A method for predicting tobacco moisture content by fusing feature screening and dynamic optimization of BP neural network

Abstract

The application discloses a kind of fusion feature screening and dynamic optimization's BP neural network tobacco moisture prediction method, it is related to cigarette production intelligent control technical field.The method includes: collecting the multiple-source production data related to moisture regaining process in silk production line;Using Relief algorithm carries out feature screening, selects the key feature variable that influences moisture regaining export moisture;The normalized processing is carried out to the feature data screened out, and is divided into training set, test set and verification set;BP neural network prediction model with specific structure is constructed and trained, the model adopts the structure of linear layer and activation function alternation, and introduces one-dimensional convolution layer to reduce parameter quantity, uses Mini-batch gradient descent method and cosine annealing learning rate scheduling strategy;Using the model trained, according to real-time production data, the moisture value of moisture regaining export is predicted, and according to this, water addition control is guided.The application can realize the accurate prediction of tobacco moisture, effectively improve the stability of moisture regaining export moisture, to improve product quality.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology in cigarette production, and in particular to a BP neural network method for predicting tobacco leaf moisture by integrating feature screening and dynamic optimization. Background Technology

[0002] The tobacco processing process is the core of cigarette production, and its quality directly determines the cigarette quality. The loosening and rehumidifying machine, as a key piece of equipment in the tobacco processing line, is responsible for increasing the temperature and humidity of the rehumidified tobacco leaves to improve their processing resistance, achieve a uniform and loose state, and enhance sensory quality. In actual production, the moisture content of the rehumidified tobacco leaves fluctuates significantly due to variations in process parameters, equipment conditions, and the characteristics of the incoming tobacco leaves (grade, year, origin, part of the plant, etc.), severely impacting the quality of subsequent processing steps and leading to inconsistent product quality. Therefore, achieving accurate prediction and control of the moisture content at the rehumidification outlet is crucial for improving the tobacco processing technology and ensuring stable product quality; developing reliable moisture prediction methods is key.

[0003] To address the issue of predicting moisture content at the outlet of loose rehydration processes, various algorithm-based approaches have emerged in existing technologies. These intelligent algorithms enhance the accuracy of moisture prediction, providing support for water addition control. For example, Chinese invention patent application CN117131987A discloses a real-time water addition prediction method for the loose rehydration process based on a deep Q-network, which uses a deep reinforcement learning algorithm to predict the correlation between water addition and outlet moisture content. Other related technologies propose a BP neural network-based model for predicting tobacco leaf rehydration moisture addition, utilizing various process data recorded during production to construct a neural network model to improve the accuracy of moisture addition prediction. Furthermore, some methods employ linear regression or correlation analysis to achieve simple moisture prediction by exploring the linear correlation between process parameters and outlet moisture content.

[0004] However, the above methods still have significant shortcomings: deep neural networks are prone to gradient explosion or vanishing, are difficult to train, and require a large amount of data; standard backpropagation (BP) neural networks are prone to overfitting or underfitting under high-dimensional, noisy industrial data, have slow convergence with a fixed learning rate, are prone to getting trapped in local optima, lack effective feature selection mechanisms, have poor model interpretability, and are difficult to implement; linear methods cannot capture complex nonlinear relationships and have insufficient prediction accuracy. Moreover, most existing methods do not consider the practical industrial needs such as lightweight model deployment, dynamic adaptation to process drift, and closed-loop continuous updates. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides a BP neural network method for predicting tobacco leaf moisture content that integrates feature screening and dynamic optimization. Without altering the existing tobacco processing line structure or adding additional measuring points, this method uses intelligent algorithms to accurately predict the moisture content at the re-moistening outlet, thereby providing a basis for precise control of water addition and achieving the goals of stabilizing product quality and reducing consumption.

[0006] This invention is implemented as follows:

[0007] A BP neural network method for predicting tobacco leaf moisture by integrating feature selection and dynamic optimization includes the following steps:

[0008] Step S1, Data Acquisition and Preprocessing:

[0009] Collect historical production data related to the rehumidification process, including ambient temperature of the rehumidification section, ambient humidity of the drying section, moisture content before entering the cabinet after feeding, moisture content at the HT (drying) inlet, ambient temperature of the drying section, and ambient humidity of the rehumidification section.

[0010] Step S2, Feature Filtering:

[0011] The Relief feature weighting algorithm is used to calculate the correlation weights between each feature variable in the original dataset and the target value of moisture content at the reabsorption outlet.

[0012] The features are sorted according to their weights, and the top n features with the largest weights are selected as the input variables of the BP neural network model.

[0013] Step S3, Data Normalization and Partitioning:

[0014] The selected key feature data is normalized, and the normalized dataset is divided into training set, test set and validation set according to a predetermined ratio.

[0015] Step S4, BP neural network model construction:

[0016] Construct a BP neural network model, including an input layer, multiple hidden layers, and an output layer;

[0017] The hidden layer incorporates a one-dimensional convolution operation to extract features and reduce the number of model parameters.

[0018] Alternate ReLU activation functions between linear layers and introduce Dropout layers to prevent overfitting;

[0019] Following the first linear layer, a data rearrangement layer and a one-dimensional convolutional layer are sequentially arranged. The data rearrangement layer is used to reorganize the feature vector output by the linear layer into a multi-dimensional matrix format with channel and sequence dimensions to adapt to the input requirements of the one-dimensional convolutional layer. The one-dimensional convolutional layer is used to extract local correlation information of the input features. The output of the one-dimensional convolutional layer is sequentially connected to a ReLU activation function layer and a Dropout layer.

[0020] Step S5, Model Training:

[0021] The model is trained by: updating the model weights using the Mini-batch gradient descent method; adjusting the learning rate using the cosine annealing algorithm; and using the mean squared error (MSE) as the loss function during training.

[0022] Furthermore, in step S2, the difference value of the feature in the Relief algorithm is calculated according to the difference rule based on the variable type. The calculation logic for the difference value between nominal variables and numerical variables is as follows: when the nominal variables have the same value, the difference value is 0, and when the values ​​are different, the difference value is 1. The difference value of the numerical variable is calculated by combining the absolute value of the difference with the normalization factor. The normalization factor is the difference between the maximum and minimum values ​​of the feature in the sample set (i.e., the range), ensuring that the result is normalized to the [0,1] interval.

[0023] Furthermore, in step S3, the normalization process employs the Min-Max standardization method, and the calculation formula is as follows:

[0024]

[0025] Where x is the original feature value, and min(x) and max(x) are the minimum and maximum values ​​of the feature in all samples, respectively. This is the normalized value.

[0026] Furthermore, in step S4, a data rearrangement layer is set between the first linear layer and the one-dimensional convolutional layer to reorganize the feature vector output by the linear layer into a multi-dimensional matrix format with channel dimension and sequence dimension, so as to adapt to the input requirements of the one-dimensional convolutional layer.

[0027] Furthermore, in step S4, the kernel size of the one-dimensional convolutional layer is set to 3, and the stride is set to 1, which matches the sequence length constructed by the data rearrangement layer.

[0028] Furthermore, in step S4, the one-dimensional convolutional layer uses a boundary padding method to ensure that the sequence length remains consistent before and after convolution, so that local correlation information between adjacent features can still be effectively extracted even when the sequence length is short.

[0029] Furthermore, in step S5, the cosine annealing algorithm adjusts the learning rate, and the calculation formula is as follows:

[0030]

[0031] in, The learning rate for the current training round. and Minimum and maximum learning rates This is the current training round number. This represents the total number of training rounds.

[0032] Furthermore, in step S5, the mean squared error (MSE) is used as the loss function during the training process, and the calculation formula is as follows:

[0033]

[0034] in, These are the model's predicted values. is the true value, and N is the number of samples.

[0035] Furthermore, the method also includes evaluating the model performance after the model training is completed, using the root mean square error (RMSE) between the predicted value and the true value as the evaluation index; if the RMSE on the validation set is less than a preset threshold, the model training is deemed qualified; and further testing the model's generalization ability on an independent test set to ensure that the model still has high-precision prediction performance on unseen new data.

[0036] Furthermore, the method also includes collecting new production data monthly, mixing it with historical data at a preset ratio to construct a training dataset, and incrementally updating and training the deployed BP neural network model to ensure that the model continuously maintains optimal prediction performance.

[0037] The beneficial effects of this invention are as follows: By combining Relief feature selection, optimizing the BP neural network and training strategy, this invention solves the problems of insufficient accuracy, difficult deployment, and poor robustness of existing methods, improves the accuracy and practicality of moisture prediction at the reabsorption outlet, and provides support for water addition control. The specific beneficial effects are as follows:

[0038] 1. By using the Relief algorithm to filter key features, the dimensionality of the model input is reduced, redundant feature interference is decreased, and the model training efficiency and prediction accuracy are improved.

[0039] 2. An unconventional BP neural network structure is proposed, placing the linear layer (fully connected layer) before the one-dimensional convolutional layer. Through rearrangement operations, a channel-sequence dimension is constructed for discrete features. This breaks with the conventional approach of convolution before fully connected layers, extending the application of one-dimensional convolution from traditional time-series signal processing to discrete process parameter modeling. A weight-sharing mechanism achieves model lightweighting while controlling the number of parameters, facilitating deployment and operation in industrial settings. The local connectivity characteristics of convolution operations effectively extract local correlation information between discrete features. Multiple mechanisms, including ReLU nonlinear transformation and Dropout random discarding, effectively suppress overfitting and improve the model's generalization ability.

[0040] 3. A training method combining dynamic learning rate scheduling and mini-batch gradient descent is adopted to improve the model's convergence stability, avoid getting trapped in local optima, and enhance the model's robustness.

[0041] 4. Construct a closed-loop prediction and update mechanism to enable the model to adapt to process drift and ensure prediction stability during long-term use.

[0042] 5. It can accurately predict the moisture content at the re-moistening outlet without changing the existing production line structure or adding additional measuring points, providing a basis for precise control of water addition, stabilizing product quality and reducing production consumption.

[0043] The present invention will be explained in detail below with reference to the accompanying drawings and specific embodiments. Attached Figure Description

[0044] Figure 1 This is a flowchart of the BP neural network method for predicting tobacco leaf moisture by integrating feature screening and dynamic optimization, as described in this invention.

[0045] Figure 2 This is the propagation diagram of the BP neural network in Example 1. Detailed Implementation

[0046] Example 1:

[0047] This embodiment presents a BP neural network method for predicting tobacco leaf moisture by integrating feature selection and dynamic optimization. Figure 1 As shown, it includes the following steps:

[0048] A BP neural network method for predicting tobacco leaf moisture by integrating feature selection and dynamic optimization includes the following steps:

[0049] Step S1, Data Acquisition and Preprocessing:

[0050] Historical production data related to the rehumidification process were collected, including ambient temperature of the rehumidification section, ambient humidity of the drying section, moisture content before entering the cabinet after feeding, moisture content at the HT (drying) inlet, ambient temperature of the drying section, and ambient humidity of the rehumidification section, to form the original dataset.

[0051] Step S2, Feature Filtering:

[0052] The Relief feature weighting algorithm is used to calculate the correlation weight between each feature variable in the original dataset and the target value of moisture content at the outlet.

[0053] In the Relief algorithm, the difference values ​​of features are calculated using a differentiation rule based on variable type. The logic for calculating the difference values ​​between nominal and numerical variables is as follows: the difference value is 0 when the nominal variables have the same value, and the difference value is 1 when the values ​​are different. For numerical variables, the difference value is calculated by combining the absolute value of the difference with a normalization factor. The normalization factor is the difference between the maximum and minimum values ​​of the feature in the sample set (i.e., the range), ensuring that the result is normalized to the [0,1] interval, as detailed below:

[0054] When the feature is a nominal variable:

[0055]

[0056] When the feature is a numerical variable:

[0057]

[0058] in, As the normalization factor (usually the eigenrangement), The values ​​are normalized to the interval [0,1].

[0059] The features are sorted according to their weights, and the top n features with the largest weights are selected. In this embodiment, four features are selected as input variables for the BP neural network model.

[0060] Step S3, Data Normalization and Partitioning:

[0061] The selected key feature data were normalized using the Min-Max normalization method to accelerate model convergence. Subsequently, the normalized dataset was divided into training, testing, and validation sets according to a predetermined ratio (e.g., 70:15:15).

[0062] The normalization process uses the Min-Max standardization method, and the calculation formula is as follows:

[0063]

[0064] Where x is the original feature value, and min(x) and max(x) are the minimum and maximum values ​​of the feature in all samples, respectively. This is the normalized value.

[0065] Step S4, BP neural network model construction:

[0066] Construct a custom backpropagation (BP) neural network model. This model includes an input layer, multiple hidden layers, and an output layer. One-dimensional convolution operations are incorporated into the hidden layers to extract features and reduce the number of model parameters. The ReLU activation function is used alternately between linear layers, and a Dropout layer is introduced to prevent overfitting. Figure 2 As shown, a specific example is as follows:

[0067] The hierarchical structure is as follows: Input layer → First hidden layer (linear layer → rearranged layer → one-dimensional convolutional layer → ReLU → Dropout) → Subsequent hidden layers (linear layer → one-dimensional convolution → ReLU → Dropout) → Flattened layer → Output layer.

[0068] A one-dimensional convolutional layer with a kernel size of 3 and a stride of 1 is placed after the first linear layer to extract local correlation information of the input features. The output of this convolutional layer is then connected to a ReLU activation function layer and a Dropout layer with a deactivation probability of 0.5.

[0069] The input layer contains 4 nodes, which correspond to 4 key process parameters that affect the moisture content: ambient temperature of the rehydration section (x1), ambient humidity of the drying section (x2), moisture content before entering the cabinet after feeding (x3), and moisture content at the HT (drying) inlet (x4).

[0070] The first hidden layer contains 12 neurons, each fully connected to 4 nodes in the input layer, and outputs a 12-dimensional feature vector;

[0071] A data rearrangement layer is set between the first linear layer and the one-dimensional convolutional layer, which rearranges the 12-dimensional feature vector into a 6-row, 2-column matrix, where 6 is the number of feature channels and 2 is the sequence length.

[0072] The one-dimensional convolutional layer contains multiple convolutional kernels, each with a size of 3 and a stride of 1, and uses boundary padding to maintain the length of the output sequence. This layer performs convolution operations on two sequence positions across six channels, outputting multiple new feature maps, each retaining two sequence positions.

[0073] The activation function layer connects to the ReLU activation function, introducing a nonlinear transformation.

[0074] The dropout layer has a deactivation probability of 0.5, randomly discarding 50% of the neuron outputs during training to prevent overfitting.

[0075] The structure of subsequent hidden layers is similar to that of the first hidden layer, also using an alternating pattern of linear layer → one-dimensional convolution → ReLU → Dropout. The specific number of neurons and convolution kernels can be adjusted according to the task.

[0076] Multiple feature maps (each with 2 sequence positions) are flattened into feature vectors.

[0077] The output layer contains one neuron, which outputs the predicted value of moisture content at the refluxing outlet.

[0078] Step S5, Model Training:

[0079] The model is trained using a training set. During training, Mini-batch gradient descent is used for weight updates and optimization to help escape local optima. Cosine annealing is used to dynamically adjust the learning rate; a larger learning rate initially accelerates convergence, while a smaller learning rate is used later to fine-tune parameters and improve model performance. The mean squared error (MSE) loss function is used for training.

[0080] The cosine annealing algorithm adjusts the learning rate using the following formula:

[0081]

[0082] in, The learning rate for the current epoch. and Minimum and maximum learning rates This is the number of the currently executing epoch. This represents the total number of epochs for training.

[0083] The loss function used for model training is the mean squared error (MSE) loss function, and the calculation formula is as follows:

[0084]

[0085] in, These are the model's predicted values. is the true value, and N is the number of samples.

[0086] Step S6, Model Validation and Application:

[0087] After model training is complete, the model performance is evaluated on the validation set, using the root mean square error (RMSE) between the predicted and true values ​​as the evaluation metric. If the RMSE on the validation set is less than a preset threshold (e.g., less than 0.15%), the model is considered to have passed training and possesses good prediction accuracy and generalization ability. The model's generalization ability is further tested on an independent test set to ensure that the model still has high-accuracy prediction performance on new, unseen data.

[0088] The trained model is solidified and packaged as an inference service, deployed to the production system control platform, receives key feature variable inputs in real time, and outputs predicted moisture content at the reabsorption outlet for on-site guidance and automatic control of water addition.

[0089] Step S7, Model Update:

[0090] To adapt to the slow changes in production conditions, new production data is collected monthly and mixed with historical data in a preset ratio to build a training dataset. The deployed BP neural network model is then incrementally updated and trained to ensure that the model continues to maintain optimal prediction performance.

[0091] Example 2:

[0092] This embodiment presents a BP neural network method for predicting tobacco leaf moisture by integrating feature selection and dynamic optimization, as detailed below:

[0093] First, relevant data on the rehumidification process over the past year is extracted from the factory database, which can contain dozens of data items. Then, the Relief algorithm is applied to score the importance of these features, and the top 6 features are selected, such as: ambient temperature in the rehumidification section (x1), ambient humidity in the drying section (x2), moisture content before entering the cabinet after feeding (x3), moisture content at the HT (drying) inlet (x4), ambient temperature in the drying section (x5), and ambient humidity in the rehumidification section (x6).

[0094] Next, the data for these 6 key features were normalized, and the dataset was randomly divided into training, testing, and validation sets in a ratio of 70:20:10.

[0095] Next, a backpropagation (BP) neural network model is constructed. The input layer of this model has 6 neurons, corresponding to 6 input features. The hidden layer consists of 3 layers, where the first hidden layer introduces a one-dimensional convolution after a linear transformation, with a kernel size of 3 and a stride of 1, to extract local features. This is followed by a ReLU activation function and a Dropout layer (with a dropout rate of 0.2). Subsequent hidden layers have a similar structure, but the number of neurons can be adjusted according to the actual situation. The output layer has one neuron and outputs the predicted water content value.

[0096] During the training phase, set the initial maximum learning rate ( The minimum learning rate is 0.01. The value is 0.0001, and the total number of training rounds is ( The learning rate is set to 200. The batch size is set to 32, using a mini-batch approach. The optimizer is Adam. The loss function is mean squared error loss (MSE). The learning rate is adjusted using a cosine annealing strategy.

[0097] After the model is trained, its performance is evaluated on the validation set. If the root mean square error (RMSE) between the predicted and true values ​​on the validation set meets a predetermined standard (e.g., less than 0.15%), the model is considered successfully trained and can be packaged as a service and deployed to the control platform of the production system.

[0098] During production, the system collects the latest data on the above six key characteristics in real time and inputs it into the deployed prediction model. The model outputs the moisture prediction value in real time. Operators or automatic control systems can adjust the water addition in a timely manner based on the deviation between this prediction value and the target moisture content, thereby achieving precise control.

[0099] In addition, each month, the system automatically collects new production data generated in the current quarter, mixes it with 50% of the original training data to form a new training set, and updates and trains the model to keep its predictive ability matching the current production status.

[0100] The above description is only used to illustrate the technical solutions of the present invention and is not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention (such as the application of various formulas, the order of steps, etc.) without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for predicting tobacco leaf moisture content using a BP neural network that integrates feature selection and dynamic optimization, characterized in that, Includes the following steps: Step S1, Data Acquisition and Preprocessing: Collect historical production data related to the rehumidification process, including ambient temperature of the rehumidification section, ambient humidity of the drying section, moisture content before entering the cabinet after feeding, moisture content at the HT drying inlet, ambient temperature of the drying section, and ambient humidity of the rehumidification section. Step S2, Feature Filtering: The Relief feature weighting algorithm is used to calculate the correlation weights between each feature variable in the original dataset and the target value of moisture content at the reabsorption outlet. The features are sorted according to their weights, and the top n features with the largest weights are selected as the input variables of the BP neural network model. Step S3, Data Normalization and Partitioning: The selected key feature data is normalized, and the normalized dataset is divided into training set, test set and validation set according to a predetermined ratio. Step S4, BP neural network model construction: Construct a BP neural network model, including an input layer, multiple hidden layers, and an output layer; The hidden layer incorporates a one-dimensional convolution operation to extract features and reduce the number of model parameters. Alternate ReLU activation functions between linear layers and introduce Dropout layers to prevent overfitting; Following the first linear layer, a data rearrangement layer and a one-dimensional convolutional layer are sequentially arranged. The data rearrangement layer is used to reorganize the feature vector output by the linear layer into a multi-dimensional matrix format with channel and sequence dimensions to adapt to the input requirements of the one-dimensional convolutional layer. The one-dimensional convolutional layer is used to extract local correlation information of the input features. The output of the one-dimensional convolutional layer is sequentially connected to a ReLU activation function layer and a Dropout layer. Step S5, Model Training: The model is trained by: updating the model weights using the Mini-batch gradient descent method; adjusting the learning rate using the cosine annealing algorithm; and using the mean squared error (MSE) as the loss function during training.

2. The BP neural network method for predicting tobacco leaf moisture content based on fusion feature screening and dynamic optimization as described in claim 1, characterized in that, In step S2, the Relief algorithm calculates the difference values ​​of features according to the variable type using a differentiation rule. The logic for calculating the difference values ​​of nominal variables and numerical variables is as follows: when the nominal variables have the same value, the difference value is 0, and when they have different values, the difference value is 1. For numerical variables, the difference value is calculated by combining the absolute value of the difference with a normalization factor. The normalization factor is the difference between the maximum and minimum values ​​of the feature in the sample set, ensuring that the result is normalized to the [0,1] interval.

3. The BP neural network method for predicting tobacco leaf moisture content based on fusion feature screening and dynamic optimization as described in claim 1, characterized in that, In step S3, the normalization process uses the Min-Max standardization method, and the calculation formula is as follows: Where x is the original feature value, and min(x) and max(x) are the minimum and maximum values ​​of the feature in all samples, respectively. This is the normalized value.

4. The BP neural network method for predicting tobacco leaf moisture content based on fusion feature screening and dynamic optimization as described in claim 1, characterized in that, In step S4, a data rearrangement layer is set between the first linear layer and the one-dimensional convolutional layer to reorganize the feature vector output by the linear layer into a multi-dimensional matrix format with channel dimension and sequence dimension to adapt to the input requirements of the one-dimensional convolutional layer.

5. The BP neural network method for predicting tobacco leaf moisture content based on fusion feature screening and dynamic optimization as described in claim 4, characterized in that, In step S4, the kernel size of the one-dimensional convolutional layer is set to 3, and the stride is set to 1, which matches the sequence length constructed by the data rearrangement layer.

6. The BP neural network method for predicting tobacco leaf moisture content based on fusion feature screening and dynamic optimization as described in claim 1, characterized in that, In step S4, the one-dimensional convolutional layer uses boundary padding to ensure that the sequence length remains consistent before and after convolution, so that local correlation information between adjacent features can still be effectively extracted even when the sequence length is short.

7. The BP neural network method for predicting tobacco leaf moisture content based on fusion feature screening and dynamic optimization as described in claim 1, characterized in that, In step S5, the cosine annealing algorithm adjusts the learning rate, and the calculation formula is as follows: in, The learning rate for the current training round. and These are the minimum and maximum values ​​of the learning rate, respectively. This is the current training round number. This represents the total number of training rounds.

8. The BP neural network method for predicting tobacco leaf moisture content based on fusion feature screening and dynamic optimization as described in claim 1, characterized in that, In step S5, the mean squared error (MSE) is used as the loss function during the training process, and the calculation formula is as follows: in, These are the model's predicted values. is the true value, and N is the number of samples.

9. The BP neural network method for predicting tobacco leaf moisture content based on fusion feature screening and dynamic optimization as described in claim 1, characterized in that, The method further includes evaluating the model performance after the model training is completed, using the root mean square error (RMSE) between the predicted value and the true value as the evaluation index; if the RMSE on the validation set is less than a preset threshold, the model training is deemed qualified; and further testing the model's generalization ability on an independent test set to ensure that the model still has high-precision prediction performance on unseen new data.

10. The BP neural network method for predicting tobacco leaf moisture content based on fusion feature screening and dynamic optimization according to claim 1, characterized in that, The method also includes collecting new production data monthly, mixing it with historical data at a preset ratio to construct a training dataset, and incrementally updating and training the deployed BP neural network model to ensure that the model continues to maintain optimal prediction performance.

Images