Deep learning-based cut tobacco drying outlet moisture content stability control method

Abstract

According to the invention, based on a deep learning method, the moisture content of a cut tobacco drying outlet of a tobacco cut-tobacco making process is subjected to stability control, and a plurality of sensors are arranged on a tobacco cut-tobacco dryer production line; a plurality of key process parameters (cut tobacco moisture content, steam flow, tobacco shred drying-anti-caking valve opening, tobacco shred oxygen content, process airflow temperature, tobacco shred drying cumulant, process flow and discharged material moisture content data) of tobacco shreds are collected in real time, and the data are preprocessed and then input into a stacked integrated learning model S-LSTM for training; the method is used for establishing a nonlinear relation model between the tobacco shred moisture content and other process parameters, an FWOA (improved whale optimization algorithm) optimization module is introduced for stable control of the moisture content of a dried tobacco shred outlet, the module minimizes an error between a prediction result of an integrated learning model S-LSTM and a target moisture content, and the moisture content of the dried tobacco shred is obtained. And process parameters (process airflow temperature and process flow) are automatically adjusted, so that accurate control of the moisture content at the outlet of the dried cut tobacco is realized.

Description

Technical Field

[0001] The present invention designs a method for controlling the moisture content stability of dried tofu at an outlet, and in particular designs a method for controlling the moisture content stability of dried tofu at an outlet based on deep learning. Background Art

[0002] The existing methods for controlling the moisture content stability of dried silk strips at the outlet usually adopt the following methods:

[0003] 1. PID + manual control method to adjust key parameters such as process flow or process hot air temperature.

[0004] 2. Based on machine learning prediction, the moisture content of the dried tofu outlet is adjusted using the exponentially weighted moving average (EWMA).

[0005] 3. Make predictions and adjust parameters based on deep learning methods.

[0006] The existing method for controlling the stability of moisture content at the outlet of dried tofu has problems such as lag in adjustment, low manual adjustment accuracy, and reliance on historical experience. The main problems are as follows:

[0007] 1. The prediction accuracy is not high, the system dynamic response is not sensitive enough when adjusting parameters, the adjustment is delayed, and the control effect is poor.

[0008] 2. The deep learning method has low interpretability for the model, and the adjusted parameters are difficult to be accepted by production practice. Summary of the invention

[0009] In view of the shortcomings of the above conventional technologies, the present invention proposes a method for controlling the moisture content stability of dried silk strips at the outlet based on deep learning. The method uses FWOA (Fast Whale optimization algorithm) and a stacked ensemble learning model S-LSTM built based on an ensemble learning stacking framework for controlling the moisture content stability of dried silk strips at the outlet, and is applied to production practice. The ensemble learning model S-LSTM is stacked by a base classifier composed of six machine learning algorithms (CatBoost, Decision Tree, Random Forest, LightGBM, ExtraTrees and KNN) and a meta-classifier composed of LSTM with an attention mechanism. The stacked ensemble learning model S-LSTM is used to predict the moisture content of dried silk strips at the outlet, and the error between the predicted value and the standard value is used as the objective function of the optimization algorithm. FWOA is used to optimize and adjust the process parameters, thereby achieving stability control of the moisture content of dried silk strips at the outlet.

[0010] In order to achieve the invention purpose of this application, this application adopts the following technical solutions:

[0011] The present invention discloses a method for controlling the moisture content stability of tobacco leaf drying outlet based on deep learning. A plurality of sensors are arranged on the production line of a tobacco leaf drying machine. The sensors are used to collect the moisture content of cut tobacco leaves, steam flow rate, opening degree 1 of a leaf drying-anti-agglomeration valve, oxygen content of the leaf leaves, process air flow temperature, cumulative amount of leaf drying, process flow rate and moisture content of the discharge in real time. The method for controlling the moisture content stability of tobacco leaf drying outlet based on deep learning includes the following steps:

[0012] 1. Data Collection

[0013] The above multiple sensors are used to collect the moisture content of cut tobacco leaves, steam flow, the opening of the leaf drying-anti-agglomeration valve 1, the oxygen content of the leaf leaves, the process air flow temperature, the cumulative amount of leaf drying, the process flow and the moisture content of the discharge in real time, and the above eight data at each moment are combined into a group of original parameters, and multiple groups of original parameters are collected;

[0014] (II) Data preprocessing

[0015] The above multiple groups of original parameters are processed as follows

[0016] (1) Fill in missing parameters

[0017] When a parameter at a certain moment is missing, the missing parameter is filled with the average of the parameters of the first five moments and the parameters of the last five moments;

[0018] (2) Processing outliers on original parameters

[0019] Calculate the average value of each parameter of each group of original data, detect the above parameters, and when the above parameters are within the range of ±3 times the standard deviation of their average values, they are regarded as constant values ​​and retained; otherwise, they are regarded as abnormal values ​​and the group of original parameters including the abnormal values ​​are deleted;

[0020] (3) Standardize the data processed in step (2)

[0021] When the above parameters are within their range, they are processed according to the following formula to normalize the above parameters to between 0 and ±1.

[0022]

[0023] Where: Z is the standardized value (Z-Score); X is the original data; μ is the mean of each parameter within its range; σ is the standard deviation of each parameter within its range;

[0024] The first seven parameters at the same time are used as the input data of the model, and the moisture content of the discharge material is used as the output data of the model;

[0025] (III) Training the stacked ensemble learning model S-LSTM

[0026] The integrated learning model S-LSTM is composed of a base classifier consisting of six machine learning algorithms (CatBoost, Decision Tree, RandomForest, LightGBM, ExtraTrees and KNN) and a meta-classifier consisting of LSTM integrated with attention mechanism.

[0027] 80% of the above-mentioned corresponding input data and output data constitute a training set, 10% of the above-mentioned corresponding input data and output data constitute a validation set, and 10% of the above-mentioned corresponding input data and output data constitute a test set. The input data of the training set is used as the input of the six sub-modules of the base classification module, and the original input data and the output data of the six sub-modules are used as the input of the meta-classification module. The output of the meta-classification module is the final prediction value. During the training process, the above operations are performed on all the corresponding input data and output data in the training set in sequence to complete a complete round of iterations. In each round of training of the training set, the data of the training set is first input into the S-LSTM model in batches, and the model output is calculated by forward propagation. Then, the model parameters are updated by back propagation, and the loss value of the training set is recorded at the same time. After completing the traversal of all training set data, enter the verification phase, input all the data of the verification set into the S-LSTM model, calculate the loss value of the verification set, and introduce a learning rate scheduler to optimize the training process. The learning rate is dynamically adjusted according to the change of the loss value of the verification set. If the loss value of the verification set in this round is lower than the lowest value of the loss value of the verification set in the previous round, the current S-LSTM model is saved as the best candidate model. The above process is repeated for at least 50 rounds to fully train the S-LSTM model. During the whole process, the parameters of the S-LSTM model are gradually optimized. The learning rate scheduler ensures that the weight parameters are adjusted more finely in the later stage of training, thereby improving the performance of the S-LSTM model. Finally, the S-LSTM model with the lowest loss value of the verification set is selected as the final trained S-LSTM model.

[0028] (IV) Use the FWOA optimization module to calculate the process air flow temperature and process flow rate to achieve the predetermined moisture content of the dried tofu.

[0029] (a) Randomly generate 20 sets of process air flow temperature and process flow

[0030] Within the range of process air flow temperature and process flow rate, 20 sets of process air flow temperature and process flow rate are randomly generated;

[0031] (b) Use the above random process air flow temperature and process flow to replace the real-time monitoring data

[0032] The 20 sets of process air flow temperatures and process flow rates randomly generated above are used to replace the real-time detected moisture content of blade, steam flow rate, blade drying-anti-agglomeration valve opening 1, blade oxygen content, process air flow temperature, blade drying cumulative amount and process flow rate in the process air flow temperature and process flow rate to form 20 new sets of input data, and the number of iterations t=1;

[0033] (c) Call the stacked ensemble learning model S-LSTM to predict the moisture content of the output of the above 20 sets of input data, and compare it with the set moisture content of the output

[0034] The above 20 groups of input data are used as the input of the stacked ensemble learning model S-LSTM to obtain 20 predicted discharge moisture contents. The above 20 predicted discharge moisture contents are compared with the set discharge moisture contents. If one of the 20 predicted discharge moisture contents is the same as the set discharge moisture content, then the process air flow temperature and process flow rate corresponding to the discharge moisture content are the process air flow temperature and process flow rate to be regulated at this time; if the 20 predicted discharge moisture contents are different from the set discharge moisture content, find the group of input data with the smallest absolute value between the discharge moisture content and the set discharge moisture content among the 20 groups of input data. And detect whether the number of iterations t reaches the predetermined number of iterations. If t reaches the predetermined number of iterations, then the above set of input data The process gas flow temperature and process flow rate in are the process gas flow temperature and process flow rate to be regulated at this time; otherwise, proceed to step (d);

[0035] (d) Update the input data of step (b) to form 20 new input data sets

[0036] Substituting t+1 for t, we use the formula

[0037] Calculate 20 new sets of input data, among which, are 20 sets of input data of generation t, It is the set of input data with the smallest absolute value between the moisture content of the discharge material and the set moisture content of the discharge material among the 20 sets of input data of generation t. is the contraction coefficient, Set to 0.5, is a random number in [0,2] It is the new 20 sets of input data after updating to generation t+1. is the distance between each set of data and the best data;

[0038] 20 sets of input data are updated and returned to step (c) until the process air flow temperature and process flow rate to be regulated are calculated.

[0039] The present invention provides a method for controlling the stability of moisture content at the outlet of dried strips based on deep learning, wherein: in the step (b), 20 randomly generated groups of process air flow temperature, process flow rate, real-time detected moisture content of cut strips, steam flow rate, strip drying-anti-agglomeration valve opening 1, strip oxygen content, process air flow temperature, strip drying cumulative amount and process flow rate parameters of process air flow temperature and process flow rate are all processed through step (3).

[0040] The present invention provides a method for controlling the moisture content stability of dried tofu at the outlet based on deep learning, wherein, in step (iii), it also includes using a test set to test the trained stacked ensemble learning model S-LSTM.

[0041] The present invention provides a method for controlling the moisture content stability of dried tofu outlet based on deep learning, wherein: the standard deviation Where: N is the total number of data points; x i is the data of the ith data point; μ is the mean of the population, and the calculation formula is:

[0042] The present invention provides a method for controlling the moisture content stability of dried tofu outlet based on deep learning, wherein: the set outlet moisture content is 14.9.

[0043] The present invention provides a method for controlling the moisture content stability of dried tofu at the outlet based on deep learning, wherein: t is 100.

[0044] The present invention provides a method for controlling the moisture content stability of dried tofu at the outlet based on deep learning, wherein: the learning rate scheduler is ReduceLROnPlateau.

[0045] The present invention provides a method for controlling moisture content stability at a dried tofu outlet based on deep learning to solve the following technical problems:

[0046] 1. Improve the adjustment response speed: Optimize the process parameters through the FWOA algorithm, reduce the adjustment lag and improve the response speed.

[0047] 2. Improve control accuracy: Use the stacked integrated learning model S-LSTM for real-time prediction and control to improve the control accuracy of outlet moisture content.

[0048] 3. Realize intelligent control: Through machine learning and optimization algorithms, realize intelligent and systematic control of the moisture content of dried tofu outlet, reducing dependence on manual experience.

[0049] The method for controlling moisture content stability of dried tofu outlet based on deep learning of the present invention has the following advantages:

[0050] 1. Improved regulation accuracy: The modified system can automatically control and optimize process parameters, significantly improving regulation accuracy and reducing manual intervention. This is mainly due to the high prediction accuracy of the stacked ensemble learning model S-LSTM and the precise optimization capability of the FWOA algorithm.

[0051] 2. Optimize convergence speed: The improved Fast Whale Optimization Algorithm (FWOA) improves convergence speed and global search capability by dynamically adjusting parameters and adaptive inertia weights. Experimental results show that the FWOA algorithm exhibits faster convergence speed during the optimization process, can find the optimal parameter combination in a shorter time, and improves production efficiency.

[0052] 3. Enhanced stability: The modified system shows significant advantages in the stability of outlet moisture content. Compared with the traditional PID+manual control method, the improved control method reduces the average deviation of outlet moisture content from 0.1301 to 0.0838, a decrease of about 35.6%. This means that the outlet moisture content is more stable and the production quality has been significantly improved.

[0053] 4. Reduce human intervention: The automated control system reduces the reliance on operator experience and reduces the uncertainty caused by human factors. Through the application of intelligent algorithms, the process parameters can be automatically adjusted to ensure the consistency and stability of the production process.

[0054] 5. Improve production efficiency: Due to the fast convergence characteristics of the FWOA algorithm and the high prediction accuracy of the stacked ensemble learning model S-LSTM, the system can adjust process parameters in the shortest time, reducing adjustment time and production stagnation, and improving overall production efficiency.

[0055] 6. The improved model has strong adaptability: The stacked ensemble learning model S-LSTM combines LSTM and attention mechanism, can process complex time series data, has strong adaptability, can be applied in different production environments, and improves the generalization ability of the model.

[0056] 7. Improve tobacco quality: By real-time monitoring and adjustment of process parameters, the system can achieve higher control accuracy, reduce the fluctuation of outlet moisture content, and improve the consistency and reliability of product quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 It is a flow chart of the method for controlling moisture content stability of dried tofu outlet based on deep learning of the present invention;

[0058] Figure 2 This is a schematic diagram of the S-LSTM module structure of the stacked ensemble learning model;

[0059] Figure 3Flowchart of the FWOA optimization module;

[0060] Figure 4 This is the result of FWOA parameter adjustment;

[0061] Figure 5 Schematic diagram of the stacked ensemble learning model S-LSTM module and FWOA optimization module structure. DETAILED DESCRIPTION

[0062] like Figure 1 As shown, a plurality of sensors are arranged on the production line of the tobacco leaf drying machine, which are used to collect the moisture content of the cut tobacco leaves, steam flow, the opening degree 1 of the leaf drying-anti-agglomeration valve, the oxygen content of the leaf leaves, the process air flow temperature, the cumulative amount of leaf drying, the process flow and the moisture content of the discharge in real time. The method for controlling the moisture content stability of the drying outlet based on deep learning of the present invention comprises the following steps:

[0063] 1. Data Collection

[0064] The above multiple sensors are used to collect the moisture content of cut tobacco leaves, steam flow, the opening of the leaf drying-anti-agglomeration valve 1, the oxygen content of the leaf leaves, the process air flow temperature, the cumulative amount of leaf drying, the process flow and the moisture content of the discharge in real time, and the above eight data at each moment are combined into a group of original parameters, and multiple groups of original parameters are collected;

[0065] (II) Data preprocessing

[0066] The above multiple groups of original parameters are processed as follows

[0067] (1) Fill in missing parameters

[0068] When a parameter at a certain moment is missing, the missing parameter is filled with the average of the parameters of the first five moments and the parameters of the last five moments;

[0069] (2) Processing outliers on original parameters

[0070] Calculate the average value of each parameter of each group of original data, detect the above parameters, and when the above parameters are within the range of ±3 times the standard deviation of their average values, they are regarded as constant values ​​and retained; otherwise, they are regarded as abnormal values ​​and the group of original parameters including the abnormal values ​​are deleted;

[0071] (3) Standardize the data processed in step (2)

[0072] When the above parameters are within their range, they are processed according to the following formula to normalize the above parameters to between 0 and ±1.

[0073]

[0074] Where: Z is the standardized value (Z-Score); X is the original data; μ is the mean of each parameter within its range; σ is the standard deviation of each parameter within its range; standard deviation Where: N is the total number of data points; x i is the data of the ith data point; μ is the mean of the population, and the calculation formula is:

[0075] The first seven parameters at the same time are used as the input data of the model, and the moisture content of the discharge material is used as the output data of the model;

[0076] (III) Training the stacked ensemble learning model S-LSTM

[0077] like Figure 2 As shown in Figure 2, the ensemble learning model S-LSTM is composed of six machine learning algorithms (CatBoost, DecisionTree, Random Forest, LightGBM, ExtraTrees, and KNN) as the base classifier (see Figure 5 ①) and a meta-classifier consisting of an LSTM with an attention mechanism (see Figure 5 ②) are stacked together.

[0078] (1) Base classification module: CatBoost, Decision Tree, Random Forest, LightGBM, ExtraTrees and KNN are used as base classification models. The base classifier is responsible for the preliminary prediction of the moisture content of the dried tofu outlet.

[0079] (2) Meta-classification module: A meta-classifier combining LSTM and attention mechanism is used to combine the prediction results of the base classifier and the original process parameters to make the final moisture content prediction. The meaning of each structure is as follows:

[0080] Input layer: maps the input data from the original dimension to the hidden layer dimension of LSTM. This step preliminarily processes the input data to make it meet the input requirements of the LSTM layer.

[0081] LSTM layer is used to process time series data. It contains three parameter settings: input_dim is the dimension of the input feature, hidden_dim is the dimension of the hidden layer, and num_layers is the number of LSTM layers.

[0082] AttentionLayer: Used to calculate the attention score. The input is the hidden layer output hidden_dim of LSTM, and the output is a scalar.

[0083] LayerNormalization: Normalize the hidden layer output of LSTM to stabilize the training process.

[0084] Flatten (fully connected layer): maps the output of the attention layer to the final output dimension.

[0085] Stacked ensemble learning model S-LSTM module structure diagram ( Figure 2 ), showing the relationship between the base classification module and the meta classification module. After the process parameters are input into the base classification module, the prediction results of the base classification module and the process parameters are simultaneously input into the meta classification module LSTM-AT, so that when the meta classification module predicts the outlet moisture content, it can not only benefit from the prediction of the base classification module, but also directly use the process parameters for supplementary analysis, instead of relying solely on the prediction results of the base classification module, thereby achieving a more accurate and stable control effect. This ensures that the meta classification module can comprehensively consider the real-time changes of the process parameters and the prediction results of the base classification module, improving the overall prediction performance and stability of the model.

[0086] 80% of the above-mentioned corresponding multiple groups of input data and output data constitute a training set, 10% of the above-mentioned corresponding multiple groups of input data and output data constitute a validation set, and 10% of the above-mentioned corresponding multiple groups of input data and output data constitute a test set. The input data of the training set is used as the input of the six sub-modules of the base classification module, and the original input data and the output data of the six sub-modules are used as the input of the meta-classification module. The output of the meta-classification module is the final predicted value. During the training process, the above operations are performed in sequence on all the corresponding input data and output data in the training set to complete a complete round of iterations. In each round of training of the training set, the data of the training set is first input into the S-LSTM model in batches, and the model output is calculated by forward propagation, and then the model parameters are updated by back propagation. At the same time, the loss value of the training set is recorded. After completing the traversal of all the training set data, the verification phase is entered, and the verification All the data of the validation set are input into the S-LSTM model, and the loss value of the validation set is calculated. In order to optimize the training process, the learning rate scheduler is introduced to dynamically adjust the learning rate according to the change of the loss value of the validation set. The learning rate scheduler is ReduceLROnPlateau. If the loss value of the validation set in this round is lower than the lowest value of the loss value of the validation set in the previous round, the current S-LSTM model is saved as the best candidate model. The above process is repeated for 100 rounds to fully train the S-LSTM model. During the whole process, the parameters of the S-LSTM model are gradually optimized. The learning rate scheduler ensures that the weight parameters are adjusted more finely in the later stage of training, thereby improving the performance of the S-LSTM model. Finally, the S-LSTM model with the lowest loss value of the validation set is selected as the final trained S-LSTM model, and the trained S-LSTM model is tested using the test set.

[0087] (IV) Figure 3 As shown in the figure, the FWOA optimization module is used to calculate the process air flow temperature and process flow rate to achieve the predetermined moisture content of the dried tofu.

[0088] (a) Randomly generate 20 sets of process air flow temperature and process flow

[0089] Within the range of process air flow temperature and process flow rate, 20 sets of process air flow temperature and process flow rate are randomly generated, and the parameters of the process air flow temperature and process flow rate are processed through step (3);

[0090] (b) Use the above random process air flow temperature and process flow to replace the real-time monitoring data

[0091] The 20 sets of process air flow temperatures and process flow rates randomly generated above are used to replace the process air flow temperatures and process flow rates in the shredded leaf moisture content, steam flow rate, shredded leaf drying-anti-clumping valve opening 1, shredded leaf oxygen content, process air flow temperature, shredded leaf drying cumulative amount and process flow rate detected in real time at this time, to form 20 new sets of input data, and the parameters of the shredded leaf moisture content, steam flow rate, shredded leaf drying-anti-clumping valve opening 1, shredded leaf oxygen content, process air flow temperature, shredded leaf drying cumulative amount and process flow rate detected in real time are all processed through step (3), and the number of iterations is set to t=1;

[0092] (c) Call the stacked ensemble learning model S-LSTM to predict the moisture content of the output of the above 20 sets of input data, and compare it with the set moisture content of the output

[0093] The above 20 groups of input data are used as the input of the stacked ensemble learning model S-LSTM to obtain 20 predicted discharge moisture contents. The above 20 predicted discharge moisture contents are compared with the set discharge moisture content of 14.9. If one of the 20 predicted discharge moisture contents is the same as the set discharge moisture content, then the process air flow temperature and process flow rate corresponding to the discharge moisture content are the process air flow temperature and process flow rate to be regulated at this time; if the 20 predicted discharge moisture contents are different from the set discharge moisture content, find the group of input data with the smallest absolute value between the discharge moisture content and the set discharge moisture content among the 20 groups of input data. And check whether the number of iterations t reaches the predetermined number of iterations. If t reaches 100 rounds, then the above set of input data The process gas flow temperature and process flow rate in are the process gas flow temperature and process flow rate to be regulated at this time; otherwise, proceed to step (d);

[0094] (d) Update the input data of step (b) to form 20 new input data sets

[0095] Substituting t+1 for t, we use the formula

[0096] Calculate 20 new sets of input data, among which, are 20 sets of input data of generation t, It is the set of input data with the smallest absolute value between the moisture content of the discharge material and the set moisture content of the discharge material among the 20 sets of input data of generation t. is the contraction coefficient, Set to 0.5, is a random number in [0,2] It is the new 20 sets of input data after updating to generation t+1. is the distance between each set of data and the best data;

[0097] 20 sets of input data are updated and returned to step (c) until the process air flow temperature and process flow rate to be regulated are calculated.

[0098] According to the optimal parameters provided by the FWOA optimization module, the process air flow temperature and process flow rate of the tofu drying machine are adjusted in real time to ensure the stability of the outlet moisture content, such as Figure 4 The curves around 14.9 shown are the outlet moisture content after model adjustment.

[0099] Figure 4 The optimization control effect diagram shown is a comparison between the outlet moisture content obtained by the deep learning-based control method of the present invention and the traditional control method. It can be intuitively seen that the control effect is better in terms of outlet moisture content stability before and after optimization. Before optimization, the outlet moisture content fluctuated greatly and deviated greatly from the standard value of 14.9; after applying the improved FWOA and the stacked integrated learning model S-LSTM, the fluctuation of the outlet moisture content was significantly reduced and closer to the standard value.

[0100] like Figure 5 As shown in Figure 1, the model first extracts the time series data of the drying process from the sensor from the manufacturing execution system (MES), and then performs the data outlier processing and normalization preprocessing steps. The processed data is input into the base classification module (see Figure 5 In ①), CatBoost, Decision Tree, Random Forest, LightGBM, ExtraTrees and KNN are used to preliminarily predict the moisture content of dried tofu outlet. The output features of the base classification module are combined with the original data and input into the meta-classification module (see Figure 5 ②). The meta-classification module maps the input data from the original dimension to the hidden layer dimension of LSTM through the Inputlayer. The LSTM layer is used to process time series data and capture the long-term and short-term dependencies in the time series to make full use of the information in the time dimension. After LSTM, the AttentionLayer is usually used to dynamically adjust the input weights of different time steps. It can help the model focus on important time steps and effectively reduce the interference of irrelevant information, thereby improving the prediction effect of the model. When processing time series data, the attention mechanism can capture the important connections between different time points and give higher weights to time steps that contribute more to the prediction results. LayerNormalization normalizes the output of the attention layer to stabilize the training process.

[0101] Finally, the output is mapped to the target prediction value through Flatten (fully connected layer), that is, the predicted moisture content of the dried tofu outlet. When the prediction result does not meet the production index, the original data is input into the FWOA module. After FWOA adjusts the original data, it calls the stacked ensemble learning model S-LSTM for prediction, and dynamically adjusts the process parameters according to the prediction result and the error (see step (IV) for the specific process). Through iterative optimization of FWOA, the optimal parameter combination is found to ensure the stability of the outlet moisture content.

[0102] The method for controlling moisture content stability at the outlet of dried tofu based on deep learning of the present invention has the following innovative features:

[0103] 1. Moisture content prediction model based on deep learning

[0104] This invention uses the stacking strategy to combine machine learning with deep learning models for the first time to predict the moisture content of dried tofu outlet. Compared with traditional control methods, the stacked ensemble learning model S-LSTM can capture the time series characteristics and nonlinear relationships in process data, thereby achieving accurate prediction of moisture content changes during the drying process, significantly improving the prediction accuracy and response speed of the system.

[0105] 2. Application of FWOA optimization algorithm

[0106] The present invention introduces an improved whale optimization algorithm (FWOA) for optimizing the process parameters in the tofu drying process. The FWOA algorithm optimizes the moisture content error predicted by the stacked ensemble learning model S-LSTM, and automatically adjusts key process parameters (such as process air flow temperature and process flow rate), thereby achieving precise control of the moisture content at the tofu drying outlet. Compared with traditional optimization methods, the FWOA algorithm has stronger global search capabilities and higher optimization efficiency, can effectively avoid the problem of local optimal solutions, and improve the stability and optimization effect of the system.

[0107] 3. Applicable to intelligent control of complex tobacco drying process

[0108] The deep learning and optimization algorithm method adopted by the present invention not only adapts to the complexity of the traditional tobacco drying process, but also can cope with changing environmental factors and production conditions, and has strong adaptability and flexibility. The method can be widely used in modern tobacco processing factories to realize the intelligence and automation of the drying process.

[0109] 4. Effect of this application

[0110] pass Figure 4The comparison of experimental results can intuitively show the control effect before and after optimization. Before optimization, the outlet moisture content fluctuated greatly and deviated greatly from the standard value of 14.9; after applying the improved FWOA and stacked ensemble learning model S-LSTM, the fluctuation of the outlet moisture content was significantly reduced and closer to the standard value.

[0111] The above is only one embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes that can be easily thought of by any technician familiar with the technical field within the technical scope disclosed in the present invention should be covered within the protection scope of the present invention.

Claims

1. A method for controlling moisture content stability at the tobacco leaf drying outlet based on deep learning, wherein a plurality of sensors are arranged on the production line of a tobacco leaf drying machine, and are used to collect moisture content of cut tobacco leaves, steam flow, opening of a leaf drying-anti-agglomeration valve 1, oxygen content of the leaf leaves, process air flow temperature, cumulative amount of leaf drying, process flow and moisture content of the discharge in real time, characterized in that: The method for controlling moisture content stability of dried tofu outlet based on deep learning includes the following steps:

1. Data Collection The above multiple sensors are used to collect the moisture content of cut tobacco leaves, steam flow, the opening of the leaf drying-anti-agglomeration valve 1, the oxygen content of the leaf leaves, the process air flow temperature, the cumulative amount of leaf drying, the process flow and the moisture content of the discharge in real time, and the above eight data at each moment are combined into a group of original parameters, and multiple groups of original parameters are collected; (II) Data preprocessing The above multiple groups of original parameters are processed as follows (1) Fill in missing parameters When a parameter at a certain moment is missing, the missing parameter is filled with the average of the parameters of the first five moments and the parameters of the last five moments; (2) Processing outliers on original parameters Calculate the average value of each parameter of each group of original data, detect the above parameters, and when the above parameters are within the range of ±3 times the standard deviation of their average values, they are regarded as constant values ​​and retained; otherwise, they are regarded as abnormal values ​​and the group of original parameters including the abnormal values ​​are deleted; (3) Standardize the data processed in step (2) When the above parameters are within their range, they are processed according to the following formula to normalize the above parameters to between 0 and ±1. Where: Z is the standardized value (Z-Score); X is the original data; μ is the mean of each parameter within its range; σ is the standard deviation of each parameter within its range; The first seven parameters at the same time are used as the input data of the model, and the moisture content of the discharge material is used as the output data of the model; (III) Training the stacked ensemble learning model S-LSTM The integrated learning model S-LSTM is composed of a base classifier consisting of six machine learning algorithms (CatBoost, Decision Tree, RandomForest, LightGBM, ExtraTrees and KNN) and a meta-classifier consisting of LSTM integrated with attention mechanism. 80% of the above-mentioned corresponding input data and output data constitute a training set, 10% of the above-mentioned corresponding input data and output data constitute a validation set, and 10% of the above-mentioned corresponding input data and output data constitute a test set. The input data of the training set is used as the input of the six sub-modules of the base classification module, and the original input data and the output data of the six sub-modules are used as the input of the meta-classification module. The output of the meta-classification module is the final prediction value. During the training process, the above operations are performed on all the corresponding input data and output data in the training set in sequence to complete a complete round of iterations. In each round of training of the training set, the data of the training set is first input into the S-LSTM model in batches, and the model output is calculated by forward propagation. Then, the model parameters are updated by back propagation, and the loss value of the training set is recorded at the same time. After completing the traversal of all training set data, enter the verification phase, input all the data of the verification set into the S-LSTM model, calculate the loss value of the verification set, and introduce a learning rate scheduler to optimize the training process. The learning rate is dynamically adjusted according to the change of the loss value of the verification set. If the loss value of the verification set in this round is lower than the lowest value of the loss value of the verification set in the previous round, the current S-LSTM model is saved as the best candidate model. The above process is repeated for at least 50 rounds to fully train the S-LSTM model. During the whole process, the parameters of the S-LSTM model are gradually optimized. The learning rate scheduler ensures that the weight parameters are adjusted more finely in the later stage of training, thereby improving the performance of the S-LSTM model. Finally, the S-LSTM model with the lowest loss value of the verification set is selected as the final trained S-LSTM model. (IV) Use the FWOA optimization module to calculate the process air flow temperature and process flow rate to achieve the predetermined moisture content of the dried tofu. (a) Randomly generate 20 sets of process air flow temperature and process flow Within the range of process air flow temperature and process flow rate, 20 sets of process air flow temperature and process flow rate are randomly generated; (b) Use the above random process air flow temperature and process flow to replace the real-time monitoring data The 20 sets of process air flow temperatures and process flow rates randomly generated above are used to replace the cut leaf moisture content, steam flow rate, cut leaf drying-anti-agglomeration valve opening 1, cut leaf oxygen content, process air flow temperature, cut leaf drying cumulative amount and process flow rate detected in real time at this time, to form 20 new sets of input data, and the number of iterations t=1; (c) Call the stacked ensemble learning model S-LSTM to predict the moisture content of the output of the above 20 sets of input data, and compare it with the set moisture content of the output The above 20 groups of input data are used as the input of the stacked ensemble learning model S-LSTM to obtain 20 predicted discharge moisture contents. The above 20 predicted discharge moisture contents are compared with the set discharge moisture contents. If one of the 20 predicted discharge moisture contents is the same as the set discharge moisture content, then the process air flow temperature and process flow rate corresponding to the discharge moisture content are the process air flow temperature and process flow rate to be regulated at this time; if the 20 predicted discharge moisture contents are different from the set discharge moisture content, find the group of input data with the smallest absolute value between the discharge moisture content and the set discharge moisture content among the 20 groups of input data. And detect whether the number of iterations t reaches the predetermined number of iterations. If t reaches the predetermined number of iterations, then the above set of input data The process gas flow temperature and process flow rate in are the process gas flow temperature and process flow rate to be regulated at this time; otherwise, proceed to step (d); (d) Update the input data of step (b) to form 20 new input data sets Substituting t+1 for t, using the formula Calculate 20 new sets of input data, among which, are 20 sets of input data of generation t, It is the set of input data with the smallest absolute value between the moisture content of the discharge material and the set moisture content of the discharge material among the 20 sets of input data of generation t. is the contraction coefficient, Set to 0.5, is a random number in [0,2] It is the new 20 sets of input data after updating to generation t+1. is the distance between each set of data and the best data; 20 sets of input data are updated and returned to step (c) until the process air flow temperature and process flow rate to be regulated are calculated.

2. The method for controlling moisture content stability of dried tofu at an outlet based on deep learning according to claim 1, characterized in that: In the step (b), the 20 randomly generated groups of process air flow temperature, process flow rate, real-time detected moisture content of cut leaf shreds, steam flow rate, leaf shred drying-anti-agglomeration valve opening 1, leaf shred oxygen content, process air flow temperature, leaf shred drying cumulative amount and process flow rate parameters of process air flow temperature and process flow are all processed through step (3).

3. The method for controlling moisture content stability of dried tofu at outlet based on deep learning according to claim 2, characterized in that: In step (iii), it also includes: using a test set to test the trained stacked ensemble learning model S-LSTM.

4. The method for controlling moisture content stability of dried tofu at an outlet based on deep learning according to claim 3, characterized in that: The standard deviation Where: N is the total number of data points; x i is the data of the ith data point; μ is the mean of the population, and the calculation formula is:

5. The method for controlling moisture content stability of dried tofu at outlet based on deep learning according to claim 4, characterized in that: The set discharge moisture content is 14.

9.

6. The method for controlling moisture content stability of dried tofu at outlet based on deep learning according to claim 5, characterized in that: The t is 100.

7. The method for controlling moisture content stability of dried tofu outlet based on deep learning according to claim 6, characterized in that: The learning rate scheduler is ReduceLROnPlateau.

Images