Cigarette cut tobacco production process quality prediction method based on DCTCN-informer-TSMIX model

Abstract

The invention discloses a cigarette cut tobacco production process quality prediction method based on a DCTCN-informer-TSMIX model, and the method comprises the steps: collecting the time series data of a loosening and moisture regaining process production line of a cigarette cut tobacco production workshop at a preset time, and forming a sample data set; preprocessing the sample data set to obtain a processed data set; according to a time sequence, dividing the preprocessed data into a training set, a verification set and a test set for model training and performance evaluation; constructing a prediction model fusing a discrete cosine transform enhancement module (DCT), a time sequence convolutional network module (TCN), an Informer model and a time sequence mixer TSMixer; according to the data of the training set and the data of the verification set, the constructed DCTCN-Informer-TSMIX model is trained, and hyper-parameters are adjusted; and predicting the test set data / to-be-tested time sequence data of the loosening and moisture regaining process production line by using the trained model, and outputting a predicted value of the water content of the discharged material at a target time point. The model has the capabilities of frequency domain feature extraction, short-term time sequence modeling, long-term dependence capture and multi-dimensional feature interaction, can effectively improve the accuracy and the stability of prediction of the water content of the discharged material, and provides decision support for intelligent control and quality adjustment in the tobacco shred making process.

Description

Technical Field

[0001] The present invention relates to the field of intelligent manufacturing, and in particular to a method for predicting the quality of a cigarette shredded tobacco process based on a DCTCN-informer-TSMixer model. Background Art

[0002] With the advancement of global economic integration, competition in the manufacturing industry is becoming increasingly fierce. my country is in a critical period of transformation and upgrading in its manufacturing sector. Improving product quality is the core of enhancing the competitiveness of the nation's manufacturing industry. In recent years, the country has attached great importance to the development of intelligent manufacturing and industrial big data technologies. The process industry is a key pillar of the national economy, encompassing sectors such as chemical engineering, metallurgy, energy, and pharmaceuticals. Under the background of Industry 4.0, the process industry is developing towards green, efficient, and integrated development, and is trending towards smart enterprises.

[0003] High-quality products are paramount to the survival and development of my country's manufacturing industry. In this pursuit of high-quality products, accurate quality prediction is crucial. This is not only a requirement for product development but also a crucial driving force in the advancement of high-quality manufacturing in my country. Unlike traditional discrete manufacturing, process manufacturing is characterized by strong production continuity, multiple production equipment, complex coupling between variables, and cumbersome process modeling. This makes quality prediction difficult, and optimizing operations even more challenging. Therefore, the stability and accuracy of industrial process quality predictions are key factors in ensuring production efficiency and product quality. Predicting product quality trends and outcomes in advance is crucial. Summary of the Invention

[0004] The present invention provides a cigarette shred process quality prediction method based on a DCTCN-informer-TSMixer model, which is used to predict the moisture content of the output material, a product quality indicator in a process manufacturing production line.

[0005] The present invention comprises the following steps: collecting time series data of loose moisture regain in a production line of a cigarette shred workshop to form a data sample set, including process parameters and quality index data; performing data preprocessing on the data set, including material head and tail processing, missing value processing, outlier processing, data dimensionality reduction and normalization processing, and dividing the data set into input sequences of fixed length according to a time window; dividing the preprocessed data set into a training set, a test set and a validation set according to a certain ratio in chronological order for model training and performance evaluation; designing a frequency domain enhanced temporal network (DCTCN), combining an informer model with a time series mixer TSMixer to construct a DCTCN-informer-TSMixer prediction model; training the constructed DCTCN-informer-TSMixer model based on training set data and validation set data, and adjusting hyperparameters until the model meets expectations and then saving the model; using the trained DCTCN-informer-TSMixer prediction model to predict time series data of a loose moisture regain process production line; and evaluating the model effect through four quality indicators and statistical test methods.

[0006] The sample data set specifically comprises: determining process parameters and quality indicators based on the specific process of the loose rehydration process in the cigarette silk making workshop; collecting time series data at preset times based on the determined process parameters and quality indicators to form a sample data set; wherein, there are multiple process parameters and one quality indicator.

[0007] The preprocessing includes: first, setting upper and lower limits based on the export material moisture threshold and time window to process the head and tail of the material; processing missing values, if the instance has a large proportion of missing values, deleting the instance; if the missing data proportion is small, filling it through forward filling and backward filling methods; then using the 3σ criterion to further detect and delete outliers, and clean up gross errors and noise in the data; finally, using the Z-score method to normalize all process parameters and quality indicators.

[0008] The dimension reduction process includes: using Pearson correlation analysis to calculate the correlation coefficient between the process parameters and the quality index, and eliminating the process parameters below the threshold.

[0009] The proposed DCTCN-informer-TSMixer prediction model takes the sample data in the dimensionality-reduced dataset as input and first performs a discrete cosine transform (DCT) on each channel data to enhance its frequency domain feature expression capability. Subsequently, the DCT-enhanced sequence is input into the channel-level feature adjustment module, where a fully connected network is used to generate adjustment weights for each channel to achieve feature compression and enhancement in the frequency domain. The processed sequence is then input into the temporal convolutional network (TCN) module, where short-term temporal dependency information is extracted through multiple layers of one-dimensional dilated convolution operations to form local feature expressions. At the same time, the original normalized sequence is input into the informer module, where a sparse attention mechanism is used to capture long-term dependencies and obtain a global semantic feature representation. The informer output and the TCN output are additively fused according to the channel dimension and input as mixed features into the TSMixer module, where feature cross-transformations of the time and channel dimensions are sequentially performed to enhance the interaction and coupling representation of multi-dimensional information. Finally, a decoder and linear projection layer are used to output the multi-step moisture content prediction result corresponding to the prediction step size and restore it to the actual physical value through denormalization.

[0010] The evaluation indicators include mean absolute percentage error (MAPE), mean absolute error (MAE), mean square error (MSE) and determination coefficient (R2), and the statistical test method is Diebold-Mariano (DM) test.

[0011] According to the second aspect of the present invention, a method for predicting the moisture content of loose rehydration discharge material in a cigarette shred process based on DCTTCN-Informer-TSMixer is provided, comprising: a collection module for collecting time series data of a loose rehydration process production line in a cigarette shred workshop under a preset time to form a sample data set; wherein the time series data includes a plurality of process parameters and corresponding discharge moisture content quality index data; a preprocessing module for preprocessing the sample data set, including material head and tail processing, missing value processing, outlier processing, data dimensionality reduction and normalization processing, to obtain preprocessed data; a partitioning module for partitioning the preprocessed sample data set into a training set, a validation set and a test set in chronological order for model training and parameter Tuning and effect evaluation; a construction module for building a deep neural network structure that integrates a discrete cosine transform channel enhancement module (DCT), a temporal convolutional network module (TCN), an informer attention module, and a time-feature mixing module (TSMixer) based on the data to predict the moisture content of loose and regained cigarette output; a training module for training the constructed DCTTCN-Informer-TSMixer model based on the training set and the validation set, optimizing the network parameters and hyperparameter settings, and improving the prediction accuracy and generalization ability; a prediction module for using the trained model to predict the test set data or the time series data to be tested in actual production, and outputting the predicted value of the moisture content of the export material at the target time point.

[0012] According to a third aspect of the present invention, there is provided a terminal comprising a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the processor is configured to execute any one of the above-mentioned methods for predicting the moisture content of loose rehydration discharge material in the cigarette shred process based on DCTTCN-Informer-TSMixer.

[0013] According to a fourth aspect of the present invention, a computer-readable storage medium is provided, in which a program is stored. When the program is executed by a processor, the processor implements any of the above-mentioned methods for predicting the moisture content of loose rehydration discharge from the cigarette shredding process based on DCTTCN-Informer-TSMixer.

[0014] The beneficial effects of the present invention are:

[0015] 1. The loosening and rehydration process in the cigarette shred production process involves multiple dynamically changing process parameters and key quality indicators, with significant nonlinear correlations and complex temporal coupling between these parameters. Traditional prediction methods struggle to effectively capture these multidimensional features, resulting in poor stability and low accuracy in prediction results under complex conditions, making them difficult to meet actual process control requirements.

[0016] 2. This paper introduces the discrete cosine transform (DCT) to perform frequency domain processing on the input sequence, which can effectively extract the dominant trend features in the sequence and suppress high-frequency noise, thereby improving the model's ability to perceive the changing patterns of the original signal. Through the temporal convolutional network (TCN) module, a multi-scale dilated convolution structure is utilized to effectively capture the changing trends of process parameters in a short time range. The sparse attention mechanism in the informer module is utilized to focus on modeling long-term dependencies between key time nodes, making up for the shortcomings of traditional models in long-term pattern recognition. The TSMixer structure is introduced to perform deep mixing between time and channel dimensions. Finally, the fused multi-dimensional features are output as single-step prediction results through the decoder. The model structure maintains high efficiency while taking into account prediction accuracy and stability.

[0017] 3. Through parameter optimization and generalization adjustment during the training process, the prediction model constructed by the present invention can achieve accurate prediction of the moisture content of the discharge material at the target moment, providing decision support for online monitoring and dynamic regulation in the cigarette shred process, and helping to improve product quality consistency and process intelligence. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a schematic diagram of the process of the present invention;

[0019] Figure 2 Schematic diagram of the DCTCN-informer-TSMixer prediction model structure constructed by the present invention;

[0020] Figure 3 It is the DCTCN module structure;

[0021] Figure 4 This is a schematic diagram of the Informer module;

[0022] Figure 5 This is a schematic diagram of the TSMixer module;

[0023] Figure 6 A comparison chart of the predicted value and the actual value of the discharge moisture content prediction implemented in the present invention;

[0024] Figure 7 This is a scatter plot of the fitting of the predicted value and the true value for the discharge moisture content prediction implemented in the present invention. DETAILED DESCRIPTION

[0025] The present invention will be described in detail below with reference to the accompanying drawings.

[0026] In order to more clearly illustrate the technical effects of the present invention, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings. Based on the embodiments of the present invention, various forms of modifications, equivalent replacements and improvements that can be made by those skilled in the art on this basis should all be deemed to fall within the scope of protection of the present invention. The technical features between the multiple embodiments involved in this specification can also be used in combination to form new implementation methods without constituting a technical conflict.

[0027] Example 1: Figure 1-7 As shown, according to the first aspect of the present invention, a method for predicting the quality of cigarette shredded tobacco process based on the DCTCN-informer-TSMixer model is provided, comprising: collecting time series data of loose moisture regain in the production line of a cigarette shredded tobacco workshop to form a data sample set, including process parameters and quality index data; performing data preprocessing on the data set, including material head and tail processing, missing value processing, outlier processing, data dimensionality reduction and normalization processing, and dividing the data set into input sequences of fixed length according to the time window; dividing the preprocessed data set into a training set, a test set and a validation set according to a certain ratio in chronological order for model training. Training and performance evaluation: Designing a frequency-domain enhanced temporal network (DCTCN) and combining it with the informer model and the time series mixer (TSMixer) to construct a DCTCN-informer-TSMixer prediction model. Training the constructed DCTCN-informer-TSMixer model based on training and validation data sets, adjusting hyperparameters until the model meets expectations and saving the model. The trained DCTCN-informer-TSMixer prediction model was used to predict test data sets. Model effectiveness was evaluated using four quality metrics and statistical tests. Furthermore, the trained model was used to predict time series data from a loose tempering process production line.

[0028] Furthermore, based on the specific process of the loose rehydration process in the cigarette silk workshop, the process parameters and quality indicators are determined; based on the determined process parameters and quality indicators, time series data under preset time periods are collected to form a sample data set; wherein, there are multiple process parameters and one quality indicator.

[0029] Furthermore, the preprocessing first sets upper and lower limits based on the export material moisture threshold and time window to perform material head and tail processing; missing values ​​are processed, if the instance has a large proportion of missing values, the instance is deleted; if the missing data proportion is small, it is filled by forward filling and backward filling methods; then the 3σ criterion is used to further detect and delete outliers, and clean up gross errors and noise in the data; finally, the Z-score method is used to normalize all process parameters and quality indicators.

[0030] The calculation formula of the Z-score normalization method is:

[0031]

[0032] Where X is the normalized variable, x is the original data, μ is the mean, and σ is the standard deviation.

[0033] The dimension reduction process includes: using Pearson correlation analysis to calculate the correlation coefficient between the process parameters and the quality index, and eliminating the process parameters below the threshold.

[0034] Furthermore, the DCTCN-informer-TSMixer prediction model takes sample data from the preprocessed dataset as input. The model first takes preprocessed time series sample data as input, which includes historical sequence information of multiple process parameters. To enhance the discriminability of the input sequence, the model first performs a frequency domain transform on each channel (i.e., each process parameter) using a discrete cosine transform (DCT) module to capture the dominant frequency characteristics of the variable changes. Subsequently, the DCT output is passed through a channel enhancement network to generate channel weights, which are used to suppress redundant features and enhance effective features in the frequency domain, thereby improving the model's ability to represent characteristic signals.

[0035] The calculation formula of the DCT change is as follows:

[0036]

[0037] where k∈{0,1,…,N-1}, N is the time step, and α is the scaling factor determined by the following equation:

[0038]

[0039] Next, the DCT-transformed and channel-weighted input data is fed into a temporal convolutional network (TCN) module to construct the DCTCN module. The TCN module is composed of multiple layers of one-dimensional dilated convolutional units. Convolution operations in different layers employ different dilation rates and receptive fields, enabling modeling of changing trends in sequences at different time scales. Through continuous convolutional activations and residual connections, the TCN module effectively captures short-term dynamic dependency information and extracts representative local temporal features. The TCN module consists of three main elements: causal convolution, dilated convolution, and residual connections.

[0040] The calculation formula of the dilated convolution is as follows:

[0041]

[0042] Where d is the dilation rate, k is the filter size, and x_(sd·i) represents the convolution of the past state. Therefore, dilated convolution greatly reduces the network complexity and improves computational efficiency.

[0043] The calculation formula of the residual connection is as follows:

[0044] y t =x t +F(x t )

[0045] where y t is the output of the residual block at time step, x t is the input sequence, F(x t ) represents a one-dimensional convolution operation on the input sequence.

[0046] At the same time, the normalized raw input sequence is fed into the Informer module. Leveraging its encoder-only structure, the ProbSparse attention mechanism rapidly extracts global dependencies at key moments in long sequences. The Informer module employs a multi-head attention mechanism, with each attention head independently learning key temporal patterns in the input sequence. Ultimately, the attention outputs from each head are combined to form a global feature representation. This module effectively avoids the computational redundancy and precision degradation of traditional self-attention mechanisms in long sequences, and enhances the modeling of temporal order features by introducing timestamp embedding.

[0047] The calculation formula of the ProbSparse attention mechanism is as follows:

[0048]

[0049] Where Q, K, and V are the query matrix, key matrix, and value matrix, respectively; dk is the dimension of the key matrix, and KT is the transpose of the K matrix.

[0050] The local and global features extracted by the DCTTCN module and the Informer module are fused in the channel dimension to form a comprehensive information expression. The fused features are fed into the Time Series Mixer (TSMixer) module. The TSMixer module consists of two alternating submodules: a time dimension mixer and a feature dimension mixer. The time mixer performs stride fusion on the time axis through nonlinear mapping, enhancing the coupling of contextual information between time series. The channel mixer uses a fully connected network to reconstruct and filter features in the variable dimension, strengthening the dependency modeling between variables. This module achieves a joint expression of spatiotemporal features through an alternating structure, improving the discriminative power of the final feature vector.

[0051] The features output by the TSMixer module are fed into the decoder's linear projection layer, which outputs the single-step discharge moisture content prediction for the corresponding time point. The prediction results are then denormalized to return to real physical quantities, serving as the model's final output. This structure efficiently completes the complete information processing pipeline, from frequency domain enhancement, time series modeling, global attention, to spatiotemporal fusion. This makes the model suitable for quality prediction tasks in industrial-grade loose moisture conditioning processes, improving prediction stability and accuracy.

[0052] The calculation formula of the denormalization method is:

[0053] x=X·σ+μ

[0054] Where X is the normalized value, x is the original data, μ is the mean, and σ is the standard deviation.

[0055] Specifically, the constructed DCTCN-Informer-TSMixer model was trained using the mean squared error (MSE) loss function as the optimization objective, aiming to minimize the error between the model's predictions and the true moisture content. During training, the Adam optimizer was used to adaptively adjust the network weights, and the ReduceLROnPlateau learning rate scheduler was employed.

[0056] Furthermore, the final parameters of the model are set as follows: the model input step size is 10, the prediction step size is 1, the number of training rounds is 100, the learning rate is initially set to 0.0001, and a batch size of 32 is used for small batch training. The model uses a nested TCN module to perform time series modeling on the features after frequency domain enhancement, the hidden layer width is set to 64 dimensions, and the activation function is ReLU. In the Informer module, the representation dimension dmodel of the model is set to 64, the number of attention heads is 4, the attention factor is 5, the dropout rate (Dropout) is 0.1, the number of encoder layers and decoder layers is 3, the feedforward neural network dimension is set to 128 dimensions, and the activation function is GeLU. In the feature fusion stage, the TSMixer module adopts an alternating stacking structure of time mixers and feature mixers, with the intermediate dimensions set to twice the input length and twice the feature dimension, respectively, to achieve multi-dimensional feature interaction through linear mapping.

[0057] Specifically, the evaluation indicators include absolute percentage error (MAPE), mean absolute error (MAE), mean square error (MSE), and coefficient of determination (R2).

[0058] Furthermore, the evaluation index calculation formula is as follows:

[0059]

[0060] in, is the predicted value; Mean value; y i is the actual value; n is the total number of samples. The smaller the MAPE, MSE, and MAE, the closer the predicted value is to the actual value; the closer R2 is to 1, the better the model fit.

[0061] The following is an explanation based on experimental data:

[0062] First, the experimental data comes from 8 batches of production line data collected at a sampling frequency of 1 Hz. The first 30,000 data points of the processed data are selected as the training dataset. An example of the dataset is shown in Table 1.

[0063] Table 1 Part of the original data of the dataset

[0064]

[0065] Then, the head and tail data in the sample data set are removed first. The removal rules are shown in Table 2. Then, the missing values ​​in the sample data are filled or deleted with the mean, and finally the 3 sigma criterion is used to further delete the outliers. The Z-score method is used to normalize the proposed sample set.

[0066] Table 2 Rules for extracting data of discharge moisture content, head and tail

[0067]

[0068] Next, we split the preprocessed dataset into a training set (18,000 records), a validation set (6,000 records), and a test set (6,000 records) in a 6:2:2 ratio. We kept the data in chronological order to simulate actual production prediction scenarios.

[0069] Finally, the DCTCN-Informer-TSMixer prediction model is constructed, which includes: frequency domain enhanced time network, Informer module and time series mixer, as shown in the attached Figure 2 The constructed DCTCN-Informer-TSMixer model is trained based on the training and validation sets, and hyperparameters are adjusted. The trained model is used to predict test set data or actual time series data to be tested in production. The model is validated using multiple evaluation metrics, and its performance compared to other models is evaluated using statistical tests.

[0070] To verify the model's predictive performance, the DCTCN-Informer-TSMixer model provided by the present invention was compared with five traditional deep learning algorithms: TCN, Informer, TCN-Transformer, and TSMixer. The deep learning model parameter configurations were all set according to the model of the present invention. Finally, the percentage error (MAPE), mean absolute error (MAE), mean square error (MSE), and coefficient of determination (R2) were used as comparative analysis indicators, as shown in Table 3. A data plot of the predicted and true values ​​of the test set of the model of the present invention was also plotted.

[0071] Table 3 Experimental prediction results

[0072]

[0073] Overall, the DCTCN-Informer-TSMixer model demonstrated the best prediction results, outperforming other comparison models in terms of MAE, MSE, MAPE, and R². This demonstrates the model's strong adaptability to complex, multi-scale process industry data. In the specific scenario of predicting the moisture content of the discharge from the loose rehydration process, it can more comprehensively model the relationship between features and time series, thereby achieving more accurate quality predictions.

[0074] To better distinguish the superiority of the models, the Diebold-Mariano (DM) test was used to further evaluate the performance differences between the proposed model and the comparison models. The DM test results are shown in Table 7. S0 represents the statistical value, and P0 represents the P value at the 95% confidence interval. When S0 < 0 and P0 < 0.05, the proposed model outperforms the comparison models. It can be clearly seen that the S0 of all comparison models is less than 0 and the P0 is less than 0.05, indicating that the proposed DCTCN-TimesNet-LSTMixer model has a significant performance advantage.

[0075] Table 4DM test results

[0076]

[0077] According to the second aspect of the present invention, a method for predicting the moisture content of loose rehydration discharge material in a cigarette shred process based on DCTTCN-Informer-TSMixer is provided, comprising: a collection module for collecting time series data of a loose rehydration process production line in a cigarette shred workshop under a preset time to form a sample data set; wherein the time series data includes a plurality of process parameters and corresponding discharge moisture content quality index data; a preprocessing module for preprocessing the sample data set, including material head and tail processing, missing value processing, outlier processing, data dimensionality reduction and normalization processing, to obtain preprocessed data; a partitioning module for partitioning the preprocessed sample data set into a training set, a validation set and a test set in chronological order for model training and parameter Tuning and effect evaluation; a construction module for building a deep neural network structure that integrates a discrete cosine transform channel enhancement module (DCT), a temporal convolutional network module (TCN), an informer attention module, and a time-feature mixing module (TSMixer) based on the data to predict the moisture content of loose and regained cigarette output; a training module for training the constructed DCTTCN-Informer-TSMixer model based on the training set and the validation set, optimizing the network parameters and hyperparameter settings, and improving the prediction accuracy and generalization ability; a prediction module for using the trained model to predict the test set data or the time series data to be tested in actual production, and outputting the predicted value of the moisture content of the export material at the target time point.

[0078] According to a third aspect of the present invention, there is provided a terminal comprising a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the processor is configured to execute any one of the above-mentioned methods for predicting the moisture content of loose rehydration discharge material in the cigarette shred process based on DCTTCN-Informer-TSMixer.

[0079] According to a fourth aspect of the present invention, a computer-readable storage medium is provided, in which a program is stored. When the program is executed by a processor, the processor implements any of the above-mentioned methods for predicting the moisture content of loose rehydration discharge from the cigarette shredding process based on DCTTCN-Informer-TSMixer.

[0080] The specific embodiments of the present invention are described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Various changes can be made within the knowledge of ordinary technicians in this field without departing from the scope of the present invention.

Claims

1. A method for predicting cigarette shred process quality based on the DCTCN-informer-TSMixer model, characterized in that: The following steps are involved: Step (1) collecting time series data of loose moisture regain in the production line of the cigarette making workshop to form a data sample set, including process parameters and quality index data; Step (2) preprocessing the sample data set to obtain a preprocessed data set; Step (3) divide the preprocessed data set into training set data, validation set data and test set data in chronological order for model training and performance evaluation; Step (4) by designing the frequency domain enhanced time network (DCTCN), combining the informer model with the time series mixer TSMixer to build the DCTCN-informer-TSMixer prediction model; Step (5) Train the constructed DCTCN-informer-TSMixer model based on the training set data and the validation set data, and adjust the hyperparameters until the model meets the expectations and then save it; Step (6) The trained DCTCN-informer-TSMixer prediction model is used to predict the time series data of the loose tempering process production line; Step (7) evaluates the model effect through four quality indicators and statistical test methods.

2. The method for predicting cigarette shred process quality based on the DCTCN-informer-TSMixer model according to claim 1, characterized in that: The sample data set is formed by: determining the process parameters and quality indicators based on the specific process of the loose rehydration process in the cigarette silk workshop; based on the determined process parameters and quality indicators, collecting time series data at preset time periods to form the sample data set; wherein, there are multiple process parameters and one quality indicator.

3. The method for predicting cigarette shred process quality based on the DCTCN-informer-TSMixer model according to claim 1, characterized in that: The preprocessing includes: first, setting upper and lower limits based on the export material moisture threshold and time window to process the head and tail of the material; processing missing values, if the instance has a large proportion of missing values, deleting the instance; if the missing data proportion is small, filling it through forward filling and backward filling methods; then using the 3σ criterion to further detect and delete outliers, and clean up gross errors and noise in the data; finally, using the Z-score method to normalize all process parameters and quality indicators.

4. The method for predicting cigarette shred process quality based on the DCTCN-informer-TSMixer model according to claim 1, characterized in that: The dimension reduction process includes: using Pearson correlation analysis to calculate the correlation coefficient between process parameters and quality indicators, and eliminating process parameters below a threshold.

5. The method for predicting cigarette shred process quality based on the DCTCN-informer-TSMixer model according to claim 1, characterized in that: The DCTCN-informer-TSMixer prediction model takes the sample data in the reduced-dimensional dataset as input and first performs a discrete cosine transform (DCT) on each channel data to enhance its frequency domain feature expression capability. Subsequently, the DCT-enhanced sequence is input into the channel-level feature adjustment module, and the fully connected network is used to generate adjustment weights for each channel to achieve feature compression and enhancement in the frequency domain; then, the processed sequence is input into the temporal convolutional network (TCN) module, and short-term temporal dependency information is extracted through multi-layer one-dimensional dilated convolution operations to form a local feature expression; at the same time, the original normalized sequence is input into the Informer module, and the long-term dependency is captured through the sparse attention mechanism to obtain a global semantic feature representation; the Informer output and the TCN output are additively fused according to the channel dimension and input into the TSMixer module as a mixed feature, in which the feature cross-transformations of the time dimension and the channel dimension are performed in turn to enhance the multi-dimensional information interaction and coupling representation; finally, the multi-step moisture content prediction result corresponding to the prediction step is output through the decoder and linear projection layer, and restored to the actual physical value through denormalization.

6. A method for predicting cigarette shred process quality based on the DCTCN-informer-TSMixer model, characterized in that: Includes the following modules: Module (1) is a collection module, which is used to collect time series data of a loose rehydration process production line in a cigarette shred workshop at a preset time to form a sample data set; wherein the time series data includes multiple process parameters and corresponding output moisture content quality index data; Module (2) preprocessing module, used to preprocess the sample data set, including data head and tail processing, missing value processing, outlier processing, data dimension reduction and normalization processing, to obtain preprocessed data; Module (3) is a partitioning module, which is used to divide the preprocessed sample data set into a training set, a validation set and a test set in chronological order for model training, parameter tuning and effect evaluation; Module (4) is a construction module for building a deep neural network structure based on the data, which integrates a discrete cosine transform channel enhancement module (DCT), a temporal convolutional network module (TCN), an informer attention module and a time-feature mixing module (TSMixer), so as to predict the moisture content of loose and rehydrated cigarettes; Module (5) training module, used to train the constructed DCTTCN-Informer-TSMixer model based on the training set and the validation set, optimize the network parameters and hyperparameter settings, and improve the prediction accuracy and generalization ability; Module (6) prediction module uses the trained model to predict the test set data or the time series data to be tested in actual production, and outputs the predicted value of the moisture content of the export material at the target time point.

7. The method for predicting cigarette shred process quality based on the DCTCN-informer-TSMixer model according to claim 6, characterized in that: The terminals involved in each of the above modules include a processor, a memory, and a computer program stored in the memory and runnable on the processor. The processor is configured to execute any one of the above-mentioned methods for predicting cigarette shred process quality based on the DCTCN-informer-TSMixer model.

8. The method for predicting cigarette shred process quality based on the DCTCN-informer-TSMixer model according to claim 7, characterized in that: The computer-readable storage medium stores a program, characterized in that when the program is executed by a processor, the processor implements the cigarette shred process quality prediction method based on the DCTCN-informer-TSMixer model according to any one of claims 1 to 5.

Images