A method, system, medium and terminal for predicting and controlling process parameters in a tobacco leaf filling process

Abstract

The application discloses a kind of tobacco filling process process parameter prediction control method, system, medium and terminal, the method is by obtaining the process parameter data of current time step and the continuous multiple historical time steps before current time step in tobacco filling process and carries out data preprocessing conversion corresponding time series sample, process parameter data of current time step and historical time step are as input, input to the process parameter prediction model of construction and complete training, output obtains the predicted value of target process parameter of next time step.The method solves the problems of insufficient feature modeling, weak time series dependence capture ability and lack of online updating mechanism in the existing technology in the prediction control of tobacco filling process, so as to realize efficient and accurate prediction and dynamic regulation of key process parameters in tobacco filling process.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of predictive control of tobacco leaf feeding process, and particularly relates 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 procedures, the tobacco leaf feeding process is one of the key steps that affect the internal quality of cut tobacco. The core goal is to achieve precise control of the instantaneous moisture content of the outlet material by accurately controlling the hot air temperature, moisture removal 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:

[0004] 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.

[0005] 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.

[0006] 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.

[0007] Fourthly, factors such as fluctuations in tobacco raw material properties and equipment aging can 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.

[0008] Therefore, there is an urgent need for a predictive control method that can deeply integrate multi-scale features, capture long-term time 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 removal opening) in the tobacco leaf feeding process, thereby improving the intelligent level of the entire production line and the consistency of product quality. SUMMARY

[0009] In view of the deficiencies of the prior art in the background art, the present application provides a method, system, medium and terminal for predicting and controlling process parameters 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 for predicting and controlling the tobacco leaf dosing process, thereby achieving efficient and accurate prediction and dynamic regulation of key process parameters in the tobacco leaf dosing process.

[0010] In a first aspect, the present application provides a method for predicting and controlling process parameters in a tobacco leaf dosing process, comprising:

[0011] The method comprises the following steps:

[0012] S1: performing feature extraction on the time sequence samples of the process parameters of the historical time steps by using multi-scale one-dimensional convolution, and then fusing the extracted features of each scale to obtain process parameter feature sequences of the multiple historical time steps;

[0013] S2: inputting the process parameter feature sequences of the multiple historical time steps obtained in S1 into a selective state space model module (i.e., 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;

[0014] 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;

[0015] 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;

[0016] 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 frequency domain features;

[0017] S6: concatenating the time domain comprehensive features obtained in S4 and the frequency domain features obtained in S5 to obtain fusion features;

[0018] S7: Linearly transforming the integrated features through a fully connected layer to output the predicted value of the target process parameter at the next time step.

[0019] Further, the training process of the process parameter prediction model is specifically:

[0020] S001: Collecting multi-source process parameter data at different sampling periods in the historical tobacco process to construct a high-dimensional feature space matrix;

[0021] S002: Using t-distributed stochastic neighbor embedding algorithm to reduce the dimension of the high-dimensional data in the high-dimensional feature space matrix to two-dimensional space to obtain a two-dimensional scatter plot;

[0022] S003: Using K-Means clustering algorithm to cluster and analyze the scatter data in the two-dimensional scatter plot, cleaning the abnormal data, and standardizing the high-dimensional data corresponding to the cleaned scatter data;

[0023] 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;

[0024] S005: Based on the model training data set, using the time series sample of the process parameter data at the historical time step and the current time step as input, and the target process parameter at the next time step as output, training and updating the process parameter prediction model until the loss of the process parameter prediction model converges within a continuous preset number of epochs.

[0025] 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;

[0026] The local time window uses two stacked one-dimensional convolution layers, and the convolution kernel size is :

[0027] ;

[0028] 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;

[0029] Point-by-point mapping introduces a convolution branch:

[0030] ;

[0031] in, The feature vector is extracted through point-by-point mapping; It is a one-dimensional convolutional layer, including one-dimensional convolution, dimension transformation, linear layer, and batch normalization layer;

[0032] The global view is achieved by using a convolutional kernel with a width equal to the input dimension and the historical time step, i.e., the kernel size is... Map all features of the historical time step:

[0033] ;

[0034] in, It is the feature vector extracted from the global view; LargeConv is a one-dimensional convolutional layer, which includes one-dimensional convolution, dimension transformation, linear layer, and batch normalization layer.

[0035] Furthermore, the specific process of obtaining the hidden state sequence of process parameters for each historical time step in S2 is to take the feature sequence of process parameters for multiple historical time steps as input and input it sequentially into the Mamba module. This module 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 process parameters between each historical time step is obtained, and the hidden state sequence of process parameters for each historical time step is obtained accordingly.

[0036] First, the feature sequence fused by S1 multi-scale one-dimensional convolution is... (in, For batch size, For historical time steps, (As the feature dimension), it is mapped to the state space dimension adapted by the Mamba module through a linear projection layer. (usually taken) To balance feature representation capability and computational efficiency, the projection formula is as follows:

[0037] ;

[0038] in, This is the weight matrix of the linear projection layer, used to achieve the transformation of feature dimensions; This is the bias term for the linear projection layer, used to fine-tune the projected eigenvalues; This is the input for the projected state space.

[0039] Then, It is split into three parallel branches: the state transition kernel branch and the state transition kernel branch. Input gate branch Selecting weighted branches Each branch dimension is related to Consistency, that is These are used for subsequent state transition control, input contribution adjustment, and key time step selection, respectively.

[0040] For state transition kernel branch Take the first Feature data of time steps (Indicates all batches in the ) (Time step, features in all dimensions), apply an exponential function to generate the first... The state transition kernel for each time step is defined by the following formula:

[0041] ;

[0042] in, For the first The state transition kernel at each time step controls the decay rate of historical states. A smaller value indicates that historical states decay faster, and the model focuses more on the current input; a larger value indicates that historical states are preserved more completely, and the model focuses more on capturing long-range dependencies.

[0043] Input gate branch Take the first Feature data of time steps Applying the sigmoid activation function to generate the first The input gate weights for each time step are calculated using the following formula:

[0044] ;

[0045] in, Using the sigmoid activation function, the output is constrained to... interval; For the first Input gate weights at time steps.

[0046] Set initial state (Zero vector, dimension is) ), combined with state transition kernel Input gate Passing the exam Input of time step ( In the The state is updated using element-wise multiplication and linear combination of the feature data at each time step, as shown in the following formula:

[0047] ;

[0048] in, The hyperbolic tangent activation function is used to constrain the output to... The interval is used to ensure the stability of the state value; represents element-wise multiplication; is the state vector of the t-th time step; is the state vector of the t-th time step.

[0049] 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.

[0050] Further, the specific process of S4 is:

[0051] 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, and calculate the attention weight :

[0052] ;

[0053] wherein q is the process parameter feature of the current time step; is the hidden state of the t-th time step; H is the historical hidden state matrix output by the Mamba module;

[0054] S402: Based on the calculated attention weight, weight sum all historical hidden states to generate a context vector ;

[0055] S403: Concatenate the context vector with the current feature representation in the feature dimension to form the integrated feature after fusion .

[0056] Further, it also includes that in actual use process, when the prediction deviation of the target process parameter prediction value and the actual value in the tobacco feeding process does not satisfy the preset prediction deviation condition, the process parameter data in the latest 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.

[0057] In the second aspect, the present application provides a system of a process parameter prediction control method in a tobacco feeding process, comprising:

[0058] A data acquisition and preprocessing module is used to acquire process parameter data in the 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:

[0059] The process parameter prediction module is configured to input time series samples of process parameter data of the current time step and historical time steps into a trained process parameter prediction model, and output a predicted value of a target process parameter of a next time step.

[0060] The process parameter prediction model predicts the following process:

[0061] S1: performing feature extraction on time series samples of process parameters of multiple 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 multiple historical time steps;

[0062] S2: inputting the process parameter feature sequences of the multiple historical time steps obtained in S1 into a Mamba module to perform time sequence feature extraction, and 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;

[0063] S3: mapping a time series sample of a process parameter of a current time step to a hidden space by using a linear transformation layer to obtain a process parameter feature of the current time step;

[0064] S4: inputting the time-domain historical hidden state matrix and the process parameter feature of the current time step into an attention module to obtain a time-domain comprehensive feature;

[0065] S5: performing 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;

[0066] S6: concatenating the time-domain comprehensive feature obtained in S4 and the frequency-domain feature obtained in S5 to obtain a fusion feature;

[0067] S7: performing linear transformation on the comprehensive feature by using a fully connected layer to output a predicted value of a target process parameter of a next time step.

[0068] 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.

[0069] 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.

[0070] Advantages

[0071] The application provides a process parameter prediction control method, system, medium and terminal in a tobacco filling process.

[0072] 1. Enhanced multi-scale feature modeling capability: Traditional methods such as multiple 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, making it difficult to effectively capture local fluctuations and global trends. The present application designs a multi-scale one-dimensional convolution module to model process parameters from different feature granularities in parallel, fusing local, point-by-point and global features, significantly improving the model's perception ability and prediction accuracy for complex dynamic processes.

[0073] 2. Improve long-term dependence modeling efficiency and accuracy: The present application discards the sequential computation paradigm of traditional RNN (such as GRU, LSTM) and adopts the advanced Mamba model. This model not only fundamentally solves the gradient vanishing problem, but also processes 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 an attention mechanism, ensuring high accuracy while greatly improving the efficiency and ability to process long sequence data in industry, making it more suitable for deployment needs in real-time control scenarios in industry.

[0074] 3. Realize key historical information focusing: Traditional neural networks cannot identify which historical time steps have a key impact on the current output, resulting in insufficient stability of the prediction results. The present application 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 key historical moments that truly affect the current output, thereby improving the accuracy and robustness of the prediction.

[0075] 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 present application proposes an online updating mechanism based on prediction error feedback. When detecting that the prediction error of consecutive time steps exceeds the threshold, the model retraining process is automatically triggered, incorporating recent data into the training set and updating the model parameters, ensuring the stability and practicality of the model during long-term operation. BRIEF DESCRIPTION OF DRAWINGS

[0076] 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 only constitute some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0077] Figure 1 is a two-dimensional scatter plot before data cleaning provided by an embodiment of the present application;

[0078] Figure 2 is a two-dimensional scatter plot after data cleaning provided by an embodiment of the present application;

[0079] Figure 3 is a test set outlet moisture true value and predicted value line chart provided by an embodiment of the present application;

[0080] Figure 4 is a test set moisture removal air volume opening true value and predicted value line chart provided by an embodiment of the present application;

[0081] Figure 5 is a test set hot air actual temperature true value and predicted value line chart provided by an embodiment of the present application;

[0082] Figure 6 is a process parameter prediction control method flowchart in a tobacco leaf dosing process provided by an embodiment of the present application. DETAILED DESCRIPTION

[0083] In order to make the objectives, technical solutions and advantages of the present application clearer, 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 the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of the present application.

[0084] Embodiment 1

[0085] As shown in Figure 6 , the present embodiment provides a process parameter prediction control method in a tobacco leaf dosing process, which comprises:

[0086] obtaining process parameter data of a current time step and a plurality of historical time steps before the current time step in a 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 to obtain predicted values of target process parameters of a next time step.

[0087] Due to the problems of insufficient feature modeling, weak time sequence dependence capturing ability, and lack of online updating mechanism in the existing technology in the prediction control of the tobacco feeding process, in the embodiment, a selective state space model module (i.e. Mamba module) is added on the basis of CNN (the type of the model is not limited to it, and specific implementation can be adjusted according to actual conditions, such as a Transformer encoder model, a Wavenet (waveform network), etc.), and finally the dynamics between the historical hidden state and the current input feature are associated through an attention mechanism, so that the result predicted by the process parameter prediction model is more accurate.

[0088] Specifically, the training process of the process parameter prediction model is specifically:

[0089] S001: Collecting multi-source process parameter data in multiple sampling periods in the historical tobacco feeding process to construct a high-dimensional feature space matrix. In the embodiment, the multi-source process parameter data in 3 historical sampling periods in the tobacco primary processing feeding process is collected to construct a high-dimensional feature space matrix , wherein is the feature dimension (including hot air actual temperature, cylinder temperature, moisture removal opening, outlet material instantaneous moisture, etc. parameter), is the total number of samples.

[0090] S002: Taking each sampling period as a batch, and sequentially using the t-distributed stochastic neighbor embedding algorithm to reduce the dimension of the high-dimensional data in the high-dimensional feature space matrix to two-dimensional space to obtain a two-dimensional scatter plot, as shown in Figure 1 .

[0091] 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, that is, the high-dimensional feature space is compressed from to dimension (taking ) so as to be visualized, and a low-dimensional mapping is obtained:

[0092] In the high-dimensional feature space, the similarity probability of sample and sample is defined as:

[0093] ;

[0094] 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; The denominator represents the square of the Euclidean distance between sample i and sample j; the denominator represents all non-Euclidean distances. Samples and samples The sum of similarities is used for normalization; For the sample The local density parameters are obtained through binary search. By using a binary search, the entropy of the joint probability distribution in the high-dimensional space is approximated to a preset value.

[0095] In low-dimensional space, define samples and similarity probability for:

[0096] ;

[0097] in, Represents samples in low-dimensional space With sample The probability of similarity; , , , Dimensionality reduction of high-dimensional data samples dimension( The mapping vector after () has a denominator representing the sum of similarities between any two samples in the low-dimensional space, used for normalization.

[0098] In the dimensionality reduction process, based on the similarity probability between samples in the high-dimensional feature space and the similarity probability between sample spaces in the low-dimensional space, the objective function for optimizing t-SNE dimensionality reduction is to minimize the Kullback-Leibler divergence:

[0099] F=min ;

[0100] Where F is the objective function for t-SNE dimensionality reduction optimization; High-dimensional distribution With low-dimensional distribution The sum of KL divergences.

[0101] Iterative updates of the low-dimensional mapping are performed using gradient descent. ,make Minimization ultimately generates a two-dimensional scatter plot to visualize the distribution characteristics of samples in a low-dimensional space. The two-dimensional scatter plot corresponding to each batch of data includes two types of data: one is a densely distributed region, representing stable operating conditions under normal production conditions; the other is a sparsely distributed region, corresponding to abnormal or non-steady-state process data, such as... Figure 1As shown, abnormal or non-steady-state process data has a negative impact on the model, where in the two-dimensional scatter plot after dimension mapping, the abscissa represents dimension 1 after t-SNE processing, and the ordinate represents dimension 2 after t-SNE processing.

[0102] S003: Adopting K-Means clustering algorithm to perform clustering analysis on the scatter data in the two-dimensional scatter plot, cleaning abnormal data, and performing standardization processing on the high-dimensional data corresponding to the cleaned scatter data.

[0103] Specifically, in the two-dimensional space after t-SNE dimension reduction, K-Means clustering algorithm is applied to identify data clusters. The number of clusters is set to (the value is taken according to the data distribution characteristics, ), the initialization of the cluster center is updated by minimizing the within-cluster sum of squares (where is an indicator function, if the sample belongs to cluster , then , otherwise 0). In the clustering result, the cluster with larger sample number is defined as the "main data cluster", representing normal production conditions; the other cluster is the "outlier cluster", corresponding to abnormal or non-steady-state 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 Figure 2 , ensuring the consistency of the data and the representativeness of the process, and providing high-quality input for subsequent modeling. The scatter plot of the cleaned data set is drawn, as shown in , from which it can be observed that the outlier points are removed, where in the two-dimensional scatter plot after dimension mapping, the abscissa represents dimension 1 after t-SNE processing, and the ordinate represents dimension 2 after t-SNE processing.

[0104] Based on the high-dimensional data corresponding to the scatter 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:

[0105] , ;

[0106] wherein is the total number of samples; represents the th feature dimension of the th sample in the cleaned data set .

[0107] Based on the calculated mean and standard deviation, the sample data of the cleaned data set is standardized according to the original feature dimension before dimension reduction:

[0108] ;

[0109] Wherein, is the standardized data.

[0110] S004: For each batch of data after standardization processing, set the time window length, extract the time series sample from the sequence in the form of sliding window to get the historical time step-current time step-next time step, and construct the model training data set.

[0111] Specifically, for each batch of cleaned data, set the time window length , in the embodiment , extract the "historical time step-current time step-next time step" from the time series in the form of sliding window, and the specific form is as follows:

[0112] ;

[0113] ;

[0114] ;

[0115] Wherein, represents the historical input data in the first time window, which contains continuous time step sample data, with the dimension of rows columns; represents the sample data of the first current time step, with the dimension of 1 row column; 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 sample), and the shape of the array is , wherein is the total number of samples, points to the sample of time step, the sample of the current time step and the sample data of the corresponding 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, randomly shuffle the training set to avoid time-dependent bias.

[0116] S005: Based on the model training data set, the time series sample of the process parameter data of the historical time step and the current time step 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.

[0117] Specifically, the specific process of the process parameter prediction model prediction is:

[0118] 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 distinguish scales by using fixed size convolution kernels to extract features from 3 data, 5 data, etc. The core is to distinguish scales by using different numbers of continuous time step data. The multi-scale one-dimensional convolution of the embodiment distinguishes scales by kernel size, and for each data in the same group of continuous historical time step data, 1x1, 1x3 and historical time step (such as 1x10) convolution kernels are used in turn to extract features of each data from three dimensions of point-by-point mapping, local time window and global view, 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:

[0119] 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.

[0120] Since the one-dimensional convolution operation has requirements for the shape of the input tensor, the dimensions of the input are transformed:

[0121] ;

[0122] wherein, represents data after cleaning operation and normalization; represents dimension transformation operation, used to transform the data dimension; is the data after dimension transformation, the dimension is .

[0123] More specifically, 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 historical time steps of the input process parameters;

[0124] The local time window adopts two stacked one-dimensional convolution layers, both with a convolution kernel size of , and both use to ensure the consistency of the output length, represents the padding of the convolution kernel, represents the convolution kernel size:

[0125] ;

[0126] wherein, is the feature vector extracted by the local time window; is the data after data preprocessing; and are one-dimensional convolution layers, each containing one-dimensional convolution ( ) + dimension transformation (Reshape) + linear layer (Linear) + batch normalization layer (batchNorm1d), with an output dimension of , is the hidden layer dimension ( );

[0127] The point-by-point mapping introduces a convolution branch, mainly used to enhance the feature expression capability:

[0128] ;

[0129] wherein, 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 ;

[0130] In order to capture longer span feature information, the global view adopts a convolution kernel with a width equal to the input dimension and the historical time steps (i.e. a convolution kernel with a width equal to the input dimension ), ), i.e. a convolution kernel size of Map all features of the historical time step:

[0131]

[0132] wherein, is the feature vector extracted by the global view; LargeConv contains a one-dimensional convolution ( ) + dimension transformation (Reshape) + linear layer (Linear) + batch normalization layer (batchNorm1d), and the output dimension is .

[0133] The outputs of the above three paths are spliced to obtain the fused high-dimensional features: , ,

[0134] ;

[0135] represents the high-dimensional features obtained by fusing , , , and the dimension is .

[0136] 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:

[0137] ;

[0138] wherein, is a sequence of process parameter features 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 perception ability of the model to features and between features is enhanced, more abundant 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.

[0139] 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 the process parameter of each historical time step, and then obtain the time-domain historical hidden state matrix composed of the hidden state sequence. Specifically as follows:

[0140] ​​To further explore the dynamic rules hidden in the evolution of each process parameter over time in the tobacco filling process, a Selective Structured State Space Model (Mamba) was introduced to the feature sequence S1 obtained in S1 Deep time series modeling was performed.

[0141] Specifically, the output feature tensor after multi-scale convolution fusion, i.e., the feature sequence S1 obtained in S1 was input into the Mamba module, which includes a state space layer, a selection mechanism layer, and an output projection layer. Through linear state space conversion and adaptive selection mechanism, the time series feature information of the process parameters between each historical time step was obtained, and the hidden state sequence of the process parameters at each historical time step was obtained. The specific process is as follows:

[0142] First, the feature sequence after S1 multi-scale one-dimensional convolution fusion (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 computational efficiency) through a linear projection layer, and the projection formula is as follows:

[0143] ;

[0144] 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.

[0145] 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.

[0146] For the state transition kernel branch , the feature data at the time step (indicating all batch features at the time step and all dimensions) is applied to the exponential function to generate the The state transition kernel for each time step is defined by the following formula:

[0147] ;

[0148] in, For the first The state transition kernel at each time step controls the decay rate of historical states. A smaller value indicates that historical states decay faster, and the model focuses more on the current input; a larger value indicates that historical states are preserved more completely, and the model focuses more on capturing long-range dependencies.

[0149] Input gate branch Take the first Feature data of time steps Applying the sigmoid activation function to generate the first The input gate weights for each time step are calculated using the following formula:

[0150] ;

[0151] in, Using the sigmoid activation function, the output is constrained to... interval; For the first Input gate weights at time steps.

[0152] Set initial state (Zero vector, dimension is) ), combined with state transition kernel Input gate Passing the exam Input of time step ( In the The state is updated using element-wise multiplication and linear combination of the feature data at each time step, as shown in the following formula:

[0153] ;

[0154] in, The hyperbolic tangent activation function is used to constrain the output to... The interval is used to ensure the stability of the state value; This represents element-wise multiplication; For the first The state vector at each time step.

[0155] Traverse all historical time steps Then, the state vectors at each time step are... Stacked along the time dimension, the hidden state matrix of the time-domain history is constructed. The selective state-space mechanism employed by this Mamba module significantly outperforms traditional recurrent neural networks (RNNs) and their variants (such as GRU and LSTM) in capturing long-sequence dynamic dependencies. This mechanism not only effectively avoids the vanishing gradient problem but also enhances the model's ability to model nonlinear changes under complex operating conditions through selective information processing. Compared to the sequential computation mode of GRU / LSTM, Mamba's parallel computing architecture greatly improves training efficiency, giving it a significant advantage in processing industrial big data.

[0156] When the input sequence is processed by the Mamba module, a corresponding hidden state representation is generated at each time step (t=1,...,T). These hidden state sequences constitute a complete temporal feature representation matrix.

[0157] This invention introduces an advanced Mamba model to achieve efficient modeling of process-level timing information. Unlike the traditional GRU / LSTM gating mechanism, this model employs selective state-space theory, achieving highly parallel training while efficiently capturing long-range dependencies. This not only maintains fast inference capabilities suitable for industrial real-time control scenarios but also significantly reduces the computational cost of model training, making it suitable for industrial applications.

[0158] 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, thereby obtaining the process parameter features at the current time step. Details are as follows:

[0159] To enable the features at 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 transform the input feature vector at the current time step. ( The model maps all features (except for the instantaneous moisture content of the effluent) to the latent space, allowing it to learn the relationship between production parameters and the instantaneous moisture content of the effluent; the linear transformation is achieved by a learnable weight matrix. Achieve (where, (random initialization, and updating the weight matrix using backpropagation gradients).

[0160] ;

[0161] Where q is the output vector, containing key information about the features at the current time step, with dimensions of . .

[0162] S4: Input the hidden state matrices of process parameters from multiple historical time steps and the process parameter features of the current time step into the attention module to obtain the fused temporal comprehensive features. Details are as follows:

[0163] In the tobacco feeding process, there is a complex dependence between the process parameters of different time steps, and the current time step is often affected by various factors in the past period. To further explore the correlation between the current time step input features and the historical state, and improve the response ability of the model to the change of key process parameters, the attention mechanism is introduced to weight and fuse the historical hidden state and the current time step feature.

[0164] S401: The process parameter features of the current time step are taken as Query, and the historical hidden state matrix output by the Mamba module As Key and Value, the attention weight is calculated

[0165]

[0166] where q is the process parameter feature of the current time step; is the hidden state of t time step; H is the historical hidden state matrix output by the Mamba module. The attention weight reflects the importance of each historical time step hidden state to the current time step, the higher the weight, the stronger the relevance of the information of the historical time step to the current situation.

[0167] S402: Based on the calculated attention weight, all historical hidden states are weighted and summed to generate a context vector

[0168]

[0169] The context vector represents the most relevant past information of the current time step, which integrates the most matched part of the historical hidden state to the current feature.

[0170] S403: The context vector is concatenated with the current feature representation in the feature dimension to form the fused time domain comprehensive feature

[0171]

[0172] The time domain fusion feature is then 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 historical process. It fully utilizes the correlation between historical information and current features, so that the model can consider the past influence when predicting the outlet moisture and related control parameters, thereby improving the prediction accuracy and robustness. In the actual tobacco cut tobacco feeding process, this 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.​​​​​​

[0173] S5: Perform a 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. Perform global average pooling on the amplitude sequence to obtain the frequency domain features. Specifically, as follows:

[0174] The result of S1 Each feature dimension Perform independent FFT transformations on the time series. For the ... The sample, the first Time series of features Its FFT transform can be expressed as:

[0175] ;

[0176] in For the frequency domain complex number representation, The imaginary unit, This is the frequency index. Because the FFT results of the time-domain signal are symmetric, only the first few digits need to be retained. Each frequency component is denoted as... The frequency domain data of all samples and feature dimensions constitute a tensor. .

[0177] The amplitude of the complex sequence in the frequency domain is calculated to characterize the energy intensity of each frequency component:

[0178] ;

[0179] in For the first The amplitude of each frequency component, Represents the real part of a complex number. The imaginary part of a complex number is used to calculate the amplitude data. The amplitude distribution reflects the contribution of different periodic components in the characteristic sequence; for example, low-frequency components may correspond to the slow changing trend of process parameters.

[0180] To compress frequency domain amplitude information into global features, global average pooling is performed on the amplitude sequences of each feature dimension:

[0181] ;

[0182] in The number of effective frequency components, For the first The sample, the first the global feature of each feature dimension in the frequency domain. Finally, the pooling results of all feature dimensions constitute the frequency domain feature matrix The physical meaning of which is the average energy distribution of each feature dimension in the frequency domain, which can effectively capture the periodic dynamic mode of the process parameters.

[0183] 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:

[0184] 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:

[0185] ;

[0186] Where GlobalAvgPool represents the global average pooling operation, and the time dimension is averaged, and the formula is:

[0187] ;

[0188] The aligned time domain features and the frequency domain features are spliced along the feature dimension to obtain fusion features :

[0189] ;

[0190] The mathematical expression of the splicing operation is:

[0191] ;

[0192] Wherein, The dimension of , while retaining the time domain dynamic information and frequency domain periodic characteristics.

[0193] S6: The comprehensive features are linearly transformed through a fully connected layer to output the predicted value of the target process parameter at the next time step. Specifically as follows:

[0194] The spliced and fused comprehensive features are represented 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:

[0195] ;

[0196] Wherein, represents the output of the model, and the output dimension is , is a learnable weight matrix, It is the bias vector, used to fine-tune the prediction results. It is continuously optimized through learning. and The value of allows the model to better capture the complex relationship between input features and the prediction target. In this embodiment, the output vector It includes three key predictive objectives:

[0197] The first dimension of the output vector represents the predicted instantaneous moisture content of the material at the outlet in the next time step. The outlet moisture content is a critical quality indicator in the tobacco processing stage, directly affecting the quality of the tobacco leaves and the effectiveness of subsequent processing. Accurately predicting the outlet moisture content helps to adjust relevant parameters in the feeding process in a timely manner, ensuring that the moisture content of the tobacco leaves meets production requirements.

[0198] The second dimension represents the recommended actual hot air temperature setpoint for the next time step. Hot air temperature significantly impacts the drying degree of tobacco leaves and the uniformity of feeding. By predicting appropriate hot air temperature setpoints, the tobacco drying process can be optimized, the feeding effect improved, and ultimately, the quality of the entire tobacco processing step enhanced.

[0199] The third dimension is the recommended adjustment value for the actual opening of the exhaust air volume in the next time step. The exhaust opening controls the humidity environment within the process. A suitable exhaust opening helps maintain stable process conditions and avoids affecting the quality and processing efficiency of tobacco leaves due to abnormal humidity.

[0200] Through this multi-objective prediction module, the model can simultaneously model and accurately predict the core quality indicators (outlet moisture) and key control variables (hot air temperature, dehumidification opening) in the feeding process. This capability provides the control system with forward-looking adjustment suggestions. Figure 3 , Figure 4 , Figure 5 These are line graphs showing the actual and predicted values ​​of outlet moisture, exhaust air volume opening, and hot air temperature for the test set data, illustrating the model's prediction results. The number and type of the predicted target process parameters can be changed according to the actual scenario and are not limited, such as actual steam flow rate, actual steam pressure, actual water temperature in the water tank, and cylinder temperature.

[0201] This embodiment provides a method for predicting and controlling process parameters during tobacco feeding. 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 correlation 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 dehumidification opening) under complex operating conditions.

[0202] Preferably, when the prediction deviation of the target process parameter during the tobacco feeding process does not meet the preset prediction deviation condition during actual use, the process parameter data and the corresponding true label in the latest preset period are selected to join the training data set, the model parameters are reinitialized, the model is trained to update the parameters until the prediction deviation meets the preset prediction deviation condition.

[0203] In the implementation of the present embodiment, during the model deployment and operation, the system continuously records the deviation between the predicted value and the actual measured value of the model output: for the three key variables predicted by the model: the instantaneous moisture of the outlet material at the next time step, the actual temperature of the hot air, and the actual opening degree of the moisture removal air volume, the relative errors thereof are calculated respectively:

[0204]

[0205] Among them: represents the actual measured value at the th time step; represents the corresponding model predicted value; respectively correspond to the three prediction targets of outlet moisture, hot air temperature, and moisture removal opening degree. By calculating the relative error, the deviation between the model predicted value and the actual value is quantified, which provides a basis for judging whether the model performance has decreased.

[0206] When the relative errors of the above three predicted values are all greater than 1% in continuous 30 time steps (5 minutes), it is determined that the current model performance has significantly decreased, triggering the online updating mechanism:

[0207] The input feature data and the corresponding true label of the last 10 days 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 according to the conventional training process, the predicted value is calculated by forward propagation, and then the loss function is calculated according to the predicted 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. In the training process, the parameters are continuously optimized through multiple iterations, so that the prediction results of the model can better fit the new data, and the adaptability of the model to the current process state is improved. After the training is completed, the updated model is loaded into the inference module to replace the original model to continue 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.

[0208] Through the above mechanism, the model can perceive the changes in the process state in real time and automatically adjust the internal parameters, thereby effectively dealing with changes in real working conditions such as equipment aging and raw material fluctuations, so that the model can continuously and stably operate in a changing production environment.

[0209] Embodiment 2​

[0210] The embodiment provides a system of process parameter prediction control in a tobacco feeding process, comprising:

[0211] 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 feeding process and perform data preprocessing to convert the process parameter data into corresponding time sequence samples:

[0212] 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 process parameter prediction model that is constructed and trained, and output a predicted value of a target process parameter of a next time step;

[0213] The process parameter prediction model predicts a specific process as follows:

[0214] S1: multi-scale one-dimensional convolution is used to extract features from the time sequence samples of the process parameters of the plurality of historical time steps, and then the features of different scales are fused to obtain process parameter feature sequences of the plurality of historical time steps;

[0215] S2: the process parameter feature sequences of the plurality of historical time steps in S1 are input into an Mamba module to extract time sequence features, 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;

[0216] S3: a linear transformation layer is used to map the time sequence samples of the process parameters of the current time step to a hidden space to obtain process parameter features of the current time step;

[0217] S4: the time-domain historical hidden state matrix and the process parameter features of the current time step are input into an attention module to obtain time-domain comprehensive features;

[0218] S5: fast Fourier transform is performed on each feature dimension of the process parameter feature sequence output by S1 to obtain a frequency domain complex sequence, the amplitude of each frequency component in the frequency domain complex sequence is calculated to obtain an amplitude sequence, and global average pooling is performed on the amplitude sequence to obtain frequency domain features;

[0219] S6: the time-domain comprehensive features obtained in S4 and the frequency domain features obtained in S5 are spliced to obtain fusion features;

[0220] S7: a fully connected layer is used to perform linear transformation on the comprehensive features to output the predicted value of the target process parameter of the next time step.

[0221] Embodiment 3

[0222] 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.

[0223] Embodiment 4

[0224] The embodiment provides an electronic terminal which comprises a processor and a memory, the memory stores a computer program, the processor calls the computer program to execute steps of the method. It can be understood that same or similar parts in the above-mentioned embodiments can be mutually referred, and contents which are not described in detail in some embodiments can be referred to same or similar contents in other embodiments.

[0225] 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.

[0226] 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.

[0227] Based on such understanding, the technical solutions of the present application, essentially or in other words, the part that contributes to the prior art, 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 a number of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present application. The aforementioned readable storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0228] 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 are not to be construed as limiting the present application, and those of ordinary skill 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 during tobacco leaf feeding, characterized in that, include: The process parameter data of the current time step and multiple consecutive historical time steps before the current time step during the tobacco feeding process are acquired and preprocessed to convert them into corresponding time series samples. These time series samples of the process parameter data from the current and historical time steps are then input into a constructed and trained process parameter prediction model. The output is the predicted value of the target process parameter for the next time step. The specific prediction process of the process parameter prediction model is as follows: S1: Multi-scale one-dimensional convolution is used to extract features from time series samples of process parameters at multiple historical time steps, and then the extracted features at each scale are fused to obtain a feature sequence 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 process parameters for each historical time step, and then obtain the time domain historical hidden state matrix composed of the hidden state sequence. 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. S4: Input the time-domain historical 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 a 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. Perform global average pooling on the amplitude sequence to obtain frequency domain features. S6: Combine the time-domain integrated features obtained in S4 and the frequency-domain features obtained in S5 to obtain the fused features; S7: The fused features are linearly transformed by a fully connected layer to output the predicted values ​​of the target process parameters for the next time step; Specifically, the hidden state sequence of process parameters for each historical time step in S2 involves taking the feature sequences of process parameters from multiple historical time steps as input and sequentially inputting them into a selective state space model module. This module includes a state space layer, a selection mechanism layer, and an output projection layer. Through linear state space transformation and an adaptive selection mechanism, the temporal feature information of process parameters between each historical time step is obtained, corresponding to the hidden state sequence of process parameters for each historical time step. The specific process is as follows: First, the feature sequence fused by S1 multi-scale one-dimensional convolution is... , For batch size, For historical time steps, The feature dimension is mapped to the state space dimension adapted by the Mamba module through a linear projection layer. The projection formula is as follows: ； in, The weight matrix of the linear projection layer; For the bias term of the linear projection layer; Input to the projected state space; Then It is split into three parallel branches: the state transition kernel branch and the state transition kernel branch. Input gate branch Selecting weighted branches Each branch dimension is related to Consistency, that is ; For state transition kernel branch Take the first Feature data of time steps Applying the exponential function to generate the first The state transition kernel for each time step is defined by the following formula: ； in, For the first The state transition kernel for each time step; Input gate branch Take the first Feature data of time steps Applying the sigmoid activation function to generate the first The input gate weights for each time step are calculated using the following formula: ； in, It is the sigmoid activation function; For the first Input gate weights at each time step; Set initial state Combined with state transition kernel Input gate weights Passing the exam Input of time step , In the The feature data at each time step is used to update the state through element-wise multiplication and linear combination, as shown in the following formula: ； in, The hyperbolic tangent activation function is used to constrain the output to... The interval is used to ensure the stability of the state value; This represents element-wise multiplication; For the first The state vector at each time step; Traverse all historical time steps Then, the state vectors at each time step are... Stacked along the time dimension, they form a time-domain historical hidden state matrix.

2. The method according to claim 1, characterized in that, The training process of the process parameter prediction model is as follows: S001: Collect multi-source process parameter data from different sampling periods during the historical tobacco leaf production process and construct a high-dimensional feature space matrix; S002: Using the sampling period as the batch, the t-distribution random neighborhood embedding algorithm is used to reduce the dimensionality of the high-dimensional data in the high-dimensional feature space matrix to a two-dimensional space, resulting in a two-dimensional scatter plot. S003: The K-Means clustering algorithm is used to perform cluster analysis on the scatter data in the two-dimensional scatter plot, clean up the abnormal data, and standardize the high-dimensional data corresponding to the cleaned scatter 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 time series samples, and build the model training dataset; S005: Based on the model training dataset, the process parameter prediction model is trained and updated by taking time series samples of process parameter data from historical time steps and the current time step as input and the target process parameter of the next time step as output, 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, In S1, the multi-scale one-dimensional convolution extracts features from the input historical time step process parameters from the local time window, point-by-point mapping, and global view, respectively. The local time window is achieved by using two stacked one-dimensional convolutional layers, both with a kernel size of [missing value]. : ； in, The feature vector extracted from the local time window; This refers to the data after preprocessing. and All are one-dimensional convolutional layers, including one-dimensional convolution, dimension transformation, linear layers, and batch normalization layers; Pointwise mapping introduces a Convolutional branches: ； in, The feature vector is extracted through point-by-point mapping; It is a one-dimensional convolutional layer, including one-dimensional convolution, dimension transformation, linear layer, and batch normalization layer; The global view is achieved by using a convolutional kernel with a width equal to the input dimension and the historical time step, i.e., the 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 convolutional layer, which includes one-dimensional convolution, dimension transformation, linear layer, and batch normalization layer.

4. The method according to claim 1, characterized in that, The specific process of S4 is as follows: S401: Using 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, calculate the attention weights. : ； Where q represents the process parameter characteristics of the current time step; Let H be the hidden state at time step t; H is the historical hidden state matrix output by the Mamba module. S402: Based on the calculated attention weights, perform a weighted summation of all historical hidden states to generate a context vector. ; S403: Transfer the context vector With current feature representation The features are combined along the feature dimension to form a fused comprehensive feature. .

5. The method according to any one of claims 1-4, characterized in that, In addition, during actual use, when the predicted value of the target process parameter does not meet the preset prediction deviation condition during the tobacco feeding process, the latest preset time period of process parameter data is selected and added to the training dataset, the model parameters are re-initialized, and the model is trained and updated until the prediction deviation meets the preset prediction deviation condition.

6. A system based on the method for predicting and controlling process parameters during tobacco feeding as described in any one of claims 1-5, characterized in that, include: Data acquisition and preprocessing module: Used to acquire process parameter data of the current time step and multiple consecutive historical time steps before the current time step during the tobacco feeding process, and perform data preprocessing to convert it into corresponding time series samples. Process parameter prediction module: It takes the process parameter data of the current time step and the historical time step as input, inputs them into the process parameter prediction model that has been built and trained, and outputs the predicted value of the target process parameter for 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 time series samples of process parameters at multiple historical time steps, and then the extracted features at each scale are fused to obtain a feature sequence 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 process parameters for each historical time step, and then obtain the time domain historical hidden state matrix composed of the hidden state sequence. 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. S4: Input the time-domain historical 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 a 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. Perform global average pooling on the amplitude sequence to obtain frequency domain features. S6: Combine the time-domain integrated features obtained in S4 and the frequency-domain features obtained in S5 to obtain the fused features; S7: The fused features are linearly transformed by a fully connected layer to output the predicted values ​​of the target process parameters for the next time step.

7. A readable storage medium, characterized in that: A computer program is stored, which, when invoked by a processor, performs the steps of the method according to any one of claims 1-5.

8. An electronic terminal, characterized in that: It includes a processor and a memory, the memory storing a computer program, the processor calling the computer program to perform the steps of the method according to any one of claims 1-5.

Images