Method and device for dynamically predicting moisture of cut tobacco at outlet of drum cut-tobacco dryer based on transformer model
Through the dynamic prediction method of moisture in the outlet leaf wire of the drum wire dryer based on the transformer model, the problem of insufficient prediction accuracy in the prior art is solved, and higher accuracy and stable moisture prediction are achieved, which improves industrial process control and product quality.
Patent Information
- Application Number
- CN202410143303.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-02-01
- Publication Date
- 2025-08-01
AI Technical Summary
The existing soft measurement technology is insufficient in the prediction of moisture in the outlet leaf wire of the drum dryer, especially the long and short-term memory network (LSTM) algorithm is difficult to describe the complex characteristics of the wire dryer, resulting in insufficient modeling accuracy.
The dynamic prediction method of outlet leaf wire of the drum dryer based on the transformer model is adopted. By obtaining the timing data of process variables, the position encoding module, encoder and decoder of the transformer model are used, and the attention mechanism, position encoding function, mask mechanism, residual connection and layer normalization are combined to realize customized sequence-to-sequence learning.
It improves prediction accuracy and stability, enhances the adaptability and reliability of the model, and promotes the improvement of industrial process control and product quality.
Smart Images
Figure CN120409171A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of deep learning, and particularly relates to a method and device for dynamically predicting the moisture content of cut tobacco at the outlet of a drum dryer based on a Transformer model. Background Art
[0002] In the field of cigarette production technology, a drum dryer is an important device in the tobacco processing process. The drum dryer is used to expand and dry cut tobacco so that the cut tobacco reaches a certain moisture content and quality requirements. The moisture content of the cut tobacco at the outlet of the dryer is an important factor affecting the quality and production efficiency of the cut tobacco. Therefore, it is very necessary to monitor and dynamically predict the moisture content of the cut tobacco at the outlet of the dryer in real time. However, since it is difficult to directly measure the moisture content of the cut tobacco at the outlet of the dryer, it is usually necessary to indirectly estimate it through soft sensor technology.
[0003] Soft sensor technology is a method that uses a mathematical model between process variables and moisture mass variables to estimate the moisture mass variables. The key to soft sensor technology is to establish an accurate, stable, and interpretable mathematical model. However, the current soft sensor technology is not accurate for predicting and evaluating the moisture content of the cut tobacco at the outlet of the dryer. Among them, the long short-term memory network (LSTM) algorithm performs well in sequence data prediction and is therefore widely used in variable prediction. However, the long short-term memory network (LSTM) algorithm has some drawbacks in processing long sequences, especially for the complex characteristics of the dryer, and it is difficult to accurately describe the dynamic characteristics, resulting in insufficient modeling accuracy.
[0004] In view of this, based on years of production design experience in this field and related fields, the inventor of the present invention has designed through repeated experiments a method and device for dynamically predicting the moisture content of cut tobacco at the outlet of a drum dryer based on a Transformer model, in order to solve the problems existing in the prior art. Summary of the Invention
[0005] The purpose of the present invention is to provide a method and device for dynamically predicting the moisture content of cut tobacco at the outlet of a drum dryer based on a Transformer model, which can effectively improve the prediction performance of the moisture content of the cut tobacco at the outlet of the dryer.
[0006] To achieve the above-mentioned invention object, the present invention proposes a method and device for dynamically predicting the moisture content of cut tobacco at the outlet of a drum cut tobacco dryer based on a transformer model. Among them, the method for dynamically predicting the moisture content of cut tobacco at the outlet of a drum cut tobacco dryer includes: obtaining the time series data of process variables during the drying process of the drum cut tobacco dryer; using the time series data of process variables as input, and using a trained dynamic prediction model of cut tobacco moisture content to predict the moisture content of cut tobacco at the outlet of the drum cut tobacco dryer; wherein, the dynamic prediction model of cut tobacco moisture content is constructed based on the transformer model and includes a position encoding module, an encoder and a decoder. The position encoding module is used to add a position encoding vector to the time series data of process variables and then transmit it to the encoder; the encoder is used to map the multi-dimensional time series data into a hidden state sequence, and the decoder is used to generate an output sequence according to the hidden state sequence, and the output sequence is used as the moisture content of cut tobacco at the outlet of the drum cut tobacco dryer.
[0007] Compared with the prior art, the present invention has the following characteristics and advantages:
[0008] The method and device for dynamically predicting the moisture content of cut tobacco at the outlet of a drum cut tobacco dryer based on the transformer model proposed by the present invention utilize the powerful non-linear fitting ability and long-term dependence capture ability of the transformer model. This solution has successfully achieved customized sequence-to-sequence learning for soft measurement problems. By combining key technologies such as the attention mechanism, position encoding function, masking mechanism, residual connection, and layer normalization, this solution is significantly superior to other classical models in terms of prediction accuracy and stability. In particular, in the problem of dynamically predicting the moisture content of cut tobacco at the outlet of the cut tobacco dryer, this solution demonstrates excellent performance and adaptability, providing a new idea and method for this field. This innovative application not only improves the prediction accuracy but also enhances the stability and reliability of the model, which is of great significance for promoting industrial process control and product quality improvement. Description of the Drawings
[0009] The drawings described herein are for illustrative purposes only and are not intended to limit the scope of the present invention disclosure in any way. Additionally, the shapes and proportional dimensions of the components in the drawings are only schematic and are used to assist in understanding the present invention, rather than specifically limiting the shapes and proportional dimensions of the components of the present invention. Those skilled in the art can select various possible shapes and proportional dimensions according to specific circumstances to implement the present invention under the teaching of the present invention.
[0010] Figure 1 It is a schematic flow chart of the prediction method of the present invention;
[0011] Figure 2 It is a schematic diagram of the iteration of different training times of the model of the present invention;
[0012] Figure 3This is a comparison chart of the prediction results of the model of the present invention and other different algorithms. Detailed implementation manners
[0013] Combined with the description of the accompanying drawings and the specific implementation manners of the present invention, the details of the present invention can be more clearly understood. However, the specific implementation manners of the present invention described herein are only for the purpose of explaining the present invention and cannot be understood in any way as a limitation of the present invention. Under the teaching of the present invention, those skilled in the art can conceive any possible variations based on the present invention, and these should be regarded as belonging to the scope of the present invention.
[0014] As Figure 1 shown, the present invention proposes a method and device for dynamically predicting the moisture content of cut tobacco at the outlet of a drum cut tobacco dryer based on a transformer model. The method for dynamically predicting the moisture content of cut tobacco at the outlet of a drum cut tobacco dryer includes: obtaining the time-series data of the process variables during the drying process of the drum cut tobacco dryer; using the time-series data of the process variables as input, and using a trained dynamic prediction model for the moisture content of cut tobacco to predict the moisture content of the cut tobacco at the outlet of the drum cut tobacco dryer; wherein, the dynamic prediction model for the moisture content of cut tobacco is constructed based on a transformer model and includes a position encoding module, an encoder, and a decoder. The position encoding module is used to add a position encoding vector to the time-series data of the process variables and then transmit it to the encoder; the encoder is used to map the multi-dimensional time-series data into a hidden state sequence, and the decoder is used to generate an output sequence according to the hidden state sequence, and use the output sequence as the moisture content of the cut tobacco at the outlet of the drum cut tobacco dryer.
[0015] The method and device for dynamically predicting the moisture content of cut tobacco at the outlet of a drum cut tobacco dryer proposed by the present invention utilize the powerful non-linear fitting ability and long-term dependence capture ability of the transformer model. This solution has successfully achieved customized sequence-to-sequence learning for soft measurement problems. By combining key technologies such as the attention mechanism, position encoding function, masking mechanism, residual connection, and layer normalization, this solution is significantly superior to other classical models in terms of prediction accuracy and stability. In particular, in the problem of dynamically predicting the moisture content of cut tobacco at the outlet of the cut tobacco dryer, this solution demonstrates excellent performance and adaptability, providing a new idea and method for this field. This innovative application not only improves the prediction accuracy but also enhances the stability and reliability of the model, which is of great significance for promoting industrial process control and product quality improvement.
[0016] The dynamic prediction method and device for the moisture content of cut tobacco at the outlet of a drum cut tobacco dryer based on the Transformer model proposed by the present invention aims to provide a dynamic prediction method for the moisture content of cut tobacco at the outlet of a cut tobacco dryer based on the Transformer model. Compared with the long short-term memory network (LSTM) algorithm, Transformer has advantages such as the ability to model long-range dependencies, parallel computing acceleration, and multi-head self-attention in time series prediction problems. This makes Transformer a powerful tool for dealing with time series prediction problems, capable of effectively capturing the relationships and features in the sequence and improving the prediction performance.
[0017] In an alternative embodiment of the present invention, the expression form of the dynamic prediction model for the moisture content of cut tobacco is:
[0018] y(t) = F(X(t),..., X(t - τ))
[0019] where t represents the sampling time, τ represents the length of the input time series, X = (x1′,..., x n ′) are the selected n process variables, and y(t) represents the output at the sampling time t, that is, the moisture content of the cut tobacco at the outlet of the cut tobacco dryer. The process of drying cut tobacco in a drum is a large-delay, nonlinear dynamic system with complex physical and chemical reactions and multi-phase and multi-field coupling. The moisture content at the outlet of the cut tobacco dryer can be described by this nonlinear model expression. By using historical time series data to predict the moisture content of cut tobacco at future moments, it provides strong support for quality control and process adjustment in the production process.
[0020] In an alternative embodiment of the present invention, the process variables include the moisture content of the inlet cut tobacco, the outlet temperature, the cylinder wall temperature, the steam pressure, the hot air temperature, and the opening degree of the moisture exhaust damper. Specifically, the moisture content of the inlet cut tobacco is the moisture content of the cut tobacco expansion at the inlet at the current moment (%); the outlet temperature is the outlet temperature of the cut tobacco expansion at the current moment (°C); the cylinder wall temperature is the cylinder wall temperature of the drum drying at the current moment (°C); the steam pressure is the steam pressure of the drum drying at the current moment (Pa); the hot air temperature is the hot air temperature of the drum drying at the current moment (°C); the opening degree of the moisture exhaust damper is the opening degree of the moisture exhaust damper of the drum drying at the current moment (%). By comprehensively considering multiple key process variables, multiple effects such as improved prediction accuracy, enhanced real-time adjustment ability, increased energy utilization efficiency, and improved equipment maintenance convenience are achieved.
[0021] In an alternative embodiment of the present invention, before inputting the process variables into the dynamic prediction model of cut tobacco moisture, denoising preprocessing and normalization processing are performed on the process variables. During the operation of the cut tobacco dryer, the collected data is always accompanied by noise signals. To ensure the accuracy of the model obtained through subsequent learning, it is necessary to denoise the collected data of the faulty batches. In this solution, a data denoising preprocessing method using wavelet transform is adopted. From the site of the cut tobacco dryer and in combination with the drying production process of cut tobacco, a total of 6 variables are collected. According to Table 1, it can be found that there are obvious differences between the values and dimensions of each process variable. To avoid adverse effects of such differences on the convergence performance of the model, normalization processing is required.
[0022] Table 1 Description of Process Input Variables
[0023]
[0024] Performing denoising preprocessing and normalization processing on the process variables in the dynamic prediction method of cut tobacco moisture at the outlet of the drum cut tobacco dryer can significantly improve the prediction accuracy, enhance the model stability, improve the calculation efficiency, and simplify the model tuning process.
[0025] In an alternative embodiment of the present invention, the position encoding module represents the position encoding vector using two sine functions and cosine functions with different frequencies, and the encoding vector is calculated. The main task of the position encoding module is to generate a unique position encoding vector for each position in the input sequence so that the model can capture the position information in the sequence. The position encoding module uses two sine functions and cosine functions with different frequencies to represent the position encoding vector. This method is called sine wave and cosine wave position encoding, which is one of the commonly used position encoding methods in the transformer model.
[0026] In an alternative example of this embodiment, the calculation formula for the position encoding vector of each position is:
[0027]
[0028] In the formula: pos represents the position of the current sample in the input sequence, and i is the current feature dimension of the input data. d model is the feature dimension of the sample, pos represents the position of the current sample in the input sequence, P is the position encoding matrix obtained by alternately encoding the sample order and feature position with sin(.) and cos(.) functions, and the dimension is pos * d model。The position encoding vector for each position is calculated through a specific formula. The formula combines sine and cosine functions to capture the information of each position in the sequence. Each element of the position encoding matrix is calculated through sine or cosine functions, which take X as the input and output a unique encoding matrix X+P. The alternating use of sine and cosine functions can ensure the uniqueness and continuity of the position encoding. P(pos,2i) and P(pos,2i+1) respectively represent the elements in the position encoding matrix corresponding to the current sample position pos and the 2i-th or 2i+1-th feature. The exponential term is used to adjust the frequency to ensure sufficient distinguishability of the encodings at different positions. By introducing position encoding, the performance of the model has been significantly improved. Position encoding provides additional context information to help the model better capture long-term dependencies and periodic patterns in time series data.
[0029] In an alternative embodiment of the present invention, the leaf moisture dynamic prediction model represents the output of the internal dependencies of the sequence through the self-attention mechanism, where the specific expression of the self-attention output is:
[0030]
[0031] where A(.) is the self-attention output, Q is the query vector, K is the key vector, V is the content vector, and D K is a factor used to scale the attention weights, and its value is the dimension of the K vector. The self-attention mechanism calculates the weights by taking the dot product of the query vector Q and the key matrix K, and then activates it using the Softmax function to obtain the weights corresponding to each position. These weights are used to weight-sum with V to obtain the self-attention output A(.). Among them, D K is a factor used to scale the attention weights, and its value is the dimension of the Q or K vector. Using Softmax to map the weights to (0,1), and using serves to alleviate the vanishing gradient problem, thereby making the model more stable and effective. The self-attention mechanism is highly robust to changes in the length and content of the input sequence. Regardless of how the length of the input sequence changes, or whether there are noises or outliers in the sequence, the model can effectively handle these situations through the self-attention mechanism and output stable prediction results.
[0032] In an alternative example of this embodiment, the query vector Q, the key vector K, and the content vector V have the following formulas:
[0033]
[0034] where X+P is the output matrix of the position encoding module, and W q 、W K and WV They are weight matrices respectively. In the self-attention mechanism, the query vector Q, the key vector K, and the content vector V are three key components that jointly determine the calculation method of the self-attention output. These vectors are usually obtained from the input data through linear transformation. The output matrix X + P is respectively multiplied by three weight matrices W q , W K and W V and linearly modified to obtain three corresponding vectors (matrices), namely the query vector Q, the key vector K, and the content vector V.
[0035] In this solution, by optimizing the calculation methods of the query vector Q, the key vector K, and the content vector V, the model can process large-scale data more efficiently. At the same time, this mechanism also makes it easier for the model to converge to a better solution during training, thereby improving the performance of the model. Among them, through the interaction between the query vector Q and the key vector K, the model can flexibly retrieve information related to the current position. This mechanism allows the model to automatically focus on the most important information when processing complex data, and then make more accurate decisions. The content vector V provides richer and more specific information for the model. By combining the information of the query vector Q and the key vector K, the content vector V can more accurately express the actual meaning of each position in the sequence, thereby improving the information expression ability of the model.
[0036] In an alternative example of this embodiment, the attention mechanism is extended to a multi-head attention mechanism, and the calculation formula of the multi-head attention mechanism is as follows:
[0037]
[0038] Among them, QWi i Q , and W o represent the multi-head attention weight matrix, h i refers to the i-th self-attention output, and the final output is the concatenation of all heads. When using the self-attention mechanism for experiments, it is found that using different linear functions to map Q, K, and V to h-dimensional vectors respectively can extract different attention content information. Among them, QWi i Q , and W o represent the multi-head attention weight matrix, h i refers to the i-th self-attention output, and the final output For the connection of all headers. The multi-head attention mechanism is an extension of the self-attention mechanism, which allows the model to simultaneously focus on information from different subspaces. By dividing the input sequence into multiple "heads" (i.e., multiple subspaces) and independently applying the self-attention mechanism to each head, the multi-head attention mechanism can capture richer and more diverse context information. Through the multi-head attention mechanism, the model can simultaneously focus on information from different aspects of the input sequence, thereby improving its ability to model complex contexts. By simultaneously focusing on information from multiple subspaces, the multi-head attention mechanism can capture richer and more diverse context information. This is crucial for processing complex natural language processing tasks, image recognition tasks, or time series data analysis tasks, as these tasks typically require understanding and modeling multi-level and multi-aspect information in the input data.
[0039] In an alternative example of this embodiment, each sublayer in the encoder and decoder contains a fully connected feed-forward network, and the formula representation of the feed-forward network is as follows:
[0040] FNN(x) = ReLU(xW1 + b1)W2 + b2
[0041] where W1 and W2 are weight matrices, and b1 and b2 are biases. Each sublayer in the encoder and decoder contains a fully connected feed-forward network (FFN), which is applied separately and identically to each position. The network consists of two linear transformations with a rectified linear unit (ReLU) activation function in the middle. Although the self-attention mechanism may not be able to effectively capture complex processes, the FNN can enhance the model's representation ability through dimensional transformation.
[0042] In an alternative embodiment of the present invention, residual connections and layer normalization are used in the encoder and the decoder. Among them, the encoder is responsible for encoding the input data (such as the historical moisture content of cut tobacco, environmental parameters, etc.) into an internal representation, while the decoder generates a prediction of the future moisture content based on this internal representation. The residual connection directly passes the original input information to the next layer of the network by adding the input sequence to the output of the layer. In this method, residual connections are introduced into the network structure and can be applied between each network layer to ensure the effective transmission of information between layers. By introducing residual connections, the model can better capture the long-term dependencies in the time-series data, thereby improving the prediction accuracy of the future moisture content. This method also applies the layer normalization (LayerNormalization) technique to the output of each layer of the encoder and the decoder. Layer normalization is a normalization technique used to normalize the input of each layer so that it has a zero mean and a unit variance. In this method, by normalizing the output of each layer of the neural network, its mean and variance are kept within a relatively stable range, which helps to reduce the internal covariate shift and improve the training speed and stability of the model. By introducing these two techniques, this method aims to improve the performance, stability, and training efficiency of the prediction model.
[0043] In an alternative example of this embodiment, the specific formula for layer normalization is as follows:
[0044] LayerNorm(x+Sublayer(x))
[0045] In the formula, LayerNorm represents the layer normalization operation, x represents the input of the current layer, and Sublayer(x) represents the output of the sublayer. Layer normalization helps the model better generalize to unseen data, thereby improving the prediction accuracy and stability.
[0046] In an alternative embodiment of the present invention, the Adam optimizer is used to train the dynamic prediction model of cut tobacco moisture. The Adam optimizer is an optimization algorithm with an adaptive learning rate, which combines the advantages of the AdaGrad and RMSProp optimization algorithms. Compared with traditional optimization algorithms (such as stochastic gradient descent), Adam can converge faster and has better performance, automatically adjusting the learning rate to adapt to different data and models, thereby improving the training efficiency and performance of the model. Moreover, it is applicable to most deep learning tasks and usually does not require manual adjustment of the learning rate parameter.
[0047] In an alternative embodiment of the present invention, the standard root mean square error is used to evaluate the prediction accuracy of the dynamic prediction model of cut tobacco moisture. The root mean square error is a commonly used index to measure the accuracy of a prediction model, and it evaluates the performance of the model by calculating the deviation between the predicted value and the true value.
[0048] In an alternative example of this embodiment, the calculation formula for the root mean square error is as follows:
[0049]
[0050] In the formula, RMSE represents the root mean square error, n represents the number of observed values, and y are the true value and predicted value of the quality variable at time t, respectively. represents the square of the difference between the predicted value and the actual value, and this difference can reflect the accuracy of the prediction. The advantage of using the root mean square error as an evaluation index is that it is more sensitive to large errors in the prediction because the errors are squared and then averaged. This makes the root mean square error an effective tool for identifying and improving parts of the model that may cause significant deviations.
[0051] The present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the dynamic prediction method for the moisture content of the cut tobacco at the outlet of the drum tobacco dryer based on the transformer model as described above are implemented.
[0052] The present invention also provides a computer device, including a memory and a processor. A computer program capable of running on the processor is stored on the memory. When the processor executes the computer program, the steps of the method for dynamically predicting the moisture content of the cut tobacco at the outlet of the drum tobacco dryer based on the transformer model as described above are implemented.
[0053] The specific implementation process of the present invention will now be described in detail in combination with an embodiment:
[0054] The experimental test environment is as follows: OS: Windows 10 Education Edition (64-bit); CPU: 11 th Gen IntelICoreI i7-11700@2.50GHz; RAM: 16.0GB; Python version is 3.6.13; TensorFlow-CPU version is 1.15.5.
[0055] Experiments were carried out on the data collected at the production site of the tobacco dryer from 10:54 on April 16, 2021 to 11:50 on September 1, 2021. During the experiment, the data sampling interval was 10 seconds. The first 35164 samples collected were used to train the model, the subsequent 5872 samples were used to verify the model, and the subsequent 969 samples were used to test the model.
[0056] The constructed training set and test set are input into the Transformer model for training. The mean squared error (MSE) is used to evaluate the control accuracy of the Transformer model until the MSE of the test set does not decrease within 60 consecutive training steps, at which point the training stops. Figure 2 It can be seen that the Transformer can reach convergence after 100 training steps, and the convergence speed can meet the control requirements. The MSE of the test set is lower than that of the training set after 60 training steps, indicating that the Transformer has good generalization ability and can accurately control different production batches.
[0057] Model parameter settings for LSTM and the model combined with the time attention mechanism (TA-LSTM): The number of training times is 100, the batch size is 128, the learning rate is 0.001, the number of lstm nodes is 128, the time window length is set to 12, and the dropout rate is 0.3. Model parameter settings for the Transformer: The number of training times is 100, the batch size is 256, the learning rate is 0.0008, the number of attention heads is 3, the time series window length is set to 12, the dropout rate is 0.3, and the prediction results are as follows Figure 3 shown. The root mean square error RMSE is shown in Table 2. In the comparison of prediction indicators under different algorithms, when using the Transformer model to predict the moisture content at the outlet of the cut tobacco dryer, the root mean square error (RMSE) is reduced by about 0.011 compared with the TA-LSTM model, and the coefficient of determination (R 2 ) is increased by about 0.074. Compared with the LSTM model, the root mean square error (RMSE) is reduced by about 0.02, and the coefficient of determination (R 2 ) is increased by about 0.200. Through Figure 3 it can be shown that by using the Transformer model, both the accuracy and stability of the moisture content at the outlet of the cut tobacco dryer are significantly improved. It can be seen that the Transformer of the present invention can more accurately predict the true value of the quality variable, verifying the effectiveness of the present invention.
[0058] Table 2 Comparison of prediction indicators under different algorithms
[0059]
[0060] Regarding the detailed explanations of the above various embodiments, their purposes are only to explain the present invention so that it can be better understood. However, these descriptions cannot be construed as limitations on the present invention for any reason. In particular, the various features described in different embodiments can be arbitrarily combined with each other to form other embodiments. Unless there are express contrary descriptions, these features should be understood to be applicable to any one of the embodiments and not limited only to the described embodiments.
Claims
1. A dynamic prediction method for the moisture content of cut tobacco at the outlet of a drum dryer based on a transformer model, characterized in that, The dynamic prediction method for the moisture content of cut tobacco at the outlet of the drum cut tobacco dryer includes: Obtaining the time series data of process variables during the drying process of the drum cut tobacco dryer; Using the time series data of the process variables as input, and predicting the moisture content of the cut tobacco at the outlet of the drum cut tobacco dryer by using the trained dynamic prediction model for cut tobacco moisture; Among them, the dynamic prediction model for cut tobacco moisture is constructed based on the Transformer model, and includes a position encoding module, an encoder and a decoder. The position encoding module is used to add a position encoding vector to the time series data of the process variables and then transfer it to the encoder; the encoder is used to map the multi-dimensional time series data into a hidden state sequence, and the decoder is used to generate an output sequence according to the hidden state sequence, and use the output sequence as the moisture content of the cut tobacco at the outlet of the drum cut tobacco dryer.
2. The dynamic prediction method for the moisture content of the cut tobacco at the outlet of the drum-type cut tobacco dryer according to claim 1, wherein The expression form of the dynamic prediction model for cut tobacco moisture is: y(t) = F(X(t),..., X(t - τ)) where \(t\) represents the sampling time, \(\tau\) represents the length of the input time series, \(X=(x'_1,\cdots,x' n )\) are the selected \(n\) process variables, and \(y(t)\) represents the output at the sampling time \(t\), that is, the moisture content of the cut tobacco leaves at the outlet of the cut tobacco dryer.
3. The dynamic prediction method for the moisture content of the cut tobacco at the outlet of the drum cut tobacco dryer according to claim 1, characterized in that, The process variables include the moisture content of the inlet cut tobacco, the outlet temperature, the cylinder wall temperature, the steam pressure, the hot air temperature and the opening degree of the moisture exhaust damper.
4. The dynamic prediction method for the moisture content of cut tobacco at the outlet of a drum cut tobacco dryer according to claim 1, wherein, Before inputting the process variables into the dynamic prediction model for cut tobacco moisture, denoising preprocessing and normalization processing are performed on the process variables.
5. The method for dynamically predicting the moisture content of the cut tobacco at the outlet of the drum-type cut tobacco dryer according to claim 1, wherein, The position encoding module uses two sine functions and cosine functions with different frequencies to represent the position encoding vector.
6. The method for dynamically predicting the moisture content of cut tobacco at the outlet of the drum cut tobacco dryer according to claim 4, wherein, The calculation formula for the position encoding vector at each position is: In the formula: pos represents the position of the current sample in the input sequence, and i is the current feature dimension of the input data; d model is the feature dimension of the samples, and the feature value dimension d of each sample model is 6. P is the position encoding matrix obtained by alternately encoding the sample order and feature positions with the sin(.) and cos(.) functions, and its dimension is pos * d model .
7. The method for dynamically predicting the moisture content of the cut tobacco at the outlet of the drum-type cut tobacco dryer according to claim 1, wherein, The dynamic prediction model for cut tobacco moisture represents the output of the internal dependencies of the sequence through the self-attention mechanism. Among them, the specific expression of the self-attention output is: where A(.) is the output of self-attention, Q is the query vector, K is the key vector, V is the content vector, and D K is a factor used to scale the attention weights, and its value is the dimension of the Q or K vector.
8. The moisture dynamic prediction method for the cut tobacco at the outlet of the drum cut tobacco dryer according to claim 7, wherein, Query vector Q, key vector K and content vector V, and their formulas are as follows: Among them, X + P is the output matrix of the position encoding module, and W Q , W K and W V are weight matrices respectively.
9. The method for dynamically predicting the moisture content of the cut tobacco at the outlet of the drum-type cut tobacco dryer according to claim 7, characterized in that, The attention mechanism is extended to a multi-head attention mechanism, and the calculation formula of the multi-head attention mechanism is as follows: Among them, QW i Q , W i K , W i V and W o represent the multi-head attention weight matrix, h i refers to the output of the i-th self-attention, and the final output X is the concatenation of all heads.
10. The moisture dynamic prediction method for cut tobacco at the outlet of the drum-type cut tobacco dryer according to claim 7, characterized in that, Each sub-layer in the encoder and the decoder contains a fully connected feed-forward network, and the formula representation of the feed-forward network is as follows: FNN(x) = ReLU(xW1 + b1)W2 + b2 Where W1 and W2 are weight matrices, and b1 and b2 are biases.
11. The moisture dynamic prediction method for the cut tobacco at the outlet of the drum cut tobacco dryer according to claim 1, wherein, Residual connections and layer normalization are used in the encoder and the decoder.
12. The moisture dynamic prediction method for the cut tobacco at the outlet of the drum cut tobacco dryer according to claim 11, characterized in that The specific formula for the layer normalization is as follows: LayerNorm(x + Sublayer(x)) In the formula, LayerNorm represents the layer normalization operation, x represents the input of the current layer, and Sublayer(x) represents the output of the sub-layer.
13. The moisture dynamic prediction method for the cut tobacco at the outlet of the drum cut tobacco dryer according to claim 1, wherein, The Adam optimizer is used to train the dynamic prediction model for cut tobacco moisture.
14. The method for dynamically predicting the moisture content of the cut tobacco at the outlet of the drum cut tobacco dryer according to claim 1, wherein, The prediction accuracy of the dynamic prediction model for cut tobacco moisture is evaluated by using the standard root mean square error.
15. The method for dynamically predicting the moisture content of cut tobacco at the outlet of the drum-type cut tobacco dryer according to claim 13, wherein The calculation formula for the root mean square error is as follows: wherein, and y are the true value and the predicted value of the mass variable at time t, respectively.
16. A computer-readable storage medium having a computer program stored thereon, wherein, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 15.
17. A computer device, comprising a memory and a processor, and a computer program capable of running on the processor is stored on the memory, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 15.