Intelligent control method of tea withering parameters based on convolutional network attention mechanism

Abstract

The present invention discloses an intelligent control method for tea withering parameters based on a convolutional network attention mechanism, including steps such as building a tea processing database, dividing data samples, building a multi-factor tea withering parameter intelligent control model, and model training and testing. The multi-factor tea withering parameter intelligent control model extracts tea features based on a convolutional layer, then inputs two stacked attention mechanism layers to adjust the feature importance, and outputs multi-factor withering parameters through two layers of full connection. Among them, the attention mechanism layer first introduces different pooling functions to mine the features of tea factors, and uses a multi-layer perceptron to calculate the weights of different convolution channels, and then weights the weights to the tea features after convolution, so as to achieve the purpose of feature importance adjustment and improve network performance. The present invention pays attention to the important features of tea processing factors through an attention mechanism and a convolutional neural network, constructs the relationship between tea features and processing parameters, and realizes intelligent control of tea processing parameters.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent tea processing, and in particular relates to an intelligent control method for tea withering parameters based on a convolutional network attention mechanism. Background Art

[0002] From the stage of picking raw materials to the final sale, tea needs to go through a series of processing stages. Most of the modern tea processing processes are completed by processing machinery on automated production lines. Withering is a very critical link in the tea processing process, which can make the final tea produce have good color, aroma and taste properties. The withering process utilizes the characteristic that the oxidative activity of polyphenol enzymes is reduced under high temperature environment, and controls the temperature conditions in the processing link to inhibit its activity in order to achieve the purpose of withering.

[0003] Since the tea withering process depends on the characteristics of the tea leaves and the movement generated by the mechanical structure, it is now mostly done by experienced workers operating the host computer on the digital production line to adjust the relevant parameters for processing control. However, manual judgment has a certain degree of subjectivity, experience is difficult to pass on, and the level of intelligence of processing equipment is low, all of which have seriously hindered the development of tea processing.

[0004] In green tea processing, the key is to correctly grasp the principles of "high temperature killing, high temperature first and low temperature later", "combination of throwing and steaming, more throwing and less steaming" and "old leaves killing young leaves, young leaves killing old leaves". Translated to the requirements on the production line, it is to control the temperature and the movement parameters of the machine, etc. How to intelligently adjust the processing parameters according to the characteristics of small amounts of tea is the difficulty in current tea processing.

[0005] For example, CN113100304A discloses a dynamic heating drying, dehydration and aroma extraction process for high-quality Tieguanyin tea, which uses a vibrating conveyor belt to complete the sun-drying, air-drying and shaking of tea leaves, and uses a vibrating conveyor belt to replace traditional manual labor, thereby achieving controllable process parameters. At the same time, an LED plant fill light is required to be installed on the upper part of the conveyor belt to supplement the light when the illumination is insufficient; the circular vibrating conveyor belt is also provided with a heating device to supplement the temperature when it is insufficient. In this way, it is possible to save labor while ensuring the quality of the tea, and to use the device to compensate for environmental defects; during the kneading and baking process of the tea leaves, an automatic kneading unit, a vibrating conveyor belt and a far-infrared heating drying chamber are used to complete the kneading and baking of the tea leaves.

[0006] Although the above technical solution can control the process parameters of tea, it is not intelligent enough and cannot intelligently adjust the optimal processing parameters of tea according to the characteristics of tea.

[0007] In view of the problems existing in the prior art, there is an urgent need for a method that can realize intelligent regulation of processing equipment parameters according to the characteristics of tea, so as to solve the problem that the existing processing equipment is not intelligent enough in adjusting the processing parameters when processing tea; and to solve the accuracy problems and experience inheritance problems caused by the fact that the existing processing equipment requires staff to subjectively judge the processing parameters based on their experience. Summary of the invention

[0008] In view of the problems and defects of the prior art, the purpose of the present invention is to provide a multi-factor tea withering intelligent control method based on attention mechanism and convolutional neural network to establish the relationship between tea processing parameters and tea characteristics, extract the deep features of tea through the convolution layer, and then construct the attention mechanism layer to focus on important features, and establish the relationship between tea characteristics and processing parameters, so as to realize the intelligent control of tea withering parameters.

[0009] In order to achieve the above object, the present invention proposes an intelligent control method for tea withering parameters based on a convolutional network attention mechanism, comprising the following steps:

[0010] Step 1: Collect processing parameter data, tea state data and other data during tea processing (tea state data and other data together constitute tea multi-factor characteristics), where processing parameters include zone 1 temperature, zone 2 temperature, zone 3 temperature, zone 4 temperature, zone 5 temperature, drum speed, feeding speed, dehumidification fan variable speed; tea state includes tea leaf size, tea grade, moisture content before and after withering, and other data includes processing date and image G component, and build a database;

[0011] Step 2: Preprocess the collected tea data and divide it into training set and test set;

[0012] Step 3: Construct a multi-factor tea withering parameter intelligent control model based on attention mechanism and convolutional neural network, the model includes convolution layer, BN layer, convolution channel attention module, and fully connected layer;

[0013] Step 4: Put the prepared data into the model for training. The model can predict and intelligently adjust the parameters of withering according to the multi-factor characteristics of tea leaves;

[0014] The root mean square error (RMSE) was used as the loss function for training. After hyperparameter optimization, Adam was selected as the optimizer, the learning rate was set to 0.001, and the batch size was 4.

[0015] The expression of the root mean square error is as follows:

[0016]

[0017] Among them, f(x i ) is the predicted processing parameter value, yi is the actual processing parameter value;

[0018] Among them, the optimization process of the Adam optimizer is as follows:

[0019]

[0020] Among them, m t and v t are the first-order matrix and second-order matrix of RMSE gradient, respectively. and They are respectively t and v t The correction of t are the parameters of the model, θ t+1 are the parameters updated by the model, and are the first-order and second-order matrix exponential decay rates respectively, ε is set to 10e-8, and η is the learning rate.

[0021] The core of the intelligent control method of tea withering parameters based on the convolutional network attention mechanism proposed in the present invention is to improve the attention mechanism to calculate the weights of multiple convolution channels and adjust the importance of features. In the actual production process, using the attention mechanism and convolutional neural network model to intelligently control the withering parameters when the tea features are limited is of great significance for the intelligent processing of tea.

[0022] Preferably, in step 1, there are several samples in the database, and the samples are composed of multi-factor characteristics of tea and processing parameters, wherein the multi-factor feature set is {processing date, image G component, tea leaf size, tea grade, moisture content before withering, moisture content after withering}, and the processing parameter set is {zone one temperature, zone two temperature, zone three temperature, zone four temperature, zone five temperature, drum speed, feeding speed, dehumidification fan variable speed}.

[0023] Among them, the G component refers to the average value of G in the RGB image component after the background of the tea image is removed. The size of the tea leaves is divided into three levels: large, medium and small based on manual judgment. The tea grade is obtained based on experienced manual scoring. The moisture content is the result of taking 5g of tea leaves and measuring them on a rapid moisture meter.

[0024] Preferably, in step 2, the grade and processing date of each batch of tea are encoded. In order to eliminate the influence of different feature dimensions and accelerate the convergence rate of the model, each data in the database is normalized and preprocessed. The specific formula of the normalized preprocessing is as follows:

[0025]

[0026] Among them, x is the parameter value in each sample, min refers to the minimum value of the same parameter in each sample, max is the maximum value of the same parameter in each sample, and x' is the normalized value;

[0027] The tea grades are coded from 1 to 6 from low to high, and the tea processing date is coded as the number of days from the tea growing date to the processing date.

[0028] Preferably, in step 3, the number of convolutional layers, convolutional channel attention modules, and fully connected layers and their corresponding positions in the entire model are determined to construct the model.

[0029] Preferably, the structure of the convolutional neural network in step 3 is: a convolutional layer with a convolution kernel of 1x3, a BN layer, two convolutional channel attention modules connected in series, and two fully connected layers with a ReLU activation layer in the middle, wherein the function expression of ReLU is:

[0030] f(x)=max(0,x).

[0031] Preferably, when the training set data in step 2 is input into the model of step 3, it is first expanded from 6 to generate 1×6, where the step size, convolution kernel dimension, and padding of the convolution layer are 1, 1, and 0 respectively, the number of convolution kernels is 64, and the data dimension after the convolution layer is 64×4.

[0032] Preferably, the features of the convolutional layer in step 3 are input into the convolutional channel attention module after passing through the BN layer. The mathematical model of the module is as follows:

[0033]

[0034] M c (F)=σ(MLP(AvgPool(F))+MLP(VarPool(F))+MLP(MaxPool(F)))

[0035] Where F is the input feature of the module, F' is the output feature of the module, and M c is the output of the convolutional channel attention module, AvgPool, VarPool, and MaxPool are average pooling, variance pooling, and maximum pooling, respectively. MLP is a multi-layer perceptron, σ is a Sigmoid activation layer, and the expression of the Sigmoid function is as follows:

[0036]

[0037] The input features are first subjected to parallel average pooling, maximum pooling, and variance pooling. The specific method is to calculate the average value, maximum value, and variance of each channel. After pooling, three parallel 64×1 pooling features are obtained as the input of MLP. MLP has two layers. The number of neurons in the first layer is 8, and the number of neurons in the second layer is 1. The three parallel pooling layer data pass through the same MLP to obtain three 64×1 features, which are added together to obtain 64×1 features. Then, the channel weights with a data dimension of 64×1 are obtained through a Sigmoid activation layer. Finally, the channel weights are multiplied by the input features to obtain tea features with a data dimension of 64×4.

[0038] The maximum pooling layer can retain the maximum value of the tea features, the average pooling layer takes into account each feature of tea processing, and the variance pooling layer takes into account the differences in tea features. After passing through the same multi-layer perceptron, the features are added together to obtain the importance of each channel of the convolution output. In order to strengthen the adjustment of feature importance, the two attention mechanism modules are stacked to output the tea features after importance adjustment, and then input into the last two layers of full connection to accurately output the multi-factor withering parameters.

[0039] Preferably, the tea features with a data dimension of 64×4 are outputted through a second convolutional channel attention module to have features with an output dimension of 64×4. When passing through the second convolutional channel attention module, the weight distribution of the 64 channels can be made more accurate. After flattening, the feature dimension becomes 256, which serves as the input of the fully connected layer.

[0040] Preferably, the feature with a dimension of 256 becomes 64 after passing through the first fully connected layer, and then passes through the second fully connected layer to obtain the tea withering processing parameters with a dimension of 8.

[0041] Preferably, when the model in step 4 is trained, the best multi-factor tea withering parameter intelligent control model based on attention mechanism and convolutional neural network can be selected according to the performance of the test set.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] 1. The present invention proposes a tea withering control method based on attention mechanism and convolutional neural network to establish the relationship between tea processing parameters and tea characteristics, extract the deep features of tea through the convolution layer, and then construct the attention mechanism layer to focus on important features, and establish the relationship between tea characteristics and processing parameters, so as to realize the intelligent control of tea withering parameters.

[0044] 2. The present invention reduces the workload of workers in adjusting the parameters of tea-withering processing, builds a model using small sample data with fewer features, fully utilizes the image, tenderness, and size information of tea leaves, extracts the features of tea leaves well, and adjusts the importance of features based on the convolutional channel attention mechanism. The present invention solves the problem that the experience of tea-withering processing is difficult to inherit, and is of great significance to promoting the digital and intelligent processing of tea leaves.

[0045] 3. The maximum pooling layer of the present invention can retain the maximum value of the tea features, the average pooling layer takes into account each feature of tea processing, and the variance pooling layer takes into account the differences in tea features. After passing through the same multi-layer perceptron, the features are added to obtain the importance of each channel of the convolution output; in order to strengthen the adjustment of feature importance, the two attention mechanism modules are stacked, and the tea features after importance adjustment are output, and then they are input into the last two layers of full connection to accurately output multi-factor tea withering parameters. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0047] Figure 1 It is a flow chart of the intelligent control method of tea withering parameters based on the convolutional network attention mechanism of the present invention;

[0048] Figure 2 It is a schematic diagram of the overall structure of the multi-factor tea withering parameter intelligent control model based on the attention mechanism and convolutional neural network of the present invention;

[0049] Figure 3 Schematic diagram of the structure of the convolutional channel attention module of the present invention. DETAILED DESCRIPTION

[0050] In order to make the purpose, technical scheme and advantages of the embodiments of the present invention clearer, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. The components of the embodiments of the present invention generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0051] Embodiment 1

[0052] like Figure 1-3 As shown, this embodiment proposes an intelligent control method for tea withering parameters based on a convolutional network attention mechanism, comprising the following steps:

[0053] Step 1: Collect processing parameter data, tea state data and other data during tea processing, and build a database. There are several samples in the database. The samples are composed of multi-factor characteristics and processing parameters of green tea. The multi-factor set is {processing date, image G component, tea leaf size, tea grade, moisture content before fixing, moisture content after fixing}, and the processing parameter set is {zone 1 temperature, zone 2 temperature, zone 3 temperature, zone 4 temperature, zone 5 temperature, drum speed, feeding speed, dehumidification fan variable speed};

[0054] Among them, the G component refers to the average value of G in the RGB image component after the background of the tea image is removed; the size of the tea leaves is divided into three levels: large, medium and small according to manual judgment; the tea grade is obtained based on experienced manual scoring, including grade 3, grade 2, grade 1, special grade 1, special grade A, and special grade; the moisture content is the result of taking 5g of tea leaves and measuring them on a rapid moisture meter;

[0055] Step 2: Preprocess the collected tea data and divide them into training set and test set. There are a total of 71 samples collected this time. 40 samples are randomly selected for training, and the remaining 31 are used as test sets. Encode the grade and processing date of each batch of tea. The coding rule for tea grade is {"Grade 3": 10, "Grade 2": 20, "Grade 1": 30, "Special 1": 40, "Special A": 50, "Special": 60}. The coding rule for tea processing date is to take the difference between the processing date and the number of days on March 1. The size of the tea leaves is coded from small to large as 1, 2, and 3. In addition, in order to eliminate the influence of different feature dimensions and accelerate model convergence, each data in the database is normalized and preprocessed. The specific formula is: Among them, x is the parameter value in each sample, min refers to the minimum value of the same parameter in each sample, max is the maximum value of the same parameter in each sample, and x' is the normalized value;

[0056] Step 3: Construct a multi-factor tea withering parameter intelligent control model based on attention mechanism and convolutional neural network. The model mainly includes convolution layer, convolution channel attention module and fully connected layer.

[0057] Preferably, determining the number of convolutional layers, convolutional channel attention modules, and fully connected layers and the corresponding position of each module in the entire model to construct the model;

[0058] Preferably, the neural network structure in step 3 is a convolution layer with a convolution kernel of 1x3, a BN layer, two convolution channel attention modules connected in series, and two fully connected layers with a ReLU activation layer in the middle, where the function expression of ReLU is:

[0059] f(x)=max(0,x)

[0060] Preferably, when the training set data in step 2 is input into the model in step 3, the 6-dimensionality is expanded to generate 1×6, the step size, convolution kernel dimension, and padding of the convolution layer are 1, 1, and 0 respectively, the number of convolution kernels is 64, and the feature dimension after the convolution layer is 64×4;

[0061] Preferably, the features of the convolutional layer in step 3 are input into the convolutional channel attention module in step 3 after passing through the BN layer. The mathematical model of the module is as follows:

[0062]

[0063] M c (F)=σ(MLP(AvgPool(F))+MLP(VarPool(F))+MLP(MaxPool(F)))

[0064] Where F is the input feature of the module, F' is the output feature of the module, Mc is the output of the convolutional channel attention module, AvgPool, VarPool, and MaxPool are average pooling, variance pooling, and maximum pooling, respectively, MLP is a multi-layer perceptron, σ is a Sigmoid activation layer, and the expression of the Sigmoid function is as follows:

[0065]

[0066] The input features are first subjected to parallel average pooling, maximum pooling, and variance pooling. The specific method is to calculate the average value, maximum value, and variance of each channel. After pooling, three parallel 64×1 pooling features are obtained as the input of MLP. MLP has two layers. The number of neurons in the first layer is 8, and the number of neurons in the second layer is 1. The three parallel pooling layer data pass through the same MLP to obtain three 64×1 features, which are added together to obtain 64×1 features. Then, the channel weights with a data dimension of 64×1 are obtained through a Sigmoid activation layer. Finally, the channel weights are multiplied by the input features to obtain tea features with a data dimension of 64×4.

[0067] As a preference, the 64×4 features are passed through the second convolutional channel attention module to output features with a dimension of 64×4. After passing through the second convolutional channel attention module, the weight distribution of the 64 channels is more accurate. After flattening, the data dimension becomes 256, which is used as the input of the next fully connected layer.

[0068] Preferably, the feature with a dimension of 256 becomes 64 after passing through the first fully connected layer, and then passes through the second fully connected layer to obtain the tea withering processing parameters with a dimension of 8.

[0069] The parameter settings of the multi-factor tea withering parameter intelligent control model based on attention mechanism and convolutional neural network are shown in Table 1 below. Figure 2 , 3 The network structure diagram in corresponds to:

[0070] Table 1:

[0071]

[0072]

[0073] Step 4: Put the prepared data into the model for training to achieve the goal of adjusting the parameters of the withering process according to the characteristics of the tea raw materials and the expected indicators. The root mean square error (RMSE) is used as the loss function. After hyperparameter optimization, Adam is selected as the optimizer, the learning rate is 0.001, and the batch size is 4;

[0074] The expression of the root mean square error is as follows:

[0075]

[0076] Among them, f(x i ) is the predicted processing parameter value, y i is the actual processing parameter value;

[0077] The optimization process of Adam optimizer is as follows:

[0078]

[0079] Among them, m t and v t are the first-order matrix and second-order matrix of RMSE gradient, respectively. and They are respectively t and v t The correction of t are the parameters of the model, θ t+1 are the parameters updated by the model, and are the first-order and second-order matrix exponential decay rates respectively, ε is set to 10e-8, and η is the learning rate.

[0080] After selection, the model with the best performance on the test set had a prediction error rate of [2.75%, 2.75%, 2.83%, 2.93%, 3.24%, 0.85%, 9.88%, 6.00%] for each processing parameter, which can meet the requirements of precise and intelligent control of green tea withering parameters.

[0081] Of course, the above description is only a specific embodiment of the present invention and is not intended to limit the scope of implementation of the present invention. All equivalent changes or modifications made according to the structure, characteristics and principles described in the patent application scope of the present invention should be included in the patent application scope of the present invention.

[0082] Finally, it should be noted that the above-described embodiments are only specific implementations of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The protection scope of the present invention is not limited thereto. Although the present invention is described in detail with reference to the above-described embodiments, ordinary technicians in the field should understand that any technician familiar with the technical field can still modify the technical solutions recorded in the above-described embodiments within the technical scope disclosed by the present invention, or can easily think of changes, or make equivalent replacements for some of the technical features therein; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. An intelligent control method for tea withering parameters based on a convolutional network attention mechanism, characterized in that: The steps include: Step 1: Collect processing parameter data and tea status data during tea processing and build a database; Step 2: Preprocess the collected tea data and divide it into training set and test set; Step 3: construct a multi-factor tea withering parameter intelligent control model based on attention mechanism and convolutional neural network, the model includes a convolutional layer, a convolutional channel attention module, and a fully connected layer; Step 4: Put the prepared data into the model for training. The model can predict and intelligently adjust the parameters of withering according to the multi-factor characteristics of tea leaves; The root mean square error (RMSE) was used as the loss function for training. After hyperparameter optimization, Adam was selected as the optimizer, the learning rate was set to 0.001, and the batch size was 4. The expression of the root mean square error is as follows: Among them, f(x i ) is the predicted processing parameter value, y i is the actual processing parameter value; Among them, the optimization process of the Adam optimizer is as follows: Among them, m t and v t are the first-order matrix and second-order matrix of RMSE gradient, respectively. and They are respectively t and v t The correction of t are the parameters of the model, θ t+1 are the parameters updated by the model, and are the first-order and second-order matrix exponential decay rates, ε is set to 10e-8, and η is the learning rate; In the step 1, there are several samples in the database, and the samples are composed of tea multi-factor features and processing parameters, wherein the multi-factor feature set is {processing date, image G component, tea leaf size, tea grade, moisture content before fixing, moisture content after fixing}, and the processing parameter set is {zone 1 temperature, zone 2 temperature, zone 3 temperature, zone 4 temperature, zone 5 temperature, drum speed, feeding speed, dehumidification fan variable speed}; In step 3, the model is constructed by determining the number of convolutional layers, convolutional channel attention modules, and fully connected layers and their corresponding positions in the entire model; The structure of the convolutional neural network in step 3 is: a convolutional layer with a convolution kernel of 1x3, a BN layer, two convolutional channel attention modules connected in series, and two fully connected layers with a ReLU activation layer in the middle, where the function expression of ReLU is: f(x)=max(0,x); When the training set data in step 2 is input into the model in step 3, it is first expanded from 6 dimensions to 1×6, where the step size, convolution kernel dimension, and padding of the convolution layer are 1, 1, and 0 respectively, the number of convolution kernels is 64, and the feature dimension after the convolution layer is 64×4; After the features of the convolutional layer in step 3 pass through the BN layer, they are input into the convolutional channel attention module, and the mathematical model of this module is as follows: M c (F)=σ(MLP(AvgPool(F))+MLP(VarPool(F))+MLP(MaxPool(F))) Where F is the input feature of the module, F' is the output feature of the module, and M c is the output of the convolutional channel attention module, AvgPool, VarPool, and MaxPool are average pooling, variance pooling, and maximum pooling, respectively. MLP is a multi-layer perceptron. σ is a Sigmoid activation layer. The expression of the Sigmoid function is as follows: The input features are first subjected to parallel average pooling, maximum pooling, and variance pooling. The specific method is to calculate the average value, maximum value, and variance of each channel. After pooling, three parallel 64×1 pooling features are obtained as the input of MLP. MLP has two layers, the number of neurons in the first layer is 8, and the number of neurons in the second layer is 1. The three parallel pooling layer data pass through the same MLP to obtain three 64×1 features, which are added to obtain 64×1 features. Then, the channel weights with a data dimension of 64×1 are obtained through the Sigmoid activation layer. Finally, the channel weights are multiplied by the input features to obtain tea features with a data dimension of 64×4. The tea features with a data dimension of 64×4 are outputted into features with a dimension of 64×4 after passing through the second convolutional channel attention module. After passing through the second convolutional channel attention module, the weight distribution of the 64 channels can be made more accurate; after flattening, the feature dimension becomes 256, which serves as the input of the fully connected layer.

2. The intelligent control method for tea withering parameters based on convolutional network attention mechanism according to claim 1 is characterized in that: In step 2, the grade and processing date of each batch of tea are encoded; at the same time, each data in the database is normalized and preprocessed to eliminate the influence of different feature dimensions and accelerate the model convergence rate. The specific formula of the normalization preprocessing is as follows: Among them, x is the parameter value in each sample, min refers to the minimum value of the same parameter in each sample, max is the maximum value of the same parameter in each sample, and x' is the normalized value.

3. The intelligent control method for tea withering parameters based on convolutional network attention mechanism according to claim 1 is characterized in that: The feature with a dimension of 256 becomes 64 after passing through the first fully connected layer, and then passes through the second fully connected layer to obtain the tea withering processing parameters with a dimension of 8.

4. The intelligent control method for tea withering parameters based on convolutional network attention mechanism according to claim 1 is characterized in that: When the model in step 4 is trained, the best multi-factor tea withering parameter intelligent control model based on attention mechanism and convolutional neural network can be selected according to the performance of the test set.

Images