A Multi-Step Prediction Method for Cigarette Drying Process Quality Based on MCF-TransGRU

Abstract

This invention discloses a multi-step prediction method for the quality of cigarette drying process based on MCF-TransGRU, comprising: collecting time-series data of the thin-plate drying process production line in a cigarette manufacturing workshop at a preset time to form a sample dataset; preprocessing the sample dataset to obtain a preprocessed dataset; dividing the preprocessed dataset into training set data, validation set data, and test set data according to time order for model training and performance evaluation; constructing an MCF-TransGRU multi-step prediction model through a multi-scale convolutional fusion module, an improved Transformer module, and a residual-based gated recurrent unit module; training the constructed MCF-TransGRU multi-step prediction model and adjusting hyperparameters based on the training set data and validation set data; and using the trained MCF-TransGRU multi-step prediction model to predict the test set data / time-series data of the thin-plate drying process production line under test. This invention, through multi-step prediction, can provide prediction results of quality indicators at multiple future time points, providing guidance for subsequent production control decisions.

Description

Technical Field

[0001] This invention relates to a multi-step prediction method for the quality of cigarette drying process based on MCF-TransGRU, belonging to the field of tobacco processing quality index prediction. Background Technology

[0002] The cigarette drying process is a crucial step in tobacco production. Its quality not only affects the taste and aroma of the tobacco but also directly impacts the product's market competitiveness. This process involves multiple physical and chemical processes, requiring continuous adjustments to variables such as temperature, humidity, and airflow. Even minor changes in these variables can lead to significant variations in the final product quality. Therefore, accurate prediction and control of the process are essential.

[0003] Because the parameters and environmental conditions in the process are highly time-dependent, traditional forecasting methods often struggle to fully capture this complex time-series information. Traditional time-series forecasting methods typically include autoregressive models, moving average models, and basic regression analysis. While these methods perform well with static data and linear relationships, their predictive performance and accuracy are often limited when dealing with nonlinear and complex time-series data. This is especially true when handling multi-step forecasting tasks, where traditional methods struggle to effectively capture long-term dependencies and nonlinear patterns in the data. Therefore, accurate multi-step forecasting of cigarette drying process quality has become a bottleneck restricting further improvements in process quality and resource utilization, and urgently needs to be addressed. Summary of the Invention

[0004] This invention provides a multi-step prediction method for the quality of cigarette drying process based on MCF-TransGRU, which can be used to construct an MCF-TransGRU multi-step prediction model to achieve multi-step prediction of the time series of quality indicators in the cigarette drying process.

[0005] The technical solution of this invention is:

[0006] According to a first aspect of the present invention, a multi-step prediction method for the quality of cigarette drying process based on MCF-TransGRU is provided, comprising: collecting time-series data of the thin-plate drying process production line in a cigarette manufacturing workshop at a preset time to form a sample dataset; wherein the time-series data includes process parameters and quality index data; preprocessing the sample dataset to obtain a preprocessed dataset; dividing the preprocessed dataset into training set data, validation set data, and test set data according to time order for model training and performance evaluation; constructing an MCF-TransGRU multi-step prediction model through a multi-scale convolutional fusion module, an improved Transformer module, and a residual-based gated recurrent unit module; training the constructed MCF-TransGRU multi-step prediction model based on the training set data and validation set data, and adjusting the hyperparameters; and using the trained MCF-TransGRU multi-step prediction model to predict the test set data / time-series data of the thin-plate drying process production line to be tested.

[0007] The sample dataset is constructed by: determining process parameters and quality indicators based on the specific process flow of the thin-plate drying process in the cigarette making workshop; collecting time series data at a preset time based on the determined process parameters and quality indicators to form the sample dataset; wherein there are multiple process parameters and one quality indicator.

[0008] The preprocessing includes: first, removing the beginning and end of the sample data in the sample dataset; then, filling the missing values ​​in the data with the mean; then, using the 3sigma criterion to remove the sample data; and finally, normalizing the sample dataset judged by the 3sigma criterion.

[0009] The MCF-TransGRU multi-step prediction model takes sample data from the preprocessed dataset as input, based on different prediction step sizes, and feeds it into parallel combinations of different one-dimensional convolutional kernels for multi-scale feature extraction. The features extracted at each scale are concatenated and fused as the output of the multi-scale convolutional fusion module. This output is then fed into an improved Transformer module to further extract deeper features. The output of the improved Transformer module is fed into a GRU network. The residual module adds the input of the GRU network to the output after its nonlinear transformation, enabling the GRU network to more easily fit complex mapping functions and capture global temporal features in the sequence. This sum is then used as the output of the residual-based gated recurrent unit (RGM) module. The output of the residual-based RGM module is then passed through two fully connected layers. The number of output neurons in the second fully connected layer is set according to the prediction step size, and this layer is used to output the multi-step prediction results. The final output of the model is the prediction results of the quality indicators of the cigarette drying process at multiple future time points. The prediction results are then inversely normalized to obtain the final data that meets the requirements.

[0010] The improved Transformer module includes an improved Transformer layer that uses the standard Transformer layer as a framework, on which the Decoder structure in the standard Transformer layer is replaced by a simple linear layer.

[0011] According to a second aspect of the present invention, a multi-step prediction system for cigarette drying process quality based on MCF-TransGRU is provided, comprising: a collection module for collecting time-series data of the thin-plate drying process production line in a cigarette manufacturing workshop at a preset time, forming a sample dataset; wherein the time-series data includes process parameters and quality index data; a preprocessing module for preprocessing the sample dataset to obtain a preprocessed dataset; a partitioning module for partitioning the preprocessed dataset into training set data, validation set data, and test set data according to time order, for model training and performance evaluation; a construction module for constructing an MCF-TransGRU multi-step prediction model through a multi-scale convolutional fusion module, an improved Transformer module, and a residual-based gated recurrent unit module; a training module for training the constructed MCF-TransGRU multi-step prediction model and adjusting hyperparameters based on the training set data and validation set data; and a prediction module for predicting the test set data / time-series data of the thin-plate drying process production line under test using the trained MCF-TransGRU multi-step prediction model.

[0012] According to a third aspect of the present invention, a terminal is provided, comprising a processor, a memory, and a computer program stored in the memory and executable on the processor, the processor being configured to execute the multi-step prediction method for cigarette drying process quality based on MCF-TransGRU as described above.

[0013] According to a fourth aspect of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium comprising a stored program that, when executed by a processor, causes the processor to implement the multi-step prediction method for cigarette drying process quality based on MCF-TransGRU as described above.

[0014] The beneficial effects of this invention are as follows: The cigarette drying process involves multiple process parameters and quality indicators, and these parameters and indicators have complex spatiotemporal relationships. Traditional prediction methods struggle to effectively capture these complex relationships, resulting in low multi-step prediction accuracy. Adjustments to process parameters have a direct and complex impact on the final quality indicators, and existing methods cannot simultaneously consider the comprehensive influence of multiple process parameters on quality indicators. This invention captures the temporal characteristics of process parameters through convolutional kernels of different scales, utilizes Transformer adaptive capture of long-term dependencies and key time step information, integrates process parameter features through gated recurrent units, and achieves multi-step prediction through fully connected layers. By training the model and adjusting hyperparameters to enhance the learning ability of time series relationships, the model can provide predictions of quality indicators for multiple future time points through multi-step prediction, providing guidance for subsequent production control decisions. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the process of the present invention;

[0016] Figure 2 This is a schematic diagram of the MCF-TransGRU multi-step prediction model structure constructed in this invention;

[0017] Figure 3 This is a structural diagram of the multi-scale convolutional fusion module of the present invention;

[0018] Figure 4 This is a structural diagram of the improved Transformer of the present invention;

[0019] Figure 5 This is a structural diagram of the residual-based gated loop unit module of the present invention;

[0020] Figure 6 This is a comparison chart of the predicted and actual moisture content of the outlet material when the prediction step size is 4 in an embodiment of the present invention.

[0021] Figure 7This is a comparison chart of the predicted and actual moisture content of the outlet material when the prediction step size is 20 in an embodiment of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this application can be arbitrarily combined with each other.

[0023] Example 1: As Figure 1-7 As shown, according to a first aspect of the present invention, a multi-step prediction method for the quality of cigarette drying process based on MCF-TransGRU is provided, comprising: collecting time-series data of the thin-plate drying process production line in a cigarette manufacturing workshop at a preset time to form a sample dataset; wherein, the time-series data includes process parameters and quality index data; preprocessing the sample dataset to obtain a preprocessed dataset; dividing the preprocessed dataset into training set data, validation set data, and test set data according to time order for model training and performance evaluation; constructing an MCF-TransGRU multi-step prediction model through a multi-scale convolutional fusion module MCF, an improved Transformer module, and a residual-based gated recurrent unit module RGM; training the constructed MCF-TransGRU multi-step prediction model based on the training set data and validation set data, and adjusting the hyperparameters; using the trained MCF-TransGRU multi-step prediction model to predict the test set data, and validating the model through various evaluation indicators. Further, the trained MCF-TransGRU multi-step prediction model is used to predict the time-series data of the thin-plate drying process production line under test.

[0024] Furthermore, the formation of the sample dataset specifically involves: determining process parameters and quality indicators based on the specific process flow of the thin-plate drying process in the cigarette manufacturing workshop; collecting time-series data at a preset time period based on the determined process parameters and quality indicators to form the sample dataset; wherein there are multiple process parameters and one quality indicator.

[0025] Furthermore, the process parameters include inlet material moisture content, exhaust damper opening, exhaust air volume, hot air temperature, HT vapor pressure, rear chamber damper opening, inlet thin plate steam pressure, front chamber damper opening, cylinder wall temperature and pressure, and inlet material flow rate; the quality index is outlet material moisture content.

[0026] Furthermore, the preprocessing includes: first removing the beginning and end of the sample data in the sample dataset, then filling the missing values ​​in the data with the mean, and then using the 3sigma criterion to remove the sample data; and then normalizing the sample dataset judged by the 3sigma criterion.

[0027] Furthermore, the MCF-TransGRU multi-step prediction model uses sample data from the preprocessed dataset as input, based on different prediction step sizes, and feeds them into parallel combinations of different one-dimensional convolutional kernels for multi-scale feature extraction. Each scale's one-dimensional convolutional layer uses convolutional kernels of varying sizes, allowing for simultaneous feature extraction from different time ranges, thus capturing the data's characteristics more comprehensively. The features extracted at each scale are concatenated and fused as the output of the multi-scale convolutional fusion module. After processing by the multi-scale convolutional fusion module, the one-dimensional time-series signal is converted into a multi-channel time-series signal, which is then input into the improved Transformer module to further extract deeper features. The improved Transformer... The output of the Ransformer module is input into the GRU network. The residual module adds the input of the GRU network to the output after its nonlinear transformation, enabling the GRU network to more easily fit complex mapping functions and capture global temporal features in the sequence. This output is then passed to the residual-based gated recurrent unit (RGM) module, allowing the model to learn higher-level abstract features and better adapt to the complex temporal relationships in the cigarette drying process. The output of the RGM is then passed through two fully connected layers. The number of output neurons in the second fully connected layer is set according to the prediction step size, used to output multi-step prediction results. The final output of the model is the prediction results of the quality indicators of the cigarette drying process at multiple future time points. The prediction results are then inversely normalized to obtain the final data that meets the requirements. For example, setting the prediction step size to 20 and the input step size to a multiple of the prediction step size, specifically 40, allows prediction of the quality indicators for the next 20 time points based on the process parameters and quality indicator data from the previous 40 time points.

[0028] The expression for the multi-scale convolutional fusion module is as follows:

[0029]

[0030] MC i F(x) = concat(MC1, MC2, MC3)

[0031] In the formula, This represents a one-dimensional convolution operation; MC1, MC2, and MC3 represent three scale features; x is the input; and ReLU is used as the activation function.

[0032] Furthermore, the improved Transformer module includes an improved Transformer layer. This improved Transformer layer uses the standard Transformer layer as a framework, replacing the Decoder structure in the standard Transformer layer with a simple linear layer. Because the sample dataset size of the thin-plate drying process in cigarette manufacturing workshops is orders of magnitude larger than that of text or image datasets, directly using the standard Transformer model may lead to excessive model complexity, resulting in accumulated errors and excessively biased prediction results. To address this issue, this invention improves the traditional Transformer model, such as... Figure 4 As shown, the improved Transformer module only uses the standard Transformer Encoder structure, replacing the original complex Decoder structure with linear layers. This reduces the number of model parameters and complexity, effectively controlling error accumulation and improving prediction accuracy and training speed. The improved Transformer module is shown below. Figure 4 As shown, the output of the multi-scale convolutional fusion module is used for input embedding to convert the input into a feature vector that the model can process. By adding time step information, the model can capture the sequential relationship in the sequence. The multi-head self-attention mechanism unit in the Transformer layer treats the encoded representation of the input as a set of key-value pairs, where the dimension of the key and value is the length of the process parameter input time series. The Transformer adopts a scaled dot product attention mechanism, and the output is a weighted sum of values, where the weight of each value is determined by the dot product of the query and all keys.

[0033] Furthermore, the attention mechanism is expressed as follows:

[0034]

[0035] In the formula, Q (Query) is the query matrix, K (Key) is the key matrix, V (Value) is the value matrix, and d k Let be the dimension of the key vector, and softmax be the activation function used to normalize the similarity scores into a probability distribution.

[0036] The GRU network not only inherits the advantages of LSTM but also simplifies the model structure by reducing one gating unit, thus achieving faster training speed and lower memory usage. Furthermore, introducing residual modules into the GRU further improves the model's stability and performance when processing long sequence data. Figure 5 As shown. In the GRU network, x t and h tThese are the input data and the GRU unit output, respectively; r t It's a reset door, z t To update the gate, r t and z t Joint control starts calculating and updating h from the previous hidden state. t-1 To the new hidden state h t Compared to the three gating units of LSTM—input gate, forget gate, and output gate—GRU combines the input and forget gates into an update gate and uses a reset gate to reduce model complexity, making GRU more efficient for processing time-series data. The GRU expression is as follows:

[0037]

[0038] In the formula: x t W is the input vector input to the memory cell at time t; rx W rh W zx W zh W hx W hh b is the weight matrix; r b z b h It is the bias vector; represents the candidate hidden state; ⊙ represents the dot product operation at the corresponding position of the element; σ and tanh represent the activation functions.

[0039] Specifically, the constructed MCF-TransGRU multi-step prediction model is trained, and the Optuna framework is used for automated hyperparameter optimization. The root mean square error (RMSE) at the 4th second on the validation set is used as the optimization objective, aiming to minimize the RMSE value.

[0040] Furthermore, the final parameters were set to a model input stride of 40. In the multi-scale convolutional fusion module, three scales of one-dimensional convolutional kernels were constructed, namely (3, 5, 7), with a stride of 1, 128 hidden layers, and ReLU activation function. The improved Transformer module had D_model set to 128, a multi-head attention mechanism with 4 heads, Dropout set to 0.1, and ReLU activation function. In the residual-based gated recurrent unit module, the GRU had 128 neurons, tanh activation function, and connected two fully connected layers with neurons of (128, 128) and (128, 20) respectively. The model learning rate was 0.001, the batch size was 32, and the optimizer was Adam.

[0041] Specifically, the various evaluation metrics include root mean square error (RMSE), mean absolute error (MAE), and goodness of fit (R²).2 The parameter sizes of the three evaluation criteria are used as a measure of the effectiveness of the model.

[0042] Furthermore, the calculation formula for the evaluation index is as follows:

[0043]

[0044] In the formula: n is the amount of data for the predicted moisture content of the exported material; y i This represents the actual moisture content of the exported material. y is the predicted moisture content of the exported material; y is the average of the actual moisture content of the exported material.

[0045] The following explanation is based on experimental data:

[0046] I. The data collected from the production workshop for drying wire was generated between December 15, 2023 and December 20, 2023, totaling 87,238 records, as shown in Table 1. The dataset includes 10 process parameters and 1 quality indicator: inlet material moisture, exhaust damper opening, exhaust air volume, hot air temperature, HT vapor pressure, rear chamber damper opening, inlet thin plate steam pressure, front chamber damper opening, cylinder wall temperature and pressure, inlet material flow rate, and outlet material moisture.

[0047] Table 1. Partial Original Data from the Dataset

[0048]

[0049] Second, the material head and tail data in the sample dataset are removed first, and the removal rules are shown in Table 2. Then, the missing values ​​in the sample data are filled with the mean. Finally, the sample data is removed using the 3sigma criterion. The sample dataset judged by the 3sigma criterion is then normalized.

[0050] Table 2 Rules for Extracting Moisture Content Data from Outgoing Materials (Head and Tail of Material)

[0051]

[0052] Taking the selected quality characteristic item of ≤13.5% moisture content of the outlet material as an example, in the thin plate drying process, one day is considered as a batch. The first data in the current batch sample data that meets the condition of >10% moisture content of the outlet material is taken as the head node, and the critical point 2.5 minutes after the head node is taken as the critical position of the head. The first data in the current batch sample data that meets the condition of <10% moisture content of the outlet material is taken as the tail node, and the critical point 2.5 minutes before the tail node is taken as the critical position of the tail. The data between the critical positions of the head and tail are retained.

[0053] Furthermore, the normalization process refers to mapping the sample data to the range [0, 1], and the normalization formula is as follows:

[0054]

[0055] In the formula: X′ represents the normalized data; max(X) and min(X) represent the maximum and minimum values ​​of the sample data before normalization in the sample dataset, respectively; X represents the original sample data.

[0056] Third, the preprocessed dataset was divided into a training set (52,344 records), a validation set (17,447 records), and a test set (17,447 records) in a 6:2:2 ratio. The partitioning was done in chronological order (i.e., the training set was earlier than the validation set, and the validation set was earlier than the test set) to simulate prediction scenarios in actual production.

[0057] IV. Constructing the MCF-TransGRU multi-step prediction model. The MCF-TransGRU multi-step prediction model includes: a multi-scale convolutional fusion module, a Transformer module, and a residual-based gated recurrent unit module, such as... Figure 2 As shown, the constructed MCF-TransGRU multi-step prediction model is trained based on the training and validation sets, and the hyperparameters are adjusted to enable it to deeply learn the relationships and patterns between time series, achieving efficient prediction performance. The trained MCF-TransGRU multi-step prediction model is then used to predict test set data, and the model's performance is validated through various evaluation metrics to achieve accurate multi-step prediction of cigarette drying process quality indicators.

[0058] The prediction performance of the model was verified, as shown in Table 3. The MCF-TransGRU multi-step prediction model provided by this invention was compared with five traditional deep learning algorithms: GRU, CNN-LSTM, Seq2Seq, TCN-Attention, and CNN-GRU-Attention. The parameter configurations of the deep learning models were all set according to the model of this invention. Finally, the root mean square error (RMSE), mean absolute error (MAE), and goodness of fit (R²) were used as the evaluation metrics. 2 As shown in Table 3, the optimal results of the test set data for prediction step sizes of 4, 8, 12, 16, and 20 in the model of this invention are plotted. Figure 6-7 As shown.

[0059] Table 3. Experimental prediction results for different phase lengths using different models.

[0060]

[0061]

[0062] The results show that the multi-step prediction method for cigarette drying process quality based on MCF-TransGRU of the present invention exhibits the best prediction effect at all prediction durations. Whether in short-term or long-term prediction, it achieves the best results in RMSE, MAE, and R. 2 The MCF-TransGRU multi-step prediction model outperforms other comparative models in all three metrics. This result demonstrates that the MCF-TransGRU multi-step prediction model has strong adaptability and predictive ability in handling complex time-series data. In the specific scenario of predicting the moisture content of the material exiting the thin plate drying process, its high-precision prediction capability can provide strong support for refined control in actual production processes.

[0063] To verify the rationality and effectiveness of the MCF-TransGRU multi-step prediction model design provided by this invention, the moisture content of the material exiting the thin plate drying process was used as the prediction target. An ablation experiment was established. The comparative models for the ablation experiment of this invention were multi-scale CNN-Transformer, Transformer-GRU, CNN-TransGRU, and multi-scale CNN-GRU (in the comparative models, those involving Transformer are all the improved Transformer modules proposed in this invention). The hyperparameter settings for each ablation part were the same. The above models were trained under the same experimental conditions, and the ablation experiment prediction results with prediction step lengths of 4 seconds to 20 seconds are shown in Table 4.

[0064] Table 4 Prediction Results of Ablation Experiment

[0065]

[0066]

[0067] As shown in Table 4, the multi-scale feature extraction mechanism of the MCF-TransGRU multi-step prediction model is a significant advantage. In the MCF-TransGRU multi-step prediction model, the multi-scale convolutional fusion module (MCF) is used to extract features at different scales, which is crucial for capturing the complex dynamic changes in moisture content of the material exiting the thin-plate drying process. When ordinary convolutional layers are used instead of multi-scale convolutional layers (such as in CNN-TransGRU, which is a variation of this invention, replacing the MCF with CNN), although the model still has some predictive ability, its performance is significantly reduced, especially in longer-term predictions. This indicates that the introduction of multi-scale convolution allows the model to capture multi-scale features in the data more comprehensively, thereby improving prediction accuracy. Secondly, the application of the Transformer structure in MCF-TransGRU also greatly enhances the model's ability to capture long-term dependencies. In the ablation experiment, removing the multi-scale convolutional fusion and retaining only the combination of Transformer and GRU (such as in Transformer-GRU) resulted in a significantly worse performance than the complete MCF-TransGRU. This indicates that while the Transformer has unique advantages in handling long-term dependencies, when used alone, it cannot fully capture multi-scale features in the data, especially in short-term and medium-to-long-term predictions, where its performance is relatively unstable. The addition of a multi-scale convolutional fusion module compensates for this deficiency, enabling MCF-TransGRU to maintain stable high prediction performance across all time scales. Furthermore, the introduction of GRU provides MCF-TransGRU with stronger temporal information processing capabilities. Unlike traditional RNNs, GRU effectively alleviates the gradient vanishing problem in long-term series through a gating mechanism. In ablation experiments, removing GRU (as in MCF-Transformer) significantly reduces the model's prediction performance, especially in long-term predictions (e.g., 16 seconds and 20 seconds). This demonstrates that GRU plays a crucial role in MCF-TransGRU; through its combination with the Transformer and the multi-scale convolutional fusion module, GRU effectively enhances the model's ability to handle long-term series. The MCF-GRU and CNN-TransGRU ablation models perform relatively well, but compared to MCF-TransGRU, they still have certain shortcomings when handling complex temporal data. MCF-GRU lacks a Transformer module, making it less effective at capturing long-term dependencies in the data, leading to a decrease in accuracy over long periods. CNN-TransGRU, on the other hand, lacks a multi-scale convolution module. While it excels at extracting local features, its ability to capture global features is inferior to MCF-TransGRU, which is also reflected in the difference in prediction accuracy.In summary, the significant advantage of the MCF-TransGRU multi-step prediction model lies in its comprehensive utilization of the ability of multi-scale convolution to capture multi-scale features of data, the effective handling of long-term dependencies by Transformer, and the enhancement of GRU's ability to express and remember temporal information. These factors work together to leverage their respective strengths, enabling the entire model to maintain excellent prediction performance at different time scales.

[0068] According to a second aspect of the present invention, a multi-step prediction system for the quality of cigarette drying process based on MCF-TransGRU is provided, comprising: a collection module for collecting time-series data of the thin-plate drying process production line in a cigarette manufacturing workshop at a preset time, forming a sample dataset; wherein the time-series data includes process parameters and quality index data; a preprocessing module for preprocessing the sample dataset to obtain a preprocessed dataset; a partitioning module for partitioning the preprocessed dataset into training set data, validation set data, and test set data according to time order, for model training and performance evaluation; a construction module for constructing an MCF-TransGRU multi-step prediction model through a multi-scale convolutional fusion module, an improved Transformer module, and a residual-based gated recurrent unit module; a training module for training the constructed MCF-TransGRU multi-step prediction model and adjusting hyperparameters based on the training set data and validation set data; and a prediction module for predicting the test set data / time-series data of the thin-plate drying process production line under test using the trained MCF-TransGRU multi-step prediction model.

[0069] According to a third aspect of the present invention, a terminal is provided, including a processor, a memory, and a computer program stored in the memory and executable on the processor, the processor being configured to execute the multi-step prediction method for cigarette drying process quality based on MCF-TransGRU as described above.

[0070] According to a fourth aspect of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium including a stored program, which, when executed by a processor, causes the processor to implement the multi-step prediction method for cigarette drying process quality based on MCF-TransGRU as described above.

[0071] The specific embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A multi-step prediction method for the quality of cigarette drying process based on MCF-TransGRU, characterized in that, include: Time series data of the thin-plate drying process production line in the cigarette making workshop were collected at a preset time to form a sample dataset; the time series data includes process parameters and quality index data. The sample dataset is preprocessed to obtain the preprocessed dataset; The preprocessed dataset is divided into training set data, validation set data and test set data according to time order, for model training and performance evaluation. A multi-step prediction model of MCF-TransGRU is constructed by using a multi-scale convolutional fusion module, an improved Transformer module, and a residual-based gated recurrent unit module. The constructed MCF-TransGRU multi-step prediction model was trained and its hyperparameters were adjusted based on the training set data and validation set data. The trained MCF-TransGRU multi-step prediction model is used to predict the test set data / time series data of the thin plate wire drying process production line under test; The MCF-TransGRU multi-step prediction model takes sample data from the preprocessed dataset as input, based on different prediction step sizes, and feeds it into parallel combinations of different one-dimensional convolutional kernels for multi-scale feature extraction. The features extracted at each scale are concatenated and fused as the output of the multi-scale convolutional fusion module. This output is then fed into an improved Transformer module to further extract deeper features. The output of the improved Transformer module is fed into a GRU network. The residual module adds the input of the GRU network to the output after its nonlinear transformation, enabling the GRU network to more easily fit complex mapping functions and capture global temporal features in the sequence. This sum is then used as the output of the residual-based gated recurrent unit (RGM) module. The output of the residual-based RGM module is then passed through two fully connected layers. The number of neurons in the second fully connected layer is set according to the prediction step size to output the multi-step prediction results. The final output of the model is the prediction results of the quality indicators of the cigarette drying process at multiple future time points. The prediction results are then inversely normalized to obtain the final data that meets the requirements. The improved Transformer module includes an improved Transformer layer that uses the standard Transformer layer as a framework, on which the Decoder structure in the standard Transformer layer is replaced by a simple linear layer.

2. The multi-step prediction method for cigarette drying process quality based on MCF-TransGRU according to claim 1, characterized in that, The sample dataset is constructed by: determining process parameters and quality indicators based on the specific process flow of the thin-plate drying process in the cigarette making workshop; collecting time series data at a preset time based on the determined process parameters and quality indicators to form the sample dataset; wherein there are multiple process parameters and one quality indicator.

3. The multi-step prediction method for cigarette drying process quality based on MCF-TransGRU according to claim 1, characterized in that, The preprocessing includes: first, removing the beginning and end of the sample data in the sample dataset; then, filling the missing values ​​in the data with the mean; then, using the 3sigma criterion to remove the sample data; and finally, normalizing the sample dataset judged by the 3sigma criterion.

4. A multi-step prediction system for the quality of cigarette drying process based on MCF-TransGRU, characterized in that, include: The collection module is used to collect time-series data of the thin-plate drying process production line in the cigarette making workshop at a preset time, forming a sample dataset; the time-series data includes process parameters and quality index data; The preprocessing module is used to preprocess the sample dataset to obtain the preprocessed dataset; The partitioning module is used to divide the preprocessed dataset into training set data, validation set data and test set data according to time order, for model training and performance evaluation. The building blocks are used to construct the MCF-TransGRU multi-step prediction model through a multi-scale convolutional fusion module, an improved Transformer module, and a residual-based gated recurrent unit module. The training module is used to train the constructed MCF-TransGRU multi-step prediction model and adjust the hyperparameters based on the training set data and validation set data. The prediction module uses a trained MCF-TransGRU multi-step prediction model to predict the test set data / time series data of the thin plate drying process production line under test; The MCF-TransGRU multi-step prediction model takes sample data from the preprocessed dataset as input, based on different prediction step sizes, and feeds it into parallel combinations of different one-dimensional convolutional kernels for multi-scale feature extraction. The features extracted at each scale are concatenated and fused as the output of the multi-scale convolutional fusion module. This output is then fed into an improved Transformer module to further extract deeper features. The output of the improved Transformer module is fed into a GRU network. The residual module adds the input of the GRU network to the output after its nonlinear transformation, enabling the GRU network to more easily fit complex mapping functions and capture global temporal features in the sequence. This sum is then used as the output of the residual-based gated recurrent unit (RGM) module. The output of the residual-based RGM module is then passed through two fully connected layers. The number of neurons in the second fully connected layer is set according to the prediction step size to output the multi-step prediction results. The final output of the model is the prediction results of the quality indicators of the cigarette drying process at multiple future time points. The prediction results are then inversely normalized to obtain the final data that meets the requirements. The improved Transformer module includes an improved Transformer layer that uses the standard Transformer layer as a framework, on which the Decoder structure in the standard Transformer layer is replaced by a simple linear layer.

5. A terminal, characterized in that: The method includes a processor, a memory, and a computer program stored in the memory and executable on the processor, the processor being configured to perform the multi-step prediction method for cigarette drying process quality based on MCF-TransGRU as described in any one of claims 1-3.

6. A computer-readable storage medium comprising a stored program, characterized in that: When the program is executed by the processor, the processor implements the multi-step prediction method for the quality of cigarette drying process based on MCF-TransGRU as described in any one of claims 1-3.

Images