Process parameter prediction control method and system in tobacco leaf feeding process, medium and terminal

Abstract

The invention discloses a process parameter prediction control method and system in a tobacco leaf charging process, a medium and a terminal, and the method comprises the steps: obtaining process parameter data of a current time step and a plurality of continuous historical time steps before the current time step in the tobacco leaf charging process, carrying out the data preprocessing, and converting the data into a corresponding time sequence sample; and inputting the process parameter data of the current time step and the historical time step into the constructed and trained process parameter prediction model, and outputting a predicted value of the target process parameter of the next time step. According to the method, the problems of insufficient feature modeling, weak time sequence dependence capturing capability and lack of online updating mechanism in the prediction control of the tobacco leaf feeding process in the prior art are solved, so that efficient and accurate pre-judgment and dynamic regulation and control of key process parameters in the tobacco leaf feeding process are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of predictive control of tobacco leaf feeding process, and in particular to a process parameter predictive control method and system in a tobacco leaf feeding process, a medium and a terminal. BACKGROUND

[0002] With the continuous development of the tobacco industry towards intelligent and fine production, the control of key process parameters in the tobacco leaf processing process has become an important link to ensure the stability and consistency of product quality. Among many process steps, the tobacco leaf feeding process is one of the key steps that affect the internal quality of the cut tobacco. The core goal is to accurately control the instantaneous moisture content of the outlet material by precisely controlling the hot air temperature, moisture discharge opening and other control variables under the premise of ensuring material uniformity.

[0003] However, in actual production, there are many challenges in the tobacco leaf feeding process: Firstly, the collected historical process data often contains abnormal or non-steady-state data due to factors such as equipment start-stop, raw material batch changes, etc. Traditional data cleaning methods are difficult to effectively identify complex distribution patterns in high-dimensional feature space, which may lead to incomplete noise removal or omission of normal data, affecting the reliability of modeling. Secondly, the process parameters in the time series contain multi-scale characteristics such as local short-term fluctuations and cross-time step long-term dependencies. Existing modeling methods mostly use single-scale feature extraction methods (such as fixed window moving average or single volume kernel size), which are difficult to fully capture the feature correlation of different time scales, resulting in insufficient representation ability of the model for complex working conditions. Thirdly, the strong time sequence of the dynamic evolution of the feeding process requires the model to effectively capture the dependence between the historical state and the current output. However, traditional recurrent neural networks (RNN) are easily affected by the gradient vanishing problem, and LSTM models, although they alleviate this problem, have the defects of complex structure and high computational cost, which have limited applicability in industrial real-time control scenarios. Fourthly, factors such as fluctuations in tobacco raw material properties and equipment aging may cause slow changes in production conditions. Existing models are trained based on fixed historical data and lack online updating mechanism, which makes it difficult to adapt to the dynamic drift of process parameters in real time, resulting in a decline in prediction accuracy over time, and frequent manual retraining is required, which is costly to operate and maintain.

[0004] Therefore, there is an urgent need for a predictive control method that can deeply integrate multi-scale features, capture long-term sequence dependencies, and have self-adaptive updating capability, to achieve efficient and accurate prediction and dynamic regulation of key quality indicators (outlet moisture) and related control variables (hot air temperature, moisture discharge opening) in the tobacco leaf feeding process, thereby improving the intelligent level of the entire production line and the consistency of product quality. SUMMARY

[0005] In view of the deficiencies of the prior art in the background art, the present application provides a process parameter prediction control method, system, medium and terminal in a tobacco leaf dosing process, which aims to solve the problems of insufficient feature modeling, weak time sequence dependence capturing ability and lack of online updating mechanism in the prior art in the prediction control of the tobacco leaf dosing process, so as to realize efficient and accurate prediction and dynamic regulation and control of key process parameters in the tobacco leaf dosing process.

[0006] In a first aspect, the present application provides a process parameter prediction control method in a tobacco leaf dosing process, comprising: Obtaining process parameter data of a current time step and a plurality of historical time steps before the current time step in the tobacco leaf dosing process and performing data preprocessing to convert the process parameter data into corresponding time sequence samples, taking the time sequence samples of the process parameter data of the current time step and the historical time steps as inputs, inputting the inputs into a trained process parameter prediction model, and outputting a predicted value of a target process parameter of a next time step; the specific process of the process parameter prediction model is as follows: S1: performing feature extraction on the time sequence samples of the process parameters of the historical time steps by using a multi-scale one-dimensional convolution, and then fusing the extracted features of each scale to obtain process parameter feature sequences of the plurality of historical time steps; S2: inputting the process parameter feature sequences of the plurality of historical time steps in S1 into a selective state space model module (i.e., a Mamba module) to extract time sequence features, and correspondingly obtaining hidden state sequences of the process parameters of each historical time step, and then obtaining a time domain historical hidden state matrix composed of the hidden state sequences; S3: mapping the time sequence samples of the process parameters of the current time step to a hidden space by using a linear transformation layer to obtain process parameter features of the current time step; S4: inputting the time domain historical hidden state matrix and the process parameter features of the current time step into an attention module to obtain time domain comprehensive features; S5: performing fast Fourier transform on each feature dimension of the process parameter feature sequences output by S1 to obtain a frequency domain complex sequence, calculating the amplitude of each frequency component in the frequency domain complex sequence to obtain an amplitude sequence, and performing global average pooling on the amplitude sequence to obtain a frequency domain feature; S6: splicing the time domain comprehensive features obtained by S4 and the frequency domain features obtained by S5 to obtain fusion features; S7: performing linear transformation on the comprehensive features by a fully connected layer to output a predicted value of a target process parameter of a next time step.

[0007] Further, the training process of the process parameter prediction model is as follows: S001: collecting multi-source process parameter data of different sampling periods in a historical tobacco leaf process to construct a high-dimensional feature space matrix; S002: Dimensionality reduction mapping of high-dimensional data in the high-dimensional feature space matrix to a two-dimensional space using t-distributed stochastic neighbor embedding algorithm in batch mode with a sampling period, to obtain a two-dimensional scatter plot; S003: Cluster analysis of scatter plot data in the two-dimensional scatter plot using K-Means clustering algorithm, clean abnormal data, and standardize the high-dimensional data corresponding to the cleaned scatter plot data; S004: For each batch of data after standardization, set the time window length, extract the time series sample from the sequence in the form of a sliding window to obtain the historical time step-current time step-next time step, and construct the model training data set; S005: Based on the model training data set, the time series sample of the historical time step and the current time step process parameter data is taken as the input, and the next time step target process parameter is taken as the output. Train and update the process parameter prediction model until the loss of the process parameter prediction model converges within a continuous preset number of epochs.

[0008] Further, the multi-scale one-dimensional convolution in S1 respectively extracts features from the local time window, point-by-point mapping, and global view of the input historical time step process parameters; The local time window extracts features from the local time window by using two stacked one-dimensional convolution layers, and the convolution kernel size is : ; wherein, is the feature vector extracted by the local time window; is the data after data preprocessing; and are one-dimensional convolution layers, including one-dimensional convolution, dimension transformation, linear layer, and batch normalization layer; Point-by-point mapping extracts features from the point-by-point mapping by introducing a convolution branch: ; wherein, is the feature vector extracted by the point-by-point mapping; is a one-dimensional convolution layer, including one-dimensional convolution, dimension transformation, linear layer, and batch normalization layer; The global view extracts features from the global view by using a convolution kernel with a width equal to the input dimension and the historical time step, i.e., the convolution kernel size is , which maps all features of the historical time step: ; wherein, is the feature vector extracted by the global view; LargeConv is a one-dimensional convolution layer, including one-dimensional convolution, dimension transformation, linear layer, and batch normalization layer.

[0009] Further, the specific process of the hidden state sequence of the process parameter of each historical time step in S2 is to input the process parameter feature sequence of multiple historical time steps as input into the Mamba module, which includes a state space layer, a selection mechanism layer and an output projection layer, to obtain the timing feature information of the process parameter between each historical time step through linear state space conversion and adaptive selection mechanism, and to obtain the hidden state sequence of the process parameter of each historical time step.

[0010] First, the feature sequence fused by the S1 multi-scale one-dimensional convolution (wherein, is the batch size, is the number of historical time steps, is the feature dimension), is mapped to the state space dimension adapted by the Mamba module (usually is taken to balance the feature expression ability and the calculation efficiency), and the projection formula is as follows: ; wherein, is the weight matrix of the linear projection layer, which is used to realize the conversion of the feature dimension; is the bias term of the linear projection layer, which is used to fine-tune the projected feature value; is the projected state space input.

[0011] Subsequently, is split into three parallel branches, namely the state transition kernel branch , the input gate branch , and the selection weight branch , each branch has the same dimension as , i.e. , which is used for subsequent state transition control, input contribution adjustment and key time step screening.

[0012] For the state transition kernel branch , the feature data of the first time step (indicating the features of all batches at the first time step and all dimensions) is applied to the exponential function to generate the state transition kernel of the first time step, and the formula is as follows: ; wherein, is the first State transition kernel of time step, used to control the decay rate of historical state, the smaller the value, the faster the historical state decays, the model pays more attention to the current input; the larger the value, the more complete the historical state is preserved, the model focuses more on capturing long-range dependencies.

[0013] For the input gate branch , take the feature data of the time step , apply the sigmoid activation function to generate the input gate weight of the time step, the formula is as follows: ; wherein, sigmoid activation function, the output is constrained in interval; input gate weight of the time step.

[0014] Set the initial state (zero vector, dimension ), combined with the state transition kernel , the input gate and the input of the time step (feature data of the time step), update the state through element-wise multiplication and linear combination, the formula is as follows: ; wherein, tanh activation function, the output is constrained in interval to ensure the stability of the state value; element-wise multiplication; state vector of the time step.

[0015] After traversing all historical time steps , stack the state vectors of each time step in the time dimension to form the time-domain historical hidden state matrix.

[0016] Further, the specific process of the S4 is: S401: Take the process parameter feature of the current time step as Query, the historical hidden state matrix output by the Mamba module as Key and Value, calculate the attention weight : ; wherein, q is the process parameter feature of the current time step. is the hidden state of the t time step; H is the historical hidden state matrix output by the Mamba module; S402: Weighted sum of all historical hidden states based on the calculated attention weight to generate a context vector ; S403: Concatenate the context vector with the current feature representation in the feature dimension to form the integrated feature after fusion .

[0017] Further, when the prediction deviation of the target process parameter in the tobacco feeding process does not satisfy the preset prediction deviation condition, the latest process parameter data in the preset period is selected to join the training data set, the model parameters are reinitialized, the model is trained to update the parameters until the prediction deviation satisfies the preset prediction deviation condition.

[0018] In a second aspect, the present application provides a system for process parameter prediction control in a tobacco feeding process, comprising: A data acquisition and preprocessing module is used to acquire process parameter data of a current time step and a plurality of historical time steps before the current time step in the tobacco feeding process and to convert the data into corresponding time series samples through data preprocessing: A process parameter prediction module is used to input the time series samples of the process parameter data of the current time step and the historical time steps into a trained process parameter prediction model, and to output the predicted value of the target process parameter of the next time step; The specific process of the process parameter prediction model is as follows: S1: Multi-scale one-dimensional convolution is used to extract features from the time series samples of the process parameters of the plurality of historical time steps, and then the extracted features of each scale are fused to obtain a process parameter feature sequence of the plurality of historical time steps; S2: The process parameter feature sequence of the plurality of historical time steps in S1 is input into a Mamba module for time series feature extraction, and a hidden state sequence of the process parameters of each historical time step is obtained, and then a time-domain historical hidden state matrix composed of the hidden state sequence is obtained; S3: A linear transformation layer is used to map the time series sample of the process parameter of the current time step to a hidden space to obtain a process parameter feature of the current time step; S4: The time-domain historical hidden state matrix and the process parameter feature of the current time step are input into an attention module to obtain a time-domain integrated feature; S5: performing a fast Fourier transform on each feature dimension of the process parameter feature sequence output by S1 to obtain a frequency domain complex sequence, calculating an amplitude of each frequency component in the frequency domain complex sequence to obtain an amplitude sequence, and performing global average pooling on the amplitude sequence to obtain a frequency domain feature; S6: concatenating the time domain comprehensive feature obtained by S4 and the frequency domain feature obtained by S5 to obtain a fusion feature; S7: performing linear transformation on the comprehensive feature through a fully connected layer to output a predicted value of the target process parameter at the next time step.

[0019] In a third aspect, the present application provides a readable storage medium storing a computer program, which is called by a processor to execute the steps of the method described above.

[0020] In a fourth aspect, the present application provides an electronic terminal comprising a processor and a memory, wherein the memory stores a computer program, and the processor calls the computer program to execute the steps of the method described above.

[0021] Advantages

[0022] The present application provides a process parameter prediction control method, system, medium and terminal in a tobacco filling process, which has the following advantages: 1. Enhancing multi-scale feature modeling capability: traditional methods such as multivariate linear regression and random forest have limited feature extraction capability when facing complex nonlinear relationships and multi-source heterogeneous data in the tobacco filling process, and it is difficult to effectively capture local fluctuations and global trends. The present application designs a multi-scale one-dimensional convolution module to model the process parameters from different feature granularities in parallel, fuses local, point-by-point and global features, and significantly improves the perception ability and prediction accuracy of the model for complex dynamic processes.

[0023] 2. Improving long-term dependence modeling efficiency and accuracy: the present application discards the sequential computation paradigm of traditional RNN (such as GRU, LSTM) and adopts an advanced Mamba model. This model not only fundamentally solves the gradient vanishing problem, but also can process sequence data with linear time complexity (O(N)), which is much better than traditional RNN. Its parallel scanning algorithm greatly improves the training efficiency of the model on GPU, and the selective state update mechanism enables it to focus on key information like the attention mechanism, ensuring high accuracy while greatly improving the efficiency and ability to process long sequence data in industry, and is more suitable for deployment requirements in real-time control scenarios in industry.

[0024] 3. Realize the key historical information focus: traditional neural networks are difficult to identify which historical time steps have a key impact on the current output, resulting in insufficient stability of the prediction result. The invention introduces an attention mechanism to build a dynamic association mechanism between historical hidden states and current input features, enabling the model to adaptively focus on the key historical moments that truly affect the current output, thereby improving the accuracy and robustness of the output prediction.

[0025] 4. Online adaptive updating capability: existing models generally lack response mechanisms for changes in working conditions such as raw material fluctuations and equipment aging, requiring frequent manual intervention for retraining. The invention proposes an online updating mechanism based on prediction error feedback. When detecting that the prediction error exceeds the threshold for multiple consecutive time steps, the model retraining process is automatically triggered, incorporating recent data into the training set and updating the model parameters to ensure the stability and practicality of the model during long-term operation. BRIEF DESCRIPTION OF DRAWINGS

[0026] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0027] Figure 1 is a two-dimensional scatter plot provided by the embodiment of the present application before data cleaning; Figure 2 is a two-dimensional scatter plot provided by the embodiment of the present application after data cleaning; Figure 3 is a test set outlet moisture true value and predicted value line chart provided by the embodiment of the present application; Figure 4 is a test set moisture removal air volume opening degree true value and predicted value line chart provided by the embodiment of the present application; Figure 5 is a test set hot air actual temperature true value and predicted value line chart provided by the embodiment of the present application; Figure 6 is a process parameter prediction control method flowchart provided by the embodiment of the present application in the tobacco leaf feeding process. DETAILED DESCRIPTION

[0028] In order to make the purpose, technical solutions and advantages of the present application more clear, the technical solutions of the present application will be described in detail below. Obviously, the described embodiments are only some of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of the present application.

[0029] Embodiment 1

[0030] As Figure 6 shown, the embodiment provides a process parameter prediction control method in a tobacco feeding process, comprising: obtaining process parameter data of a current time step and a plurality of historical time steps before the current time step in the tobacco feeding process and performing data preprocessing to convert the process parameter data into corresponding time sequence samples, taking the time sequence samples of the process parameter data of the current time step and the historical time steps as inputs, and inputting the inputs into a process parameter prediction model constructed and trained to output predicted values of target process parameters of a next time step.

[0031] Since the existing technology has problems such as insufficient feature modeling, weak time sequence dependence capturing ability, and lack of online updating mechanism in the prediction control of the tobacco feeding process, the embodiment adds a selective state space model module (i.e., Mamba module) on the basis of CNN (the type of the model is not limited, and can be adjusted according to actual conditions in specific implementation, such as a Transformer encoder model, a Wavenet (waveform network), etc.), and finally associates the dynamics between the historical hidden state and the current input feature through an attention mechanism, so that the result predicted by the process parameter prediction model is more accurate.

[0032] Specifically, the training process of the process parameter prediction model is specifically: S001: Collecting multi-source process parameter data of a plurality of sampling periods in a historical tobacco feeding process, and constructing a high-dimensional feature space matrix. In the embodiment, multi-source process parameter data in 3 historical sampling periods in the tobacco primary processing feeding process is collected, and a high-dimensional feature space matrix is constructed , wherein is a feature dimension (including hot air actual temperature, cylinder temperature, moisture removal opening, outlet material instantaneous moisture, etc. parameter), is the total number of samples.

[0033] S002: Taking each sampling period as a batch, and sequentially using a t-distributed stochastic neighbor embedding algorithm to reduce the dimension of the high-dimensional data in the high-dimensional feature space matrix to a two-dimensional space to obtain a two-dimensional scatter plot, as shown in Figure 1 .

[0034] Specifically, for the collected initial data, the t-SNE (i.e., t-distributed stochastic neighbor embedding) algorithm is used to reduce the dimension of the high-dimensional data in batches, i.e., the high-dimensional feature space is compressed from to dimensions (taking for example) for visualization, to obtain a low-dimensional mapping : In the high-dimensional feature space, the similarity probability of sample and sample is defined as: ; wherein, is the similarity probability of sample i and sample j in the high-dimensional space; , , is the data sample vector of the original high-dimensional feature space; represents the square of the Euclidean distance between sample i and sample j; the denominator represents the sum of the similarity of all non samples and sample , for normalization; is the local density parameter of sample , obtained by binary search , determined by binary search, so that the entropy of the joint probability distribution in the high-dimensional space is approximately a preset value.

[0035] In the low-dimensional space, the similarity probability of sample and is defined as: ; wherein, represents the similarity probability of sample and sample in the low-dimensional space; , , , is the mapping vector after the high-dimensional data sample is reduced to dimensional , and the denominator represents the sum of the similarity of any two samples in the low-dimensional space, for normalization.

[0036] In the dimension reduction process, based on the similarity probability between samples in the high-dimensional feature space and the similarity probability of the sample space in the low-dimensional space, the optimization objective function of t-SNE dimension reduction is obtained by minimizing the Kullback-Leibler divergence as: F=min ; wherein, F is the optimization objective function of t-SNE dimension reduction; is the sum of the KL divergence of the high-dimensional distribution and the low-dimensional distribution .

[0037] The gradient descent method is used to iteratively update the low-dimensional mapping , so that ​​Minimize, eventually generate two-dimensional scatter plot to visualize the sample in low-dimensional space distribution characteristics. Each batch of data corresponding to the two-dimensional scatter plot includes two categories of data: one is the data-intensive area, representing the stable working condition data under normal production state; the other is the sparse distribution area, corresponding to abnormal or non-steady process data, such as Figure 1 As shown, abnormal or non-steady process data will have a negative impact on the model, wherein in the two-dimensional scatter plot after dimension mapping, the abscissa represents the dimension 1 after t-SNE processing, and the ordinate represents the dimension 2 after t-SNE processing.

[0038] S003: Adopt K-Means clustering algorithm to perform cluster analysis on the scatter point data in the two-dimensional scatter plot, clean the abnormal data, and perform standardization processing on the high-dimensional data corresponding to the cleaned scatter point data.

[0039] Specifically, in the two-dimensional space after t-SNE dimension reduction, the K-Means clustering algorithm is applied to identify data clusters. The number of clusters is set to (the value is determined according to the data distribution characteristics, ), the initial cluster centers are updated by minimizing the within-cluster sum of squares (wherein is an indicator function, if the sample belongs to cluster , then , otherwise 0). In the clustering result, the cluster with more samples is defined as the "main data cluster", representing the normal production condition; the other cluster is the "outlier cluster", corresponding to abnormal or non-steady data. The deviation degree of the sample from the center of the main cluster is measured by the Euclidean distance, and finally the samples in the main data cluster are retained to constitute the cleaned data set , ensuring the consistency and process representativeness of the data, and providing high-quality input for subsequent modeling. The scatter plot of the cleaned data set is drawn, as shown in Figure 2 , it can be observed that the outlier points are eliminated, wherein in the two-dimensional scatter plot after dimension mapping, the abscissa represents the dimension 1 after t-SNE processing, and the ordinate represents the dimension 2 after t-SNE processing.

[0040] Based on the high-dimensional data corresponding to the scatter point data in the cleaned data set , the mean and standard deviation of each high-dimensional data feature dimension (the number of feature dimensions , representing the number of feature dimensions at each time step) are calculated: , ; wherein, is the total number of samples; Represents the cleaned dataset Middle Sample No. feature dimensions.

[0041] Based on the calculated mean and standard deviation, the cleaned dataset sample data is standardized according to the original feature dimension before dimensionality reduction: ; in, for Normalized data.

[0042] S004: For each batch of data after standardization, set the time window length, extract the historical time step-current time step-next time step from the sequence in a sliding window manner to obtain a time series sample, and build a model training data set.

[0043] Specifically, for each cleaned batch of data, set the time window length In this embodiment , extracting "historical time step - current time step - next time step" from the time series in a sliding window manner. The specific form is as follows: ; ; ; in, Indicates the The historical input data within a time window contains continuous The sample data of time steps has the dimension of OK List; Indicates the Sample data of the current time step, dimension is 1 row List; Represents the target output data of the corresponding next time step. Finally, all the extracted samples are integrated to form a unified array (i.e., time series samples), the shape of which is ,in is the total number of samples, refer to The samples of the time step, the samples of the current time step and the corresponding sample data of the future time step, represents the feature dimension of each time step. To ensure the generalization ability of model training, the constructed sample set is divided into training set, validation set and test set: set the test set size to 120 samples (calculated at the frequency of 10s sampling of a certain tobacco factory, the test set is a continuous 20-minute data); randomly select 20% of the remaining samples as the validation set, and the rest as the training set; finally, the training set is randomly shuffled to avoid time-dependent bias.

[0044] S005: Based on the model training data set, the time series sample of the historical time step and the current time step of the process parameter data is taken as the input, and the target process parameter of the next time step is taken as the output. The process parameter prediction model is trained and the parameters are updated until the loss of the process parameter prediction model converges within a continuous preset number of epochs.

[0045] Specifically, the specific process of the process parameter prediction model prediction is: S1: A multi-scale one-dimensional convolution is used to extract features from the time series sample of the historical time step of the process parameter, and then the extracted features of each scale are fused to obtain a process parameter feature sequence of multiple historical time steps. In the prior art, multi-scale is usually used to extract features from different sizes of time steps, such as 3 data, 5 data, etc. using a fixed size of convolution kernel, the core of which is to distinguish the scale by different numbers of continuous time step data; while the multi-scale one-dimensional convolution of the embodiment distinguishes the scale by the size of the convolution kernel, for each data in the same group of continuous historical time step data, 1x1, 1x3, and historical time step (such as 1x10) convolution kernel are used in turn, respectively from point-by-point mapping, local time window, and global view to extract features of each data, and then the extracted features of each scale are fused to obtain a process parameter feature sequence of multiple historical time steps. The specific process is as follows: To fully exploit the characteristics of the process parameters in the tobacco primary processing process, the multi-scale one-dimensional convolution uses a multi-scale one-dimensional convolution feature extraction module to model the features of the input historical time series (i.e. the process parameters of the historical time step). The input historical time series data is , The batch size is represents the number of time steps, represents the feature dimension of each time step.

[0046] Since the one-dimensional convolution operation has requirements for the shape of the input tensor, the dimensions of the input are transformed: ; wherein, represents the data after the cleaning operation and standardization; represents a dimension transformation operation, used to transform the input data into a new dimension dimension; is the data after dimension transformation, and the dimension is .

[0047] More specifically, the S1 multi-scale one-dimensional convolution respectively extracts features from the local time window, point-by-point mapping, and global view of the input historical time step process parameters; The local time window uses two stacked one-dimensional convolution layers, with a convolution kernel size of , both using to ensure the output length is consistent, represents the padding of the convolution kernel, represents the convolution kernel size: ; where, is the feature vector extracted by the local time window; is the data after data preprocessing; and are one-dimensional convolution layers, containing one-dimensional convolution ( ) + dimension transformation (Reshape) + linear layer (Linear) + batch normalization layer (batchNorm1d), with an output dimension of , is the hidden layer dimension ( ); Point-by-point mapping uses a convolution branch, mainly used to enhance feature expression capability: ; where, is the feature vector extracted by the point-by-point mapping; contains one-dimensional convolution ( ) + dimension transformation (Reshape) + linear layer (Linear) + batch normalization layer (batchNorm1d), with an output dimension of ; In order to capture longer span feature information, the global view uses a convolution kernel with a width equal to the input dimension and the historical time step (i.e. a convolution kernel with a width equal to the input dimension , ), i.e. the convolution kernel size is , to map all features of the historical time step: ; where, is the feature vector extracted by the global view; LargeConv contains one-dimensional convolution ( ) + Reshape + Linear + batchNorm1d, the output dimension is .

[0048] The outputs of the above three paths are spliced to obtain the fused high-dimensional features: 、 、 ; represents the high-dimensional features obtained by fusing 、 、 , the dimension is .

[0049] Then, a linear layer is used to compress the channel dimension to ( i.e., the hidden layer dimension, ), to realize adaptive fusion of multi-scale features: ; wherein is a process parameter feature sequence of multiple historical time steps, and the dimension is , representing the feature information of each time step extracted after multi-scale convolution feature extraction and fusion. Through such multi-scale convolution feature extraction and fusion process, the feature information under different feature scales can be fully mined, the model's perception ability of features and between features is enhanced, more rich and effective feature representation is provided for subsequent process-level time sequence information modeling, and the modeling effect of complex dynamic characteristics is significantly improved.

[0050] S2: input the process parameter feature sequence of multiple historical time steps of S1 into the Mamba module for time sequence feature extraction, to obtain the hidden state sequence of each historical time step of the process parameter, and then obtain the time domain historical hidden state matrix composed of the hidden state sequence. Specifically as follows: In order to further mine the dynamic rules implied by the evolution of each process parameter with time in the tobacco adding process, a selective structured state space model (Selective Structured State Space Model, Mamba) is introduced to deeply model the feature sequence obtained by S1.

[0051] Specifically, the output feature tensor after multi-scale convolution fusion, i.e., the feature sequence ​As input, it is sequentially input into the Mamba module, which includes a state space layer, a selection mechanism layer, and an output projection layer. Through linear state space transformation and adaptive selection mechanism, the temporal feature information of the process parameters between each historical time step is obtained, and the hidden state sequence of the process parameters at each historical time step is obtained. The specific process is as follows: First, the feature sequence after S1 multi-scale one-dimensional convolution fusion (in, is the batch size, is the number of historical time steps, is the feature dimension), mapped to the state space dimension adapted by the Mamba module through the linear projection layer (Usually take , to balance feature expression ability and computational efficiency), the projection formula is as follows: ; in, is the weight matrix of the linear projection layer, which is used to realize the conversion of feature dimensions; It is the bias term of the linear projection layer, which is used to fine-tune the feature value after projection; is the projected state space input.

[0052] Afterwards, Split into three parallel branches, namely the state transfer core branch , input gate branch , select weight branch , each branch dimension is consistent with Consistent, that is , which are used for subsequent state transfer control, input contribution adjustment and key time step screening respectively.

[0053] State transfer core branch , take the Feature data of time step (Indicates that all batches time steps, features of all dimensions), apply the exponential function to generate the The state transition kernel of the time step is as follows: ; in, For the The state transition kernel of the time step is used to control the decay rate of the historical state. The smaller the value, the faster the historical state decays, and the model pays more attention to the current input; the larger the value, the more complete the historical state is retained, and the model focuses more on capturing long-term dependencies.

[0054] Branch to input gate , take the Feature data of time step , the sigmoid activation function generates the input gate weight of the time step, as follows: ; wherein, is the sigmoid activation function, which constrains the output to the interval; is the input gate weight of the time step.

[0055] The initial state (zero vector, dimension ) is set, combined with the state transition kernel , the input gate , and the input of the time step (feature data at the time step), the state is updated by element-wise multiplication and linear combination, as follows: ; wherein, is the hyperbolic tangent activation function, which constrains the output to the interval to ensure stable state values; denotes element-wise multiplication; is the state vector of the time step.

[0056] After traversing all historical time steps , the state vectors of each time step are stacked in the time dimension to form a time-domain historical hidden state matrix. The selective state space mechanism adopted by this Mamba module significantly outperforms traditional recurrent neural networks (RNN) and their variants (such as GRU, LSTM) in capturing long sequence dynamic dependencies. This mechanism not only effectively avoids the gradient vanishing problem, but also enhances the model's ability to model nonlinear changes under complex conditions through selective processing of information. Compared to the sequential computation mode of GRU / LSTM, the parallel computation architecture of Mamba significantly improves training efficiency, making it a significant advantage in processing industrial big data.

[0057] When the input sequence is processed through the Mamba module, a corresponding hidden state representation is generated for each time step (t = 1,..., T). These hidden state sequences form a complete time-series feature representation matrix.

[0058] The application introduces an advanced Mamba model to realize efficient modeling of process-level timing information. Unlike the gating mechanism of traditional GRU / LSTM, the model adopts a selective state space theory, efficiently capturing long-range dependencies while achieving high parallelization in the training process. This not only maintains the fast reasoning ability suitable for industrial real-time control scenarios, but also significantly reduces the computational cost of model training, making it suitable for industrial scenarios.

[0059] S3: A linear transformation layer is used to map the time series samples of the process parameters at the current time step to the hidden space to obtain the process parameter features at the current time step. Specifically as follows: In order to enable the features of the current time step to interact with the historical hidden states output by Mamba in the same space, a linear transformation layer is used to map the input feature vector of the current time step , containing all features except the instantaneous moisture of the outlet material, to the hidden space, so that the model learns the relationship between production parameters and the instantaneous moisture of the outlet material; wherein the linear transformation is realized by a learnable weight matrix ( where is randomly initialized and the weight matrix is updated using backpropagation gradient) : ; where q is the output vector containing the key information of the current time step features, with a dimension of .

[0060] S4: The hidden state matrix of the process parameters at multiple historical time steps and the process parameter features at the current time step are input into the attention module to obtain the fused time domain comprehensive features. Specifically as follows: In the tobacco feeding process, there is a complex dependency relationship between process parameters at different time steps, and the current time step is often affected by various factors over a period of time. To further explore the relationship between the current time step input features and the historical state, and improve the model's response ability to changes in key process parameters, an attention mechanism is introduced to weight and fuse the historical hidden state and the current time step features.

[0061] S401: The process parameter features at the current time step are taken as Query, the historical hidden state matrix output by the Mamba module is taken as Key and Value, and the attention weight is calculated: ; where q is the process parameter feature at the current time step; ​is the hidden state at time step t; H is the history hidden state matrix output by the Mamba module. The attention weight reflects the importance of the hidden state of each history time step to the current time step, and the higher the weight, the stronger the relevance of the information of the history time step to the current situation.

[0062] S402: Weighted sum of all history hidden states based on the calculated attention weight to generate context vector : ;

[0063] The context vector represents the most relevant past information at the current time step, which integrates the most matched part of the history hidden state to the current feature.

[0064] S403: Concatenate the context vector with the current feature representation in the feature dimension to form the fused time domain comprehensive feature : ;

[0065] The time domain fusion feature is subsequently fused with the frequency domain feature as the input of the subsequent prediction module, which can adaptively focus on the key time points in the history process. It fully utilizes the correlation between historical information and current features, so that the model can comprehensively consider the influence of the past when predicting the outlet moisture and related control parameters, thereby improving the prediction accuracy and robustness. In the actual tobacco primary processing process, such a fusion feature can help the model better cope with various complex process situation changes and provide more reliable basis for the optimization of the production process.

[0066] S5: Perform fast Fourier transform on each feature dimension of the process parameter feature sequence output by S1 to obtain a frequency domain complex sequence, calculate the amplitude of each frequency component in the frequency domain complex sequence to obtain an amplitude sequence, and perform global average pooling on the amplitude sequence to obtain a frequency domain feature. Specifically as follows: Perform independent FFT transform on the time series of each feature dimension in the obtained by S1. For the time series of the i-th sample and the j-th feature dimension , its FFT transform can be represented as: ; ; where is the frequency domain complex representation, is the imaginary unit, is the frequency index. Since the FFT result of the time domain signal has symmetry, only the first ​a frequency component, denoted as , the frequency domain data of all samples and feature dimensions form a tensor .

[0067] The amplitude of the frequency domain complex sequence is calculated to represent the energy intensity of each frequency component: ; wherein is the amplitude of the th frequency component, represents the real part of the complex number (Real part), represents the imaginary part of the complex number (Imaginary part), and the amplitude data is obtained after amplitude calculation. The amplitude distribution reflects the contribution of different periodic components in the feature sequence, for example, low frequency components may correspond to the slow changing trend of process parameters.

[0068] To compress the frequency domain amplitude information into global features, global average pooling is performed on the amplitude sequence of each feature dimension: ;

[0069] wherein is the number of effective frequency components, is the frequency domain global feature of the th sample and the th feature dimension. Finally, the pooling results of all feature dimensions form a frequency domain feature matrix , which physically represents the average energy distribution of each feature dimension in the frequency domain, and can effectively capture the periodic dynamic mode of the process parameters.

[0070] S6: The time domain comprehensive features obtained in S4 and the frequency domain features obtained in S5 are spliced to obtain fusion features. Specifically as follows: Due to the difference in dimensions, the time domain features need to be globally averaged and pooled to eliminate the time dimension and achieve dimension alignment: ; wherein GlobalAvgPool represents the global average pooling operation, and the time dimension is averaged, and the formula is: ; The aligned time domain features are spliced with the frequency domain features along the feature dimension to obtain fusion features : ; The mathematical expression of the splicing operation is: ; wherein, the dimension of , while preserving the time-domain dynamic information and frequency-domain periodic characteristics.

[0071] S6: Perform linear transformation on the integrated features through a fully connected layer to output the predicted value of the target process parameter at the next time step. Specifically as follows: The integrated features after splicing and fusion are taken as input and passed to a fully connected layer. The fully connected layer processes the input features through linear transformation, and the mathematical expression is: ; wherein, represents the output of the model, and the output dimension is , is a learnable weight matrix, is a bias vector used to fine-tune the prediction result. By continuously learning and optimizing the values of and , the model can better capture the complex relationship between the input features and the prediction target. In this embodiment, the output vector contains three key prediction targets: The first dimension of the output vector represents the predicted value of the instantaneous moisture content of the outlet material at the next time step. Outlet moisture content is a key quality indicator in the tobacco leaf cutting process, which directly affects the quality of the tobacco leaf and the effectiveness of subsequent processing. Accurate prediction of outlet moisture content helps to adjust the relevant parameters of the cutting process in a timely manner to ensure that the moisture content of the tobacco leaf meets the production requirements.

[0072] The second dimension is the recommended hot air actual temperature set value at the next time step. Hot air temperature has an important influence on the drying degree and uniformity of the cutting process. By predicting the appropriate hot air temperature set value, the drying process of the tobacco leaf can be optimized, the cutting effect can be improved, and the quality of the entire cutting process can be improved.

[0073] The third dimension is the recommended exhaust air volume actual opening adjustment value at the next time step. The exhaust opening controls the humidity environment within the process, and appropriate exhaust opening helps to maintain stable process conditions and avoid affecting the quality and processing efficiency of the tobacco leaf due to abnormal humidity.

[0074] Through this multi-target prediction module, the model can realize simultaneous modeling and accurate prediction of core quality indicators (outlet moisture) and key control variables (hot air temperature, exhaust opening) in the cutting process, which provides the control system with forward-looking adjustment suggestions. Figure 3 、 Figure 4 、 Figure 5The following line graphs show the model's prediction results for outlet moisture, exhaust airflow, and actual hot air temperature, respectively, for the test data set. The number and type of target process parameters can be adjusted based on the actual scenario and are not limited to specific parameters. Examples include actual steam flow rate, actual steam pressure, actual tank water temperature, and cylinder temperature.

[0075] This embodiment provides a method for predicting and controlling process parameters in the tobacco leaf feeding process. It extracts local features through multi-scale convolution, captures long-term time dependencies through gated recurrent units, and introduces an attention mechanism to achieve dynamic association between historical states and current inputs. This significantly improves the model's prediction accuracy and robustness for key quality indicators (such as outlet moisture) and related control variables (such as hot air temperature and moisture drainage opening) under complex working conditions.

[0076] Preferably, it also includes the process of selecting the process parameter data within the latest preset time period and the corresponding real label to add to the training data set during actual use, reinitializing the model parameters, and training the model to update the parameters until the prediction deviation meets the preset prediction deviation condition when the prediction deviation between the target process parameter prediction value and the actual value during the tobacco leaf feeding process does not meet the preset prediction deviation condition.

[0077] During the implementation of this embodiment, during the model deployment and operation, the system continuously records the deviation between the predicted value output by the model and the actual measured value: for the three key variables predicted by the model: the instantaneous moisture content of the outlet material at the next time step, the actual temperature of the hot air, and the actual opening of the dehumidification air volume, the relative error is calculated respectively: ; in: Indicates the The actual measurement value of the time step; Indicates the corresponding model prediction value; These correspond to the three prediction targets: outlet moisture, hot air temperature, and dehumidification opening. The relative error is calculated to quantify the degree of deviation between the model's predicted value and the actual value, providing a basis for determining whether the model's performance has deteriorated.

[0078] When the relative errors of the above three prediction values ​​are greater than 1% for 30 consecutive time steps (5 minutes), the current model performance is judged to have significantly degraded, triggering the online update mechanism: The input feature data of the last 10 days and the corresponding true labels are added to the training data set. These new data reflect the current process state, and they are included in the data set. Using the updated data set, the parameters of the model are reinitialized, and then the prediction value is calculated by forward propagation according to the conventional training process, and the loss function is calculated according to the prediction value and the true label. Then the gradient is calculated using the back propagation algorithm, and the parameters of all layers of the model are updated according to the gradient. During the training process, the parameters are continuously optimized through multiple iterations, so that the prediction result of the model can better fit the new data and improve the adaptability of the model to the current process state. After the training is completed, the updated model is loaded into the inference module to replace the original model and continue to perform the prediction task. In this way, in the subsequent production process, the updated model can be used for prediction, thereby improving the accuracy and reliability of the prediction.

[0079] Through the above mechanism, the model can perceive the process state change in real time and automatically adjust the internal parameters, thereby effectively dealing with the real working condition changes such as equipment aging and raw material fluctuation, so that the model can continuously and stably run in the changing production environment.

[0080] Embodiment 2

[0081] The embodiment provides a system of a process parameter prediction control method in a tobacco leaf feeding process, comprising: A data acquisition and preprocessing module is configured to acquire process parameter data of a current time step and a plurality of historical time steps before the current time step in the tobacco leaf feeding process and perform data preprocessing to convert the process parameter data into corresponding time sequence samples: A process parameter prediction module is configured to input the time sequence samples of the process parameter data of the current time step and the historical time steps into a trained process parameter prediction model, and output a prediction value of a target process parameter of a next time step; The specific process of the process parameter prediction model is as follows: S1: performing feature extraction on the time sequence samples of the process parameters of the plurality of historical time steps by using a multi-scale one-dimensional convolution, and then fusing the extracted features of different scales to obtain process parameter feature sequences of the plurality of historical time steps; S2: inputting the process parameter feature sequences of the plurality of historical time steps in S1 into an Mamba module to perform time sequence feature extraction, and correspondingly obtaining hidden state sequences of the process parameters of the historical time steps, and then obtaining a time domain historical hidden state matrix composed of the hidden state sequences; S3: mapping the time sequence samples of the process parameters of the current time step to a hidden space by using a linear transformation layer to obtain process parameter features of the current time step; S4: input the time domain history hidden state matrix and the process parameter feature of the current time step into an attention module to obtain a time domain comprehensive feature; S5: perform fast Fourier transform on each feature dimension of the process parameter feature sequence output by S1 to obtain a frequency domain complex sequence, calculate the amplitude of each frequency component in the frequency domain complex sequence to obtain an amplitude sequence, and perform global average pooling on the amplitude sequence to obtain a frequency domain feature; S6: splice the time domain comprehensive feature obtained by S4 and the frequency domain feature obtained by S5 to obtain a fusion feature; S7: perform linear transformation on the comprehensive feature through a fully connected layer to output the predicted value of the target process parameter of the next time step.

[0082] Embodiment 3

[0083] The embodiment provides a readable storage medium which stores a computer program, the computer program is called by a processor to execute steps of the method.

[0084] Embodiment 4

[0085] The embodiment provides an electronic terminal which comprises a processor and a memory, the memory stores a computer program, and the processor calls the computer program to execute steps of the method. It can be understood that the same or similar parts in the above-mentioned embodiments can be mutually referred to, and contents not described in detail in some embodiments can be referred to the same or similar contents in other embodiments.

[0086] It should be understood that, in the embodiments of the present application, the processor can be a central processing unit (CPU), and the processor can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor. The memory can include read-only memory and random access memory, and provide instructions and data for the processor. A part of the memory can also include non-volatile random access memory. For example, the memory can also store device type information.

[0087] The readable storage medium is a computer readable storage medium, which can be an internal storage unit of the controller, such as a hard disk or a memory of the controller. The readable storage medium can also be an external storage device of the controller, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. Further, the readable storage medium can include both the internal storage unit and the external storage device of the controller. The readable storage medium is used to store the computer program and other programs and data required by the controller. The readable storage medium can also be used to temporarily store data that has been output or will be output.

[0088] Based on such understanding, the technical solutions of the present application, essentially or in other words, the part of the prior art that contributes to the present application, or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes instructions for making a computer device (which can be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the method described in the embodiments of the present application. The aforementioned readable storage medium includes a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media that can store program codes.

[0089] Although the embodiments of the present application have been shown and described above, it should be understood that the above-described embodiments are exemplary and should not be construed as limiting the present application. Those skilled in the art can make changes, modifications, replacements and variations to the above-described embodiments within the scope of the present application.

Claims

1. A method for predicting and controlling process parameters in a tobacco leaf feeding process, characterized in that: include: The process parameter data of the current time step and the multiple consecutive historical time steps before the current time step in the tobacco leaf feeding process are obtained and converted into corresponding time series samples through data preprocessing. The time series samples of the process parameter data of the current time step and the historical time step are used as input to the constructed and trained process parameter prediction model, and the predicted value of the target process parameter for the next time step is obtained as output. The specific process of the process parameter prediction model prediction is as follows: S1: Use multi-scale one-dimensional convolution to extract features from the time series samples of process parameters at multiple historical time steps, and then fuse the extracted features at each scale to obtain the feature sequences of process parameters at multiple historical time steps; S2: Input the process parameter feature sequences of multiple historical time steps of S1 into the selective state space model module for time series feature extraction, and obtain the hidden state sequence of the process parameters of each historical time step, and then obtain the time domain historical hidden state matrix composed of the hidden state sequence; S3: Use the linear transformation layer to map the time series samples of the process parameters of the current time step to the latent space to obtain the process parameter features of the current time step; S4: Input the time domain history hidden state matrix and the process parameter features of the current time step into the attention module to obtain the time domain comprehensive features; S5: Perform fast Fourier transform on each feature dimension of the process parameter feature sequence output by S1 to obtain a frequency domain complex sequence, calculate the amplitude of each frequency component in the frequency domain complex sequence to obtain an amplitude sequence, and perform global average pooling on the amplitude sequence to obtain frequency domain features; S6: Combine the time domain comprehensive features obtained in S4 and the frequency domain features obtained in S5 to obtain fused features; S7: Perform linear transformation on the comprehensive features through the fully connected layer and output the predicted value of the target process parameter in the next time step.

2. The method according to claim 1, characterized in that The training process of the process parameter prediction model is specifically as follows: S001: Collect multi-source process parameter data of different sampling periods in the historical tobacco process and construct a high-dimensional feature space matrix; S002: Using the sampling period as a batch, the t-distributed random neighbor embedding algorithm is used to reduce the dimension of the high-dimensional data in the high-dimensional feature space matrix and map it to a two-dimensional space to obtain a two-dimensional scatter plot; S003: Use the K-Means clustering algorithm to perform cluster analysis on the scattered data in the two-dimensional scatter plot, clean up abnormal data, and standardize the high-dimensional data corresponding to the cleaned scattered data; S004: For each batch of data after standardization, set the time window length, extract the historical time step - current time step - next time step from the sequence in a sliding window manner to obtain a time series sample, and build a model training data set; S005: Based on the model training data set, the time series samples of the process parameter data of the historical time step and the current time step are used as input, and the target process parameters of the next time step are used as output. The process parameter prediction model is trained and the parameters are updated until the loss of the process parameter prediction model converges within a preset number of consecutive epochs.

3. The method according to claim 1, characterized in that The multi-scale one-dimensional convolution in S1 extracts features of the process parameters of the input historical time steps from the local time window, point-by-point mapping, and global perspective respectively; The local time window is constructed by using two stacked one-dimensional convolution layers with kernel sizes of : ； in, Feature vector extracted for the local time window; The data is preprocessed; and They are all one-dimensional convolutional layers, including one-dimensional convolution, dimension transformation, linear layer, and batch normalization layer; The point-by-point mapping is achieved by introducing a Convolution branch: ； in, Feature vectors extracted for point-by-point mapping; It is a one-dimensional convolution layer, including one-dimensional convolution, dimension transformation, linear layer, and batch normalization layer; The global view is obtained by using a convolution kernel with a width equal to the input dimension and the length of the historical time step, that is, the convolution kernel size is , map all features of the historical time step: ； in, It is the feature vector extracted from the global view; LargeConv is a one-dimensional convolution layer, which includes one-dimensional convolution, dimension transformation, linear layer, and batch normalization layer.

4. The method according to claim 1, wherein The specific process of the hidden state sequence of the process parameters of each historical time step in S2 is to take the feature sequences of the process parameters of multiple historical time steps as input and sequentially input them into the selective state space model module, which includes a state space layer, a selection mechanism layer and an output projection layer. The time series feature information of the process parameters between each historical time step is obtained through linear state space transformation and adaptive selection mechanism, and the hidden state sequence of the process parameters of each historical time step is obtained accordingly. The specific process is as follows: First, the feature sequence after S1 multi-scale one-dimensional convolution fusion , is the batch size, is the number of historical time steps, is the feature dimension, which is mapped to the state space dimension adapted by the Mamba module through the linear projection layer , the projection formula is as follows: ； in, is the weight matrix of the linear projection layer; is the bias term of the linear projection layer; is the projected state space input; Afterwards, Split into three parallel branches, namely the state transfer core branch , input gate branch , select weight branch , each branch dimension is consistent with Consistent, that is ; State transfer core branch , take the Feature data of time step , apply the exponential function to generate the The state transition kernel of the time step is as follows: ； in, For the The state transition kernel of the time step; Branch to input gate , take the Feature data of time step , apply the sigmoid activation function to generate the The input gate weight of the time step is as follows: ； in, is the sigmoid activation function; For the The input gate weights at each time step; Set initial state , combined with the state transfer kernel , input gate Pass the exam Time step input , In the The feature data of the time step is updated through element-by-element multiplication and linear combination. The formula is as follows: ； in, is the hyperbolic tangent activation function, which constrains the output to interval to ensure the stability of the state value; represents element-wise multiplication; For the The state vector of the time step; Traverse all historical time steps Then, the state vector of each time step is Stacked by time dimension, it forms a time domain historical hidden state matrix.

5. The method according to claim 1, wherein The specific process of S4 is: S401: Use the process parameter features of the current time step as the query, and the historical hidden state matrix output by the Mamba module as the key and value to calculate the attention weight : ； Among them, q is the process parameter characteristic of the current time step; is the hidden state at time step t; H is the historical hidden state matrix output by the Mamba module; S402: Perform weighted summation of all historical hidden states based on the calculated attention weights to generate a context vector ; S403: context vector With the current feature representation Splicing is performed on the feature dimension to form a fused comprehensive feature .

6. The method according to any one of claims 1 to 5, characterized in that It also includes the process of selecting the process parameter data within the latest preset time period and adding it to the training data set, reinitializing the model parameters, and training the model to update the parameters until the prediction deviation meets the preset prediction deviation condition when the prediction deviation between the target process parameter predicted value and the actual value during the tobacco leaf adding process does not meet the preset prediction deviation condition during actual use.

7. A process parameter prediction and control system for tobacco leaf feeding, the system executing the method according to any one of claims 1 to 6, characterized in that: include: Data acquisition and preprocessing module: used to obtain the process parameter data of the current time step and multiple consecutive historical time steps before the current time step during the tobacco leaf feeding process and perform data preprocessing to convert it into corresponding time series samples: Process parameter prediction module: It is used to input the process parameter data of the current time step and the historical time step into the constructed and trained process parameter prediction model, and output the predicted value of the target process parameter in the next time step; The specific process of the process parameter prediction model prediction is as follows: S1: Use multi-scale one-dimensional convolution to extract features from the time series samples of process parameters at multiple historical time steps, and then fuse the extracted features at each scale to obtain the feature sequences of process parameters at multiple historical time steps; S2: Input the process parameter feature sequences of multiple historical time steps of S1 into the selective state space model module for time series feature extraction, and obtain the hidden state sequence of the process parameters of each historical time step, and then obtain the time domain historical hidden state matrix composed of the hidden state sequence; S3: Use the linear transformation layer to map the time series samples of the process parameters of the current time step to the latent space to obtain the process parameter features of the current time step; S4: Input the time domain history hidden state matrix and the process parameter features of the current time step into the attention module to obtain the time domain comprehensive features; S5: Perform fast Fourier transform on each feature dimension of the process parameter feature sequence output by S1 to obtain a frequency domain complex sequence, calculate the amplitude of each frequency component in the frequency domain complex sequence to obtain an amplitude sequence, and perform global average pooling on the amplitude sequence to obtain frequency domain features; S6: Combine the time domain comprehensive features obtained in S4 and the frequency domain features obtained in S5 to obtain fused features; S7: Perform linear transformation on the comprehensive features through the fully connected layer and output the predicted value of the target process parameter in the next time step.

8. A readable storage medium, characterized in that: A computer program is stored, and when the computer program is called by a processor, it is used to execute: the steps of the method according to any one of claims 1 to 6.

9. An electronic terminal, characterized in that: The method comprises a processor and a memory, wherein the memory stores a computer program, and the processor calls the computer program to execute the steps of the method according to any one of claims 1 to 6.

Images