Neural network multi-process quality prediction method based on fusion hierarchical graph construction
Patent Information
- Application Number
- CN202610828981.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-09
- Publication Date
- 2026-10-16
AI Technical Summary
[0005]本申请的主要目的在于提供一种融合层次图构建的神经网络多工序质量预测方法，旨在解决如何同时刻画工序内参数耦合与工序间逐级传递的问题
[0068] 1. By using the hierarchical graph structure of micro-parameter networks and macro-process chains, the coupling of variables within a process and the step-by-step transmission between processes can be characterized simultaneously, overcoming the limitation of a single model that cannot take into account multi-scale features, and is more in line with the physical logic of actual process production.
Smart Images
Figure CN122927593A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of deep learning technology, and in particular to a method for multi-process quality prediction of neural networks that integrates hierarchical graph construction. Background Technology
[0002] The production process in the process industry typically consists of multiple sequential technological stages. In multi-stage production scenarios, process parameters are highly dimensional, significantly nonlinear, and exhibit strong temporal dependencies and multi-scale coupling. Coupled with uncertainties such as fluctuations in raw material properties and equipment operating condition drift, product quality often exhibits an evolutionary pattern of multi-source disturbances, gradual transmission, and cumulative formation. Existing prediction methods based on statistical learning or shallow machine learning struggle to uncover deep nonlinear mappings, while single deep learning models, although capable of extracting local temporal features, lack the ability to collaboratively characterize the coupling of variables within a process and the information transmission between processes, thus limiting the prediction accuracy and stability in complex scenarios.
[0003] Taking tobacco processing as an example, tobacco leaves undergo key processes such as loosening and rehydration, leaf moistening and feeding, and thin-plate drying. These processes are deeply coupled due to changes in material conditions and do not operate independently. Changes in the process conditions of upstream processes are transmitted downstream along the production line, ultimately altering the properties of the material flowing into the thin-plate drying machine and affecting the moisture content of the exiting material. This indicator is a crucial parameter for evaluating the stability and consistency of tobacco quality, and its fluctuations are directly related to the combustion performance and sensory quality of cigarettes.
[0004] To address the aforementioned issues, this invention proposes a neural network-based multi-process quality prediction method that integrates hierarchical graphs, thereby solving the problem in existing technologies that it is difficult to simultaneously characterize parameter coupling within processes and the step-by-step transmission between processes. Summary of the Invention
[0005] The main objective of this application is to provide a neural network-based multi-process quality prediction method that integrates hierarchical graphs, aiming to solve the problem of how to simultaneously characterize parameter coupling within a process and the step-by-step transmission between processes.
[0006] To achieve the above objectives, this application provides a method for multi-process quality prediction using a neural network constructed from a fused hierarchical graph, the method comprising:
[0007] S1: Clean, normalize, and process missing values for the collected historical process data of production processes. The historical process data includes the process parameters and corresponding quality indicators of each process.
[0008] S2: Calculate the correlation between each process parameter and the target process quality index, and select parameters with a correlation higher than the preset threshold as model input features;
[0009] S3: Model the production process as a hierarchical graph structure of micro-parameter network and macro-process directed chain; the micro-parameter network uses the process parameters in each process as nodes to construct a fully connected candidate graph; the macro-process directed chain uses each process as a node and constructs directed edges along the actual process flow direction.
[0010] S4: Construct a deep network model, including:
[0011] The micro-encoding module is used to adaptively learn the correlation strength between nodes using a graph attention mechanism on the micro-parameter network, and compress the node-level representation into a sequence of process-level representations through a weighted aggregation and readout mechanism;
[0012] The macro-chain gating fusion module is used to perform gating propagation along the process direction on the directed chain of the macro process, adaptively balance the information of the current process with the information transmitted upstream, and output the fused process representation.
[0013] The temporal decoding module is used to improve the LSTM temporal modeling of the fused process representation sequence, and adaptively assign importance weights for different historical moments by combining a self-attention mechanism, and output the predicted value of the quality index of the target process.
[0014] S5: Divide the preprocessed dataset into training set, validation set and test set according to the proportion. Use the training set to iteratively optimize the parameters of the deep network model to minimize the loss function between the predicted value and the real quality index. Then, use the validation set to perform cross-validation and hyperparameter adjustment.
[0015] S6: Input the multi-process parameters to be predicted into the trained model, and sequentially process them through MGE micro-encoding, CGF macro-fusion and LAD temporal decoding to output the predicted value of the export quality index of the target process at the next moment.
[0016] Optionally, step S1 specifically includes:
[0017] Based on the material head and tail interception rules in the production process, data from the non-steady-state stages of equipment startup and shutdown are removed to complete data cleaning.
[0018] Min-max standardization normalizes the data to a preset interval using the following formula:
[0019]
[0020] In the formula, This is the original data. For the normalized data, and These are the maximum and minimum values of the original data, respectively. and The endpoints of the target mapping interval.
[0021] Optionally, in step S2, the mutual information value between each process parameter and the target process quality index is calculated as the correlation, and the formula is:
[0022]
[0023] In the formula, and These represent sets of variables that respectively represent process parameters and quality indicators. For joint probability distribution, and Each parameter has its own marginal probability distribution; based on the calculated mutual information value, a threshold is set, redundant parameters are eliminated, and key process parameters with mutual information values higher than the threshold are retained as model inputs.
[0024] Optionally, the construction of the hierarchical graph structure in step S3 includes:
[0025] The entire production process Each process is abstracted into macroscopic nodes, and a macroscopic directed graph is constructed. , where the node set correspond Each process, edge collection This represents the chain-like transmission relationship between processes;
[0026] For the process , and at time All key process parameter nodes are grouped into a node set. And it uses a fully connected candidate graph as the basis for the microstructure, edge set any edge in satisfy:
[0027] .
[0028] Optionally, in step S4, the MGE micro-coding module satisfies the following steps during execution:
[0029] The process is transformed by linear transformation. At any moment nodes Observations Mapping to the initial embedding representation :
[0030]
[0031] In the formula, and The weight matrix and bias vector are learnable.
[0032] For any node pair Calculate and normalize the attention score to obtain the implicit edge weights that vary over time. :
[0033]
[0034]
[0035] In the formula, , It is a linear mapping matrix. For attention vectors, This represents vector concatenation. For activation function, For nodes For nodes Contribution weight;
[0036] Update node representations through weighted aggregation:
[0037]
[0038] Attention-based aggregation is used to elevate node-level representations to process-level representations:
[0039]
[0040]
[0041] In the formula, As the importance weight of the nodes, For process At any moment The process level is represented by the process level.
[0042] Optionally, in step S4, the CGF macroscopic chain-gated fusion module satisfies the following steps during execution:
[0043] Time step The first process is integrated. For the process , The upstream process is fused and mapped to the feature space of the current process:
[0044]
[0045] In the formula, This is a cross-process mapping matrix;
[0046] The gating mechanism adaptively adjusts the upstream information injection strength, and the gating vector is calculated as follows:
[0047]
[0048] The merged process is represented by a weighted composite output:
[0049]
[0050] In the formula, This represents element-wise multiplication. , These are learnable parameters.
[0051] Optionally, in step S4, the LAD timing decoding module satisfies the following steps during execution:
[0052] Improved LSTM unit: Replace the tanh activation function in the traditional LSTM unit with the softsign activation function to stabilize the backpropagation gradient;
[0053] Self-attention mechanism: The hidden sequence obtained by processing the improved LSTM is... Input self-attention mechanism, with the hidden state at the last time step. Generate query vectors The hidden state at each time step is projected as a key vector. Calculate attention weights:
[0054]
[0055] In the formula, The dimension of the key vector;
[0056] A context vector is generated by weighting and aggregating all hidden states based on attention weights. Finally, the context vector Input the multilayer perceptron prediction block to obtain the predicted quality index value of the target process at the next time step. .
[0057] Optionally, the model training in step S5 specifically includes:
[0058] The preprocessed dataset is divided into training set, validation set and test set according to a preset ratio;
[0059] The Adam optimizer is used to iteratively optimize the model parameters, and the mean squared error (MSE) is used as the loss function.
[0060] An early stopping mechanism is used during training. Training is terminated when the validation set loss does not decrease within a preset number of consecutive rounds to prevent overfitting.
[0061] Optionally, in step S6, the mean absolute error (MAE), root mean square error (RMSE), and goodness of fit are used to measure the results. The model's effectiveness was evaluated using the following formulas:
[0062]
[0063]
[0064]
[0065] In the formula, For the number of samples, This represents the true value of the export quality indicators. For predicted values, This is the mean of the true values.
[0066] Optionally, the multi-process parameters include the loosening and rehydration, leaf feeding, and thin-plate drying processes in process industrial production, the target process is at least one of the thin-plate drying processes, and the outlet quality index includes the moisture content of the material exiting the thin-plate drying machine.
[0067] This application has at least the following beneficial effects:
[0068] 1. By using the hierarchical graph structure of micro-parameter networks and macro-process chains, the coupling of variables within a process and the step-by-step transmission between processes can be characterized simultaneously, overcoming the limitation of a single model that cannot take into account multi-scale features, and is more in line with the physical logic of actual process production.
[0069] 2. Within a process, the implicit correlation of process parameters is adaptively learned through the graph attention encoding module. Between processes, chain-gated fusion is performed along the actual process flow. The gradual accumulation process of quality fluctuations is explicitly modeled, and the dynamic evolution model of the fused process representation is performed, which solves the defects of multi-process feature extraction in traditional models.
[0070] 3. In the multi-process quality prediction experiment of silk production, the method of the present invention has low MAE and RMSE, and R2 reaches above 0.98, showing excellent prediction accuracy and fitting consistency.
[0071] 4. It can be extended to multi-process quality prediction scenarios in other process industries such as chemical and food industries, providing a reliable technical reference for quality control and production optimization in process industries. Attached Figure Description
[0072] Figure 1 This is a flowchart illustrating the neural network multi-process quality prediction method constructed from a fusion hierarchy graph, as described in the embodiments of this application.
[0073] Figure 2 This is a schematic diagram of the fusion layer graph construction and deep neural network model involved in the embodiments of this application;
[0074] Figure 3 This is a comparison chart of the predicted and actual values of the wire drying process strength involved in the embodiments of this application;
[0075] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation
[0076] To better understand the above technical solutions, exemplary embodiments of this disclosure will be described in more detail below with reference to the accompanying drawings. While exemplary embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of this disclosure to those skilled in the art.
[0077] First Embodiment
[0078] Reference Figure 1 This embodiment provides a method for multi-process quality prediction using a neural network constructed by fusing hierarchical graphs. The method includes the following steps:
[0079] S1: Clean, normalize, and process missing values for the collected historical process data of production processes. The historical process data includes the process parameters and corresponding quality indicators of each process.
[0080] Further and optionally, step S1 specifically includes:
[0081] Based on the material head and tail interception rules in the production process, data from the non-steady-state stages of equipment startup and shutdown are removed to complete data cleaning.
[0082] Min-max standardization normalizes the data to a preset interval using the following formula:
[0083]
[0084] In the formula, The original data, For the normalized data, and These are the maximum and minimum values of the original data, respectively. and The endpoints of the target mapping interval.
[0085] S2: Calculate the correlation between each process parameter and the target process quality index, and select parameters with a correlation higher than the preset threshold as model input features;
[0086] Further and optionally, in step S2, the mutual information value between each process parameter and the target process quality index is calculated as the correlation, and the formula is:
[0087]
[0088] In the formula, and These represent sets of variables that respectively represent process parameters and quality indicators. For joint probability distribution, and Each parameter has its own marginal probability distribution; based on the calculated mutual information value, a threshold is set, redundant parameters are eliminated, and key process parameters with mutual information values higher than the threshold are retained as model inputs.
[0089] S3: Model the production process as a hierarchical graph structure of micro-parameter network and macro-process directed chain; the micro-parameter network uses the process parameters in each process as nodes to construct a fully connected candidate graph; the macro-process directed chain uses each process as a node and constructs directed edges along the actual process flow direction.
[0090] Further and optionally, the construction of the hierarchical graph structure in step S3 includes:
[0091] The entire production process Each process is abstracted into macroscopic nodes, and a macroscopic directed graph is constructed. , where the node set correspond Each process, edge collection This represents the chain-like transmission relationship between processes;
[0092] For the process , and at time All key process parameter nodes are grouped into a node set. And it uses a fully connected candidate graph as the basis for the microstructure, edge set any edge in satisfy:
[0093] .
[0094] S4: Construct a deep network model, including:
[0095] Reference Figure 2 Deep network models include:
[0096] The Micrograph GAT Encoder (MGE) module is used to adaptively learn the correlation strength between nodes using a graph attention mechanism on the micro-parameter network, and compress the node-level representation into a sequence of process-level representations through a weighted aggregation and readout mechanism.
[0097] Further and optionally, the MGE micro-coding module satisfies the following steps during execution:
[0098] The process is transformed by linear transformation. At any moment nodes Observations Mapping to the initial embedding representation :
[0099]
[0100] In the formula, and These are the learnable weight matrix and bias vector;
[0101] For any node pair Calculate and normalize the attention score to obtain the implicit edge weights that vary over time. :
[0102]
[0103]
[0104] In the formula, , It is a linear mapping matrix. For attention vectors, This represents vector concatenation. For activation function, For nodes For nodes Contribution weight;
[0105] Update node representations through weighted aggregation:
[0106]
[0107] Attention-based aggregation is used to elevate node-level representations to process-level representations:
[0108]
[0109]
[0110] In the formula, As the importance weight of the nodes, For process At any moment The process level is represented by the process level.
[0111] The CGF macroscopic chain gating fusion module is used to perform gating propagation along the process direction on the directed chain of the macroscopic process, adaptively balance the information of the current process with the information transmitted upstream, and output the fused process representation.
[0112] Further and optionally, the macroscopic chain-gated fusion module (CGF) satisfies the following steps during execution:
[0113] Time step The first process is integrated. For the process , The upstream process is fused and mapped to the feature space of the current process:
[0114]
[0115] In the formula, This is a cross-process mapping matrix;
[0116] The gating mechanism adaptively adjusts the upstream information injection strength, and the gating vector is calculated as follows:
[0117]
[0118] The merged process is represented by a weighted composite output:
[0119]
[0120] In the formula, This represents element-wise multiplication. , These are learnable parameters.
[0121] The temporal decoding module (LSTM Attention Decoder, LAD) performs improved LSTM temporal modeling on the fused process representation sequence and adaptively assigns importance weights for different historical moments using a self-attention mechanism, outputting the predicted quality index value of the target process.
[0122] Further and optionally, in step S4, the LAD timing decoding module satisfies the following steps during execution:
[0123] Improved LSTM unit: Replace the tanh activation function in the traditional LSTM unit with the softsign activation function to stabilize the backpropagation gradient;
[0124] Self-attention mechanism: The hidden sequence obtained by processing the improved LSTM is... Input self-attention mechanism, with the hidden state at the last time step. Generate query vectors The hidden state at each time step is projected as a key vector. Calculate attention weights:
[0125]
[0126] In the formula, The dimension of the key vector;
[0127] A context vector is generated by weighting and aggregating all hidden states based on attention weights. Finally, the context vector Input the multilayer perceptron prediction block to obtain the predicted quality index value of the target process at the next time step. .
[0128] S5: Divide the preprocessed dataset into training set, validation set and test set according to the proportion. Use the training set to iteratively optimize the parameters of the deep network model to minimize the loss function between the predicted value and the true quality index. Then, use the validation set to perform cross-validation and hyperparameter adjustment.
[0129] Further and optionally, the model training in step S5 specifically includes:
[0130] The preprocessed dataset is divided into training set, validation set and test set according to a preset ratio;
[0131] The Adam optimizer is used to iteratively optimize the model parameters, and the mean squared error (MSE) is used as the loss function.
[0132] An early stopping mechanism is used during training. Training is terminated when the validation set loss does not decrease within a preset number of consecutive rounds to prevent overfitting.
[0133] S6: Input the multi-process parameters to be predicted into the trained model, and sequentially process them through MGE micro-encoding, CGF macro-fusion and LAD temporal decoding to output the predicted value of the export quality index of the target process at the next moment.
[0134] Further and optionally, in step S6, the mean absolute error (MAE), root mean square error (RMSE), and goodness of fit are used to measure the results. The model's effectiveness was evaluated using the following formulas:
[0135]
[0136]
[0137]
[0138] In the formula, For the number of samples, This represents the true value of the export quality indicators. For predicted values, This is the mean of the true values.
[0139] Further and optionally, the multi-process parameters include loosening and rehydration, leaf feeding, and thin-plate drying processes in process industrial production, the target process is at least one of the thin-plate drying processes, and the outlet quality index includes the moisture content of the material exiting the thin-plate drying machine.
[0140] Verification of Examples
[0141] Based on the first embodiment, in this embodiment, historical process data from October 1, 2023 to October 3, 2023 was extracted from the server of a tobacco processing production line of a certain cigarette factory, totaling 16,027 samples. This involves three key processes: loosening and rehydration, leaf moistening and feeding, and thin-plate drying. The original process parameters for the loosening and rehydration process are 19 dimensions, for the leaf moistening and feeding process they are 27 dimensions, and for the thin-plate drying process they are 23 dimensions. The quality index set according to the process flow is the moisture content of the outlet material in the thin-plate drying process.
[0142] First, based on the enterprise's material start-up and end-of-line truncation rules, non-steady-state data segments caused by equipment start-up and shutdown are removed to ensure that the training samples mainly come from stable production ranges that meet process requirements. Then, min-max standardization is used to normalize the data to the (0,1) interval to eliminate the influence of dimensions; the formula is: In the formula, The original data, For the normalized data, and These represent the maximum and minimum values of the original data, respectively.
[0143] Furthermore, the mutual information method is used to screen features and calculate the mutual information value between each process parameter and the moisture content of the material exiting the thin-plate drying process. The formula is as follows: Based on the calculated mutual information values, a threshold of 0.1 was set to eliminate redundant parameters. After screening, 12 key process parameters were retained for the loosening and rehydration process, 17 for the leaf moistening and feeding process, and 16 for the thin plate drying process, which were used as input features for subsequent models.
[0144] Furthermore, the production process is modeled as a hierarchical graph structure: Macroscopic layer graph construction: The three processes of loosening and rehydration, leaf moistening and feeding, and thin plate drying are abstracted into three macroscopic nodes, and a macroscopic directed graph is constructed. , where the node set edge set This represents the chain-like transmission relationship between processes.
[0145] Furthermore, regarding the process ( ), and at time All key process parameter nodes are grouped into a node set. Furthermore, a fully connected candidate graph is used as the basis for the microstructure, meaning that there are candidate edges between any two nodes with different parameters.
[0146] Furthermore, an MGE-CGF-LTAD deep network model was constructed, and hyperparameters were adjusted. The key hyperparameter configurations for each module are as follows: In the MGE module, the initial node embedding dimension is 32, the number of GAT layers is 1, the GAT hidden layer dimension is 128, and the GAT output dimension is 32; in the CGF module, the gate vector dimension is 32, and the fusion output dimension is 32; in the LTAD module, the number of LSTM layers is 2, the LSTM hidden layer dimension is 64, the Dropout rate is 0.2, and the number of MLP hidden layers is 1. The representations at each process level are obtained through the MGE micro-encoding module. Subsequently, the final thin-plate wire drying process fusion output is obtained through the CGF macro-chain gated fusion module. And input it into the LAD timing decoding module to generate a context vector. Finally, the data is input into the multilayer perceptron prediction block to obtain the predicted moisture content at the outlet of the thin plate drying wire at the next time step. .
[0147] Furthermore, the preprocessed data was divided into training, validation, and test sets at a ratio of 80%, 10%, and 10%, respectively. The training set was then input into the MGE-CGF-LTAD model, and the Adam optimizer was used for iterative optimization. The initial learning rate was set to 0.001, and the mean squared error (MSE) loss function was used. An early stopping mechanism was implemented during training; training was terminated when the validation set loss did not decrease within 15 consecutive epochs to prevent overfitting. The maximum number of iterations was set to 100, and the sequence length was set to 15. Simultaneously, validation data was input into the model for cross-validation to ensure the model had good generalization ability and stability.
[0148] Furthermore, the multi-process parameters of the batch to be predicted are input into the trained model, which is then sequentially processed through MGE micro-encoding, CGF macro-fusion, and LAD temporal decoding to output the predicted moisture content of the outlet material at the next moment of the thin plate drying process. The model's effectiveness is evaluated using three parameters: goodness of fit (R²), mean absolute error (MAE), and mean absolute percentage error (MAPE). The goodness of fit (R²), mean absolute error (MAE), and mean absolute percentage error (MAPE) are calculated as follows:
[0149]
[0150]
[0151] Among them, the MAE value is greater than 0, and the smaller the value, the better the model prediction effect; the RMSE value is greater than 0, and the smaller the value, the better the model prediction effect; the goodness of fit value is between 0 and 1, and the larger the value, the better the model prediction effect.
[0152] In this embodiment, the results of the three indicators, namely mean absolute error (MAE), root mean square error (RMSE), and goodness of fit (R²), are shown in Table 1 below:
[0153] Table 1. Comparison of Results for the Three Indicators
[0154]
[0155] As shown in the table, the model of this invention has a MAE of 0.0064, an RMSE of 0.0080, and an R² of 0.9816 on the test set. The relatively low MAE and RMSE values, along with the R² value close to 1, indicate a high degree of consistency between the predicted and actual values, demonstrating small prediction bias and high stability of the model. Figure 2 By comparing and analyzing the predicted values with the actual values, it can be seen that the predicted curves and the actual values are highly consistent in terms of overall trend and fluctuation response. This further proves that the multi-process quality prediction method based on MGE-CGF-LTAD of this invention has high prediction accuracy and can be effectively applied to quality prediction and process optimization in multi-process production scenarios in the process industry.
[0156] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.
[0157] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.
Claims
1. A method for multi-process quality prediction using a neural network constructed by fusing hierarchical graphs, characterized in that, The method includes the following steps: S1: Clean, normalize, and process missing values for the collected historical process data of production processes. The historical process data includes the process parameters and corresponding quality indicators of each process. S2: Calculate the correlation between each process parameter and the target process quality index, and select parameters with a correlation higher than the preset threshold as model input features; S3: Model the production process as a hierarchical graph structure of micro-parameter network and macro-process directed chain; the micro-parameter network uses the process parameters in each process as nodes to construct a fully connected candidate graph; the macro-process directed chain uses each process as a node and constructs directed edges along the actual process flow direction. S4: Construct a deep network model, including: The micro-encoding module is used to adaptively learn the correlation strength between nodes using a graph attention mechanism on the micro-parameter network, and compress the node-level representation into a sequence of process-level representations through a weighted aggregation and readout mechanism; The macro-chain gating fusion module is used to perform gating propagation along the process direction on the directed chain of the macro process, adaptively balance the information of the current process with the information transmitted upstream, and output the fused process representation. The temporal decoding module is used to improve the LSTM temporal modeling of the fused process representation sequence, and adaptively assign importance weights for different historical moments by combining a self-attention mechanism, and output the predicted value of the quality index of the target process. S5: Divide the preprocessed dataset into training set, validation set and test set according to the proportion. Use the training set to iteratively optimize the parameters of the deep network model to minimize the loss function between the predicted value and the true quality index. Then, use the validation set to perform cross-validation and hyperparameter adjustment. S6: Input the multi-process parameters to be predicted into the trained model, and sequentially process them through MGE micro-encoding, CGF macro-fusion and LAD temporal decoding to output the predicted value of the export quality index of the target process at the next moment.
2. The neural network multi-process quality prediction method based on the fusion hierarchical graph construction as described in claim 1, characterized in that, Step S1 specifically includes: Based on the material head and tail interception rules in the production process, data from the non-steady-state stages of equipment startup and shutdown are removed to complete data cleaning. Min-max standardization normalizes the data to a preset interval using the following formula: ； In the formula, The original data, For the normalized data, and These are the maximum and minimum values of the original data, respectively. and The endpoints of the target mapping interval.
3. The neural network multi-process quality prediction method based on the fusion hierarchical graph construction as described in claim 1, characterized in that, In step S2, the mutual information value between each process parameter and the target process quality index is calculated as the correlation, and the formula is as follows: ； In the formula, and These represent sets of variables that respectively represent process parameters and quality indicators. For joint probability distribution, and Each parameter has its own marginal probability distribution; based on the calculated mutual information value, a threshold is set, redundant parameters are eliminated, and key process parameters with mutual information values higher than the threshold are retained as model inputs.
4. The method for multi-process quality prediction using a neural network constructed from a fused hierarchical graph as described in claim 1, characterized in that, The construction of the hierarchical graph structure in step S3 includes: The entire production process Each process is abstracted into macroscopic nodes, and a macroscopic directed graph is constructed. , where the node set correspond Each process, edge collection This represents the chain-like transmission relationship between processes; For the process , and at time All key process parameter nodes are grouped into a node set. And it uses a fully connected candidate graph as the basis for the microstructure, edge set any edge in satisfy: 。 5. The neural network multi-process quality prediction method based on the fusion hierarchical graph construction as described in claim 1, characterized in that, In step S4, the MGE micro-coding module satisfies the following steps during execution: The process is transformed by linear transformation. At any moment nodes Observations Mapping to the initial embedding representation : ； In the formula, and These are the learnable weight matrix and bias vector; For any node pair Calculate and normalize the attention score to obtain the implicit edge weights that vary over time. : ； ； In the formula, , It is a linear mapping matrix. For attention vectors, This represents vector concatenation. For activation function, For nodes For nodes Contribution weight; Update node representations through weighted aggregation: ； Attention-based aggregation is used to elevate node-level representations to process-level representations: ； ； In the formula, As the importance weight of the nodes, For process At any moment The process level is represented by the process level.
6. The method for multi-process quality prediction of a neural network constructed by fusing hierarchical graphs as described in claim 1, characterized in that, In step S4, the CGF macroscopic chain-gated fusion module satisfies the following steps during execution: Time step The first process is integrated. For the process , The upstream process is fused and mapped to the feature space of the current process: ； In the formula, This is a cross-process mapping matrix; The gating mechanism adaptively adjusts the upstream information injection strength, and the gating vector is calculated as follows: ； The merged process is represented by a weighted composite output: ； In the formula, This represents element-wise multiplication. , These are learnable parameters.
7. The neural network multi-process quality prediction method based on the fusion hierarchical graph construction as described in claim 1, characterized in that, In step S4, the LAD timing decoding module satisfies the following steps during execution: Improved LSTM unit: Replace the tanh activation function in the traditional LSTM unit with the softsign activation function to stabilize the backpropagation gradient; Self-attention mechanism: The hidden sequence obtained by processing the improved LSTM is... Input self-attention mechanism, with the hidden state at the last time step. Generate query vectors The hidden state at each time step is projected as a key vector. Calculate attention weights: ； In the formula, The dimension of the key vector; A context vector is generated by weighting and aggregating all hidden states based on attention weights. Finally, the context vector Input the multilayer perceptron prediction block to obtain the predicted quality index value of the target process at the next time step. .
8. The method for multi-process quality prediction of neural networks constructed by fusing hierarchical graphs as described in claim 1, characterized in that, The model training in step S5 specifically includes: The preprocessed dataset is divided into training set, validation set and test set according to a preset ratio; The Adam optimizer is used to iteratively optimize the model parameters, and the mean squared error (MSE) is used as the loss function. An early stopping mechanism is used during training. Training is terminated when the validation set loss does not decrease within a preset number of consecutive rounds to prevent overfitting.
9. The neural network multi-process quality prediction method based on the fusion hierarchical graph construction as described in claim 1, characterized in that, In step S6, the mean absolute error (MAE), root mean square error (RMSE), and goodness of fit are used. The model's effectiveness was evaluated using the following formulas: ； ； ； In the formula, For the number of samples, This represents the true value of the export quality indicators. For predicted values, This is the mean of the true values.
10. The method for multi-process quality prediction of a neural network based on the fusion of hierarchical graphs as described in claim 1, characterized in that, The multi-process parameters include the loosening and rehydration, leaf feeding, and thin plate drying processes in process industrial production. The target process is at least one of the thin plate drying processes, and the outlet quality index includes the moisture content of the material exiting the thin plate drying machine.