Method for predicting moisture content of loosening and moisture regaining outlet based on graph neural network

Abstract

A plurality of sensors are arranged on a tobacco leaf moisture regaining production line and used for collecting the process flow, the water adding proportion, the gas-water mixing automatic valve opening degree, the process hot air temperature, the water adding flow, the material cumulant, the water adding cumulant, the steam automatic valve opening degree and the discharging water content of tobacco leaves in real time. The method for predicting the moisture content of the loosening and moisture regaining outlet based on the graph neural network comprises the following steps: acquiring data, preprocessing the data, performing Spearman correlation analysis on the parameters, constructing graph data, training a model by using multiple groups of associated data, and predicting the moisture content of the tobacco leaf outlet at the next moment of the loosening and moisture regaining process by using the trained model. According to the method, the GAT is used for capturing the complex incidence relation among the parameters in the production data, the time sequence data is processed through the LSTM, the spatial features and the time features are effectively combined, the prediction precision of the outlet moisture content is remarkably improved, and the experimental result shows that the model is excellent in MAE evaluation indexes.

Description

Technical Field

[0001] The invention relates to a method for predicting the moisture content of a loose rehydration outlet, in particular to a method for predicting the moisture content of a loose rehydration outlet based on a graph neural network. Background Art

[0002] At present, tobacco factories mostly use closed-loop control systems to stabilize the moisture content at the outlet. The system monitors the moisture content and temperature of tobacco leaves in real time through sensors, and feeds the measured data back to the controller. The controller compares these data with the preset target values, calculates the error, and generates adjustment instructions through algorithms. The actuator dynamically adjusts the operation of the heating and humidification equipment based on this. The sensor measures the adjusted parameters again to form a closed-loop control to ensure that the moisture content is stable within the preset range.

[0003] In closed-loop control systems, PID controllers are commonly used control algorithms. PID controllers precisely adjust the intensity of heating and humidification through the combined effects of proportion (P), integration (I) and differentiation (D). Proportional control is proportional to the error and responds quickly; integral control eliminates steady-state errors; and differential control makes advance adjustments based on the error change trend to improve system stability and response speed.

[0004] The prior art has the following disadvantages:

[0005] 1) Response speed limitation: The response speed of the PID controller is limited by the settings of the proportional, integral and differential parameters. It cannot quickly respond to sudden large changes, resulting in system adjustment lag or overshoot.

[0006] 2) Poor adaptability under complex process conditions: During the production process of a tobacco factory, environmental conditions and raw material characteristics may undergo complex changes. Traditional PID controllers are difficult to respond flexibly and perform poorly in nonlinear or multivariable systems.

[0007] 3) Difficulty in parameter adjustment: The PID controller requires precise adjustment of three parameters (Kp, Ki, Kd), which requires high professional knowledge and experience of technicians, and the parameter adjustment process is time-consuming and labor-intensive.

[0008] 4) Sensitive to noise: PID controller, especially the differential control part, is sensitive to noise. Measurement noise may introduce high-frequency oscillations, resulting in system instability.

[0009] 5) Lack of predictive capabilities: Traditional PID controllers are mainly adjusted based on current and historical errors. They lack the ability to predict future trends and are difficult to deal with potential problems in advance. Summary of the invention

[0010] The purpose of the invention of this application is to solve the problems of slow response speed, poor adaptability, difficulty in parameter adjustment, sensitivity to noise and lack of prediction ability existing in the traditional PID control system in the loose rehumidification process of tobacco production, and to provide a new loose rehumidification outlet moisture content prediction method to improve the prediction accuracy and stability, which is achieved by introducing a graph attention network (GAT) and a long short-term memory network (LSTM).

[0011] In order to achieve the invention purpose of this application, this application adopts the following technical solutions:

[0012] The present invention discloses a method for predicting the moisture content of a loose rehydration outlet based on a graph neural network. A plurality of sensors are arranged on a tobacco leaf rehydration production line. The sensors are used to collect the process flow rate, water addition ratio, gas-water mixing automatic valve opening, process hot air temperature, water addition flow rate, material cumulative amount, water addition cumulative amount, steam automatic valve opening and discharge moisture content of tobacco leaves in real time. The method for predicting the moisture content of a loose rehydration outlet based on a graph neural network includes the following steps:

[0013] 1. Data Collection

[0014] The above multiple sensors are used to collect the process flow rate, water addition ratio, gas-water mixing automatic valve opening, process hot air temperature, water addition flow rate, material cumulative amount, water addition cumulative amount, steam automatic valve opening and discharge moisture content of tobacco leaves in real time, and the above nine data at each moment are combined into a group of original parameters, and multiple groups of original parameters are collected;

[0015] (II) Data preprocessing

[0016] The above multiple groups of original parameters are processed as follows

[0017] (1) Processing outliers of original parameters

[0018] Calculate the average value of each parameter of a set of original data, detect the above parameters, and when the above parameters are within the range of ±3 times the standard deviation of their average values, they are regarded as constant values ​​and retained; otherwise, they are regarded as abnormal values ​​and the group of original parameters including the abnormal value data is deleted;

[0019] (2) Standardize the data processed in step (1)

[0020] When the above parameters are within their range, they are processed according to the following formula to normalize the above parameters to between 0 and ±1.

[0021]

[0022] Where: Z is the standardized value (Z-Score); X is the original data; μ is the mean of each parameter within its range; σ is the standard deviation of each parameter within its range;

[0023] (III) Spearman correlation analysis of the above parameters

[0024] At every moment, the Spearman coefficient between nine parameters including process flow, water addition ratio, gas-water mixing automatic valve opening, process hot air temperature, water addition flow, material cumulative amount, water addition cumulative amount, steam automatic valve opening and discharge moisture content is calculated according to the Spearman correlation analysis;

[0025] (IV) Graph Data Construction

[0026] (a) In a graph, nine points at the same time are used to represent the process flow, water addition ratio, gas-water mixing automatic valve opening, process hot air temperature, water addition flow, material cumulative amount, water addition cumulative amount, steam automatic valve opening and discharge moisture content respectively. Lines are connected according to the Spearman coefficients between the nine parameters calculated above. When the Spearman coefficient between the above two parameters is greater than the set threshold, a line is connected between the two parameters, otherwise, no line is connected; according to the above method, the parameters at all times are converted into graph data, and the graph data are arranged in chronological order;

[0027] (b) Divide all graph data into input data and predicted values

[0028] A sliding window is used to divide all the above graph data, and the size of the window is set to at least 10 graph data. One graph data is moved each time by sliding, which means that at least 10 consecutive graph data are extracted each time as input features; the moisture content of the discharge of at least 10 graph data is used as a label value, that is, the predicted target value. The at least 10 graph data extracted continuously and the label value are used as a group of associated data, and multiple groups of associated data are obtained according to this method;

[0029] (V) Training the model with multiple sets of related data

[0030] The model includes: a graph attention network layer (GAT), a long short-term memory network layer (LSTM), and a fully connected layer connected in sequence

[0031] 80% of the above multiple groups of associated data constitute a training set, 10% of the above multiple groups of associated data constitute a validation set, and 10% of the above multiple groups of associated data constitute a test set. At least 10 graph data of each group of associated data in the training set are used as input and input into the graph attention network (GAT) layer. The features output by the graph attention network (GAT) layer are input into the long short-term memory network (LSTM) layer. The features output by the long short-term memory network (LSTM) layer are input into the fully connected layer to obtain the final prediction value. During the training process, the above operations are performed on all the multiple groups of associated data in the training set in turn to complete a complete round of iterations. In each round of training of the training set, the data of the training set is first input into the model in batches, and the model output is calculated by forward propagation, and then Update the model parameters through back propagation, and record the loss value of the training set at the same time. After completing the traversal of all the training set data, enter the verification stage, input all the data of the verification set into the model, and calculate the loss value of the verification set. In order to optimize the training process, introduce the learning rate scheduler to dynamically adjust the learning rate according to the change of the loss value of the verification set. If the loss value of the verification set in this round is lower than the lowest value of the loss value of the verification set in the previous round, then save the current model as the candidate best model. Repeat the above process for at least 50 rounds to fully train the model. During the whole process, the model parameters are gradually optimized. The learning rate scheduler ensures that the weight parameters are adjusted more finely in the later stage of training, thereby improving the model performance. Finally, select the model with the lowest loss value of the verification set as the final trained model;

[0032] (VI) Use the trained model to predict the tobacco leaf outlet moisture content at the next moment of the loosening and rehumidification process

[0033] Collect in real time the tobacco process flow, water addition ratio, air-water mixing automatic valve opening, process hot air temperature, water addition flow, material cumulative amount, water addition cumulative amount, steam automatic valve opening and discharge moisture content of the tobacco leaf rehumidification production line at least 10 consecutive time points, repeat step (ii) (2) to step (iv) (b) to obtain at least 10 graph data, input the above graph data into the trained model, and predict the tobacco leaf outlet moisture content at the next moment of the loosening and rehumidification process.

[0034] The loose rehydration outlet moisture content prediction method based on graph neural network of the present invention, wherein: in step (five), it also includes: using a test set to test the optimal model.

[0035] The loose rehydration outlet moisture content prediction method based on graph neural network of the present invention, wherein: the Spearman coefficient of the step (iii) is calculated using SPSS software.

[0036] The method for predicting the moisture content of the loose rehydration outlet based on the graph neural network of the present invention, wherein: the standard deviation Where: N is the total number of data points; x i is the data of the ith data point; μ is the mean of the population, and the calculation formula is:

[0037] The method for predicting the moisture content of the loose rewetting outlet based on the graph neural network of the present invention, wherein: the loss value is the mean square error Where: n is the total number of samples; y i is the true value of the i-th sample; is the predicted value of the i-th sample.

[0038] The present invention provides a loose rehydration outlet moisture content prediction method based on a graph neural network, wherein: the node input feature dimension of the graph attention network layer is 9, the hidden layer dimension is 64, and the multi-head attention mechanism is set to 2.

[0039] The loose rehydration outlet moisture content prediction method based on graph neural network of the present invention, wherein: in step (iv), the set threshold is 0.5.

[0040] The present invention provides a method for predicting the moisture content at the loose rehydration outlet based on a graph neural network, wherein: in step (five), the learning rate scheduler is ReduceLROnPlateau.

[0041] The loose regain outlet moisture content prediction method based on graph neural network of the present invention, wherein: the size of the setting window is 15, which means that 15 consecutive graph data are extracted as input features each time, and the discharge moisture content of the 16th graph data is used as the label value; each time a graph data is moved along the arrangement of the graph data, then the consecutive 15 graph data in the window are used as the next input feature, and the discharge moisture content of the 17th graph data is used as the next label value.

[0042] The loose rehydration outlet moisture content prediction method based on graph neural network of the present invention, wherein: the number of graph data input into the model in step (six) is the same as the number of graph data input into the model in step (five).

[0043] The loose rehydration outlet moisture content prediction method based on graph neural network of the present invention solves the following technical problems:

[0044] Response speed problem: Traditional PID controllers have a slow response speed when dealing with sudden and large changes, resulting in system adjustment lag or overshoot. This invention uses a model combining GAT and LSTM, and utilizes the efficient feature extraction of graph networks and the time series information processing of LSTM to quickly respond to system changes and improve the adjustment speed.

[0045] Adaptability issues: Traditional PID controllers perform poorly in complex process conditions and multivariable systems. The GAT-LSTM model constructs a graph data structure through a variety of correlation analysis methods to capture the complex relationship between parameters and enhance the system's adaptability to nonlinear and dynamic changes.

[0046] Difficulty in parameter adjustment: The PID controller requires precise adjustment of three parameters, which is time-consuming, labor-intensive, and requires high technical staff. The GAT-LSTM model automatically adjusts and optimizes parameters through deep learning, reducing the complexity and time cost of manual adjustment.

[0047] Noise sensitivity problem: PID controllers are sensitive to noise, and measurement noise may introduce high-frequency oscillations. This paper uses the graph attention mechanism of GAT and the time series processing of LSTM to enhance the robustness of the model and reduce the impact of noise.

[0048] Lack of prediction ability: Traditional PID controllers lack the ability to predict future trends and are difficult to pre-process potential problems. This invention uses the time series prediction ability of LSTM and combines it with the structural feature extraction of GAT to achieve high-precision prediction of the outlet moisture content, adjust the process parameters in advance, and ensure the stability and efficiency of the production process.

[0049] By solving the above technical problems, the present invention improves the prediction accuracy of the moisture content at the loose rehumidification outlet, enhances the stability and adaptability of the system, reduces maintenance costs, and improves the efficiency and economic benefits of the entire tobacco production line.

[0050] The loose regain outlet moisture content prediction method based on graph attention network (GAT) and long short-term memory network (LSTM) of the present invention has the following significant advantages and effects:

[0051] 1. Improve prediction accuracy:

[0052] The present invention uses GAT to capture the complex correlation between various parameters in production data, processes time series data through LSTM, effectively combines spatial features and temporal features, and significantly improves the prediction accuracy of outlet moisture content. Experimental results show that the model is effective in MAE, MSE and R 2 The evaluation indicators such as 0.0151, 0.0011 and 0.9932 were all excellent, which were significantly better than the traditional methods.

[0053] (II) Enhance system stability:

[0054] By constructing a graph data model and utilizing the powerful feature extraction capability of the GAT network, the present invention can more stably capture and process the complex relationship between various parameters in the production process, improve the adaptability of the system under different process conditions, and ensure stable control of the outlet moisture content.

[0055] (III) Fast response speed:

[0056] The present invention combines the advantages of GAT and LSTM, can quickly respond to changes in parameters in the production process, realize real-time prediction and adjustment, avoid the problems of slow response speed and system adjustment lag of traditional PID controller, and improve production efficiency.

[0057] (IV) Reduce the complexity of parameter adjustment:

[0058] Traditional PID controllers require precise adjustment of multiple parameters (Kp, Ki, Kd), which is time-consuming and labor-intensive and requires high technical staff. The present invention automatically optimizes parameters through a deep learning model, reducing the complexity and time cost of manual adjustment and lowering the requirements for technical staff.

[0059] (V) Improve robustness:

[0060] Traditional PID controllers are sensitive to noise and are prone to high-frequency oscillations due to measurement noise. This paper uses the graph attention mechanism of GAT and the time series processing capability of LSTM to enhance the robustness to noise and reduce the impact of noise on the system.

[0061] 6. Possessing predictive capabilities:

[0062] Traditional PID controllers are mainly adjusted based on current and historical errors, and lack the ability to predict future trends. This invention uses the time series prediction capability of LSTM and combines it with the structural feature extraction of GAT to achieve high-precision prediction of outlet moisture content, which can adjust process parameters in advance, prevent potential problems, and ensure the stability and efficiency of the production process.

[0063] (VII) Improve production efficiency and tobacco leaf quality:

[0064] Through the implementation of the present invention, the moisture content at the loose rehumidification outlet can be controlled more accurately, the process parameters can be optimized, the processing resistance and hygroscopicity of tobacco leaves can be improved, thereby improving the quality of tobacco leaves and the efficiency and economic benefits of the entire tobacco production line. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 It is a flow chart of the method for predicting the moisture content at the outlet of loose rewetting based on graph neural network of the present invention;

[0066] Figure 2 A schematic diagram of a graph data based on the Manspeare coefficients between the various parameters in Table 1;

[0067] Figure 3 Schematic diagram of the model used in the prediction method of the present invention. DETAILED DESCRIPTION

[0068] like Figure 1 As shown, a plurality of sensors are arranged on the production line of tobacco leaf moisture recovery, which are used to collect the process flow rate, water addition ratio, gas-water mixing automatic valve opening, process hot air temperature, water addition flow rate, material cumulative amount, water addition cumulative amount, steam automatic valve opening and discharge moisture content of tobacco leaves in real time. The method for predicting the moisture content of loose moisture recovery outlet based on graph neural network of the present invention comprises the following steps:

[0069] 1. Data Collection

[0070] The above multiple sensors are used to collect the process flow rate, water addition ratio, gas-water mixing automatic valve opening, process hot air temperature, water addition flow rate, material cumulative amount, water addition cumulative amount, steam automatic valve opening and discharge moisture content of tobacco leaves in real time, and the above nine data at each moment are combined into a group of original parameters, and multiple groups of original parameters are collected;

[0071] (II) Data preprocessing

[0072] The above multiple groups of original parameters are processed as follows

[0073] (1) Processing outliers of original parameters

[0074] Calculate the average value of each parameter of a set of original data, test the above parameters, and when the above parameters are within the range of their average value ± 3 times the standard deviation, they are regarded as constant values ​​and retained; otherwise, they are regarded as abnormal values ​​and the group of original parameters including the abnormal value data is deleted. Where: N is the total number of data points; x i is the data of the ith data point; μ is the mean of the population, and the calculation formula is:

[0075] (2) Standardize the data processed in step (1)

[0076] When the above parameters are within their range, they are processed according to the following formula to normalize the above parameters to between 0 and ±1.

[0077]

[0078] Where: Z is the standardized value (Z-Score); X is the original data; μ is the mean of each parameter within its range; σ is the standard deviation of each parameter within its range;

[0079] (III) Spearman correlation analysis of the above parameters

[0080] At every moment, the Spearman correlation analysis is used to calculate the Spearman coefficient between the nine parameters of process flow, water addition ratio, gas-water mixing automatic valve opening, process hot air temperature, water addition flow, material cumulative amount, water addition cumulative amount, steam automatic valve opening and discharge moisture content. The Spearman coefficient is calculated using SPSS software;

[0081] Using Spearman correlation analysis, the preprocessed data was subjected to correlation analysis to clarify the mutual influence relationship between the parameters and obtain the correlation coefficients between the various factors, as shown in Table 1 below:

[0082]

[0083] Table 1

[0084] (IV) Graph Data Construction

[0085] (a) Figure 2 As shown, in one graph, nine points at the same time are used to represent the process flow, water addition ratio, gas-water mixing automatic valve opening, process hot air temperature, water addition flow, material cumulative amount, water addition cumulative amount, steam automatic valve opening and discharge moisture content, and lines are connected according to the Spearman coefficients between the nine parameters calculated above. When the Spearman coefficient between the above two parameters is greater than the set threshold value = 0.5, a line is connected between the two parameters, otherwise, no line is connected; according to the above method, the parameters at all times are converted into graph data, and the graph data are arranged in chronological order;

[0086] (b) Divide all graph data into input data and predicted values

[0087] All the above graph data are divided by a sliding window, and the window size is set to 15, which means that 15 consecutive graph data are extracted each time as input features, and the moisture content of the 16th graph data is used as the label value; each time a graph data is moved along the arrangement of the graph data, then the 15 consecutive graph data in the window are used as the next input feature, and the moisture content of the 17th graph data is used as the next label value. The above 15 consecutively extracted graph data and the above label value are used as a group of associated data, and multiple groups of associated data are obtained according to this method;

[0088] (V) Training the model with multiple sets of related data

[0089] The model includes: a graph attention network layer (GAT), a long short-term memory network layer (LSTM), and a fully connected layer connected in sequence;

[0090] The node input feature dimension of the graph attention network layer (GAT) is 9, the hidden layer dimension is 64, and the multi-head attention mechanism is set to 2 (for the graph attention network layer, see the literature: Velikovi P, Cucurull G, Casanova A, et al. GraphAttention Networks[J]. 2017.)

[0091] The GAT network performs a linear transformation on the feature vector of each node and maps it to a high-dimensional space. Then, the attention coefficient of the node and its neighboring nodes is calculated, and the attention coefficient is standardized by the LeakyReLU activation function and the softmax function to measure the importance of the neighboring nodes to the target node. Next, a weighted aggregation method is used to aggregate the features of the neighboring nodes to the target node to generate a new node feature representation. In order to enhance the stability and expression ability of the model, GAT adopts a multi-head attention mechanism to calculate the attention coefficient and aggregated features for each head, and splice the output features of all heads together to form the final node feature representation. Through this series of steps, the GAT module effectively captures and reflects the complex correlation between process parameters, providing richer and higher-quality input features to the subsequent long short-term memory network (LSTM) module, thereby improving the accuracy and performance of the overall prediction model.

[0092] Long short-term memory network layer (LSTM) (For the long short-term memory network layer, see the reference: Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long Short-Term Memory. Neural Comput. 9, 8 (November 15, 1997), 1735–1780).

[0093] The features output by the GAT module are input into the long short-term memory network (LSTM) module to further capture the long-term and short-term dependencies in the time series. The LSTM module controls the flow of information through the forget gate, input gate and output gate through its unique gating mechanism. The forget gate determines how much memory of the previous moment is retained, the input gate determines the amount of current input information added, and the output gate controls the output of the hidden state at the current moment.

[0094] Through these gating mechanisms, the LSTM module can effectively capture the long-term and short-term dependencies in the time series, ensure the effective learning of long-term dependent features, and input the obtained output into the fully connected layer and map it to the final predicted value. The loss function (mean square error MSE) is calculated according to the label value of the training set. The model parameters are updated through the back-propagation algorithm so that the prediction results gradually approach the true labels. Then the network weights are updated through gradient descent (Adam optimizer) to minimize the loss function and iterate multiple training. After the training is completed, the model is evaluated using the test set. By forward propagating the test set data, we get the final prediction result and then calculate the prediction error of the model. In order to measure the performance of the model, the mean square error (MSE) is used as the evaluation indicator.

[0095] Fully connected layer (For more information on fully connected layers, see Chapter 4 of Deep Learning for Time Series Forecasting) Fully connected layers are introduced in the article on using deep learning for time series forecasting.

[0096] 80% of the above multiple groups of associated data form a training set, 10% of the above multiple groups of associated data form a validation set, and 10% of the above multiple groups of associated data form a test set. At least 10 graph data of each group of associated data in the training set are used as input and input into the graph attention network (GAT) layer. The features output by the graph attention network (GAT) layer are input into the long short-term memory network (LSTM) layer. The features output by the long short-term memory network (LSTM) layer are input into the fully connected layer to obtain the final prediction value. During the training process, the above operations are performed on all multiple groups of associated data in the training set in turn to complete a complete round of iterations. In each round of training of the training set, the data of the training set is first input into the model in batches, and the model output is calculated by forward propagation. Then, the model parameters are updated through back propagation, and the loss of the training set is recorded at the same time. After traversing all the training set data, the validation phase begins. All the validation set data is input into the model and the validation set loss value is calculated. In order to optimize the training process, a learning rate scheduler is introduced. The learning rate scheduler is ReduceLROnPlateau. The learning rate is adjusted dynamically according to the change of the validation set loss value. If the validation set loss value of this round is lower than the lowest value of the validation set loss value of the previous round, the current model is saved as the candidate best model. The above process is repeated for at least 50 rounds to fully train the model. During the whole process, the model parameters are gradually optimized. The learning rate scheduler ensures that the weight parameters are adjusted more finely in the later stage of training, thereby improving the model performance. Finally, the model with the lowest validation set loss value is selected as the final trained model. Finally, the optimal model is tested using the test set.

[0097] The loss value is the mean square error Where: n is the total number of samples; y i is the true value of the i-th sample; is the predicted value of the i-th sample;

[0098] (VI) Use the trained model to predict the tobacco leaf outlet moisture content at the next moment of the loosening and rehumidification process

[0099] The tobacco process flow rate, water addition ratio, air-water mixing automatic valve opening, process hot air temperature, water addition flow rate, material cumulative amount, water addition cumulative amount, steam automatic valve opening and discharge moisture content on the tobacco leaf rehumidification production line are collected in real time at 15 consecutive time points, and step (ii) (2) to step (iv) (b) are repeated to obtain 15 graph data. The above graph data are input into the trained model to predict the tobacco leaf outlet moisture content at the next moment of the loosening and rehumidification process.

[0100] The specific effects of the loose regain outlet moisture content prediction method based on graph neural network of the present invention are as follows:

[0101] Experimental verification: Through experimental verification of the production data of a cigarette factory from January to December 2023, the GAT-LSTM model proposed in this paper outperforms other existing machine learning and deep learning models in multiple performance indicators, especially in MAE, MSE and R 2 The key indicators are significantly better than traditional methods.

[0102] System application: In practical applications, the present invention can monitor and predict the outlet moisture content of the loosening and regaining process in real time, and automatically adjust the process parameters according to the prediction results to ensure the stability of the outlet moisture content and improve the automation level and efficiency of the production process.

[0103] like Figure 3 As shown in the figure, the model first extracts the time series data of the loose rehydration process from the sensor from the manufacturing execution system (MES), and then performs the data outlier processing and normalization preprocessing steps. The data is analyzed by multiple correlations through Spearman correlation analysis to clarify the mutual influence relationship between the parameters. At the same time, the data is segmented, and a graph data model is constructed based on these correlations for the segmented data, with each parameter as a node and the correlation as an edge, so as to transform the data into a graph structure data that can reflect the relationship between the parameters.

[0104] The processed graph data is input into the GAT module, which processes the graph data using the graph attention mechanism to extract the features and relationships between nodes (parameters). Specifically, the GAT module performs weighted aggregation on the neighborhood of each node and learns the weights between nodes through the attention mechanism to better represent the features of the nodes. After being processed by the GAT module, the feature representation obtained contains the high-order relationship information between the parameters.

[0105] The features output by the GAT module are input into the LSTM module. The LSTM module is responsible for capturing the long-term and short-term dependencies in the time series to make full use of the information in the time dimension. Through its gating mechanism, LSTM selectively memorizes and forgets information to ensure effective learning of long-term dependencies. Under the processing of LSTM, the model can combine time series features and parameter relationships to achieve more accurate outlet moisture content prediction.

[0106] Finally, the output of LSTM is mapped to the target prediction value, that is, the outlet moisture content of the loosening and rehydration process, through the fully connected layer. Through this multi-level processing method combining GAT and LSTM, the model can fully explore and utilize the complex relationship and time series information between parameters, thereby achieving higher prediction accuracy.

[0107] It is particularly important to emphasize that the Spearman correlation analysis is performed on the pre-processed data, and the steps are as follows:

[0108] 1) Sort each parameter within its range to obtain the ranking of each parameter;

[0109] 2) Calculate the difference d between the two parameter rankings i , that is: d i =R(X i )-R(Y i ), where R(X i ) and R(Y i ) are the rankings of a parameter X i and another parameter rank Y i rank;

[0110] 3) Calculate the sum of squares of all ranking differences between the two parameters:

[0111] 4) Use the following formula to calculate the Spearman coefficient:

[0112] The value of the Spearman correlation coefficient is between -1 and 1, indicating different degrees of correlation. The larger the absolute value, the stronger the correlation. The Spearman coefficient is calculated using SPSS software.

[0113] The loose rehydration outlet moisture content prediction method based on graph neural network of the present invention has the following innovative features:

[0114] 1. The method of the present application is to apply the Spearman analysis method to the time series data analysis in the tobacco production process. By revealing the potential correlation between process parameters, a relationship graph structure is constructed, which provides richer input features for subsequent prediction models and significantly improves the prediction accuracy and stability.

[0115] 2. The method of this application proposes for the first time to combine the graph attention network (GAT) with the long short-term memory network (LSTM), and use the spatial feature capture ability of GAT to reveal the complex interactions between production parameters. GAT dynamically adjusts the importance of each parameter through the self-attention mechanism, thereby improving the ability to accurately predict the tobacco production process.

[0116] 3. The method of this application combines LSTM to capture the long-term and short-term dependencies in time series data. This model effectively handles the dynamic changes in the tobacco production process, especially in the prediction of the moisture content of the output material, significantly improving the prediction accuracy and performing better than traditional methods.

[0117] 4. Effect of this application

[0118] The MAE of the GAT-LSTM model is 0.0151, the MSE is 0.0011, and the R 2 The MAE and MSE indicators show significantly better results than other traditional machine learning and deep learning models. Compared with random forest, LightGBM and KNN, GAT-LSTM shows obvious advantages in prediction accuracy and stability, especially in processing complex data patterns. Compared with traditional RNN and GRU models, GAT-LSTM reduces MAE by 0.0117 and 0.052 respectively, reflecting its further improvement in accuracy and efficiency. Overall, GAT-LSTM performs best among all compared models, fully verifying the excellent performance of this model.

[0119] In addition, this model has been successfully applied to real-time data analysis of silk production lines. It can process and predict changes in key parameters in the production process in real time, provide accurate prediction results, and assist in the optimization of the production process and decision support.

[0120] In summary, the present invention significantly improves the prediction accuracy of loose rehumidification outlet moisture content and the stability of the system by combining the advantages of graph attention network and long short-term memory network, reduces maintenance costs and complexity of manual adjustment, has strong prediction ability and robustness, can effectively improve the efficiency of tobacco production and the quality of tobacco leaves, and has important practical application value and promotion prospects.

[0121] The above is only one embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes that can be easily thought of by any technician familiar with the technical field within the technical scope disclosed in the present invention should be covered within the protection scope of the present invention.

Claims

1. A method for predicting the moisture content of loose rehumidification outlet based on graph neural network, wherein multiple sensors are arranged on the tobacco rehumidification production line, which are used to collect the process flow rate, water addition ratio, gas-water mixing automatic valve opening, process hot air temperature, water addition flow rate, material cumulative amount, water addition cumulative amount, steam automatic valve opening and discharge moisture content of tobacco leaves in real time, characterized in that: The loose rewetting outlet moisture content prediction method based on graph neural network includes the following steps:

1. Data Collection The above multiple sensors are used to collect the process flow rate, water addition ratio, gas-water mixing automatic valve opening, process hot air temperature, water addition flow rate, material cumulative amount, water addition cumulative amount, steam automatic valve opening and discharge moisture content of tobacco leaves in real time, and the above nine data at each moment are combined into a group of original parameters, and multiple groups of original parameters are collected; (II) Data preprocessing The above multiple groups of original parameters are processed as follows (1) Processing outliers of original parameters Calculate the average value of each parameter of a set of original data, detect the above parameters, and when the above parameters are within the range of ±3 times the standard deviation of their average values, they are regarded as constant values ​​and retained; otherwise, they are regarded as abnormal values ​​and the group of original parameters including the abnormal value data is deleted; (2) Standardize the data processed in step (1) When the above parameters are within their range, they are processed according to the following formula to normalize the above parameters to between 0 and ±1. Where: Z is the standardized value (Z-Score); X is the original data; μ is the mean of each parameter within its range; σ is the standard deviation of each parameter within its range; (III) Spearman correlation analysis of the above parameters At every moment, the Spearman coefficient between nine parameters including process flow, water addition ratio, gas-water mixing automatic valve opening, process hot air temperature, water addition flow, material cumulative amount, water addition cumulative amount, steam automatic valve opening and discharge moisture content is calculated according to the Spearman correlation analysis; (IV) Graph Data Construction (a) In a graph, nine points at the same time are used to represent the process flow, water addition ratio, gas-water mixing automatic valve opening, process hot air temperature, water addition flow, material cumulative amount, water addition cumulative amount, steam automatic valve opening and discharge moisture content respectively. Lines are connected according to the Spearman coefficients between the nine parameters calculated above. When the Spearman coefficient between the above two parameters is greater than the set threshold, a line is connected between the two parameters, otherwise, no line is connected; according to the above method, the parameters at all times are converted into graph data, and the graph data are arranged in chronological order; (b) Divide all graph data into input data and predicted values A sliding window is used to divide all the above graph data, and the size of the window is set to at least 10 graph data. One graph data is moved each time by sliding, which means that at least 10 consecutive graph data are extracted each time as input features; the moisture content of the discharge of at least 10 graph data is used as a label value, that is, the predicted target value. The at least 10 graph data extracted continuously and the label value are used as a group of associated data, and multiple groups of associated data are obtained according to this method; (V) Training the model with multiple sets of related data The model includes: a graph attention network layer (GAT), a long short-term memory network layer (LSTM) and a fully connected layer connected in sequence. 80% of the above multiple groups of associated data constitute a training set, 10% of the above multiple groups of associated data constitute a validation set, and 10% of the above multiple groups of associated data constitute a test set. At least 10 graph data of each group of associated data in the training set are used as input and input into the graph attention network (GAT) layer. The features output by the graph attention network (GAT) layer are input into the long short-term memory network (LSTM) layer. The features output by the long short-term memory network (LSTM) layer are input into the fully connected layer to obtain the final prediction value. During the training process, the above operations are performed on all multiple groups of associated data in the training set in sequence to complete a complete round of iterations. In each round of training of the training set, the data of the training set is firstly processed in batches. Input into the model, perform forward propagation to calculate the model output, and then update the model parameters through back propagation, while recording the loss value of the training set. After completing the traversal of all training set data, enter the verification stage, input all the data of the verification set into the model, and calculate the loss value of the verification set. In order to optimize the training process, introduce the learning rate scheduler to dynamically adjust the learning rate according to the change of the loss value of the verification set. If the loss value of the verification set in this round is lower than the lowest value of the loss value of the verification set in the previous round, then save the current model as the candidate best model. Repeat the above process for at least 50 rounds to fully train the model. During the whole process, the model parameters are gradually optimized, and the learning rate scheduler ensures that the weight parameters are adjusted more finely in the later stage of training, thereby improving the model performance. Finally, select the model with the lowest loss value of the verification set as the final trained model; (VI) Use the trained model to predict the tobacco leaf outlet moisture content at the next moment of the loosening and rehumidification process Collect in real time the tobacco process flow, water addition ratio, air-water mixing automatic valve opening, process hot air temperature, water addition flow, material cumulative amount, water addition cumulative amount, steam automatic valve opening and discharge moisture content of the tobacco leaf rehumidification production line at least 10 consecutive time points, repeat step (ii) (2) to step (iv) (b) to obtain at least 10 graph data, input the above graph data into the trained model, and predict the tobacco leaf outlet moisture content at the next moment of the loosening and rehumidification process.

2. A method for predicting the moisture content at the loose regain outlet based on a graph neural network as claimed in claim 1, characterized in that: In step (five), it also includes: using a test set to test the optimal model.

3. A method for predicting the moisture content at the loose regain outlet based on a graph neural network as claimed in claim 2, characterized in that: The Spearman coefficient of step (iii) is calculated using SPSS software.

4. A method for predicting the moisture content at the loose regain outlet based on a graph neural network as claimed in claim 3, characterized in that: The standard deviation Where: N is the total number of data points; x i is the data of the ith data point; μ is the mean of the population, and the calculation formula is:

5. The method for predicting the moisture content at the loose regain outlet based on graph neural network according to claim 4, characterized in that: The loss value is the mean square error Where: n is the total number of samples; y i is the true value of the i-th sample; is the predicted value of the i-th sample.

6. A method for predicting the moisture content at the loose regain outlet based on a graph neural network as claimed in claim 5, characterized in that: The node input feature dimension of the graph attention network layer is 9, the hidden layer dimension is 64, and the multi-head attention mechanism is set to 2.

7. The method for predicting the moisture content at the loose regain outlet based on graph neural network according to claim 6, characterized in that: In step (iv), the threshold is set to 0.

5.

8. The method for predicting the moisture content at the loose regain outlet based on graph neural network according to claim 7, characterized in that: In step (V), the learning rate scheduler is ReduceLROnPlateau.

9. The method for predicting the moisture content at the loose rehydration outlet based on graph neural network according to claim 8, characterized in that: The size of the window is set to 15, which means that 15 consecutive graph data are extracted each time as input features, and the output moisture content of the 16th graph data is used as the label value; each time a graph data is moved along the arrangement of the graph data, then the 15 consecutive graph data in the window are used as the next input feature, and the output moisture content of the 17th graph data is used as the next label value.

10. The method for predicting the moisture content at the loose regain outlet based on graph neural network according to claim 9, characterized in that: The amount of graph data input into the model in step (six) is the same as the amount of graph data input into the model in step (five).

Images