A method for screening key process parameters associated with carbon emissions in a compounding injection molding process
By using spatiotemporal graph neural networks and deep causal feature selection methods, the problems of spatiotemporal dependence and spurious correlation in the mixing and injection molding process of traditional machine learning are solved, enabling accurate screening of key process parameters and identification of causal correlations, and supporting process optimization in low-carbon manufacturing.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHONGQING UNIV
- Filing Date
- 2026-01-29
- Publication Date
- 2026-06-05
AI Technical Summary
In existing technologies, traditional machine learning methods in the mixing and injection molding process cannot effectively capture the spatiotemporal dependence of parameters, resulting in the omission of high-contribution spatiotemporal features and susceptibility to confounding variables. This leads to poor reliability of the screening results and an inability to accurately screen for the true causal relationship with carbon emissions.
The method employs a spatiotemporal graph neural network (STGNN) model combined with a deep causal feature selection (DCFS) approach. By collecting data from the entire process, spatiotemporal features are extracted and fused to construct a causal adjacency matrix. Key features with strong causal associations are selected, and importance scores and causal importance weights are calculated through gradient backpropagation to eliminate false associations and construct a core candidate parameter set.
It effectively captures the spatiotemporal dependence of process parameters, identifies the true causal relationship with carbon emissions, eliminates false correlations, provides a reliable basis for low-carbon manufacturing parameters, and supports the optimization of compounding injection molding processes.
Smart Images
Figure CN122153632A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of low-carbon manufacturing technology, specifically a method for screening key process parameters related to carbon emissions in the mixing and injection molding process. Background Technology
[0002] The integrated mixing and injection molding machine is a core piece of equipment for molding polymer materials. Its process parameters (such as temperature, pressure, and rotation speed) are diverse and exhibit multi-source heterogeneous characteristics. Different parameters have significantly different impacts on carbon emissions, making accurate selection of key parameters a prerequisite for achieving low-carbon control. Current technologies often employ traditional machine learning methods, such as random forests and XGBoost, for process parameter selection. However, traditional machine learning methods have two major drawbacks in selecting key process parameters for low-carbon correlation in mixing and injection molding systems: 1) They cannot effectively capture the spatiotemporal dependence of parameters during injection molding, leading to the omission of high-contribution spatiotemporal features; 2) Based solely on correlation selection, they are susceptible to spurious associations due to confounding variables, resulting in poor reliability and failing to guarantee a true causal relationship with carbon emissions. With the upgrading of low-carbon manufacturing demands, there is an urgent need for a key process parameter selection method that can adapt to multi-source heterogeneous data, accurately capture spatiotemporal correlations, and eliminate spurious associations, providing a reliable parameter foundation for subsequent low-carbon optimization of the mixing and injection molding process. Summary of the Invention
[0003] The purpose of this invention is to provide a method for screening key process parameters related to carbon emissions in the mixing and injection molding process, comprising the following steps:
[0004] Step 1) Collect full-process data of the mixing and injection molding machine in one or more complete production cycles, and preprocess it to build an initial dataset;
[0005] Step 2) Extract and fuse spatiotemporal features from the time-series process parameters and spatial distribution parameters in the initial dataset to obtain a spatiotemporal fused feature vector;
[0006] Step 3) Reassemble the spatiotemporal fusion feature vectors to obtain a three-dimensional dataset that conforms to the input of the spatiotemporal graph neural network model;
[0007] Step 4) Train the spatiotemporal graph neural network model using a 3D dataset. After training, use the gradient backpropagation method to calculate the importance score of each input feature to the output, and construct a candidate key feature set based on the importance score.
[0008] Step 5) Construct the causal structure between the candidate key feature set and carbon emissions, and determine the causal adjacency matrix A;
[0009] Step 6) Using candidate key features H, causal adjacency matrix A, and carbon emission value Y as inputs to the deep causal feature selection model, a set of candidate key features with strong causal associations is obtained through causal strength screening, and the causal importance weight of the candidate key features is determined.
[0010] Step 7) Select candidate key features whose causal importance weight is greater than the preset causal strength threshold as core candidate parameters, thereby constructing a core candidate parameter set;
[0011] Step 8) Filter the core candidate parameter set and calculate the comprehensive score of the core candidate parameters after filtering; the comprehensive score is the weighted sum of the causal importance weight and the importance score;
[0012] Step 9) Sort all core candidate parameters in descending order according to the comprehensive score and output the ranking list of key process parameters.
[0013] Furthermore, the full-process process data includes the PLC system, sensor network, production logs, and full-process process parameters, raw material information, product information, and real-time energy consumption data of the mixing and injection molding integrated machine;
[0014] Process parameters include sequential process parameters, spatial distribution parameters, and discrete state parameters;
[0015] The timing process parameters include the pressure, speed, position, and time sequence data for each stage of mold closing, mold opening, injection, holding pressure, and sol-gel.
[0016] Spatial distribution parameters include the set temperature and measured temperature of each heating zone in the barrel;
[0017] Discrete state parameters include machine operation mode, raw material grade, and product model.
[0018] Furthermore, in step 1), the preprocessing steps include:
[0019] Step 1.1) Convert energy consumption data into carbon emission data using an energy consumption-carbon emission model;
[0020] The energy consumption-carbon emission model is shown below:
[0021] (1)
[0022] Where C represents carbon emissions; E represents energy consumption; E elec The carbon emission factor value for electricity;
[0023] Step 1.2) Construct the original dataset ;in, It is a continuous time series parameter matrix; It is a discrete state parameter vector; For spatial distribution tensors; Let be the carbon emissions during the t-th production cycle;
[0024] Step 1.3) Perform data cleaning, data standardization, and data fusion on the original dataset to obtain the initial dataset, i.e.:
[0025] (2)
[0026] The steps for data cleaning and standardization of continuous data in the original dataset include: using improved... The criteria are used to detect outliers, followed by robust normalization and standardization.
[0027] The steps for cleaning and standardizing discrete data in the original dataset include: filling missing values in the discrete data with the mode, and then standardizing using one-hot encoding;
[0028] The steps for data cleaning and standardization of spatiotemporal data in the original dataset include: filling missing values in the spatiotemporal data using KNN interpolation, followed by standardization; wherein, the filled missing values in the spatiotemporal data are... , α and β are the spatiotemporal distance attenuation coefficients. , For time and space data.
[0029] Furthermore, in step 2), the steps for spatiotemporal feature extraction and fusion include:
[0030] In terms of time dimension, for the time series data of each process stage, the mean, peak value, standard deviation and rate of change characteristics are calculated using a sliding window to form a time series feature vector;
[0031] In the spatial dimension, a graph structure with production stages as time steps is constructed, using process parameter collection points or process stages as nodes and mutual information or process logic relationships between parameters as edge weights. The temporal feature vectors are then combined with the graph structure to form spatiotemporal graph data.
[0032] The Pearson correlation coefficient and mutual information value between spatiotemporal fusion features and carbon emissions were calculated. Weakly correlated features with correlation degree |r| < 0.2 or mutual information value < 0.08 were removed to obtain a preliminary combined dataset of structured spatiotemporal graph data and high-dimensional spatiotemporal feature vectors: ① Structured spatiotemporal graph (including node feature matrix, adjacency matrix, and time step sequence); ② Corresponding high-dimensional spatiotemporal feature vectors (dimension matched with spatiotemporal graph nodes and time steps).
[0033] Furthermore, in step 4), the spatiotemporal graph neural network model includes two graph convolutional layers, one temporal convolutional layer, and one fully connected output layer;
[0034] The output of the first graph convolutional layer As shown below:
[0035] (3)
[0036] in, The feature matrix output from step 2 has n spatial nodes and d node feature dimensions. To add self-connected adjacency matrices; This is the first layer of trainable weight matrix; This is the bias term for the first layer GCN.
[0037] The output of the second convolutional layer As shown below:
[0038] (4)
[0039] in, This is the trainable weight matrix for the second layer; This is the bias term for the second-layer GCN.
[0040] The output of the temporal convolutional layer is shown below:
[0041] (5)
[0042] (6)
[0043] in, , These are the kernel weights; , This is the convolution bias. , This is the output of the convolution branch;
[0044] The output of the fully connected output layer is shown below:
[0045] (7)
[0046] in, The output features of the temporal convolutional layer at the last time step t; This is the weight matrix of the fully connected layer; For the input of the fully connected output layer, This is the bias term for the fully connected layer.
[0047] Furthermore, in step 4), the importance score is the absolute value of the gradient of the carbon emission prediction value;
[0048] When constructing a candidate key feature set based on importance scores, the input features are sorted in descending order according to their importance scores, and the top n input features are selected as candidate key features, thereby constructing the candidate key feature set.
[0049] Furthermore, in step 5), the steps for constructing the causal structure between the candidate key feature set and carbon emissions include:
[0050] Step 5.1) Perform deep feature encoding on the candidate key features to obtain the features. ; For encoder parameters;
[0051] Step 5.2) Construct linear structural equations for continuous candidate key features, discrete candidate key features, and carbon emissions, respectively, i.e.:
[0052] (8)
[0053] (9)
[0054] (10)
[0055] in, for The parent node; , , For causal path coefficients; , , This is exogenous noise; , Let i be the i-th continuous candidate key feature and m be the m-th discrete candidate key feature; , , It is a constant; for The parent node;
[0056] Step 5.3) Construct the objective function, namely:
[0057] (11)
[0058] Among them, constraint loss ; For SEM to X i The predicted value of the nth sample; The SEM prediction value for the nth sample of Y; acyclic constraint term. k is the number of candidate process parameters, and A is the causal adjacency matrix; , For coefficients; , This is the causal path coefficient vector; The weights are acyclic constraints.
[0059] Step 5.4) Iteratively optimize the objective function using the Adam optimizer to output the causal structure of candidate key feature sets and carbon emissions.
[0060] Furthermore, the deep causal feature selection model includes an input layer, a causal constraint layer, a fully connected hidden layer, and an output layer;
[0061] The input to the input layer includes features H, causal adjacency matrix A, and carbon emissions Y;
[0062] The output of the causal constraint layer is Learnable mask ; The learnable parameter matrix;
[0063] The causal constraint regularization used in the causal constraint layer is: ;
[0064] The fully connected hidden layer includes a first hidden layer, a second hidden layer, and a third hidden layer;
[0065] The outputs of the first hidden layer, the second hidden layer, and the third hidden layer are shown below:
[0066] (12)
[0067] (13)
[0068] (14)
[0069] In the formula, , , This is the output of the first hidden layer, the second hidden layer, and the third hidden layer; , , For bias; , , As weight;
[0070] The output of the output layer is ;
[0071] The loss function used in the training process of the deep causal feature selection model As shown below:
[0072] (15)
[0073] Among them, losses ,loss ,loss ;
[0074] Furthermore, the importance weight of causation ; The optimal causal path coefficients obtained in step 5.2) This represents the variation range of the standard process, taken as ±10% of the standard process value.
[0075] Furthermore, in step 8), a 5-fold cross-validation method is used to screen the core candidate parameter set, eliminating core candidate parameters with volatility coefficients greater than a preset volatility threshold; wherein, the volatility coefficient... ; The standard deviation of the causal importance weights during five cross-validations; This represents the average of the causal importance weights during five cross-validations.
[0076] The technical effectiveness of this invention is undeniable. It captures the spatiotemporal dependence of process parameters using an STGNN model, and then identifies the true causal relationship between parameters and carbon emissions through DCFS, effectively eliminating spurious correlation parameters and adapting to the characteristics of multi-source heterogeneous data. The screening results have been verified statistically, causally, and experimentally, demonstrating high reliability. This provides a precise parameter basis for low-carbon optimization of mixing and injection molding processes, and has significant engineering application value. Attached Figure Description
[0077] Figure 1 A flowchart illustrating the overall technical process flow of a method for screening key process parameters related to carbon emissions in a mixing and injection molding process.
[0078] Figure 2 This is a flowchart of the STGNN model involved in this invention;
[0079] Figure 3 This is a flowchart of the screening process for selecting key process parameters related to carbon emissions based on deep causal features, as involved in this invention.
[0080] Figure 4 This is a flowchart of the causal structure learning process involved in the fine screening process of this invention;
[0081] Figure 5 This is a diagram of the DCFS fusion model architecture in the fine screening process involved in this invention. Detailed Implementation
[0082] The present invention will be further described below with reference to embodiments, but it should not be construed that the scope of the present invention is limited to the following embodiments. Various substitutions and modifications made based on ordinary technical knowledge and common practices in the art without departing from the above-described technical concept of the present invention should be included within the scope of protection of the present invention.
[0083] Example 1:
[0084] See Figures 1 to 5 A method for screening key process parameters related to carbon emissions in the mixing and injection molding process includes the following steps:
[0085] Step 1) Collect full-process data of the mixing and injection molding machine in one or more complete production cycles, and preprocess it to build an initial dataset;
[0086] Step 2) Extract and fuse spatiotemporal features from the time-series process parameters and spatial distribution parameters in the initial dataset to obtain a spatiotemporal fused feature vector;
[0087] Step 3) Reassemble the spatiotemporal fusion feature vectors to obtain a three-dimensional dataset that conforms to the input of the spatiotemporal graph neural network model;
[0088] Step 4) Train the spatiotemporal graph neural network model using a 3D dataset. After training, use the gradient backpropagation method to calculate the importance score of each input feature to the output, and construct a candidate key feature set based on the importance score.
[0089] Step 5) Construct the causal structure between the candidate key feature set and carbon emissions, and determine the causal adjacency matrix A;
[0090] Step 6) Using candidate key features H, causal adjacency matrix A, and carbon emission value Y as inputs to the deep causal feature selection model, a set of candidate key features with strong causal associations is obtained through causal strength screening, and the causal importance weight of the candidate key features is determined.
[0091] Step 7) Select candidate key features whose causal importance weight is greater than the preset causal strength threshold as core candidate parameters, thereby constructing a core candidate parameter set;
[0092] Step 8) Filter the core candidate parameter set and calculate the comprehensive score of the core candidate parameters after filtering; the comprehensive score is the weighted sum of the causal importance weight and the importance score;
[0093] Step 9) Sort all core candidate parameters in descending order according to the comprehensive score and output the ranking list of key process parameters.
[0094] Example 2:
[0095] A method for screening key process parameters related to carbon emissions in a mixing and injection molding process, with the same technical content as in Example 1, further including the full-process process data including the PLC system of the mixing and injection molding integrated machine, sensor network, production logs, and full-process process parameters, raw material information, product information, and real-time energy consumption data on the energy consumption monitor;
[0096] Process parameters include sequential process parameters, spatial distribution parameters, and discrete state parameters;
[0097] The timing process parameters include the pressure, speed, position, and time sequence data for each stage of mold closing, mold opening, injection, holding pressure, and sol-gel.
[0098] Spatial distribution parameters include the set temperature and measured temperature of each heating zone in the barrel;
[0099] Discrete state parameters include machine operation mode, raw material grade, and product model.
[0100] Example 3:
[0101] A method for screening key process parameters related to carbon emissions in a compounding and injection molding process, with the same technical content as any one of Examples 1-2, further comprising the following pretreatment steps in step 1):
[0102] Step 1.1) Convert energy consumption data into carbon emission data using an energy consumption-carbon emission model;
[0103] The energy consumption-carbon emission model is shown below:
[0104] (1)
[0105] Where C represents carbon emissions; E represents energy consumption; E elec The carbon emission factor value for electricity;
[0106] Step 1.2) Construct the original dataset ;in, It is a continuous time series parameter matrix; It is a discrete state parameter vector; For spatial distribution tensors; Let be the carbon emissions during the t-th production cycle;
[0107] Step 1.3) Perform data cleaning, data standardization, and data fusion on the original dataset to obtain the initial dataset, i.e.:
[0108] (2)
[0109] The steps for data cleaning and standardization of continuous data in the original dataset include: using improved... The criteria are used to detect outliers, followed by robust normalization and standardization.
[0110] The steps for cleaning and standardizing discrete data in the original dataset include: filling missing values in the discrete data with the mode, and then standardizing using one-hot encoding;
[0111] The steps for data cleaning and standardization of spatiotemporal data in the original dataset include: filling missing values in the spatiotemporal data using KNN interpolation, followed by standardization; wherein, the filled missing values in the spatiotemporal data are... , α and β are the spatiotemporal distance attenuation coefficients. , For time and space data.
[0112] Example 4:
[0113] A method for screening key process parameters related to carbon emissions in a compounding and injection molding process, with the same technical content as any one of Examples 1-3, further comprising, in step 2), the steps of extracting and fusing spatiotemporal features, including:
[0114] In terms of time dimension, for the time series data of each process stage, the mean, peak value, standard deviation and rate of change characteristics are calculated using a sliding window to form a time series feature vector;
[0115] In the spatial dimension, a graph structure with production stages as time steps is constructed, using process parameter collection points or process stages as nodes and mutual information or process logic relationships between parameters as edge weights. The temporal feature vectors are then combined with the graph structure to form spatiotemporal graph data.
[0116] The Pearson correlation coefficient and mutual information value between spatiotemporal fusion features and carbon emissions were calculated. Weakly correlated features with correlation degree |r| < 0.2 or mutual information value < 0.08 were removed to obtain a preliminary combined dataset of structured spatiotemporal graph data and high-dimensional spatiotemporal feature vectors: ① Structured spatiotemporal graph (including node feature matrix, adjacency matrix, and time step sequence); ② Corresponding high-dimensional spatiotemporal feature vectors (dimension matched with spatiotemporal graph nodes and time steps).
[0117] Example 5:
[0118] A method for screening key process parameters related to carbon emissions in a mixing and injection molding process, the technical content of which is the same as any one of Examples 1-4, further wherein, in step 4), the spatiotemporal graph neural network model includes two graph convolutional layers, one temporal convolutional layer and one fully connected output layer.
[0119] The output of the first graph convolutional layer As shown below:
[0120] (3)
[0121] in, The feature matrix output from step 2 has n spatial nodes and d node feature dimensions. To add self-connected adjacency matrices; This is the first layer of trainable weight matrix; This is the bias term for the first layer GCN.
[0122] The output of the second convolutional layer As shown below:
[0123] (4)
[0124] in, This is the trainable weight matrix for the second layer; This is the bias term for the second-layer GCN.
[0125] The output of the temporal convolutional layer is shown below:
[0126] (5)
[0127] (6)
[0128] in, , These are the kernel weights; , This is the convolution bias. , This is the output of the convolution branch;
[0129] The output of the fully connected output layer is shown below:
[0130] (7)
[0131] in, The output features of the temporal convolutional layer at the last time step t ( , (fusion characteristics) This is the weight matrix of the fully connected layer; For the input of the fully connected output layer, This is the bias term for the fully connected layer.
[0132] Example 6:
[0133] A method for screening key process parameters related to carbon emissions in a mixing and injection molding process, with the same technical content as any one of Examples 1-5, further wherein, in step 4), the importance score is the absolute value of the gradient of the predicted carbon emissions.
[0134] When constructing a candidate key feature set based on importance scores, the input features are sorted in descending order according to their importance scores, and the top n input features are selected as candidate key features, thereby constructing the candidate key feature set.
[0135] Example 7:
[0136] A method for screening key process parameters related to carbon emissions in a compounding and injection molding process, with the same technical content as any one of Examples 1-6, further comprising, in step 5), the step of constructing a causal structure between the candidate key feature set and carbon emissions, including:
[0137] Step 5.1) Perform deep feature encoding on the candidate key features to obtain the features. ; For encoder parameters;
[0138] Step 5.2) Construct linear structural equations for continuous candidate key features, discrete candidate key features, and carbon emissions, respectively, i.e.:
[0139] (8)
[0140] (9)
[0141] (10)
[0142] in, for The parent node; , , For causal path coefficients; , , This is exogenous noise; , Let i be the i-th continuous candidate key feature and m be the m-th discrete candidate key feature; , , It is a constant; for The parent node;
[0143] Step 5.3) Construct the objective function, namely:
[0144] (11)
[0145] Among them, constraint loss ; For SEM to X i The predicted value of the nth sample; The SEM prediction value for the nth sample of Y; acyclic constraint term. k is the number of candidate process parameters, and A is the causal adjacency matrix; , For coefficients; , This is the causal path coefficient vector; The weights are acyclic constraints.
[0146] Step 5.4) Iteratively optimize the objective function using the Adam optimizer to output the causal structure of candidate key feature sets and carbon emissions.
[0147] Example 8:
[0148] A method for screening key process parameters related to carbon emissions in a mixing and injection molding process, with the same technical content as any one of Examples 1-7, further comprising an input layer, a causal constraint layer, a fully connected hidden layer, and an output layer;
[0149] The input to the input layer includes features H, causal adjacency matrix A, and carbon emissions Y;
[0150] The output of the causal constraint layer is Learnable mask ; The learnable parameter matrix;
[0151] The causal constraint regularization used in the causal constraint layer is: ;
[0152] The fully connected hidden layer includes a first hidden layer, a second hidden layer, and a third hidden layer;
[0153] The outputs of the first hidden layer, the second hidden layer, and the third hidden layer are shown below:
[0154] (12)
[0155] (13)
[0156] (14)
[0157] In the formula, , , This is the output of the first hidden layer, the second hidden layer, and the third hidden layer; , , For bias; , , As weight;
[0158] The output of the output layer is ;
[0159] The loss function used in the training process of the deep causal feature selection model As shown below:
[0160] (15)
[0161] Among them, losses ,loss ,loss ;
[0162] Example 9:
[0163] A method for screening key process parameters related to carbon emissions in a compounding and injection molding process, with the same technical content as any one of Examples 1-8, further including causal importance weighting. ; The optimal causal path coefficients obtained in step 5.2) This represents the variation range of the standard process, taken as ±10% of the standard process value.
[0164] Example 10:
[0165] A method for screening key process parameters related to carbon emissions in a mixing and injection molding process, with the same technical content as any one of Examples 1-2, further comprising the following step 8): using a 5-fold cross-validation method to screen the core candidate parameter set, eliminating core candidate parameters with fluctuation coefficients greater than a preset fluctuation threshold; wherein, the fluctuation coefficient... ; The standard deviation of the causal importance weights during five cross-validations; This represents the average of the causal importance weights during five cross-validations.
[0166] Example 11:
[0167] A method for screening key process parameters related to carbon emissions in a compounding and injection molding process, comprising the following steps:
[0168] S1: Acquisition and preprocessing of multi-source heterogeneous process data and carbon emission data. Collect full-process data from the integrated mixing and injection molding machine throughout one or more complete production cycles to construct an initial dataset;
[0169] S2: Spatiotemporal Feature Extraction and Fusion. Spatiotemporal feature extraction and fusion are performed on time-series process parameters and spatial distribution parameters.
[0170] S3: Preliminary feature importance screening based on spatiotemporal graph neural network. Using the spatiotemporal fusion feature vector described in S2 as input and carbon emissions as output, a spatiotemporal graph neural network model is trained. After training, the importance score of each input feature to the model output is calculated using gradient backpropagation. Based on a preset first threshold, the feature set with the highest scores is selected as the candidate key feature set.
[0171] S4: Causal correlation screening based on deep causal feature selection. Using the candidate process parameter set obtained in S3 as input and carbon emissions as the result variable, a deep causal feature selection model based on causal graph model and neural network is constructed to establish causal relationships between variables, and then the core candidate parameter set is obtained by screening using causal strength.
[0172] S5: Screening Result Verification and Key Parameter Ranking. Cross-validation is performed on the core candidate parameter set to assess stability, eliminating parameters with excessive fluctuations. Then, a weighted calculation is performed by combining the feature importance score obtained in step S3 and the causal importance weight obtained in step S4 to obtain the final comprehensive score for each parameter. All parameters are then sorted in descending order based on the comprehensive score, outputting a ranking list of key process parameters.
[0173] (S1) Acquisition and preprocessing of multi-source heterogeneous process data and carbon emission data
[0174] The data collection scope includes: the PLC system of the mixing and injection molding machine, sensor network, production logs, and full-process process parameters, raw material information, product information, and real-time energy consumption data on the energy consumption monitor.
[0175] Among them, process parameters are divided into time-series parameters, spatial distribution parameters, discrete state parameters, and result variables.
[0176] The timing process parameters include: pressure, speed, position, and time sequence data for each stage of mold closing, mold opening, injection, holding pressure, and sol-gel.
[0177] Spatial parameters include: the set temperature and the measured temperature of each heating zone in the barrel;
[0178] Discrete state parameters include: machine operation mode, raw material grade, and product model;
[0179] Outcome variable: Carbon emissions corresponding to the production cycle. Energy consumption data is converted into carbon emission data using an energy consumption-carbon emission model.
[0180] The preprocessing steps include: data cleaning, data standardization, and data fusion.
[0181] 1) Let the collected original dataset be:
[0182] in, It is a continuous time series parameter matrix; It is a discrete state parameter vector; For spatial distribution tensors; Let be the carbon emissions during the t-th production cycle.
[0183] Cleaning and standardization processes.
[0184] For continuous data, an improved method is used. After outlier detection, robust normalization is performed according to the criteria.
[0185] Where MAD is the median absolute deviation:
[0186] Then, standardization processing is performed: Q p The p-th quantile enhances robustness against outliers.
[0187] For missing values in discrete data, the mode is used for imputation; then one-hot encoding is used for standardization.
[0188] For missing values in spatiotemporal data, KNN interpolation is used to fill in the missing values: , , where α and β are the spatiotemporal distance attenuation coefficients.
[0189] 3) Finally, the standardized data from different classes are merged into a single matrix:
[0190]
[0191] (S2) Spatiotemporal Feature Extraction and Fusion
[0192] Time series feature extraction: For the time series data of each process stage, the mean, peak value, standard deviation and rate of change are calculated using a sliding window to form a time series feature vector.
[0193] Construction of the spatiotemporal graph: Using process parameter acquisition points or process stages as nodes, and mutual information or process logic relationships between parameters as edge weights, a spatially related graph structure with production stages as time steps is constructed; the temporal feature vectors are combined with the graph structure to form spatiotemporal graph data.
[0194] Spatiotemporal feature fusion: The spatiotemporal graph data is embedded and learned using a graph neural network encoder to generate a high-dimensional spatiotemporal fusion feature vector.
[0195] 3. (S3) Preliminary screening of feature importance based on spatiotemporal graph neural network
[0196] 1. The constructed STGNN model contains a fusion architecture consisting of two graph convolutional layers, one temporal convolutional layer, and one fully connected output layer.
[0197] 1) First GCN convolutional layer:
[0198] in, Let be a feature matrix with n spatial nodes and d node feature dimensions; To add self-connected adjacency matrices; This is the first layer of trainable weight matrix; This is the bias term for the first layer GCN.
[0199] 2) Second GCN convolutional layer:
[0200] in, This is the trainable weight matrix for the second layer; This is the bias term for the second-layer GCN.
[0201] 3) Third TCN convolutional layer:
[0202]
[0203] in, , These are the kernel weights; , This is the convolution bias.
[0204] 4) Fully connected output layer:
[0205]
[0206] in, The output features of TCN at the last time step t; This is the weight matrix of the fully connected layer; This is the bias term for the fully connected layer.
[0207] 2. Spatiotemporal feature importance calculation based on gradient backpropagation: After the model is trained, all parameters are fixed, the gradient of the carbon emission prediction value is calculated, the absolute value of the gradient is used to represent the feature importance score, and then the features are sorted in descending order. The top 40% of features are selected as the candidate spatiotemporal feature set.
[0208] 4. (S4) Causal association screening based on deep causal feature selection
[0209] Design process:
[0210] 1. The input is the set of candidate process parameters obtained from S3 screening. , where k is the number of candidate key parameters; the output is the result variable Y, which is the carbon emissions of the mixing and injection molding process.
[0211] 2. Causal Structure Learning
[0212] 1) Construct the feature encoding layer of DAG-GNN:
[0213] Deep feature encoding is performed on candidate parameters, and the spatiotemporal residual information implicit in the parameters is fused to improve the robustness of causal structure learning.
[0214] 2) Constructing a structural equation model:
[0215] Linear structure equations with continuous parameters:
[0216] Linear structural equation for carbon emissions Y:
[0217] Nonlinear structure equations with discrete parameters:
[0218] in, for The parent node, i.e., the direct cause; , For causal path coefficients; , , This is exogenous noise.
[0219] 3) Implement acyclic constraint optimization for NOTEARS, with the objective function being:
[0220] The fitting loss is:
[0221] Acyclic constraint terms
[0222] 4) The objective function is iteratively optimized using the Adam optimizer to output the optimal causal structure.
[0223] 3. DCFS fusion model, which consists of a four-layer structure:
[0224] 1) Input layer: encoded high-dimensional features H, causal adjacency matrix W, and carbon emissions Y.
[0225] 2) Causal constraint layer: The learnable mask is
[0226] in, is the learnable parameter matrix; W is the causal prior adjacency matrix.
[0227] Then apply the causal mask to the encoded features: To ensure the rationality of causality, causal constraint regularization is...
[0228] 3) Fully connected hidden layer: First hidden layer:
[0229] Second hidden layer:
[0230] Third hidden layer:
[0231] 4) Output layer:
[0232] 5) Comprehensive forward propagation calculation formula:
[0233] The total loss consists of three parts:
[0234] in: , ,
[0235] 4. Further quantify the strength of causal association using the Average Treatment Effect (ATE) calculation, retaining parameters with ATE values greater than 0.3 and removing weakly associated parameters. The calculation formula is:
[0236] 5. (S5) Screening Result Validation and Key Parameter Ranking
[0237] 1. Perform 5-fold cross-validation to eliminate unstable parameters.
[0238] 1) Randomly divide the dataset D into 5 equal parts, and select 4 parts as the training set and 1 part as the test set. Repeat the training of the DCFS model 5 times to obtain each parameter X. i The scores of the 5 groups are recorded as follows: .
[0239] 2) Then calculate the volatility coefficient:
[0240] in, Five ATE groups i Standard deviation; For five groups of ATE i The mean.
[0241] Unstable parameters with a fluctuation coefficient greater than 0.2 are removed, and the core candidate parameter set is finally obtained.
[0242] 2. Screening Results Validation and Ranking. The significance of the core parameters' impact on carbon emissions was tested using analysis of variance. Finally, the causal effect value of DCGS and the feature importance score of STGNN were weighted and summed to obtain the final ranking of the process parameters.
Claims
1. A method for screening key process parameters related to carbon emissions in a mixing and injection molding process, characterized in that, Includes the following steps: Step 1) Collect full-process data of the mixing and injection molding machine in one or more complete production cycles, and preprocess it to build an initial dataset; Step 2) Extract and fuse spatiotemporal features from the time-series process parameters and spatial distribution parameters in the initial dataset to obtain a spatiotemporal fused feature vector; Step 3) Reassemble the spatiotemporal fusion feature vectors to obtain a three-dimensional dataset that conforms to the input of the spatiotemporal graph neural network model; Step 4) Train the spatiotemporal graph neural network model using a 3D dataset. After training, use the gradient backpropagation method to calculate the importance score of each input feature to the output, and construct a candidate key feature set based on the importance score. Step 5) Construct the causal structure between the candidate key feature set and carbon emissions, and determine the causal adjacency matrix A; Step 6) Using candidate key features H, causal adjacency matrix A, and carbon emission value Y as inputs to the deep causal feature selection model, a set of candidate key features with strong causal associations is obtained through causal strength screening, and the causal importance weight of the candidate key features is determined. Step 7) Select candidate key features whose causal importance weight is greater than the preset causal strength threshold as core candidate parameters, thereby constructing a core candidate parameter set; Step 8) Filter the core candidate parameter set and calculate the comprehensive score of the core candidate parameters after filtering; the comprehensive score is the weighted sum of the causal importance weight and the importance score; Step 9) Sort all core candidate parameters in descending order according to the comprehensive score and output the ranking list of key process parameters.
2. The method for screening key process parameters related to carbon emissions in the mixing and injection molding process according to claim 1, characterized in that, The full-process process data includes the PLC system, sensor network, production logs, and full-process process parameters, raw material information, product information, and real-time energy consumption data of the mixing and injection molding integrated machine; Process parameters include sequential process parameters, spatial distribution parameters, and discrete state parameters; The timing process parameters include the pressure, speed, position, and time sequence data for each stage of mold closing, mold opening, injection, holding pressure, and sol-gel. Spatial distribution parameters include the set temperature and measured temperature of each heating zone in the barrel; Discrete state parameters include machine operation mode, raw material grade, and product model.
3. The method for screening key process parameters related to carbon emissions in the mixing and injection molding process according to claim 1, characterized in that, Step 1) includes the following preprocessing steps: Step 1.1) Convert energy consumption data into carbon emission data using an energy consumption-carbon emission model; The energy consumption-carbon emission model is shown below: （1） Where C represents carbon emissions; E represents energy consumption; E elec The carbon emission factor value for electricity; Step 1.2) Construct the original dataset ;in, It is a continuous time series parameter matrix; It is a discrete state parameter vector; For spatial distribution tensors; Let be the carbon emissions during the t-th production cycle; Step 1.3) Perform data cleaning, data standardization, and data fusion on the original dataset to obtain the initial dataset, i.e.: （2） The steps for data cleaning and standardization of continuous data in the original dataset include: using improved... The criteria are used to detect outliers, followed by robust normalization and standardization. The steps for cleaning and standardizing discrete data in the original dataset include: filling missing values in the discrete data with the mode, and then standardizing using one-hot encoding; The steps for data cleaning and standardization of spatiotemporal data in the original dataset include: filling missing values in the spatiotemporal data using KNN interpolation, followed by standardization; wherein, the filled missing values in the spatiotemporal data are... , α and β are the spatiotemporal distance attenuation coefficients; , For time and space data.
4. The method for screening key process parameters related to carbon emissions in the mixing and injection molding process according to claim 1, characterized in that, Step 2) includes the following steps for spatiotemporal feature extraction and fusion: In terms of time dimension, for the time series data of each process stage, the mean, peak value, standard deviation and rate of change characteristics are calculated using a sliding window to form a time series feature vector; In the spatial dimension, a graph structure with production stages as time steps is constructed, using process parameter collection points or process stages as nodes and mutual information or process logic relationships between parameters as edge weights. The temporal feature vectors are then combined with the graph structure to form spatiotemporal graph data. The Pearson correlation coefficient and mutual information value between spatiotemporal fusion features and carbon emissions were calculated. Weakly correlated features with correlation degree |r| < 0.2 or mutual information value < 0.08 were removed to obtain a preliminary combined dataset of structured spatiotemporal graph data and high-dimensional spatiotemporal feature vectors.
5. The method for screening key process parameters related to carbon emissions in a mixing and injection molding process according to claim 1, characterized in that, In step 4), the spatiotemporal graph neural network model includes two graph convolutional layers, one temporal convolutional layer, and one fully connected output layer; The output of the first graph convolutional layer As shown below: （3） in, The feature matrix output from step 2 has n spatial nodes and d node feature dimensions. To add self-connected adjacency matrices; This is the first layer of trainable weight matrix; This is the bias term for the first layer GCN; The output of the second convolutional layer As shown below: （4） in, This is the trainable weight matrix for the second layer; This is the bias term for the second-layer GCN; The output of the temporal convolutional layer is shown below: （5） （6） in, , These are the kernel weights; , For convolution bias; , This is the output of the convolution branch; The output of the fully connected output layer is shown below: （7） in, The output features of the temporal convolutional layer at the last time step t; This is the weight matrix of the fully connected layer; For the input of the fully connected output layer, This is the bias term for the fully connected layer.
6. The method for screening key process parameters related to carbon emissions in a mixing and injection molding process according to claim 1, characterized in that, In step 4), the importance score is the absolute value of the gradient of the carbon emission prediction value; When constructing a candidate key feature set based on importance scores, the input features are sorted in descending order according to their importance scores, and the top n input features are selected as candidate key features, thereby constructing the candidate key feature set.
7. The method for screening key process parameters related to carbon emissions in a mixing and injection molding process according to claim 1, characterized in that, Step 5) involves constructing the causal structure between the candidate key feature set and carbon emissions, including: Step 5.1) Perform deep feature encoding on the candidate key features to obtain the features. ; For encoder parameters; Step 5.2) Construct linear structural equations for continuous candidate key features, discrete candidate key features, and carbon emissions, respectively, i.e.: （8） （9） （10） in, for The parent node; , , For causal path coefficients; , , This is exogenous noise; , Let i be the i-th continuous candidate key feature and m be the m-th discrete candidate key feature; , , It is a constant; for The parent node; Step 5.3) Construct the objective function, namely: （11） Among them, constraint loss ; For SEM to X i The predicted value of the nth sample; The SEM prediction value for the nth sample of Y; acyclic constraint term. k is the number of candidate process parameters, and A is the causal adjacency matrix; , For coefficients; , This is the causal path coefficient vector; The weights are acyclic constraints. Step 5.4) Iteratively optimize the objective function using the Adam optimizer to output the causal structure of candidate key feature sets and carbon emissions.
8. The method for screening key process parameters related to carbon emissions in a mixing and injection molding process according to claim 7, characterized in that, The deep causal feature selection model consists of an input layer, a causal constraint layer, a fully connected hidden layer, and an output layer. The input to the input layer includes features H, causal adjacency matrix A, and carbon emissions Y; The output of the causal constraint layer is Learnable mask ; The learnable parameter matrix; The causal constraint regularization used in the causal constraint layer is: ; The fully connected hidden layer includes a first hidden layer, a second hidden layer, and a third hidden layer; The outputs of the first hidden layer, the second hidden layer, and the third hidden layer are shown below: （12） （13） （14） In the formula, , , This is the output of the first hidden layer, the second hidden layer, and the third hidden layer; , , For bias; , , As weight; The output of the output layer is ; The loss function used in the training process of the deep causal feature selection model As shown below: （15） Among them, losses ,loss ,loss .
9. The method for screening key process parameters related to carbon emissions in a mixing and injection molding process according to claim 8, characterized in that, Causal Importance Weight ; The optimal causal path coefficients obtained in step 5.2) This represents the variation range of the standard process.
10. The method for screening key process parameters related to carbon emissions in the injection molding process according to claim 1, characterized in that: In step 8), a 5-fold cross-validation method is used to filter the core candidate parameter set, eliminating core candidate parameters with volatility coefficients greater than a preset volatility threshold; among which, the volatility coefficient... ; The standard deviation of the causal importance weights during five cross-validations; This represents the average of the causal importance weights during five cross-validations.