Cut tobacco dryer cooling bed outlet moisture control method, storage medium and system

Abstract

The invention discloses a cut tobacco dryer cooling bed outlet moisture control method, a storage medium and a cut tobacco dryer cooling bed outlet moisture control system. The cut tobacco dryer cooling bed outlet moisture control method comprises the steps of preprocessing historical data, aligning feature data, filtering broken material batches, screening model feature parameters, and constructing data required by training; building a temperature prediction model network structure; performing data division on the preprocessed data, dividing the preprocessed data into a training set and a test set, and inputting the training set into a network for training; inputting the test set into the network verification model performance; dynamically predicting the temperature of the cylinder wall by using the trained model; a linear model is constructed according to historical data, and Pearson correlation analysis is carried out to obtain the linear weight and correlation coefficient of the cylinder wall temperature and the cooling bed outlet moisture. And performing feedback regulation on the predicted value of the cut tobacco dryer cylinder wall temperature prediction model in combination with the deviation between the current cooling bed outlet moisture and the standard value during production to obtain a final cylinder wall temperature recommended value. According to the invention, the control on the water at the outlet of the cooling bed is well realized, and the manual intervention is reduced.

Description

Technical Field

[0001] The invention relates to a moisture control method, storage medium and system for a cooling bed outlet of a tobacco drying machine, and belongs to the field of controlling the moisture content of tobacco products. Background Art

[0002] In the tobacco processing industry, moisture content at the cooling bed outlet of the cut tobacco dryer is a key indicator for evaluating cut tobacco quality. Precisely controlling moisture content at the cooling bed outlet is crucial for improving the overall quality, stability, and consistency of cigarette products. However, the cut tobacco drying process involves complex physical and chemical changes and is influenced by numerous process parameters and environmental factors, making it difficult for traditional control methods to achieve optimal results.

[0003] Data-driven modeling does not require knowledge of complex tobacco drying mechanisms. Instead, it relies primarily on industrial data and theoretical methods such as machine learning and deep learning to build models that link easily measurable process variables with difficult-to-detect variables. This approach can address the difficult problem of predicting key process or quality indicators, and has therefore attracted significant attention from scholars and technicians. In the tobacco drying process, ANGANG CHEN et al. proposed a two-layered model predictive control strategy (SSTO-MPC) for the non-square, nonlinear cut tobacco drying process (Chen, A., Ren, Z., Fan, Z., & Feng, X. (2020). Two-Layered Model Predictive Control Strategy of the Cut Tobacco Drying Process. IEEE Access, 8, 155697-155709.). This two-layered control structure incorporates a steady-state target optimization layer (SSTO) to optimize the drum temperature, hot air temperature, and tobacco drying outlet temperature to meet process constraints and achieve tobacco moisture content requirements. Zhiping Fan et al. established a mathematical model for the tobacco drying process and proposed a model-based multi-objective predictive control (MOMPC) algorithm (Fan, Z., Ren, Z. & Chen, A. Multi-objective predictive control based on the cutting tobacco outlet moisture priority. Sci Rep 13, 199(2023). https: / / doi.org / 10.1038 / s41598-022-26694-x). Compared with traditional predictive control, this strategy showed significant advantages in optimizing control output variables and improving product quality and yield.

[0004] Although these models have achieved positive research results, the following problems still exist: on the one hand, it is difficult to effectively mine the deep structural information hidden in tobacco industry data through mathematical algorithms alone, resulting in limited model generalization ability; on the other hand, due to factors such as differences in raw material quality grades and unstable steam pressure, the tobacco drying process has significant non-steady-state characteristics, and traditional static network structures cannot describe the dynamic changes in the tobacco drying process.

[0005] Time series-based forecasting is the process of predicting future changes based on existing time series data. Time series data is formed under the influence of various factors, so forecasting using historical time series data indirectly takes into account various influencing factors.

[0006] Traditional time series forecasting models, such as the autoregressive moving average (ARMA) and ARIMA models, can effectively capture temporal correlations in time series under strict assumptions and constraints. However, they struggle to capture nonlinear relationships and effectively handle the nonstationarity of sequence data. Data in real-world applications often exhibit complex nonlinear and nonstationary relationships. In recent years, deep learning technology has made significant progress in the field of time series forecasting. Deep learning models, particularly recurrent neural networks (RNNs) and their variants (such as long short-term memory (LSTM) networks and gated recurrent units (GRUs), can effectively handle long-term dependencies in time series data and possess powerful nonlinear modeling capabilities. These characteristics give deep learning models great potential for the prediction and control of complex systems.

[0007] LSTM (Long Short-Term Memory) is a special type of RNN (Recurrent Neural Network) that has superior long-term memory compared to standard RNNs, effectively solving the vanishing gradient problem. The LSTM network structure is similar to that of a standard RNN, with the main difference being the internal structure of the neurons. While a standard RNN neuron consists of only a single tanh function, the LSTM introduces three gated activation functions to capture longer-term dependencies: the input gate, the output gate, and the forget gate.

[0008] A CNN (convolutional neural network) is a type of neural network composed of an input layer, hidden layers, and an output layer. It incorporates convolutional computations and is a feedforward neural network. The network structure is shown in the figure. Convolutional neural networks can exploit local correlations in features, and their convolutional structure can reduce the memory usage of deep networks. Convolutional neural networks feature weight sharing and local perception, while downsampling layers effectively reduce the number of network parameters and mitigate overfitting. CNNs can effectively identify simple patterns in data, which are then used in higher-level networks to form more complex patterns. 1D-CNNs can capture interesting features in sequence segments and identify local patterns that contribute to the target. These key local patterns are fed into an LSTM, capturing their temporal dependencies while reducing the computational cost of the network model.

[0009] The attention mechanism, also known as the Attention mechanism, was inspired by human attention. In sequence prediction models, the Attention mechanism is often used to connect the encoder and decoder, allowing the model to focus on key information. Seq2Seq models can be used for many tasks, such as machine translation and speech recognition. In machine translation, the encoder corresponds to the source language text sequence, while the decoder corresponds to the target language text sequence or text summary.

[0010] The encoder-decoder model with an attention mechanism introduces a structural mechanism to the basic encoder-decoder model. This mechanism calculates the contribution of the encoder's output to the decoder and then calculates the weighted average, known as the context. By introducing contextual information between the encoder's output and the next step's input, the decoder and encoder's outputs can be correlated. Summary of the Invention

[0011] The present invention aims to solve the problems of moisture fluctuation at the cooling bed outlet and a relatively high dry head and dry tail rate caused by manual intervention in the wire drying process. The overall concept is to propose a wire drying machine barrel wall temperature prediction model based on a long short-term memory neural network and a one-dimensional convolutional neural network combined with an attention mechanism. During the production process, the predicted value of the barrel wall temperature prediction model is fed back and adjusted based on the moisture at the cooling bed outlet at that time to obtain a recommended barrel wall temperature value, thereby reducing manual intervention and controlling the moisture at the cooling bed outlet within the process standard range.

[0012] The technical solution adopted in the present invention is:

[0013] A method for controlling moisture at a cooling bed outlet of a tofu drying machine comprises the following steps:

[0014] 1) Preprocess the historical production data of the tofu drying machine, align the feature data, filter the broken batches, select the model feature parameters, and construct the data required for training;

[0015] 2) Build a network structure for the prediction model of the barrel wall temperature of the tofu drying machine;

[0016] 3) Divide the preprocessed data into a training set and a test set, and input the training set into the network for training;

[0017] 4) After model training is completed, the test set is input into the network to verify the model performance;

[0018] 5) Using the trained prediction model for the temperature of the tofu drying machine barrel wall to dynamically predict the temperature of the tofu drying machine barrel wall;

[0019] 6) A linear model was constructed based on historical data, and Pearson correlation analysis was performed to obtain the linear weight and correlation coefficient between the cylinder wall temperature and the moisture at the cooling bed outlet.

[0020] 7) During production, the predicted value of the tofu drying machine barrel wall temperature prediction model is feedback-adjusted based on the deviation between the moisture at the cooling bed outlet and the standard value to obtain the final recommended value of the barrel wall temperature.

[0021] A computer-readable storage medium stores a computer program, which can be executed by a processor to implement the steps of a method for controlling moisture at a cooling bed outlet of a tofu drying machine according to the present invention.

[0022] A moisture control system for a cooling bed outlet of a tofu drying machine comprises a computer and a computer-readable storage medium according to the present invention.

[0023] The beneficial effects of the present invention are:

[0024] (1) The tofu drying machine barrel wall temperature prediction model of the present invention can accurately predict the barrel wall temperature during the production process, reducing manual intervention;

[0025] (2) The feedback adjustment method of the present invention combining the moisture at the cooling bed outlet with the predicted drum wall temperature can ensure that the moisture at the cooling bed outlet is within the standard range. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is a structural schematic diagram of the tofu drying machine barrel wall temperature prediction model in the present invention.

[0027] Figure 2 This is a flow chart of the training of the tofu drying machine barrel wall temperature prediction model in the present invention.

[0028] Figure 3 The present invention is a flow chart of a method for controlling moisture at a cooling bed outlet of a tofu drying machine. DETAILED DESCRIPTION

[0029] See also Figure 1-Figure 3As shown, a kind of based on deep learning time series prediction moisture control method of drying machine cooling bed outlet, including the following steps:

[0030] step1, since in actual production process may appear material breakage phenomenon, therefore need to clean up historical data, eliminate material breakage batch.

[0031] step2, historical data are aligned, and the alignment mode one is aligned according to timestamp, and data set data1 is obtained;Mode two: since there is electric control time, conveying belt conveying time and the like in the production process, process alignment is carried out for these times, and data set data2 is obtained;Finally, according to the evaluation index of prediction model, the final alignment mode used is selected.

[0032] step3, normalization and data division, the preprocessed data is divided into training data set and test data set.

[0033] The normalization calculation formula is as follows:

[0034]

[0035] Wherein x is the original characteristic value, min x is the minimum value of the characteristic in the entire data set, max x is the maximum value of the characteristic in the entire data set, and x' is the normalized characteristic value.

[0036] step4, screening feature parameters. The final feature parameters are screened out by using Pearson correlation analysis and XGBoost limit gradient boosting algorithm comprehensive analysis.

[0037] According to the drying machine cylinder wall temperature prediction model, the screened feature data points include: DB3253 process flow, drying DB3253 cumulative amount, cut tobacco moisture, SX3256 steam pressure, SX3256 steam valve opening, SX3256 expanded steam flow (mass), SX3256 expanded steam flow (volume), TT3259 roller speed, TT3259 hot air temperature, TT3259 hot air speed actual value, TT3259 circulating air valve opening, TT3259 circulating air steam valve opening, TT3259 cover pressure actual value, dehydration, TT3259 drying outlet temperature, drying outlet moisture, cooling bed outlet moisture, moisture removal air pipe temperature, moisture removal air pipe relative humidity, SIROX_steam dryness, SIROX_steam temperature, drying machine equipment state, drying machine moisture removal valve opening, drying machine steam pressure.

[0038] Step 5. Construct the feature data set according to the filtered features. The feature data set structure is: featureset = (m, d, n); the target data set structure is targetset = (m, 2): where m is the number of samples, d is the time window size, and n is the number of feature parameters. The predicted parameters in the target data set are the cylinder wall temperatures in zones 1 and 2, so the number is 2.

[0039] Step 6: Construct the neural network architecture for the tofu drying machine barrel wall temperature prediction model. First, a two-layer 1D-CNN is built. CNNs can effectively identify simple patterns in data, which can then be used in higher-level networks to form more complex patterns. 1D-CNNs can capture interesting features in sequence segments and identify local patterns that contribute to the target. These key local patterns are fed into an LSTM to capture their temporal dependencies while reducing the computational cost of the network model. A multi-layer LSTM is then used to capture the complex temporal dependencies of the tofu drying sequence. An attention mechanism is employed in the second LSTM layer to calculate the weights for each time step, allowing the model to focus more on the time steps that contribute to the prediction target and effectively capturing the dependency of the target output on each time step.

[0040] The prediction model of the wall temperature of the tofu drying machine can be defined as: Assume that X={x1,x2,…,x d}∈R d It represents the time series of the barrel wall temperature of the tofu drying machine at the past d moments, where d is the length of the time window, and predicts the set values ​​of the barrel wall temperature of zone 1 and zone 2 at the next moment.

[0041] The cylinder wall temperature prediction model can be expressed as:

[0042] y(t)=F(X(t))

[0043]

[0044] Where y(t) represents the cylinder wall temperature at time t, X(t) represents the time window constructed by the past d groups of feature sequences at time t, d is the length of the time window, x(t) represents the feature data at time t, and n is the number of feature parameters.

[0045] 1D-CNN calculation process:

[0046] In a 1D-CNN, the convolution kernel is a one-dimensional sliding window that performs a convolution operation on the input data. At each location, the kernel performs a dot product with a portion of the input data, a bias is added, and an activation function (such as Reluctant Unit) is applied to produce the output. This process generates a new feature map containing information about the local features of the input data. The kernel slides along the input data, moving one step at a time. The step size determines the speed at which the kernel slides across the input data; a larger step size results in a smaller output feature map. A 1D-CNN typically consists of multiple convolutional layers, each of which generates one or more feature maps. These feature maps serve as input to the next convolutional layer. Multiple layers of convolution progressively extract higher-level features from the input data. Pooling layers are added between convolutional layers to reduce the dimensionality of the feature maps and the computational overhead. Pooling reduces dimensionality by selecting the maximum or average value in the feature map. After multiple layers of convolution and pooling, the feature maps are flattened and passed to the LSTM.

[0047] For the jth element in the i-th feature map in the one-dimensional convolutional layer, the calculation formula can be expressed as:

[0048]

[0049] Where: y i [j] is the output value of position j in the i-th feature map. f is the activation function (such as ReLU, Sigmoid, etc.). x[j+m] is the value of position j+m in the input sequence. w i [m] is the weight value of position m in the i-th convolution kernel (filter). K is the size (i.e. width) of the convolution kernel. i is the bias term for the i-th feature map.

[0050] LSTM calculation process:

[0051] The core of the long short-term memory network is controlled based on the unit state c. It introduces three gate structures: forget gate, input gate, and output gate to store and control information. These three gate structures are mainly completed through a sigmoid neural layer and a point-by-point multiplication operation.

[0052] The forget gate can decide which information to continue to pass through the neuron, which is determined by the s at the previous moment. t-1 with c t-1 Status, determine whether the information inside can be retained to the current status c t Among them. The input gate is based on the output s of the previous moment t-1 and the current input x t , after the sigmoid function transformation, the internal output f at the current moment is obtained t , the formula is as follows:

[0053] f t =σ(W f [s t-1 ,x t ]+b f )

[0054] Among them, W f Represents the weight matrix of the forget gate, [s t-1 ,x t ] is the vertical connection operation of two vectors, b f is the bias term of the input. The input gate can determine the current input x t How much information can be saved to the current state c t The update formula is as follows:

[0055] i t =σ(W i [s t-1 ,x t ]+b i )

[0056]

[0057] Where "⊙" indicates element-by-element multiplication. t and , different weight matrices and bias terms need to be used for separate training.

[0058] After completing the selective memory and update of information, the output gate will determine the control unit state c t How much information can be output by the current output value s of the long short-term memory network t In the formula, the following is:

[0059] o t =σ(W o [s t-1 ,x t ]+b o )

[0060] s t =o t ⊙tanh(c t )

[0061] In the above formula, b o is the bias term of the output gate. σ is the sigmoid function. The formulas for tanh and σ are as follows:

[0062]

[0063] The calculation process of introducing the attention mechanism:

[0064] First, the output from the LSTM is resized so that attention can be applied to each time step (as a column in the new dimension) later. Next, a fully connected layer with a softmax activation function is used on the resized tensor. This step is the core of the attention mechanism, which generates a weight value for each time step. This means that a time-step weight vector is calculated for each input feature (this time as a row, corresponding to all possible input features at each time step), so that the sum of the weights of all time steps is 1, achieving weighted treatment of different time steps. Finally, the dimensional order of the attention weights is adjusted to match the dimensionality of the original input data, and the original input tensor is multiplied element-by-element by the calculated attention weight tensor. This is actually applying different weights to each time step and each feature dimension in the input sequence, emphasizing the parts that the model considers important and weakening the unimportant parts. The attention mechanism calculation process is as follows:

[0065] For the i-th element in the input sequence, calculate its energy value e i This energy value represents the degree of attention of the current decoding position to the i-th position in the input sequence. The formula is as follows:

[0066]

[0067] in, is a weight vector used to weight the attention weight, W1 and W2 are two weight matrices used to weight the hidden state h of the input sequence i and the current state s of the decoder are linearly transformed.

[0068] The energy value is converted into attention weight α through the Softmax function i , to ensure that the sum of the weights of all positions is 1, the formula is as follows:

[0069]

[0070] Where exp is the natural exponential function and N is the length of the input sequence.

[0071] The context vector s is obtained by weighting the hidden states of the input sequence using the attention weights. This vector contains the most relevant information about the current decoding position in the input sequence. The formula is as follows:

[0072]

[0073] Among them, h i is the hidden state at position i in the input sequence.

[0074] Step 7: Input the training data set into the network for training and verify it on the test set. According to the root mean square error (RMSE), mean absolute error (MAE), coefficient of determination (R 2 ) and other evaluation indicators to comprehensively judge the accuracy of model prediction. The smaller the RMSE and MAE indicators are, the better the prediction effect is. 2 The closer it is to 1, the better the prediction effect.

[0075] The RMSE calculation formula is as follows:

[0076]

[0077] The MAE calculation formula is as follows:

[0078]

[0079] R 2 The calculation formula is as follows:

[0080]

[0081] In the above formula, n represents the sample size of the test set, y i represents the true value of the cylinder wall temperature in the test set, It represents the predicted value of the cylinder wall temperature obtained by the cylinder wall temperature prediction model when predicting the test set data. It represents the average value of the cylinder wall temperature in the test set.

[0082] Step 8. Adjust hyperparameters such as sliding window size, number of neurons, regularization coefficient, etc. to obtain the optimal hyperparameters.

[0083] Step 9. Build a linear model based on the preprocessed data, and perform Pearson correlation analysis to obtain the linear weights and correlation coefficients of the cylinder wall temperature and the moisture at the cooling bed outlet. Multiply the linear weights and the correlation coefficients to obtain the subsequent feedback adjustment proportional control coefficient w.

[0084] Step 10: Construct a feedback adjustment algorithm. After the prediction model of the tofu drying machine wall temperature gives the predicted value p, the predicted value needs to be feedback-adjusted based on the moisture m at the cooling bed outlet and the set process standard s to obtain the final recommended value of the wall temperature res. The feedback adjustment control algorithm is as follows:

[0085] .

Claims

1. A method for controlling moisture at the outlet of a cooling bed of a tofu drying machine, characterized in that: The steps include: Step 1: Preprocess the historical production data of the tofu drying machine, align the feature data, filter the broken batches, select the model feature parameters, and construct the data required for training; Step 2: Build a network structure for predicting the temperature of the tofu drying machine barrel wall, including: (1) Build a two-layer 1D-CNN; (2) It is then used in higher-level networks to form more complex patterns; (3) A multi-layer LSTM is used to obtain the complex temporal correlation of the tofu drying sequence. At the same time, an attention mechanism is used on the second layer of LSTM to calculate the weight of each time step, so that the model pays more attention to the time step that plays a role in the prediction target, and is used to obtain the dependency information of the target output on each time step; (4) The prediction model of the drying machine barrel wall temperature is y(t) = F(X(t)) Where: y(t) represents the cylinder wall temperature at time t, X(t) represents the time window constructed by the past d groups of feature sequences at time t, d is the length of the time window, x(t) represents the feature data at time t, and n is the number of feature parameters; Step 3: Divide the preprocessed data into a training set and a test set, and input the training set into the network for training; Step 4: After model training is completed, the test set is input into the network to verify the model performance; Step 5, using the trained tow-cutter drying machine barrel wall temperature prediction model to dynamically predict the tow-cutter drying machine barrel wall temperature; Step 6: Build a linear model based on historical data and perform Pearson correlation analysis to obtain the linear weight and correlation coefficient between the drum wall temperature and the moisture at the cooling bed outlet; Step 7: During production, the predicted value of the tofu drying machine barrel wall temperature prediction model is feedback-adjusted based on the deviation between the moisture at the cooling bed outlet and the standard value to obtain the final recommended value of the barrel wall temperature.

2. The method according to claim 1, characterized in that The 1D-CNN calculation process includes: The 1D-CNN consists of multiple convolutional layers, each of which generates one or more feature maps. These feature maps serve as input to the next convolutional layer. Multiple layers of convolution are used to gradually extract higher-level features from the input data. Pooling layers are added between convolutional layers to reduce the dimensionality and computational complexity of the feature maps. Pooling achieves dimensionality reduction by selecting the maximum or average value in the feature maps. After multiple layers of convolution and pooling, the feature maps are flattened and passed to the LSTM. For the jth element in the i-th feature map in the one-dimensional convolutional layer, the calculation formula can be expressed as: Where: y i [j] is the output value of position j in the i-th feature map, f is the activation function, x[j+m] is the value of position j+m in the input sequence, w i [m] is the weight value of position m in the i-th convolution kernel, K is the width of the convolution kernel, b i is the bias term for the i-th feature map.

3. The method according to claim 2, characterized in that The LSTM calculation process includes: LSTM is controlled based on the cell state c and introduces three gate structures: forget gate, input gate, and output gate to store and control information. These three gate structures are implemented through a sigmoid neural layer and a point-by-point multiplication operation. The forget gate determines which information continues to pass through the neuron, which is determined by the s at the previous moment. t-1 with c t-1 Status, determine whether the information inside can be retained to the current status c t Among them; the input gate is based on the output s of the previous moment t-1 and the current input x t , after the sigmoid function transformation, the internal output f at the current moment is obtained t , the formula is as follows: f t =σ(W f [s t-1 ,x t ]+b f ) Among them, W f Represents the weight matrix of the forget gate, [s t-1 ,x t ] is the vertical connection operation of two vectors, b f is the input bias term; The input gate determines the current input x t How much information can be saved to the current state c t The update formula is as follows: I t =σ(W i [s t-1 ,x t ]+b i ) Among them: "⊙" means element-by-element multiplication. t and When , different weight matrices and bias items need to be used for separate training; After completing the selective memory and update of information, the current output value s of the output gate t for: the t =σ(W o [s t-1 ,x t ]+b o ) s t =o t ⊙tanh(c t ) Where: b o is the bias term of the output gate, σ is the sigmoid function, and the tanh and σ formulas are as follows:

4. The method according to claim 2, characterized in that The attention mechanism calculation process includes: (1) Adjust the dimension of the output obtained from LSTM; (2) Apply a fully connected layer with a softmax activation function to the resized tensor. (3) Adjust the dimensional order of the attention weights to make them consistent with the dimensions of the original input data, and multiply the original input tensor by the calculated attention weight tensor element by element. The calculation process is as follows: For the i-th element in the input sequence, calculate its energy value e i , this energy value represents the degree of attention of the current decoding position to the i-th position in the input sequence. The formula is as follows: in, is a weight vector used to weight the attention weight, W1 and W2 are two weight matrices used to weight the hidden state h of the input sequence i Perform a linear transformation on the current state s of the decoder; The energy value is converted into attention weight α through the Softmax function i , to ensure that the sum of the weights of all positions is 1, the formula is as follows: Where exp is the natural exponential function and N is the length of the input sequence; The context vector s is obtained by weighted summing the hidden states of the input sequence using the attention weights: Among them, h i is the hidden state at position i in the input sequence.

5. The method according to claim 1, wherein The model feature parameters screened in step 1 include: DB3253 process flow, DB3253 cumulative amount of wire drying, wire cutting moisture, SX3256 steam pressure, SX3256 steam valve opening, SX3256 expansion steam mass flow, SX3256 expansion steam volume flow, TT3259 drum speed, TT3259 hot air temperature, TT3259 hot air speed actual value, TT3259 circulating air valve opening, TT3259 circulating air steam valve opening, TT3259 hood pressure actual value, dehydration, TT3259 wire drying outlet temperature, wire drying outlet moisture, cooling bed outlet moisture, dehumidification air duct temperature, dehumidification air duct relative humidity, SIROX_steam dryness, SIROX_steam temperature, wire drying machine equipment status, wire drying machine dehumidification valve opening and wire drying machine steam pressure.

6. The method according to claim 1, characterized in that Step 1 also includes: Align the historical data. Alignment method 1: align based on timestamps to obtain data set data1; alignment method 2: for the electronic control time and conveyor belt transmission time in the production process, process alignment is performed on these times to obtain data set data2; finally, the final alignment method to be used is selected based on the prediction model evaluation indicators. The feature data set is constructed according to the screened features. The feature data set structure is: featureset = (m, d, n); the target data set structure is targetset = (m, 2), where: m is the number of samples, d is the time window size, and n is the number of feature parameters. The predicted parameters in the target data set are the cylinder wall temperatures in zones 1 and 2.

7. The method according to claim 1, characterized in that Step 3 also includes: According to the root mean square error RMSE, mean absolute error MAE, determination coefficient R 2 Comprehensively evaluate the accuracy of model prediction; the smaller the RMSE and MAE indicators are, the better the prediction effect is. 2 The closer it is to 1, the better the prediction effect; The RMSE is: The MAE is: The R 2 for: In the above formula, n represents the sample size of the test set, y i represents the true value of the cylinder wall temperature in the test set, It represents the predicted value of the cylinder wall temperature obtained by the cylinder wall temperature prediction model when predicting the test set data. It represents the average value of the cylinder wall temperature in the test set.

8. The method according to any one of claims 1 to 7, characterized in that Step 7 also includes: After the prediction model of the tofu drying machine barrel wall temperature gives the predicted value p, the predicted value needs to be feedback-adjusted based on the moisture m at the cooling bed outlet and the set process standard s to obtain the final recommended value of the barrel wall temperature res. The feedback adjustment control algorithm is as follows:

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: The computer program can be executed by a processor to implement the steps of a method for controlling moisture at the outlet of a cooling bed of a tow dryer as described in any one of claims 1 to 8.

10. A moisture control system for a cooling bed outlet of a tofu drying machine, the system comprising a computer and a computer-readable storage medium according to claim 9.

Images