sparsity while ensuring model reconstruction accuracy, h r () is the encoding function of the autoencoder, h r (x r,i ) represents the encoded output of the autoencoder layer for the i-th sample.

[0112] To enhance the sparsity of feature encoding in the autoencoder, the L1 regularization term is calculated as follows:

[0113]

[0114] In the formula, n r α is the number of nodes in the encoding layer, j is the index of the first autoencoder neuron, and α is the number of nodes in the encoding layer. r α is the first dynamic adjustment factor of the autoencoder. r Set to 0.3, β r β is the second dynamic adjustment factor of the autoencoder. r Set to 0.5, h r,j (x r,i ) represents the encoded output of the j-th neuron of the autoencoder for the i-th sample.

[0115] To analyze the effect of feature sparsification on improving classification accuracy, a heatmap was used to compare the differences in noise features before and after dimensionality reduction. This shows that the feature sparsification method has a good noise suppression capability, such as... Figure 1 and Figure 2 As shown.

[0116] S204 uses dynamic tensor coupling decomposition to reduce the dimensionality and reconstruct the hidden layer features, eliminating noise and redundant information and highlighting the key structures in carbon emission data. First, tensor decomposition is performed, treating the encoding layer output as a tensor to expand and decompose it, then folding it into a low-dimensional spatial representation, resulting in a more compact feature representation. Dynamic tensor coupling decomposition reduces the dimensionality of high-dimensional carbon emission features while retaining core information. For example, multimodal features (energy consumption, vibration, dust concentration, etc.) are treated as high-order tensors, and after Tucker decomposition, low-dimensional key feature representations are extracted. The model retains the main features directly related to emissions, such as construction equipment energy consumption, while minimizing the secondary influence of features like noise. After dimensionality reduction, the model can more accurately represent the main influencing factors of construction emissions. The calculation method is expressed as follows:

[0117] H r =fold(tdc(unfold(h) r (Xr )),[R1,R2,R3]))

[0118] In the formula, H r The feature representation is after dimensionality reduction. fold() represents the tensor folding operation, h r (X r ) represents the encoded output of the autoencoder layer on the sample, tdc() represents the Tucker decomposition operation, unfold() represents the tensor expansion operation, and X r The input data for the autoencoder encoding layer is [R1,R2,R3], which represents the kernel tensor dimensions. R1 is the first kernel tensor dimension, R2 is the second kernel tensor dimension, and R3 is the third kernel tensor dimension. R1 is set to 0.2 times the input feature dimension of the autoencoder and rounded to the nearest integer. R2 is set to 10 and R3 is set to 5.

[0119] The calculation method of Tucker decomposition is expressed as follows:

[0120]

[0121] In the formula, G r For the core tensor, The first mode matrix, This is the second mode matrix. The third modal matrix is ​​obtained by singular value decomposition. ×1 represents the first step of matrix multiplication, ×2 represents the second step of matrix multiplication, and ×3 represents the third step of matrix multiplication.

[0122] S205 employs a dynamic adjustment layer in the autoencoder. This layer dynamically selects and adjusts the activation function and model parameters based on the characteristics of the carbon emission data. This addresses carbon emission data distributions of varying complexity and prevents overfitting. For example, when the variance in energy consumption data is large (e.g., high-energy-consuming and low-energy-consuming equipment coexist), the dynamic adjustment layer can dynamically adjust its scaling factor, allowing the dimensionality reduction of features to more accurately reflect carbon emission levels. This prevents the model from overfitting or underfitting complex data distributions. This can be represented as:

[0123] σ r (x rdc ) = c r tanh(α r x rdc )+(1-c r ReLU(α) r x rdc )

[0124] In the formula, σ r () is the activation function of the dynamic adjustment layer, x rdc c is the input to the dynamic adjustment layer. rα is the adjustment coefficient for the activation function ratio under the current data distribution, tanh() is the hyperbolic tangent function, and α r α is the scaling factor for the dynamically adjusted layer. r The activation function of the dynamically adjusted layer is set to 0.2, using the Sigmoid activation function and ReLU() as the ReLU activation function.

[0125] This coefficient is automatically adjusted based on the average activity value of carbon emission data. The adjustment coefficient is set as follows:

[0126] c r =Sig(γ) r mean(abs(h r (X r ))))

[0127]

[0128] In the formula, Sig() is the Sigmoid activation function, and γ r This is a dynamic adjustment factor used to ensure the stability of dynamic adjustment when the ratio of data variance to standard deviation changes significantly. mean() is the function to calculate the mean, abs() is the function to take the absolute value, std() is the function to calculate the standard deviation, and var() is the function to calculate the variance.

[0129] like Figure 3 As shown, this paper analyzes the adaptability of the dynamic adjustment layer to complex data distributions and demonstrates the higher robustness of the dynamically adjusted model in handling complex data distributions.

[0130] S206 eliminates redundant features, simplifies the model, and improves inference speed. It avoids situations where interconnected features remain after dimensionality reduction. Correlation analysis is performed on the reduced features to remove redundant features (such as those highly correlated with material consumption). By setting a correlation threshold, the model can weaken or delete highly correlated features, such as those between energy consumption and vibration index, retaining energy consumption as the primary indicator. This further simplifies the model, improves inference speed, and reduces computational complexity. The calculation method is expressed as follows:

[0131]

[0132] In the formula, For the i-th a The dimensionality reduction feature and the j-th feature a The correlation coefficients between the reduced-dimensional features, and cov() performs covariance calculations. Features standard deviation Features standard deviation For the i-th aA dimensionality reduction feature, For the j-th a A dimensionality reduction feature, i a j is the index of the first dimensionality reduction feature. a This is the index of the second dimensionality reduction feature.

[0133] The correlation threshold θ r Set to 0.5, when If a feature is considered highly redundant, one of the features is randomly removed, and the calculation is repeated until the condition is met. conditions.

[0134] S207 employs an adaptive feature weight adjustment method. This method combines local and global statistical information during dimensionality reduction to dynamically weight the correlation between features and the target variable. This increases the focus on key features and suppresses irrelevant or noisy features. For carbon emission target optimization (such as low-carbon emission classification), features are assigned weights through dynamic weighting. For example, energy consumption features have higher mutual information with emissions, so their weights are dynamically adjusted to higher values, while the weights of noise level features are weakened due to lower mutual information. This allows the model to focus more on features with the greatest impact on emissions during dimensionality reduction, improving the expressive power of key features. The calculation method is as follows:

[0135]

[0136] In the formula, For the i-th a The mutual information between each feature and the target variable. Let be the target variable, representing the optimal value of the eigenvalues ​​after dimensionality reduction, and p() be the probability density function.

[0137] Dynamically adjusting the weights of each feature in the model enhances the model's sensitivity to key features while suppressing the influence of noise or irrelevant features. For example, based on multimodal features of the construction environment, the model prioritizes updating the weights of features closely related to low-carbon emission categories, reducing the interference of bias on feature representation and improving the dimensionality reduction quality of the model. By dynamically adjusting feature weights, the model more accurately characterizes the feature patterns of low-carbon emissions. The calculation method for feature weights is as follows:

[0138]

[0139] In the formula, For the i-th dimension after dimensionality reduction a The weights of each feature, where exp() is an exponential function. For the i-th a The mutual information between each feature and the target variable. For the j-th a The mutual information between each feature and the target variable, λ rTo adjust the hyperparameter of mutual information influence, λ r Set it to 0.1.

[0140] Feature weights are embedded into the autoencoder update process. By calculating the parameter increments of the autoencoder, adaptive adjustment of feature weights is achieved during autoencoder training. Fine-tuning the entire network optimizes reconstruction error and feature retention, focusing on the parameters that have the greatest impact on dimensionality reduction quality, thereby improving model performance. The calculation method is expressed as follows:

[0141]

[0142] In the formula, η represents the update amount of the weights in the l-th layer of the autoencoder. r (t) represents the learning rate of the autoencoder in the t-th iteration, where t is the index of the iteration number. J is the sign of the partial derivative. r is the loss function of the autoencoder.

[0143] The update amount of the autoencoder bias is calculated using the same update amount calculation method as the autoencoder weight. The weight and bias of the autoencoder are then updated based on the update amount of the autoencoder weight and bias. The update method is to add the update amount of the autoencoder weight and bias of the current iteration to the autoencoder weight and bias of the next iteration.

[0144] S208 employs a dynamic energy dissipation control method, enabling each neuron in the autoencoder to flexibly adjust its energy dissipation rate based on the complexity of the input carbon emission data during feature extraction and dimensionality reduction. The goal of this dynamic energy dissipation control method is to dynamically control the activation threshold of each neuron in the autoencoder during feature extraction, thereby optimizing the network's feature processing capabilities, reducing the influence of irrelevant features, and improving the overall efficiency and accuracy of the model. Specifically, the dynamic energy dissipation control method dynamically adjusts the energy dissipation rate of each neuron in the network according to the complexity and distribution of the input carbon emission data. By controlling the propagation speed and range of information in the neural network, it enhances the model's ability to extract key features. For example, during high-energy-consuming equipment operation, the energy dissipation rate of neurons is dynamically increased to emphasize the extraction of key features, while when there is a lot of noise or minor features, the energy dissipation rate is reduced to minimize the influence of irrelevant features. The energy dissipation rate is calculated as follows:

[0145]

[0146] In the formula, The energy dissipation rate of the l-th layer of the self-encoder. This represents the number of nodes in the (l-1)th layer of the autoencoder. For the weights of the l-th layer of the autoencoder, and the weights of the i-th layer... bThe element corresponding to each input, i b Here is the index of the second autoencoder neuron, and Sig() is the sigmoid activation function. For the i-th b Dimensionality reduction features.

[0147] To dynamically control the information propagation speed during autoencoder training and to measure the activation threshold within network layers, the threshold constant is defined as follows:

[0148]

[0149] In the formula, γ is the information propagation threshold constant of the l-th layer of the autoencoder. rec γ is the threshold adjustment factor for the autoencoder, var() is the variance calculation function, and γ rec Set it to 0.2.

[0150] The feature output of the autoencoder is updated based on a threshold constant, and the calculation method is expressed as follows:

[0151]

[0152] In the formula, This is the output of the l-th layer of the autoencoder. This is the output of the (l-1)th layer of the autoencoder.

[0153] S209, repeat the above steps until the preset stopping iteration condition is met. The preset stopping iteration condition is reaching the preset maximum number of iterations, which is set to 1000 times, indicating that the model training is complete.

[0154] After the autoencoder model is trained, the dimensionality-reduced feature vectors are classified using a preset Softmax function to obtain the classification prediction categories, which include: low carbon emissions, medium carbon emissions, and high carbon emissions, for a total of 3 categories.

[0155] To verify the effectiveness of feature dimensionality reduction, experimental data are as follows: Figure 4 As shown in the figure, by analyzing the changes in classification accuracy before and after feature sparsification, it is demonstrated that feature sparsification can effectively improve classification accuracy.

[0156] S3 uses a neural network based on quantum superposition as a feature extraction model to extract features from carbon emission data. During training, a gradient search mechanism under quantum superposition is used to improve optimization efficiency and generalization ability, thereby enhancing the neural network's ability to process complex nonlinear carbon emission data.

[0157] like Figure 5 As shown, the training process for a neural network based on quantum superposition states includes the following steps:

[0158] S301 employs a quantum state superposition method to randomly initialize the parameters of the neural network model. The nondeterminism of quantum state superposition enhances the network's exploration capability in unknown spaces, improving its adaptability and convergence speed in complex nonlinear carbon emission data environments. Carbon emission data feature extraction typically involves complex nonlinear relationships (such as the nonlinear correlation between energy consumption and emissions, and material usage and environmental factors). Through random initialization using quantum state superposition, the model can explore potential global optimal solutions more extensively in the parameter space, improving the adaptability and convergence efficiency of the neural network model in complex feature spaces. This can be expressed as:

[0159]

[0160] In the formula, These are the initial weights of the neural network. For the i-th c The quantum probability amplitude of a quantum ground state. For the i-th c A quantum ground state, i c is the index of the quantum ground state, and np is the total number of quantum ground states.

[0161] To verify the effectiveness of the initialization of neural network parameters in this invention, the influence of different gradient search strategies on the search for the global optimum was analyzed through experiments. The experimental results are shown in Table 1, with the probability of finding the global optimum as the unit. The quantum superposition gradient search showed a higher probability of finding the global optimum under data of different complexities.

[0162] Table 1. Impact of different gradient search strategies on the search for the global optimum.

[0163]

[0164] In S302, dynamic information bottleneck regularization is used in each iteration to constrain the complexity of the information flow. This allows the model to focus more on features that are decisive for the final output during the learning process, reducing the interference of redundant information. This results in more discriminative features in complex nonlinear carbon emission data analysis, improving the model's robustness and generalization ability. A dynamic regularization method based on differential privacy is employed to balance carbon emission data privacy protection and model performance when handling sensitive data. Carbon emission data typically contains a large amount of redundant information (such as the secondary impact of environmental humidity on emissions) and sensitive data (such as energy-consuming equipment models). Dynamic information bottleneck regularization limits the mutual information of the hidden layers, making the model focus more on the decisive role of key features (such as energy consumption and material usage) in the emission results, while reducing dependence on redundant features. For example, dynamic regularization can compress the complexity of environmental humidity features while maintaining a complete representation of energy consumption features. Simultaneously, the differential privacy regularization method protects data privacy by adding Laplace noise, as shown below:

[0165] L p =L0(W p ,X p )+β p ·I(Y p Z p |X p )+λ pe ·Lap(μ pe ,b pe )

[0166] In the formula, L p Let W be the total loss function of the neural network, L0 be the basic loss function of the neural network, and W be the total loss function of the neural network. p X represents the weights of the neural network. p β is the carbon emission data input to the neural network. p Y is the regularization parameter. p Z is the label for the carbon emission data input to the neural network. p For the feature representation of the intermediate layer of the neural network, I(Y) p Z p |X p ) represents the state of a given input X p Output Y under the condition p With hidden representation Z p Mutual information between them, I() is the mutual information function, λ pe λ is the privacy protection strength coefficient. pe Set to 0.2, Lap(μ) pe ,b pe Characterized by μ pe Centered on, b pe The Laplace noise distribution at scale μ peLet b be the mean of the current neural network input feature vector. pe b is the Laplace noise scale. pe Set to 1.

[0167] The regularization coefficient is dynamically adjusted based on the uncertainty of the hidden layer information, thereby achieving a balance between information compression and preservation. Simultaneously, considering different privacy cost requirements, the regularization parameter is dynamically adjusted, and the calculation method is expressed as follows:

[0168]

[0169] In the formula, σ p σ is the regularization intensity adjustment factor. p Set to 0.1, kc p To adjust the steepness of the response curve, kc p Set to 0.3, H p (Y p |Z p Given a hidden representation Z p Time tag Y p The conditional entropy, τ p τ is the target threshold for entropy. p Set to 0.1, ∈ pe For privacy cost parameters, d p (X p ,X′ p The input carbon emission data X to the neural network is... p Carbon emission data X′ with added Laplace noise p Distance metric.

[0170] S303 employs a quantum superposition gradient search strategy to update parameters. By exploring multiple gradient directions simultaneously, it improves the efficiency and accuracy of finding the global optimum, avoids the network getting trapped in local optima, and enhances the model's generalization performance. Carbon emission data features, due to their diversity and nonlinearity, can easily lead to local optima during training. The quantum superposition gradient search strategy explores multiple gradient directions simultaneously. For example, in construction site energy consumption data, high-energy-consuming equipment may form a strong contrast with low-energy-consuming equipment, causing traditional optimization algorithms to favor one type of data. The quantum superposition search strategy can achieve balanced optimization among multiple features, extracting more representative key carbon emission features, represented as:

[0171]

[0172] In the formula, Let be the weights of the neural network in the (t+1)th iteration. Let η be the weight of the neural network in the t-th iteration. pThe learning rate used to update the weight parameters of the neural network. For the i-th c The probability amplitude of each quantum state Let be the gradient of the total loss function of the neural network with respect to the weights.

[0173] S304, to improve the network's structural adaptability during training, employs nonlinear dynamic topology adjustment. This treats the neural network's topology as a system state influenced by nonlinear dynamic equations, and combines quantum superposition feedback to achieve real-time adjustment of inter-layer connections and node activation functions. The complexity of carbon emission characteristics necessitates dynamic adaptation of the network structure to optimize model performance. For example, when energy consumption and noise level features are highly correlated, the network topology can dynamically adjust inter-layer connections to reduce information redundancy and enhance the extraction of independent features. The calculation method is expressed as follows:

[0174]

[0175] In the formula, Let G be the topological change in the t-th iteration. p () is a nonlinear adjustment function. Let be the network state in the t-th iteration. Let be the loss value in the t-th iteration. Characterize the quantum state information in the t-th iteration.

[0176] The nonlinear adjustment function incorporates a feedback integration mechanism for the quantum state, and its calculation method is expressed as follows:

[0177]

[0178] In the formula, ν pe To adjust the amplitude coefficient, ν pe Set to 0.2, K pc To adjust the linear factor in the function, K pc If is set to 2, Sig() is the Sigmoid activation function, and tanh() is the hyperbolic tangent function.

[0179] S305, repeat the above steps until the preset stopping iteration condition is met. The preset stopping iteration condition is reaching the preset maximum number of iterations, which is set to 1000 times, indicating that the model training is complete.

[0180] After the neural network model is trained, the feature vectors after feature extraction are classified using a preset Softmax function to obtain the classification prediction category. In one embodiment, the categories include: low carbon emissions, medium carbon emissions, and high carbon emissions, for a total of three categories.

[0181] By analyzing the performance comparison of different methods on the carbon emission classification task, the experimental data are shown in Table 2. The quantum superposition state neural network performed best in the main indicators of the carbon emission classification task, demonstrating high nonlinear feature extraction ability and generalization ability.

[0182] Table 2. Performance comparison of different methods for carbon emission classification tasks

[0183]

[0184] S4 automates the training, validation, and updating of the prediction model, ensuring that the model can adapt to the dynamic changes in the construction process.

[0185] S5 uses a trained model to predict carbon emissions at each stage of the construction process in real time, providing a basis for optimizing and adjusting the construction process.

[0186] The carbon emission prediction system for construction processes in this embodiment for intelligent manufacturing includes a carbon emission data acquisition module, a carbon emission data feature compression module, a machine learning modeling module, a data management module, a model training and update module, and a real-time prediction module.

[0187] The carbon emission data acquisition module is responsible for real-time acquisition and preprocessing of multi-source heterogeneous carbon emission data during the construction process to support the input data requirements for subsequent modeling and analysis.

[0188] The carbon emission data feature compression module is used to reduce and compress high-dimensional carbon emission features to reduce redundant features and improve the efficiency of model training and prediction.

[0189] The features of high-dimensional carbon emissions should have more than 50 dimensions, meaning that the collected carbon emission data should have more than 50 attributes. Otherwise, the collected carbon emission data should be directly input into the machine learning modeling module for machine learning modeling.

[0190] The feature compression module uses an autoencoder to perform nonlinear dimensionality reduction on the multidimensional carbon emission data of the construction process. The encoder part extracts the potential low-dimensional feature representation of the carbon emission data, and the decoder part verifies whether the compressed features can reconstruct the original carbon emission data, ensuring the effectiveness of feature compression, removing redundant information and noise, and retaining the key features that contribute most to carbon emission prediction.

[0191] The machine learning modeling module is used to process carbon emission data with 50 or fewer dimensions.

[0192] The data management module provides the system with efficient data storage, retrieval, and security management functions, ensuring system reliability and data utilization efficiency.

[0193] The data management module's functions include: building a multi-dimensional database for the construction process, storing historical data, real-time collected data, and model-generated data; employing distributed storage technology to ensure the reliability and scalability of data storage; ensuring data security based on access control and encryption technology to prevent unauthorized access; and providing an efficient data query interface to support rapid data retrieval by time, process, equipment, and other dimensions.

[0194] The model training and update module is used to automate the training, validation, and updating of the prediction model to ensure that the model can adapt to the dynamic changes in the construction process.

[0195] The model training and update module includes the following functions: preliminary feature selection and engineering based on collected carbon emission data, removing redundant information and enhancing key features; comprehensive evaluation of model performance through cross-validation technology to select the optimal combination of model parameters; enabling the model to quickly adapt to new scenarios or processes using transfer learning methods; and regularly updating and training the model to cope with technological upgrades or environmental changes during construction.

[0196] The real-time prediction module predicts carbon emissions at each stage of the construction process based on a trained model, providing a basis for optimizing and adjusting the construction process.

[0197] The real-time prediction module has the following functions: receiving real-time carbon emission data transmitted from the data acquisition module and quickly generating carbon emission prediction results; supporting individual carbon emission assessments of construction equipment, processes, or materials to help identify high-carbon emission links; providing early warning information by comparing predicted values ​​with set carbon emission targets; and visualizing the prediction results in real time to provide decision support for construction managers.

[0198] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention.

Claims

1. A construction process carbon emission prediction method for smart manufacturing, characterized in that, The method comprises the following steps: S1, collecting carbon emission data, and manually labeling the collected carbon emission data, the labeling types including low carbon emission, medium carbon emission and high carbon emission; S2, performing feature dimension reduction by using a sparse feature-based autoencoder to improve the effective representation and compression capability of the features, and enabling the autoencoder model to mine key features and effectively suppress redundant information by performing sparse constraint on the feature vectors reduced by the autoencoder; The training process of the sparse feature-based autoencoder in S2 comprises the following steps: S201, initializing the parameters of the autoencoder, including the weights and bias of the autoencoder, the weight initialization mode of the autoencoder being represented as: wherein, is the weight of the l-th layer of the autoencoder, and randn() is a function for generating a random number following a normal distribution, is the number of nodes of the l-1-th layer of the autoencoder, is the number of nodes of the l-th layer of the autoencoder, and l is an index of the layer of the autoencoder. The bias initialization mode of the autoencoder is represented as: wherein is the bias of the l-th layer of the autoencoder, and zeros() is a function that generates a vector of all zeros. S202, dynamically configuring the learning rate parameter of the autoencoder according to the preliminary analysis result of the carbon emission data, which is represented as: where η r (t) is the learning rate of the tth iteration of the autoencoder, η r,init is the initial learning rate of the autoencoder, δ r is the decay rate of the autoencoder, and int(t) is the current iteration number. S203, performing sparse coding learning on the carbon emission data, realizing sparse representation of the features by L1 regularization, and improving the capture capability of the model on key features, wherein when processing the dimension reduction of the carbon emission data, the sparse feature-based autoencoder realizes the capture of key features by L1 regularization constraint, which is represented as: In the formula, J r (W r , b r ) is the total loss function of the autoencoder, m r is the number of training samples input to the autoencoder, i is the index of the autoencoder input sample, x r,i is the i-th input sample of the autoencoder, is the reconstruction output of the autoencoder, |||| is the L2 norm, ||||1 is the L1 norm, λ r is the sparse regularization coefficient, h r () is the encoding function of the autoencoder, h r (x r,i ) is the encoding output of the i-th sample by the encoding layer of the autoencoder; In order to strengthen the sparsity of the feature coding of the autoencoder, the calculation mode of the L1 regularization term is represented as: where n r is the number of encoding layer nodes, j is the index of the first autoencoder neuron, a r is the first dynamic adjustment factor of the autoencoder, b r is the second dynamic adjustment factor of the autoencoder, h r,j (x r,i ) is the encoding output of the jth neuron of the autoencoder for the ith sample; S204, performing dimension reduction and reconstruction on the hidden layer features by dynamic tensor coupling decomposition to eliminate noise and redundant information, the calculation mode being represented as: H r = fold(tdc(unfold(h r (X r )),[R1,R2,R3])) where H r is the reduced dimension feature representation, fold() represents the folding operation of the tensor, h r (X r ) is the encoding output of the sample by the autoencoder encoding layer, tdc() represents the Tucker decomposition operation, unfold() represents the unfolding operation of the tensor, X r is the input data of the autoencoder encoding layer, [R1, R2, R3] is the core tensor dimension, R1 is the first core tensor dimension, R2 is the second core tensor dimension, and R3 is the third core tensor dimension. The calculation mode of the Tucker decomposition is represented as: In the formula, G r is a core tensor, is a first modal matrix, is a second modal matrix, is a third modal matrix, the first modal matrix, the second modal matrix and the third modal matrix are obtained through singular value decomposition, ×1 is a first step matrix multiplication, ×2 is a second step matrix multiplication, and ×3 is a third step matrix multiplication. S205, adopting a dynamic adjustment layer in the autoencoder, dynamically selecting and adjusting the activation function and model parameters according to the features of the carbon emission data, coping with carbon emission data distribution of different complexities, and preventing model overfitting, which is represented as: σ r (x rdc ) = c r tanh(α r x rdc ) + (1 - c r ) ReLU(α r x rdc ) In the formula, σ r () is an activation function of the dynamic adjustment layer, x rdc is an input of the dynamic adjustment layer, c r is an adjustment coefficient of the activation function under the current data distribution, tanh() is a hyperbolic tangent function, and α r is a scaling factor of the dynamic adjustment layer, and ReLU() is a ReLU activation function. The coefficient is automatically adjusted according to the average activity value of the carbon emission data, and the adjustment coefficient is set as: c r = Sig(γ r mean(abs(h r (X r )))) In the formula, Sig() is a Sigmoid activation function, γ r is a dynamic adjustment factor, used to ensure stability of dynamic adjustment when the ratio of data variance to standard deviation changes greatly, mean() is an average value calculation function, abs() is an absolute value function, std() is a standard deviation calculation function, and var() is a variance calculation function. S206, eliminating redundant features, simplifying the model and improving the reasoning speed, avoiding the case that the reduced features still contain interrelated features, performing correlation analysis on the reduced features, removing redundant features, and the calculation mode being represented as: In the formula, For the i-th a The dimensionality reduction feature and the j-th feature a The correlation coefficients between the reduced-dimensional features, and cov() performs covariance calculations. Features standard deviation Features standard deviation For the i-th a A dimensionality reduction feature, For the jth a A dimensionality reduction feature, i a j is the index of the first dimensionality reduction feature. a The index for the second dimensionality reduction feature; S207, adopting an adaptive feature weight adjustment method, combining local and global statistical information in the dimension reduction process of the model, dynamically weighting the correlation between features and target variables, improving the attention to key features and suppressing irrelevant or noise features, and the calculation mode being represented as: In the formula, is the mutual information amount of the i a th feature and the target variable, is the target variable, represents the optimal value of the feature value after dimension reduction, and p() is a probability density function. The weight of each feature in the model is dynamically adjusted to improve the sensitivity of the model to key features, and the calculation mode of the feature weight is represented as: In the formula, For the i-th dimension after dimensionality reduction a The weights of each feature, where exp() is an exponential function. For the i-th a The mutual information between each feature and the target variable. For the jth a The mutual information between each feature and the target variable, λ r Hyperparameters used to adjust the degree of influence of mutual information; The feature weight is embedded in the update process of the autoencoder, the parameter increment of the autoencoder is calculated to realize adaptive feature weight adjustment in the training process of the autoencoder, the reconstruction error and feature retention degree are optimized by fine-tuning the entire network, and the calculation mode is represented as: wherein, is the update quantity for the l-th layer weight of the autoencoder, η r (t) is the learning rate for the t-th iteration of the autoencoder, t is an index of the iteration number, is the partial derivative sign, J r is the loss function of the autoencoder; The update amount of the autoencoder bias is calculated in the same way as the update amount of the autoencoder weight, and the autoencoder weight and bias are updated based on the update amount of the autoencoder weight and bias, and the update mode is that the autoencoder weight and bias of the current iteration are added to the update amount, which are used as the autoencoder weight and bias of the next iteration. S208, using a dynamic energy dissipation control method, the energy dissipation rate of each layer of neurons of the autoencoder is adjusted flexibly according to the complexity of the input carbon emission data during feature extraction and dimension reduction, and the calculation method of the energy dissipation rate is represented as: In the formula, The energy dissipation rate of the l-th layer of the self-encoder. This represents the number of nodes in the (l-1)th layer of the autoencoder. For the weights of the l-th layer of the autoencoder, and the weights of the i-th layer... b The element corresponding to each input, i b Here is the index of the second autoencoder neuron, and Sig() is the sigmoid activation function. For the i-th b One dimensionality reduction feature; In order to dynamically control the information propagation speed in the training process of the autoencoder and measure the activation threshold in the network layer, the calculation method of the threshold constant is defined as: wherein is the information propagation threshold constant of the self-encoder, lth layer, γ rec is the threshold adjustment factor of the self-encoder, var() is the variance calculation function; Based on the threshold constant, the feature output in the encoding process of the autoencoder is updated, and the calculation method is represented as: wherein is the output of the l-th layer of the autoencoder, is the output of the l-1-th layer of the autoencoder; S209, repeat the above steps until the preset stopping iteration condition is met, that is, the model training is completed; S3, using a neural network based on quantum superposition state as a feature extraction model, the carbon emission data is extracted, and in the training process, the gradient search mechanism under quantum superposition state is used to improve the optimization efficiency and generalization ability, and the processing ability of the neural network for complex nonlinear carbon emission data is improved; S4, automatically training, verifying and updating the prediction model to ensure that the model can adapt to the dynamic changes of the construction process; S5, based on the trained model, the carbon emission of each link in the construction process is predicted in real time, which provides a basis for the optimization and adjustment of the construction process.

2. The smart manufacturing-oriented construction process carbon emission prediction method according to claim 1, characterized in that: The carbon emission data in S1 includes equipment running time Da, material consumption Ca, environmental temperature Ma, environmental humidity Ka, dust concentration Pa, noise level Qa, vibration index Sa, overtime work Ua and energy saving mode time Ha. 3.The smart manufacturing oriented construction process carbon emission prediction method according to claim 1, characterized in that, The training process of the neural network based on quantum superposition state in S3 includes the following steps: S301, the parameters of the neural network model are randomly initialized by using the method of quantum state superposition, the non-deterministic of quantum state superposition is used to improve the exploration ability of the network in unknown space, and the adaptability and convergence speed of the model in complex nonlinear carbon emission data environment are enhanced, which is represented as: wherein are initial weights of the neural network, are quantum probability amplitudes of the i c quantum ground state, are quantum probability amplitudes of the i c quantum ground state, i c is an index of the quantum ground state, and npis the total number of quantum ground states. S302, in each iteration, the complexity of information flow is constrained by using dynamic information bottleneck regularization, so that the model pays more attention to the features that have decisive effect on the final output in the learning process, reduces the interference of redundant information, and obtains more discriminative features in complex nonlinear carbon emission data analysis, improves the robustness and generalization ability of the model to data, and uses the dynamic regularization method based on differential privacy to balance the carbon emission data privacy protection and model performance when processing sensitive data, the dynamic information bottleneck regularization limits the mutual information of hidden layer, so that the model pays more attention to the decisive effect of key features on emission results, and reduces the dependence on redundant features, represented as: L p = L0(W p ,X p )+ β p · I(Y p ; Z p | X p )+ λ pe · Lap(μ pe ,b pe ) In the formula, L p Let W be the total loss function of the neural network, L0 be the basic loss function of the neural network, and W be the total loss function of the neural network. p X represents the weights of the neural network. p β is the carbon emission data input to the neural network. p Y is the regularization parameter. p Z is the label for the carbon emission data input to the neural network. p For the feature representation of the intermediate layer of the neural network, I(Y) p Z p |X p ) represents the state of a given input X p Output Y under the condition p With hidden representation Z p Mutual information between them, I() is the mutual information function, λ pe Lap(μ) is the privacy protection strength coefficient. pe ,b pe Characterized by μ pe Centered on, b pe The Laplace noise distribution at scale μ pe Let b be the mean of the current neural network input feature vector. pe The Laplace noise scale; Based on the measurement of the uncertainty of the hidden layer information, the regularization coefficient is dynamically adjusted, so as to balance the compression and reservation of information, and the regularization parameter is dynamically adjusted, and the calculation method is represented as: In the formula, σ p kc is the regularization intensity adjustment factor. p To adjust the steepness of the response curve, H p (Y p |Z p Given a hidden representation Z p Time tag Y p The conditional entropy, τ p ε is the target threshold for entropy. pe For privacy cost parameters, d p (X p ,X′ p The input carbon emission data X to the neural network is... p Carbon emission data X′ with added Laplace noise p Distance metric; S303, the gradient search strategy of quantum superposition state is used to complete parameter updating, the efficiency and accuracy of global optimal solution search are improved by simultaneously exploring in multiple gradient directions, network is prevented from falling into local optimum, and the generalization performance of the model is strengthened. The characteristics of carbon emission data may be diversified and nonlinear, so that the training process is prone to fall into local optimum. The gradient search strategy of quantum superposition state simultaneously explores multiple gradient directions, and is represented as follows: wherein are the weights of the neural network for the t+1th iteration, are the weights of the neural network for the tth iteration, η p is a learning rate for the neural network weight parameter update, is the probability amplitude of the ith c quantum state, is the gradient of the total loss function of the neural network with respect to the weights; S304, in order to improve the structural adaptive ability of the network in the training process, a nonlinear dynamic topology adjustment is used, the topology structure of the neural network is regarded as a system state affected by a nonlinear dynamic equation, and real-time adjustment of interlayer connection and node activation function is realized in combination with quantum superposition state feedback. The calculation mode is represented as follows: In the formula, is the topology variation under the tth iteration, G p () is a nonlinear adjustment function, is the network state under the tth iteration, is the loss value under the tth iteration, characterizes the quantum state information in the tth iteration; The nonlinear adjustment function contains a feedback integration mechanism for quantum state, and the calculation mode is represented as follows: In the formula, ν pe is an adjustment amplitude coefficient, K pc is a linear factor in the adjustment function, Sig() is a Sigmoid activation function, and tanh() is a hyperbolic tangent function. S305, the above steps are repeatedly iterated until a preset stop iteration condition is met, that is, the model training is completed.

Images