Space-time mechanism guided roller cut tobacco drying temperature and humidity variable step length prediction method

Abstract

The invention relates to a space-time mechanism guided roller cut tobacco drying outlet cut tobacco temperature and humidity variable step length prediction method, which comprises the following steps: collecting operation data in a stable state in a roller cut tobacco drying process as a training set, and carrying out sliding window division according to the time duration of a single batch of cut tobacco passing through the cut tobacco drying process; variables related to tobacco shred temperature and humidity prediction are selected according to expert experience, and local spatial-temporal feature extraction is performed on K control units in the tobacco shred drying process by using a long-short-term memory network and a graph convolutional network; for other process variables which are not selected, effective information in the process variables is extracted by using a sparse auto-encoder. According to the method, the equipment control principle and space-time mechanism information in the roller cut tobacco drying process are fully utilized, compared with a traditional data driving prediction method, the model convergence speed is higher, better accuracy and interpretability are achieved, and a better prediction algorithm is provided for temperature and humidity prediction of outlet cut tobacco in the roller cut tobacco drying process.

Description

Technical Field

[0001] The present invention belongs to the technical field of industrial process soft measurement, and particularly relates to a variable step length prediction method for temperature and humidity of a drum drying wire guided by a time-space mechanism. Background Art

[0002] In the tobacco industry, roller drying is one of the key steps in cigarette manufacturing, and its main task is to process tobacco leaves into tobacco suitable for cigarette production. The roller drying process includes humidification, softening, and drying. Specifically, the SIROX leaf expander is used to expand and humidify the tobacco slices to restore the appropriate moisture content and increase the softness for subsequent processing. This is the humidification process and softening process. The KLD thin plate drying machine uses a dual temperature zone control method to control the wall temperature of the drum, which is divided into zone 1 and zone 2. During the roller drying process, the temperature and humidity of the tobacco are affected by the KLD thin plate drying machine adjusting the hot air and the wall temperature, and are also affected by the expansion and humidification of SIROX. The control of the above two processes has strong nonlinear and dynamic change characteristics.

[0003] In actual production, there are many unstable factors in the environment, such as uneven moisture distribution and uneven heat conduction of tobacco, which make it extremely difficult to control the temperature and humidity of tobacco. The manual control process of tobacco has slow response and sometimes incorrect adjustment of parameters. The above factors ultimately affect the quality of cigarettes. Summary of the invention

[0004] The present invention proposes a variable step prediction method for temperature and humidity of drum drying of silk threads guided by a time-space mechanism to solve the above technical problems.

[0005] A method for predicting the temperature and humidity of tobacco at the outlet of a roller drying process with variable step length guided by a spatiotemporal mechanism, the method comprising: collecting the operating data of the roller drying process in a steady state as a training set, and performing sliding window division according to the length of time a single batch of tobacco passes through the drying process; selecting variables related to the prediction of tobacco temperature and humidity according to expert experience, and using a long short-term memory network and a graph convolutional network to extract local spatiotemporal features for K control units of the drying process; using a sparse autoencoder to extract effective information from other process variables that are not selected; using a long short-term memory network to extract temporal features from historical temperature and humidity data; using an adjacency matrix to express the spatial relationship between the K control units of the roller drying process, adding other data features and historical temperature and humidity features to expand the adjacency matrix, and using a graph convolutional network to perform global feature fusion; and selecting a suitable prediction step length for prediction according to the actual needs of the roller drying process.

[0006] In some embodiments, the operation data of the drum drying process in a steady state is collected as a training set, and sliding window division is performed according to the length of time a single batch of tobacco shreds undergoes the drying process, including: collecting M samples of the drum drying process in a steady state as a training set , N samples as the test set , and obtain the time length T1 for a single batch of tobacco to pass through K control units, and the interval T2 for the system to collect data; according to the mechanism information, the temperature information of the exported tobacco is determined by the information generated when the tobacco passes through K control units, and the sliding window length is set , the data are divided into three categories, namely the variables V selected by the experts 1 , other variables V 2 , temperature and humidity data variable V 3 , the input and label of the model at time t are defined as: , , for large-scale data sets, the input matrix X contains the input data segments at each moment, and the output matrix Y contains the corresponding predicted data segments, defined as: , .

[0007] In some embodiments, the method of selecting variables related to the prediction of tobacco temperature and humidity according to expert experience, and using a long short-term memory network and a graph convolution network to extract local spatiotemporal features for K control units in the tobacco drying process, comprises: 1 According to the number of control units, they are divided into K categories: ,Will Divide , respectively The samples in are input into the long short-term memory network in chronological order, and the temporal features of K control units are extracted. For samples of different control units Calculate and get the final state , is the time series feature extracted from K control units; for samples of different control units , calculate the K group correlation coefficient matrix , for different variables within each control unit, the Pearson correlation coefficient calculated for two variables as a group is defined as: , where i and j represent different variables, and Represents the mean of different variables, and sets the threshold according to the Pearson correlation coefficient , define the adjacency matrix A: , design a graph convolutional network, the input is the adjacency matrix A and the samples of different control units of the feature matrix : ,in, is the adjacency matrix plus the identity matrix, yes The degree matrix of is the weight matrix of the first layer, is the activation function, is the weight matrix of the second layer, for samples of different control units Calculate and get the final state , The spatial information features extracted by the K control units at this time; the temporal features of the K control units and spatial information features Merge to form K comprehensive features ; The features of K control units are connected through the fully connected layer Mapped into K features: ,in, is the weight matrix of the fully connected layer, is the bias vector, is the activation function, and finally we get , that is, the local spatiotemporal characteristics of K control units.

[0008] In some embodiments, for other unselected process variables, using a sparse autoencoder to extract valid information therein includes: for other data , design a sparse autoencoder whose input layer has the same number of nodes as the feature dimension n of the time series data. For time t, the encoder will input data Mapping to a low-dimensional sparse representation , the decoder outputs the original data estimate , let the parameters of the sparse autoencoder be ; Use the back-propagation algorithm to train the parameters of the sparse autoencoder by minimizing the loss function , where the loss function is ,in, is the input data, is an estimate of the decoder output, is the regularization coefficient, express After training, the encoder is used to obtain the sparse features of other data. .

[0009] In some embodiments, the use of a long short-term memory network to extract temporal features from historical temperature and humidity data includes: The samples in the LSTM network are input into the LSTM network in chronological order to extract the historical temperature and humidity features. Calculate and get the final state , The historical temperature and humidity characteristics at this time are recorded as .

[0010] In some embodiments, for the K control units of the drum drying process, the spatial relationship between them is expressed using an adjacency matrix, other data features and historical temperature and humidity features are added to expand the adjacency matrix, and a graph convolution network is used for global feature fusion, including:

[0011] According to the interaction mechanism and physical position relationship between K control units in the drum drying process, the spatial mechanism information matrix A is constructed, that is, ,in, Indicates the control unit and control unit Whether they are connected, the value is 0 or 1; expand the matrix A, and get , connect K+2 features in parallel to get the input matrix F: , through the operation of two layers of GCN, the fused feature matrix is ​​obtained ,Right now ,in, is the adjacency matrix plus the identity matrix, yes The degree matrix of is the weight matrix of the first layer, is the activation function, is the weight matrix of the second layer, Each row of represents a fused feature vector, that is ,in, Represents the fused feature vector i.

[0012] In some embodiments, according to the actual requirements of the drum drying process, a suitable prediction step size is selected for prediction, including: the calculation formula of the self-attention mechanism is: ,in, , , , , , is a trainable weight, D is the dimension of the input feature, and for predictions at different times, use The calculation formula of the multi-head self-attention mechanism is: ,in, is the attention prediction result corresponding to the future i-th moment; The calculation of the attention head is prediction results, corresponding to the future The predicted value of temperature and humidity for each time step is ,in, Represents the prediction result of the tth time step; the back propagation algorithm is used to train the overall network by minimizing the loss function, where the loss function is After the training is completed, according to the actual need to control the tobacco quality in the drum drying process or other task requirements, the appropriate prediction step size is selected from the prediction results.

[0013] The beneficial effects of the present disclosure are as follows: the temperature and humidity of tobacco at the outlet of the drum drying process can be predicted with high precision. The variable step-size prediction method for the temperature and humidity of tobacco at the outlet of the drum drying guided by the spatiotemporal mechanism solves the problem that the temperature and humidity of tobacco are difficult to predict accurately due to the complex nonlinearity and spatiotemporal coupling in the drum drying process. This method integrates the spatiotemporal mechanism characteristics of the drying process on the basis of data-driven, and can realize the variable step-size prediction of temperature and humidity according to the needs of online applications. Specifically, the data is first divided into sliding windows according to the time series characteristics, and then the local spatiotemporal characteristics of K control units are extracted respectively according to the process flow of the process using a neural network, and the sparse characteristics of other process data and the time series characteristics of historical temperature and humidity are extracted. According to the mechanism relationship and physical position relationship between the control units in the drying process, the above features are fused with global spatial features using a graph convolutional network. In order to improve the convenience of user use, a variable step-size prediction function is also developed to output the prediction results of tobacco temperature and humidity with a specified step-size. The present invention makes full use of the equipment control principles and spatiotemporal mechanism information of the drum drying process. Compared with the traditional data-driven prediction method, the model converges faster and has better accuracy and interpretability, providing a better prediction algorithm for the temperature and humidity prediction of the outlet tobacco in the drum drying process.

[0014] Other features and advantages of the present disclosure will become apparent from the following detailed description of exemplary embodiments of the present disclosure with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] The accompanying drawings, which are incorporated in and constitute a part of the specification, illustrate embodiments of the present disclosure and, together with the description, serve to explain the principles of the present disclosure.

[0016] Figure 1 Schematic diagram of the four control units in the drum drying process.

[0017] Figure 2 This is the model structure diagram for predicting the outlet temperature of tobacco in the drum drying process. DETAILED DESCRIPTION

[0018] Various exemplary embodiments of the present disclosure will now be described in detail with reference to the accompanying drawings. It should be noted that the relative arrangement of components and steps, numerical expressions and numerical values ​​set forth in these embodiments do not limit the scope of the present disclosure unless otherwise specifically stated.

[0019] The following description of at least one exemplary embodiment is merely illustrative in nature and is in no way intended to limit the present disclosure, its application, or uses.

[0020] Technologies, methods, and equipment known to ordinary technicians in the relevant art may not be discussed in detail, but where appropriate, the technologies, methods, and equipment should be considered as part of the specification.

[0021] In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limiting. Therefore, other examples of the exemplary embodiments may have different values.

[0022] This disclosure uses the prediction of outlet tobacco temperature and humidity in a drum drying process (including 4 control units, SIROX unit, zone 1 drum wall temperature unit, zone 2 drum wall temperature unit, KLD hot air unit) of a tobacco factory in East China as an application scenario for explanation:

[0023] In this embodiment, a total of 5000 sample data were collected during the tobacco drying process, of which 4500 samples were used as a training set to train the model, and 500 samples were used as a test set to test the model. Each sample contained 42 variables, of which the outlet tobacco temperature and outlet tobacco humidity were target variables; the variables were all collected when the tobacco drying process was running smoothly.

[0024] like Figure 1 and Figure 2 As shown, the variable step length prediction method of temperature and humidity of cut tobacco at the outlet of roller drying guided by the spatiotemporal mechanism includes the following steps:

[0025] Step 1: (1) Collect normal operation data of the drum drying process: 4500 samples of the drum drying process in a stable state are collected as training sets, and 500 samples are collected as test sets. The time length of a single batch of tobacco passing through the four control units of SIROX unit, KLD1 zone drum wall temperature unit, KLD2 zone drum wall temperature unit, and KLD hot air unit is 270 seconds, and the interval of system data collection is 10 seconds.

[0026] (2) Classify the data and divide the data set into sliding windows: According to the mechanism information, the temperature and humidity information of the exported tobacco can be determined by the information generated when the tobacco passes through the four control units, so the sliding window length T is set to 270s / 10s=27. Then the data is divided into three categories: (Expert-selected variables) with a dimension of 14, (Other variables) dimension is 26, (Temperature and humidity data variables) dimension is 2. The input of the model at time t can be defined as: , the label is defined as: .

[0027] For large datasets, the input matrix contains the input data segment at each moment, and the output matrix Contains the corresponding prediction data segment, which is defined as , .

[0028] Step 2: (1) Use LSTM to extract the time series features of the four control units respectively: First, Divided into 4 categories: (SIROX unit), (KLD1 zone cylinder wall temperature unit), (KLD2 zone cylinder wall temperature unit), (KLD hot air unit). Specific division , , , . Then respectively , , , The samples in are input into the long short-term memory network in chronological order to extract the temporal features of the four control units. , , , , the final state obtained by calculation , , , , , , , These are the timing features extracted from the four control units at this time.

[0029] (2) Use GCN to extract spatial information features of the four control units: First, use the Pearson correlation coefficient to help define the adjacency matrix. For samples of different control units , , , , 4 sets of correlation coefficient matrices are calculated , , ,For different variables within each control unit, the Pearson correlation coefficient calculated for a group of two variables is defined as: , where i and j represent different variables, and They represent the means of different variables.

[0030] Setting thresholds based on Pearson correlation coefficient , define the adjacency matrix ; Design a graph convolutional network (GCN) with adjacency matrix as input Samples of different control units with characteristic matrices , , , : ,in, is the adjacency matrix plus the identity matrix, yes is the weight matrix of the first layer, is the activation function. In this embodiment, all activation functions are ReLU functions. is the weight matrix of the second layer.

[0031] Samples for different control units , , , The final state obtained by calculation , , , These are the spatial information features extracted by the four control units at this time.

[0032] (3) Combining time series features and spatial information features to obtain four unit features: The time series features of the four control units , , , and spatial information features , , , Merge to form 4 comprehensive features , , , ;Finally, the features of the four control units are connected through the fully connected layer , , , Mapped into 4 features: ,in, is the weight matrix of the fully connected layer, b is the bias vector, is the activation function. Finally, we get , , , 4 control unit features.

[0033] Step 3: Use sparse autoencoders to extract other data features: For other data , design a sparse autoencoder whose input layer has the same number of nodes as the feature dimension of the time series data At time t, the encoder will input data Mapping to a low-dimensional sparse representation , the decoder reconstructs the sparse representation back to the original data .

[0034] Using the back-propagation algorithm, the parameters of the sparse autoencoder are trained by minimizing the following loss function : ,in, is the regularization coefficient, which is set to 0.1 in this embodiment. Represents the 1-norm in W. After training, the encoder is used to obtain the sparse features of other data. .

[0035] Step 4: Use LSTM to extract historical temperature and humidity features: The samples in the LSTM network are input into the LSTM network in chronological order to extract the historical temperature and humidity features. The final state obtained by calculation The historical temperature and humidity characteristics at this time are recorded as .

[0036] Step 5: (1) Analyze the spatial position relationship to obtain the spatial mechanism information matrix: Construct the spatial mechanism information matrix based on the interaction mechanism and physical position relationship between the four control units, such as Figure 1 shown.

[0037] The spatial mechanism information matrix is ​​constructed according to the mutual influence relationship between the four control units, namely .

[0038] (2) Expand the matrix and use the graph convolutional network to fuse features according to the matrix: Because other data features and historical temperature and humidity features have an impact on the four control units, the matrix Expand and finally get the matrix ,Right now . Concatenate the six features to get the input matrix F: , then, through the operation of two layers of GCN, the fused feature matrix is ​​obtained : ,in, is the adjacency matrix plus the identity matrix, yes is the weight matrix of the first layer, is the activation function. is the weight matrix of the second layer.

[0039] Through the operation of two layers of GCN, the fused feature matrix is ​​obtained . Each row of the matrix represents a fused feature vector: ,in, Represents the fused feature vector Finally, we get 6 features after fusion , , , , , .

[0040] Step 6: Online variable step-size temperature and humidity prediction: The fused features The multi-head self-attention algorithm is used, and the fully connected network is applied to each feature, and the temperature and humidity prediction results of the drum drying process are finally calculated. .

[0041] The calculation formula of the self-attention mechanism is as follows: ,in, , , , is a trainable weight, and D is the dimension of the input feature. For predictions at different future moments, the self-attention mechanism is used to calculate them, so we use Multi-head self-attention mechanism: ,in, is the attention prediction result corresponding to the future i-th moment. The calculation of the attention head is prediction results, corresponding to the future Temperature and humidity prediction values ​​for time steps: ,in, Represents the prediction result of the tth time step. Use the back propagation algorithm to train the overall network by minimizing the loss function, and the loss function is: According to the actual need to control the quality of tobacco in the drum drying process or other task requirements, in this example, the results of the next 6 time steps are predicted, that is, the first 6 prediction results are output: .

[0042] The present disclosure was used to test a test set containing 500 samples. The results of the single-step prediction experiment are shown in the following table:

[0043] Experimental effect comparison table

[0044]

[0045] According to the comparison between this method and the long short-term memory network, this disclosure can significantly improve the accuracy of the prediction of the temperature and humidity of the outlet tobacco in the drum drying process. In addition, the method described in this disclosure combines the mechanism and spatial position relationship between the four control units of the drum drying process (SIROX unit, 1 zone drum wall temperature unit, 2 zone drum wall temperature unit, KLD hot air unit), effectively improving the interpretability and reliability of the data-driven model, and has higher prediction accuracy and faster convergence speed than traditional methods, which provides great help in predicting the outlet temperature and humidity of tobacco.

[0046] Although some specific embodiments of the present disclosure have been described in detail by way of examples, it should be understood by those skilled in the art that the above examples are for illustration only and are not intended to limit the scope of the present disclosure. It should be understood by those skilled in the art that the above embodiments may be modified without departing from the scope and spirit of the present disclosure. The scope of the present disclosure is defined by the appended claims.

Claims

1. A method for predicting the temperature and humidity of cut tobacco at the outlet of a drum drying machine guided by a spatiotemporal mechanism, characterized in that: The method comprises: The running data of the drum tobacco drying process in a stable state is collected as the training set, and the sliding window division is performed according to the length of time a single batch of tobacco goes through the drying process; Variables related to tobacco temperature and humidity prediction are selected based on expert experience, and local spatiotemporal features are extracted for K control units in the tobacco drying process using long short-term memory networks and graph convolutional networks. For other unselected process variables, a sparse autoencoder is used to extract effective information from them; For historical temperature and humidity data, long short-term memory network is used to extract the time series features; For the K control units in the drum drying process, the spatial relationship between them is expressed using the adjacency matrix. Other data features and historical temperature and humidity features are added to expand the adjacency matrix, and a graph convolutional network is used for global feature fusion. According to the actual needs of the drum silk drying process, select the appropriate prediction step size for prediction.

2. The method for predicting the temperature and humidity of cut tobacco at the outlet of the drum drying machine guided by the spatiotemporal mechanism according to claim 1 is characterized in that: The operation data in the steady state of the drum tobacco drying process is collected as a training set, and sliding window division is performed according to the length of time a single batch of tobacco shreds undergoes the tobacco drying process, including: Collect M samples of the drum drying process in a stable state as the training set , N samples as the test set , and obtain the time length T1 of a single batch of tobacco passing through K control units, and the interval T2 for system data collection; According to the mechanism information, the temperature information of the exported tobacco is determined by the information generated when the tobacco passes through K control units, and the sliding window length is set. , the data is divided into three categories, namely, the expert selected variable V1, other variables V2, and temperature and humidity data variable V3. The input and label of the model at time t are defined as: , , for large-scale data sets, the input matrix X contains the input data segments at each moment, and the output matrix Y contains the corresponding predicted data segments, defined as: , .

3. The method for predicting the temperature and humidity of cut tobacco at the outlet of the drum drying machine guided by the spatiotemporal mechanism according to claim 2 is characterized in that: The method selects variables related to the prediction of tobacco temperature and humidity based on expert experience, and uses a long short-term memory network and a graph convolutional network to extract local spatiotemporal features for K control units in the tobacco drying process, including: V1 is divided into K categories according to the number of control units: ,Will Divide , respectively The samples in are input into the long short-term memory network in chronological order, and the temporal features of K control units are extracted. For samples of different control units Calculate and get the final state , is the time series features extracted from K control units; Samples for different control units , calculate the K group correlation coefficient matrix , for different variables within each control unit, the Pearson correlation coefficient calculated for two variables as a group is defined as: , where i and j represent different variables, and Represents the mean of different variables, and sets the threshold according to the Pearson correlation coefficient , define the adjacency matrix A: , design a graph convolutional network, the input is the adjacency matrix A and the samples of different control units of the feature matrix : ,in, is the adjacency matrix plus the identity matrix, yes The degree matrix of is the weight matrix of the first layer, is the activation function, is the weight matrix of the second layer, for samples of different control units Calculate and get the final state , This is the spatial information feature extracted by the K control units at this time; The timing characteristics of K control units and spatial information features Merge to form K comprehensive features ; The features of K control units are transformed through the fully connected layer Mapped into K features: ,in, is the weight matrix of the fully connected layer, is the bias vector, is the activation function, and finally we get , that is, the local spatiotemporal characteristics of K control units.

4. The method for predicting the temperature and humidity of cut tobacco at the outlet of the drum drying machine guided by the spatiotemporal mechanism according to claim 3 is characterized in that: For other unselected process variables, a sparse autoencoder is used to extract effective information therein, including: For other data , design a sparse autoencoder whose input layer has the same number of nodes as the feature dimension n of the time series data. For time t, the encoder will input data Mapping to a low-dimensional sparse representation , the decoder outputs the original data estimate , let the parameters of the sparse autoencoder be ; Use the back-propagation algorithm to train the parameters of the sparse autoencoder by minimizing the loss function , where the loss function is ,in, is the input data, is an estimate of the decoder output, is the regularization coefficient, express 1-norm of ; After training is completed, the encoder is used to obtain sparse features of other data .

5. The method for predicting the temperature and humidity of cut tobacco at the outlet of the drum drying machine guided by the spatiotemporal mechanism according to claim 4 is characterized in that: For the historical temperature and humidity data, the long short-term memory network is used to extract the time series features, including: Will The samples in the LSTM network are input into the LSTM network in chronological order to extract the historical temperature and humidity features. Calculate and get the final state , The historical temperature and humidity characteristics at this time are recorded as .

6. The method for predicting the temperature and humidity of cut tobacco at the outlet of the drum drying machine guided by the spatiotemporal mechanism according to claim 5 is characterized in that: For the K control units in the drum drying process, the spatial relationship between them is expressed using an adjacency matrix, other data features and historical temperature and humidity features are added to expand the adjacency matrix, and a graph convolutional network is used for global feature fusion, including: According to the interaction mechanism and physical position relationship between K control units in the drum drying process, the spatial mechanism information matrix A is constructed, that is, ,in, Indicates control unit and control unit Whether they are connected, the value is 0 or 1; Expand the matrix A to get , connect K+2 features in parallel to get the input matrix F: , through the operation of two layers of GCN, the fused feature matrix is ​​obtained ,Right now ,in, is the adjacency matrix plus the identity matrix, yes The degree matrix of is the weight matrix of the first layer, is the activation function, is the weight matrix of the second layer, Each row of represents a fused feature vector, that is ,in, Represents the fused feature vector i.

7. The method for predicting the temperature and humidity of cut tobacco at the outlet of the drum drying machine guided by the spatiotemporal mechanism according to claim 6 is characterized in that: According to the actual needs of the drum silk drying process, a suitable prediction step is selected for prediction, including: The calculation formula of the self-attention mechanism is: ,in, , , , , , is a trainable weight, D is the dimension of the input feature, and for predictions at different times, use The calculation formula of the multi-head self-attention mechanism is: ,in, is the attention prediction result corresponding to the i-th moment in the future; pass The calculation of the attention head is prediction results, corresponding to the future The predicted value of temperature and humidity for each time step is ,in, Represents the prediction result of the tth time step; The back propagation algorithm is used to train the entire network by minimizing the loss function, where the loss function is ; After training is completed, the appropriate prediction step size is selected from the prediction results based on the actual need to control the tobacco quality during the drum drying process or other task requirements.

Images