Production monitoring and management method and system for tagatose sugar-free tea

Abstract

The present invention relates to a production monitoring and management method and system for tagatose sugar-free tea. The method: perform correlation analysis on key parameters in the production process of tagatose sugar-free tea to obtain a set of correlation production parameters including temperature, stirring speed, pH value, solubility and tagatose addition, tea extraction temperature and tea extraction time; perform quality indicator prediction through a quality indicator prediction model to obtain quality indicator prediction values; use the quality indicator prediction values ​​and the temperature-stirring speed data subset as inputs of a two-stage learner-evaluator multi-task sequential neural model for prediction, and create a temperature and stirring speed control strategy; apply the temperature and stirring speed control strategy to the tagatose sugar-free tea production process, collect the quality-related feature set of the production process, and perform quality evaluation and fusion analysis to obtain a comprehensive quality evaluation result. The implementation of the present invention improves the production efficiency and quality stability of tagatose sugar-free tea.

Description

Technical Field

[0001] The present invention relates to the technical field of production monitoring and management, and in particular to a production monitoring and management method and system for tagatose sugar-free tea. Background Art

[0002] As a new sweetener, tagatose has different solubility characteristics and stability in the tea production process from traditional sugars, which brings new challenges to production monitoring and quality control. Traditional tea production monitoring methods are difficult to accurately grasp the dissolution behavior of tagatose under different temperatures, pH values ​​and stirring conditions, resulting in large fluctuations in product quality.

[0003] In addition, the production process of tagatose sugar-free tea involves multiple key parameters, such as temperature, stirring speed, pH value, solubility and tagatose addition, tea extraction temperature and tea extraction time, etc. There is a complex nonlinear relationship between these parameters, and the traditional single parameter monitoring method is difficult to fully reflect the dynamic changes in the production process. At the same time, due to the lack of a specialized quality assessment system for the characteristics of tagatose, potential problems in the production process are difficult to discover and correct in a timely manner, affecting the consistency and taste of the product. Summary of the invention

[0004] The main purpose of the present invention is to provide a production monitoring and management method and system for tagatose sugar-free tea, so as to improve the production efficiency and quality stability of tagatose sugar-free tea.

[0005] To achieve the above object, the present invention provides a production monitoring and management method for tagatose sugar-free tea, comprising the following steps:

[0006] The key parameters in the production process of tagatose sugar-free tea were analyzed for correlation, and a set of correlation production parameters including temperature, stirring speed, pH value, solubility and tagatose addition amount, tea extraction temperature and tea extraction time were obtained;

[0007] Organizing the relevant production parameter set into a two-dimensional data matrix according to time series, and inputting the two-dimensional data matrix into a quality indicator prediction model to perform quality indicator prediction to obtain a quality indicator prediction value;

[0008] The quality index prediction values ​​and the temperature-stirring speed data subset are used as inputs of a two-stage learner-evaluator multi-task sequential neural model to predict the dissolution uniformity of tagatose and create temperature and stirring speed control strategies;

[0009] The temperature and stirring speed control strategy is applied to the tagatose sugar-free tea production process, a set of quality-related features of the production process is collected, and quality assessment and fusion analysis are performed through a multimodal feature fusion network to obtain a comprehensive quality assessment result.

[0010] The present invention also provides a production monitoring and management system for tagatose sugar-free tea, comprising:

[0011] An analysis module is used to perform correlation analysis on key parameters in the production process of tagatose sugar-free tea, and obtain a set of correlation production parameters including temperature, stirring speed, pH value, solubility and tagatose addition amount, tea extraction temperature and tea extraction time;

[0012] A prediction module, used for organizing the relevant production parameter set into a two-dimensional data matrix according to a time series, and inputting the two-dimensional data matrix into a quality index prediction model to perform quality index prediction to obtain a quality index prediction value;

[0013] Creating a module for using the quality indicator prediction value and the temperature-stirring speed data subset as inputs to a two-stage learner-evaluator multi-task sequential neural model to predict the dissolution uniformity of tagatose and create a temperature and stirring speed control strategy;

[0014] The evaluation module is used to apply the temperature and stirring speed control strategy to the tagatose sugar-free tea production process, collect the quality-related feature set of the production process, and perform quality evaluation and fusion analysis through a multimodal feature fusion network to obtain a comprehensive quality evaluation result.

[0015] The present invention also provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of any one of the above methods when executing the computer program.

[0016] The present invention also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of any of the above-mentioned methods are implemented.

[0017] In summary, the technical solution provided by the present invention realizes the accurate grasp of the complex parameter relationship in the production process of tagatose sugar-free tea by performing correlation analysis and clustering on key production parameters, which helps to optimize the production formula and process parameters. The quality index prediction model is constructed by using a multi-level convolutional neural network, which can accurately predict key quality indicators such as tagatose solubility, improve the controllability of the production process and the consistency of product quality. The multi-task sequential neural model of the two-stage learner-evaluator is introduced to realize the accurate prediction of tagatose dissolution uniformity and the intelligent control of temperature and stirring speed, effectively improving the automation level of the production process. The developed multimodal feature fusion network realizes the comprehensive evaluation of the multi-dimensional quality characteristics of tagatose sugar-free tea through the attention mechanism and multi-scale feature fusion, and improves the accuracy and comprehensiveness of quality control. The designed indicator calibration module can dynamically adjust various quality indicators according to the characteristics of tagatose, ensuring the reliability and applicability of the quality evaluation results. The present invention integrates data-driven and expert knowledge, which can not only adapt to the changes in different production environments, but also make full use of the existing knowledge of tagatose solubility characteristics, greatly improving the robustness and practicality of the method. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is a schematic diagram of the steps of a production monitoring and management method for tagatose sugar-free tea in one embodiment of the present invention;

[0019] Figure 2 This is a structural block diagram of a production monitoring and management system for tagatose sugar-free tea according to an embodiment of the present invention;

[0020] Figure 3 It is a schematic block diagram of the structure of a computer device according to an embodiment of the present invention.

[0021] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0022] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0023] Reference Figure 1 This embodiment provides a method for monitoring and managing the production of tagatose sugar-free tea, comprising the following steps:

[0024] S1, performing correlation analysis on key parameters in the production process of tagatose sugar-free tea, and obtaining a set of correlation production parameters including temperature, stirring speed, pH value, solubility and tagatose addition amount, tea extraction temperature and tea extraction time;

[0025] Among them, high-frequency data collection was performed on the key parameters in the production process of tagatose sugar-free tea to obtain a multi-dimensional original parameter data set, including temperature, stirring speed, pH value, solubility, tagatose addition, tea extraction temperature, tea extraction time and other key indicators that affect production. Data cleaning was performed on the multi-dimensional original parameter data set, including outlier detection, missing value interpolation and data standardization. Through outlier detection, the values ​​that obviously deviate from the normal range are identified and removed to reduce the interference of noise on subsequent analysis; missing value interpolation fills the missing parts in the data through appropriate methods to ensure the integrity of the data; data standardization is to eliminate the influence between different dimensions, convert all parameters to the same scale, and obtain a cleaned standardized parameter data set. The maximum information coefficient (MIC) is calculated for each pair of parameters in the cleaned standardized parameter data set to obtain the MIC correlation matrix. MIC is an indicator of the general degree of association between measuring variables, which can effectively capture linear and nonlinear relationships. By calculating MIC, the dependence between various production parameters is measured. At the same time, in order to measure the degree of linear correlation between the parameters, the Pearson correlation coefficient calculation formula is used to calculate the linear correlation between each pair of parameters to obtain the Pearson correlation matrix. The Pearson correlation coefficient is used to measure the linear relationship between two variables, and its value ranges from -1 to 1, indicating negative linear correlation, no correlation, and positive linear correlation, respectively. The MIC correlation matrix and the Pearson correlation matrix are weighted and fused to obtain a fused correlation matrix. Combining the advantages of different correlation indicators, the MIC matrix can capture various forms of associations, while the Pearson matrix provides detailed information about linear relationships. By weighted fusion of the two, a more comprehensive and accurate description of parameter correlation is obtained. Based on the fused correlation matrix, the key parameters in the production process are grouped to reveal the correlation between these parameters. The clustering method is used to group the parameters in the fused correlation matrix to obtain parameter clustering results, which usually include multiple clusters, each cluster representing a group of highly correlated parameters. Based on the parameter clustering results, combined with the pre-defined weight table of factors affecting the solubility characteristics of tagatose in the expert knowledge base, the weighted scores of the parameters in each cluster are calculated. The expert knowledge base contains important domain knowledge about the solubility characteristics of tagatose, such as the degree of influence of different parameters on the dissolution process, and this information is stored in the weight table in the form of weights. By using these weights, each parameter in each cluster is weighted and scored to determine the most representative parameters in each cluster. In order to finally select the relevant production parameter set, the parameters with the highest scores are selected from each cluster, and finally the relevant production parameter set is obtained, including temperature, stirring speed, pH value, solubility and tagatose addition, tea extraction temperature and tea extraction time.

[0026] S2, organizing the relevant production parameter set into a two-dimensional data matrix according to the time series, and inputting the two-dimensional data matrix into the quality index prediction model to predict the quality index, and obtaining the quality index prediction value;

[0027] Specifically, the relevant production parameter set is sorted according to the timestamp to form a time-ordered parameter sequence. By sampling the time-ordered parameter sequence at equal intervals, parameter samples with fixed time intervals are obtained to ensure the uniformity and representativeness of the data at different time points. The parameter samples with fixed time intervals are organized into a two-dimensional data matrix, the rows of the matrix represent different time points, and the columns represent different parameters, forming an original time series feature matrix. The original time series feature matrix is ​​divided into time windows to obtain the first time series feature submatrix, which helps to capture the time series characteristics of parameter changes in the production process. The first time series feature submatrix is ​​normalized to obtain the second time series feature submatrix, and the value ranges of different parameters are adjusted to the same scale to eliminate the influence of different numerical scales on the training of the prediction model. The second time series feature submatrix is ​​input into the quality indicator prediction model, which includes an input layer, three convolution blocks, two fully connected layers and an output layer, wherein each convolution block is composed of a convolution layer, a batch normalization layer, an activation function layer and a pooling layer. In the first convolution block of the quality indicator prediction model, the second time series feature submatrix of the input is convolved using 64 convolution kernels of size 3x3, with the step size set to 1 and the padding mode set to "same", that is, the output size is made the same as the input by padding to ensure that the feature information is not lost due to the convolution operation. After the first convolution feature vector is obtained through the convolution operation, batch normalization is performed, which helps to stabilize model training, accelerate convergence, and reduce the risk of overfitting. The normalized feature vector is nonlinearly transformed using the ReLU activation function to introduce nonlinear features so that the model can better fit complex data relationships. The feature vector is downsampled using the 2x2 maximum pooling operation to obtain the first transformed feature vector. The pooling operation can effectively reduce the dimension of the feature while retaining important feature information. In the second convolution block, the first transformed feature vector is convolved using 128 3x3 convolution kernels, with the step size still set to 1 and the padding mode set to "same", to obtain the second convolution feature vector. Batch normalization is performed to stabilize the data distribution, and then the ReLU activation function is used for nonlinear transformation to map the linear results to nonlinear results to increase the learning ability of the model. The feature vector is downsampled using a 2x2 maximum pooling operation to obtain the second transformed feature vector, which helps to compress the spatial representation of the features, reduce the amount of calculation and enhance the generalization ability of the model. In the third convolution block, the second convolution feature vector is convolved using 256 3x3 convolution kernels with a step size of 1 and a padding of "same" to obtain a feature map. Batch normalization, ReLU activation function transformation, and 2x2 maximum pooling operations are continued to obtain the third transformed feature vector. The third transformed feature vector is flattened to convert it into a one-dimensional vector for easy input into the fully connected layer.The flattened feature vector is connected to the first fully connected layer, which has 1024 neurons. The ReLU activation function is used for nonlinear transformation to further process the features. At the same time, dropout regularization is applied, and the dropout rate is set to 0.5. Some neurons are randomly discarded to prevent overfitting of the model and improve the generalization performance of the model. The obtained high-level feature representation is input to the second fully connected layer with 512 neurons. The ReLU activation function and regularization with a dropout rate of 0.5 are also used to further extract effective information at a deeper feature level, enhance the model's ability to predict quality indicators, and obtain the target feature representation. The target feature representation is input into the output layer. The number of neurons in the output layer is the same as the number of predicted quality indicators. The linear activation function is used for the final output to obtain the predicted quality indicator value.

[0028] S3, using the quality index prediction values ​​and the temperature-stirring speed data subset as inputs to a two-stage learner-evaluator multi-task sequential neural model to predict tagatose dissolution uniformity and create temperature and stirring speed control strategies;

[0029] It should be noted that the standardized parameter data set after cleaning is divided into a temperature-stirring speed data subset and a dissolution uniformity data subset. The temperature-stirring speed data subset contains important dynamic parameters controlled during the production process, while the dissolution uniformity data subset contains the uniform dissolution degree of tagatose in the solution. An evaluator network is constructed to predict dissolution uniformity. The structure of the evaluator network includes a second input layer, three first hidden layers, and a second output layer. The second input layer is used to receive the temperature-stirring speed data subset. The three first hidden layers contain 64, 128, and 256 neurons, respectively. These neuron layers effectively extract features from the data by gradually increasing their capacity. Each hidden layer uses the ReLU activation function to introduce nonlinear characteristics and enhance the model's ability to express complex data patterns. In order to ensure the stability of the model during training and accelerate convergence, batch normalization operations are used after each hidden layer. The second output layer uses the Sigmoid activation function to output the predicted value of dissolution uniformity. After the evaluator network is built, the quality index prediction value and the temperature-stirring speed data subset are used as input and input into the evaluator network for training to obtain the prediction model of dissolution uniformity. By learning the input data, the evaluator network accurately estimates the dissolution uniformity of tagatose and provides the necessary prediction information for the control strategy in the next stage. The initial learner network is constructed, which is used to generate the control strategy of temperature and stirring speed. The initial learner network consists of the third input layer, four second hidden layers and the third output layer. The third input layer receives the target dissolution uniformity, and the four second hidden layers contain 128, 256, 128 and 64 neurons respectively. These neuron layers enable the network to capture and learn the complex relationships related to the control parameters through the quantitative structure that changes layer by layer. Each second hidden layer uses the Leaky ReLU activation function, which also has a certain activation effect when the input is negative, can effectively avoid the phenomenon of neuron "death" and enhance the learning ability of the network. Dropout regularization is applied in each hidden layer to prevent overfitting of the model and ensure that the model has good generalization ability. The third output layer uses the Tanh activation function to output the control parameters of temperature and stirring speed. The output range of Tanh is -1 to 1, which provides an appropriate amplitude and range of variation for the control parameters. The target dissolution uniformity is input into the initial learner network to obtain the initial temperature and stirring speed control parameters. The initial control parameters are input into the previously trained evaluator network to obtain the predicted dissolution uniformity. Through this process, the control parameters generated by the initial learner network are combined with the prediction results of the evaluator network for iterative optimization. In order to improve the effectiveness of the control strategy, the mean square error between the predicted dissolution uniformity and the target dissolution uniformity is calculated, and the multi-task loss function is constructed in combination with the variation range of temperature and stirring speed. The error of dissolution uniformity and the variation range of control parameters are considered at the same time to achieve a comprehensive optimization effect.In the loss function, the weight of the dissolution uniformity error is 0.7, while the weights of the temperature variation and stirring speed variation are 0.15 respectively, ensuring that the model pays more attention to the accurate prediction of dissolution uniformity, while limiting the drastic fluctuations of temperature and stirring speed to ensure the stability and safety of the production process. The initial learner network is back-propagated and trained using a multi-task loss function. By continuously adjusting the network parameters and gradually reducing the value of the loss function, an optimized target learner network is obtained. The target learner network has a strong generalization ability and can predict the optimal temperature and stirring speed control parameters based on the input target dissolution uniformity, thereby achieving precise control of the tagatose dissolution process. The target dissolution uniformity is input into the target learner network, and the control parameters of temperature and stirring speed are predicted to obtain the final control strategy.

[0030] S4, the temperature and stirring speed control strategies were applied to the tagatose sugar-free tea production process, the quality-related feature set of the production process was collected, and quality assessment and fusion analysis were performed through a multimodal feature fusion network to obtain a comprehensive quality assessment result.

[0031] Specifically, the set of quality-related features is input into the attention-guided multi-scale feature fusion coordination module. The attention mechanism is used to perform weighted fusion of different quality features to achieve feature importance distinction and enhancement. Through the attention mechanism, the network adaptively identifies which features are more critical at a specific time point, so as to focus more on these important features and weaken the influence of minor features. The purpose of multi-scale feature fusion is to capture information at different stages and levels in the production process by introducing feature expressions at different scales, so as to obtain a more comprehensive quality description. The multi-scale fused features are input into the indicator calibration module for feature calibration. In the indicator calibration module, the model adjusts and optimizes the multi-scale fused features according to the preset standards to reduce the deviation between features and ensure that features from different sources are consistent and comparable at the same scale, so as to obtain a set of optimized fused features. The calibrated fused features are input into the adaptive deep fusion module. The adaptive deep fusion module contains multiple residual blocks and an adaptive weight layer, where each residual block consists of two 3x3 convolutional layers and a short-circuit connection. The residual block uses short-circuit connections (jump connections) to enable information to be transmitted more efficiently, avoid the gradient vanishing problem, and enhance the model's deep learning ability. Each convolution layer extracts deeper feature information by performing convolution operations on the features. At the same time, the short-circuit connection allows shallow features to be directly transferred to deeper layers, thereby effectively retaining the information of the original features. The residual block fusion module contains an adaptive weight layer, which uses the softmax function to dynamically adjust the importance of different deep features, so that the network automatically assigns appropriate weights according to the contribution of different features, so that the final global fusion feature can fully reflect the quality status of various aspects of the production process. The global fusion feature is input into the regression head network for quality assessment. The regression head network consists of two fully connected layers. The first fully connected layer contains 512 neurons and uses the ReLU activation function to perform nonlinear transformation on the features, effectively enhancing the model's ability to fit complex data. The second fully connected layer uses a linear activation function to output the final quality assessment score, which reflects the overall situation of the production process, including the influence of control parameters such as temperature and stirring speed on the dissolution uniformity and final quality of tagatose sugar-free tea.

[0032] Data preprocessing is performed on the quality-related feature set, including data standardization and normalization, to eliminate the numerical scale differences between features. At the same time, all features are adjusted to the same dimension to obtain the preprocessed feature matrix. The preprocessed feature matrix is ​​input into different convolution branches in the attention-guided multi-scale feature fusion coordination module. The feature matrix is ​​input into the first convolution branch, in which 64 1x1 convolution kernels are used to perform convolution operations on the input data to obtain the first-scale feature map. The purpose of using 1x1 convolution kernels is to achieve linear combinations between features, so as to capture the point features of various indicators of tagatose sugar-free tea, which mainly reflect the intensity and changes of a single feature at different positions. Through the operation of 1x1 convolution kernels, the response intensity of each feature at each position is effectively extracted. The preprocessed feature matrix is ​​input into the second convolution branch, in which 32 1x1 convolution kernels are used to reduce the dimension of the input data. Through the dimensionality reduction operation, the feature dimension is reduced, the amount of calculation is reduced, and the main feature information is retained. The features after dimensionality reduction are convolved with 64 3x3 convolution kernels to obtain the second-scale feature map. The use of 3x3 convolution kernels enables the model to capture the local correlation characteristics between the various indicators of tagatose sugar-free tea. The local correlation characteristics can reflect the correlation between different indicators in the spatial neighborhood, such as the local effect of temperature change on solubility. The preprocessed feature matrix is ​​input into the third convolution branch, and 16 1x1 convolution kernels are used to reduce the dimension to reduce the computational complexity and retain the main feature information. The features after dimensionality reduction are convolved with 64 5x5 convolution kernels to obtain the third-scale feature map. The use of 5x5 convolution kernels captures the target correlation characteristics in a larger range, which are used to describe the relationship between various indicators in the production process of tagatose sugar-free tea on a larger scale. For example, 5x5 convolution can effectively capture the macroscopic effect of overall temperature control and stirring speed on dissolution uniformity during the production process. The first scale feature map, the second scale feature map and the third scale feature map are spliced ​​in the channel dimension to obtain a multi-scale feature map, so that the model can simultaneously use point features, local correlation features and target range correlation features in subsequent processing. Based on the multi-scale feature map, the channel attention weight is calculated to optimize the feature fusion effect. Global average pooling and global maximum pooling operations are performed on the multi-scale feature map to obtain the average pooling channel descriptor and the maximum pooling channel descriptor respectively. The global average pooling and maximum pooling operations summarize the features of each channel from different angles. Average pooling is used to capture the overall trend in the channel, while maximum pooling is used to capture the significant features in the channel. Through these two operations, two complementary channel descriptors are obtained. The average pooling channel descriptor and the maximum pooling channel descriptor are added element-wise to obtain the fused channel descriptor.The fused channel descriptor is multiplied with the multi-scale feature map channel by channel to dynamically adjust the importance of each channel, emphasize important feature channels, suppress unimportant channels, and achieve weighted fusion of features. The final multi-scale fusion feature is obtained by summing the adjusted features along the channel dimension.

[0033] The multi-scale fusion features are dimensionally analyzed to obtain the dimension size of the input features, and then the feature dimension parameters are determined. Based on the feature dimension parameters, a fully connected layer is constructed, in which the number of input neurons of the fully connected layer is set to the size of the feature dimension parameters, and the number of output neurons is set to 512. In this way, an indicator mapping fully connected layer is constructed, and the multi-scale fusion features are input into the indicator mapping fully connected layer, and nonlinear transformation is performed through the ReLU activation function. The use of the ReLU activation function enables the mapped features to retain nonlinear relationships, which helps to capture complex feature interaction relationships. After the ReLU transformation, the preliminary mapping features are obtained. The preliminary mapping features are statistically analyzed, and the mean and variance of each feature dimension are calculated to obtain feature statistical parameters, which reflect the distribution status of the features in different dimensions. Based on the feature statistical parameters, a BatchNormalization layer, i.e., a feature normalization layer, is constructed. The feature normalization layer standardizes each feature dimension, adjusts the mean of the feature to zero, and adjusts the variance to unit variance, so that the distribution of the features is more concentrated and the scale differences between the features are reduced. Standardization can accelerate the convergence of the model, reduce problems such as gradient disappearance and gradient explosion, and improve the training stability of the model. The preliminary mapped features are input into the feature normalization layer and standardized to obtain standardized features. The standardized features are scaled and translated to enhance the model's adaptability to different data distributions. During the standardization process, the mean and variance of the features are uniformly adjusted to a specific range. In order to ensure that the model can effectively adapt to different input data distributions, the standardized features are appropriately transformed through scaling and translation operations to restore the original dynamic range of some features, improve the model's ability to express diverse inputs, and obtain calibrated fused features. The calibrated fused features are residually connected with the original multi-scale fused features to obtain the final calibrated fused features.

[0034] In one example, a correlation analysis is performed on key parameters in the tagatose sugar-free tea production process to obtain a set of correlation production parameters including temperature, stirring speed, pH value, solubility, tagatose addition amount, tea extraction temperature, and tea extraction time, including:

[0035] High-frequency data collection is performed on key parameters in the production process of tagatose sugar-free tea to obtain a multi-dimensional original parameter data set, and outlier detection, missing value interpolation and data standardization are performed on the multi-dimensional original parameter data set to obtain a cleaned standardized parameter data set;

[0036] The maximum information coefficient is calculated for each pair of parameters in the cleaned standardized parameter data set to obtain the MIC correlation matrix, and based on the cleaned standardized parameter data set, the linear correlation degree of each pair of parameters is calculated using the Pearson correlation coefficient calculation formula to obtain the Pearson correlation matrix;

[0037] The MIC correlation matrix and the Pearson correlation matrix are weightedly fused to obtain a fused correlation matrix, and the key parameters are grouped according to the fused correlation matrix to obtain a parameter clustering result, which includes multiple clusters;

[0038] Based on the parameter clustering results and the pre-defined weight table of factors affecting tagatose solubility characteristics in the expert knowledge base, the weighted scores of the parameters in each cluster were calculated, and the parameters with the highest scores in each cluster were selected to obtain a set of relevant production parameters, which included temperature, stirring speed, pH value, solubility and tagatose addition amount, tea extraction temperature and tea extraction time.

[0039] In this example, parameters such as temperature, stirring speed, pH value, solubility, tagatose addition, tea extraction temperature, and tea extraction time in the production process are collected in real time. Through high-frequency collection, various fluctuations and subtle changes in the production process are captured to form a multi-dimensional original parameter data set. Data cleaning is performed on the multi-dimensional original parameter data set. The cleaning process includes outlier detection, missing value interpolation, and data standardization to ensure the quality of the data. Outlier detection uses statistical methods to identify and eliminate outliers in the data set. Outliers are extreme data points caused by factors such as equipment failure, sensor errors, or improper human operation. Assuming that the numerical distribution of a parameter follows a normal distribution, the triple standard deviation principle is used to detect outliers, that is, if a data point satisfy , then the data point is considered to be an outlier, where represents the mean value of the parameter, Represents the standard deviation. After detecting the outliers, use interpolation or other reasonable methods to replace or delete them. Perform missing value interpolation. When some parameter data is missing due to sensor failure, use mean interpolation, linear interpolation and other methods to fill these missing values ​​so that the data set remains complete. Standardize the cleaned data set and convert parameters of different dimensions to the same scale. Usually, a standardization method with a mean of 0 and a variance of 1 is used to eliminate the impact of the dimension on model training. Perform correlation analysis on the cleaned standardized parameter data set. Calculate the maximum information coefficient (MIC) between each pair of parameters. MIC is an indicator used to measure the correlation between two variables. It can capture linear and nonlinear relationships, so that the model can more comprehensively understand the interaction between parameters. Assumptions and For two variables, the calculation of MIC is defined as:

[0040] ;

[0041] in, express and The mutual information between and They are and The number of partitions, Represents the normalization coefficient. The value of MIC ranges from 0 to 1, and the larger the value, the stronger the relationship between the two variables. The Pearson correlation coefficient is used to calculate the degree of linear correlation between each pair of parameters. The Pearson correlation coefficient is an indicator used to measure the strength of the linear relationship between two variables. Assuming that the two variables and , its Pearson correlation coefficient It is expressed as:

[0042] ;

[0043] in, and They are and The sample value of and They are and The mean of is the number of samples. The value of Pearson correlation coefficient is between -1 and 1, with a value of 1 indicating a complete positive correlation, a value of -1 indicating a complete negative correlation, and a value of 0 indicating no linear relationship. The MIC correlation matrix and the Pearson correlation matrix are weighted and fused to make full use of the respective advantages of MIC and Pearson correlation coefficients. MIC can capture nonlinear relationships, while Pearson can quantify linear relationships. The weighted fusion method is expressed as:

[0044] ;

[0045] in, is the fusion correlation matrix, and are MIC and Pearson correlation matrices, is the fusion weight, which ranges from 0 to 1. , balance the influence of linear and nonlinear relationships in the fusion matrix. Group the key parameters in the production process based on the fusion correlation matrix and perform cluster analysis on the parameters. Group parameters with higher correlation into the same group so that you can better understand the mutual influence between the parameters later. Divide the parameters into multiple clusters through methods such as K-means clustering algorithm or hierarchical clustering, and each cluster represents a group of parameters with strong correlation in the production process. For example, in the production process of tagatose sugar-free tea, temperature and stirring speed are divided into the same cluster. Select the most representative production parameters based on the parameters in the parameter clustering results. Combined with the pre-defined weight table of factors affecting the solubility characteristics of tagatose in the expert knowledge base, calculate the weighted score for the parameters in each cluster. The expert knowledge base provides important domain knowledge about the impact of each parameter on the dissolution process, and uses these weights to effectively calculate the importance score of each parameter. Assume that a cluster contains parameters , and their corresponding impact factor weights are , then the weighted score of each parameter It is expressed as:

[0046] ;

[0047] By calculating the weighted score of each parameter, the parameter with the highest score is selected as the representative parameter of the cluster, and finally a set of relevant production parameters is obtained. The set of relevant production parameters includes temperature, stirring speed, pH value, solubility and tagatose addition amount, tea extraction temperature and tea extraction time.

[0048] In one example, a set of related production parameters is organized into a two-dimensional data matrix according to a time series, and the two-dimensional data matrix is ​​input into a quality indicator prediction model to perform quality indicator prediction, and a quality indicator prediction value is obtained, including:

[0049] The relevant production parameter set is sorted according to the timestamp to obtain a time-ordered parameter sequence, and the time-ordered parameter sequence is sampled at equal intervals to obtain parameter samples at fixed time intervals;

[0050] Organize the parameter samples at fixed time intervals into a two-dimensional data matrix, where the rows of the matrix represent different time points and the columns represent different parameters, to obtain the original time series feature matrix, and divide the original time series feature matrix into time windows to obtain the first time series feature sub-matrix;

[0051] The first time series feature submatrix is ​​normalized to obtain a second time series feature submatrix, and at the same time, the second time series feature submatrix is ​​input into a quality indicator prediction model, the quality indicator prediction model includes a first input layer, three convolution blocks, two fully connected layers and a first output layer, wherein each convolution block includes a convolution layer, a batch normalization layer, an activation function layer and a pooling layer;

[0052] In the first convolution block of the quality indicator prediction model, the second time series feature sub-matrix of the input is convolved using 64 3x3 convolution kernels with a step size of 1 and the same padding to obtain the first convolution feature vector, which is then batch normalized and nonlinearly transformed using the ReLU activation function. Finally, 2x2 maximum pooling is used for downsampling to obtain the first transformed feature vector.

[0053] In the second convolution block of the quality indicator prediction model, the first transformed feature vector is convolved with 128 3x3 convolution kernels, with a step size of 1 and the same padding to obtain the second convolution feature vector, which is then batch normalized, nonlinearly transformed using the ReLU activation function, and finally downsampled using 2x2 maximum pooling to obtain the second transformed feature vector;

[0054] In the third convolutional block of the quality indicator prediction model, the second convolution feature vector is convolved with 256 3x3 convolution kernels, with a step size of 1 and padding of the same to obtain a feature map, which is then batch normalized and nonlinearly transformed using the ReLU activation function. Finally, 2x2 maximum pooling is used for downsampling to obtain the third transformed feature vector.

[0055] Flatten the third transformed feature vector, connect it to a fully connected layer with 1024 neurons, use the ReLU activation function, and apply regularization with a dropout rate of 0.5 to obtain a high-level feature representation, and input the high-level feature representation to a second fully connected layer with 512 neurons, use the ReLU activation function, and apply regularization with a dropout rate of 0.5 to obtain the target feature representation;

[0056] The target feature representation is input to the output layer. The number of neurons in the output layer is the same as the number of predicted quality indicators. A linear activation function is used to obtain the predicted value of the quality indicator.

[0057] In this example, the set of related production parameters is sorted by timestamp to obtain a time-ordered parameter sequence. The production parameters include temperature, stirring speed, pH value, solubility and tagatose addition, tea extraction temperature and tea extraction time, etc. By sorting the parameters in time, the dynamic changes in the production process are obtained, which is used to analyze the impact of each parameter on product quality. The time-ordered parameter sequence is sampled at equal intervals to obtain parameter samples at fixed time intervals. The data is converted into a regular time series to eliminate random errors caused by inconsistent sampling frequencies. The parameter samples at fixed time intervals are organized into a two-dimensional data matrix, where the rows of the matrix represent different time points and the columns represent different parameters, to obtain the original time series feature matrix. The original time series feature matrix is ​​divided into time windows to obtain the first time series feature submatrix, and a part of the continuous time series is intercepted as an independent input to capture the change trend of production parameters in the short term and ensure that the model can learn time correlation. The first time series feature submatrix is ​​normalized to obtain the second time series feature submatrix. The second time series feature submatrix is ​​input into the quality indicator prediction model, which consists of an input layer, three convolution blocks, two fully connected layers, and an output layer. Each convolution block includes a convolution layer, a batch normalization layer, an activation function layer, and a pooling layer. The first convolution block of the model uses 64 convolution blocks of size The convolution kernel performs a convolution operation on the second time series feature submatrix of the input, with a step size of 1 and a padding mode of "same", that is, the output size is made the same as the input by padding to ensure that the feature boundary information is preserved. Assume that the input feature matrix is ,in represents the number of time steps, Represents the number of features, and the convolution operation is expressed as:

[0058] ;

[0059] in, represents the convolution result, represents the weight of the convolution kernel, is the bias term. The first convolution feature vector is obtained through the convolution operation, and then batch normalization is performed to adjust the mean of each small batch of data to 0 and the variance to 1 to accelerate convergence and stabilize the training process. The ReLU activation function is used to perform nonlinear transformation on the feature vector. The ReLU activation function is defined as:

[0060] ;

[0061] Introducing nonlinear characteristics allows the network to fit complex functional relationships. The maximum pooling operation of is used to downsample the feature vector, reduce the size of the feature map, reduce the amount of calculation and improve the generalization ability of the model, and obtain the first transformed feature vector. In the second convolution block, 128 convolution blocks of size The convolution kernel is convolved with a step size of 1 and a padding of "same" to obtain the second convolution feature vector. Batch normalization is performed and the ReLU activation function is used for nonlinear transformation. Finally, The second convolution feature vector is obtained by downsampling the maximum pooling of . In the third convolution block, the second convolution feature vector is convolved using 256 convolution blocks of size The convolution operation is performed with the convolution kernel of , the step size is 1, and the filling mode is "same" to obtain the feature map. Batch normalization is performed, and the ReLU activation function is used for nonlinear transformation. Finally, The third transformed feature vector is obtained by downsampling the maximum pooling operation. Through the extraction of three-layer convolution blocks, the model can learn multi-level features from the time series, thereby effectively expressing the impact of various parameters on quality indicators in the production process. The third transformed feature vector is flattened into a one-dimensional vector and connected to a fully connected layer with 1024 neurons. The fully connected layer multiplies the input vector by the weight matrix and adds a bias term to achieve a linear combination of features, which is defined as:

[0062] ;

[0063] in, is the output of the fully connected layer, is the weight matrix, is the input vector, is the bias term. In this fully connected layer, the ReLU activation function is used for nonlinear transformation, and dropout regularization is applied. The dropout ratio is set to 0.5, which means that 50% of the neurons are randomly set to invalid in each training, thereby reducing the risk of overfitting and improving the generalization performance of the model. After this fully connected layer, a high-level feature representation is obtained. The high-level feature representation is input into the second fully connected layer. The second fully connected layer has 512 neurons. The ReLU activation function is used, and regularization with a dropout rate of 0.5 is applied to further extract and compress features to obtain the target feature representation. The target feature representation is the final expression of the input features, which contains information on the potential impact of all key parameters in the production process on quality indicators. The target feature representation is input into the output layer. The number of neurons in the output layer is the same as the number of quality indicators to be predicted, and the linear activation function is used to output the predicted value of the quality indicator. The linear activation function is defined as:

[0064] ;

[0065] The activation function is chosen so that the output value is a continuous real number to meet the needs of production quality indicators.

[0066] In one example, the quality indicator predictions and the temperature-stirring speed data subset were used as inputs to a two-stage learner-evaluator multi-task sequential neural model to predict tagatose dissolution uniformity and create temperature and stirring speed control strategies, including:

[0067] The cleaned standardized parameter data set is divided into a temperature-stirring speed data subset and a dissolution uniformity data subset;

[0068] Construct an evaluator network, including a second input layer, three first hidden layers and a second output layer, wherein the second input layer receives a subset of temperature-stirring speed data, the three first hidden layers contain 64, 128 and 256 neurons respectively, each first hidden layer uses a ReLU activation function and batch normalization, and the second output layer uses a Sigmoid activation function to output a predicted value of dissolution uniformity;

[0069] The quality index prediction values ​​and the temperature-stirring speed data subset are input into the evaluator network for training to obtain a dissolution uniformity prediction model;

[0070] Construct an initial learner network, including a third input layer, four second hidden layers and a third output layer, wherein the third input layer receives the target dissolution uniformity, the four second hidden layers contain 128, 256, 128 and 64 neurons respectively, each second hidden layer uses a Leaky ReLU activation function and dropout regularization, and the third output layer uses a Tanh activation function to output temperature and stirring speed control parameters;

[0071] Inputting the target dissolution uniformity into the initial learner network to obtain initial temperature and stirring speed control parameters, and inputting the initial temperature and stirring speed control parameters into the evaluator network to obtain predicted dissolution uniformity;

[0072] The mean square error between the predicted dissolution uniformity and the target dissolution uniformity was calculated, and a multi-task loss function was constructed by combining the variation range of temperature and stirring speed, where the weight of dissolution uniformity error was 0.7, the weight of temperature variation range was 0.15, and the weight of stirring speed variation range was 0.15;

[0073] The initial learner network is back-propagated and trained using a multi-task loss function to update the network parameters and obtain the target learner network. The target dissolution uniformity is input into the target learner network for prediction to obtain the temperature and stirring speed control strategies.

[0074] In this example, the cleaned standardized parameter data set is divided into a temperature-stirring speed data subset and a dissolution uniformity data subset. An evaluator network is constructed to predict dissolution uniformity. The evaluator network includes a second input layer, three first hidden layers, and a second output layer. The second input layer is used to receive the temperature-stirring speed data subset. The number of neurons in the three first hidden layers is 64, 128, and 256, respectively, to increase the network's expressive power. Each hidden layer uses the ReLU activation function and batch normalization operation. The ReLU activation function can introduce nonlinear characteristics, while batch normalization helps to stabilize network training and accelerate convergence. The second output layer uses the Sigmoid activation function to output the predicted value of dissolution uniformity. The output range of Sigmoid is 0 to 1, which is suitable for describing the proportion of dissolution uniformity. The quality indicator prediction value and the temperature-stirring speed data subset are input into the evaluator network for training to obtain a prediction model for dissolution uniformity. By learning the relationship between temperature and stirring speed and dissolution uniformity, the model can effectively predict dissolution uniformity under different conditions. In order to optimize the production parameters, an initial learner network is constructed to infer the appropriate temperature and stirring speed control parameters through the target dissolution uniformity. The initial learner network includes a third input layer, four second hidden layers and a third output layer. The third input layer receives the target dissolution uniformity, and the four second hidden layers contain 128, 256, 128 and 64 neurons respectively. Through the layer-by-layer network structure, the model can gradually extract feature information at different levels. Each second hidden layer uses the Leaky ReLU activation function, which retains a certain negative output when the input is negative, thereby effectively avoiding the "neuron death" phenomenon. At the same time, dropout regularization is applied in the hidden layer to reduce the risk of overfitting of the network and ensure the generalization ability of the model. The third output layer uses the Tanh activation function to output the temperature and stirring speed control parameters. The output range of Tanh is between -1 and 1, providing a suitable output amplitude to adjust the temperature and stirring speed. The target dissolution uniformity is input into the initial learner network to obtain the initial temperature and stirring speed control parameters, and the initial control parameters are input into the evaluator network to obtain the predicted dissolution uniformity. In order to ensure that the obtained temperature and stirring speed control parameters can make the dissolution uniformity close to the target value, the mean square error (MSE) between the predicted dissolution uniformity and the target dissolution uniformity is calculated. The MSE is expressed as:

[0075] ;

[0076] in, The target dissolution uniformity is is the predicted dissolution uniformity, is the number of samples. The smaller the mean square error, the closer the predicted value is to the target value. In actual operation, the variation range of temperature and stirring speed is considered, and a multi-task loss function is constructed to combine the dissolution uniformity error with the variation range of control parameters. The variation range of temperature and stirring speed is expressed by the following formula:

[0077] ;

[0078] in, and Respectively represent The variation of temperature and stirring speed in each sample, is the number of samples. The final multi-task loss function combines the dissolution uniformity error, temperature variation, and stirring speed variation, and is expressed as:

[0079] ;

[0080] in, and They are the loss terms for changes in temperature and stirring speed, with weights of 0.15 respectively, while the weight of the dissolution uniformity error is 0.7. The weight distribution ensures that the model pays more attention to the accuracy of dissolution uniformity during the optimization process, while also controlling the drastic changes in temperature and stirring speed to ensure the stability of the production process. The initial learner network is back-propagated and trained using a multi-task loss function, and the network parameters are gradually updated so that the learner network can output more reasonable temperature and stirring speed control parameters. Through continuous iteration and training, the target learner network is obtained. The target learner network can output the optimal temperature and stirring speed control strategy based on the input target dissolution uniformity to ensure that the dissolution uniformity in the production process can meet the expected standards.

[0081] In one example, the temperature and stirring speed control strategy was applied to the tagatose sugar-free tea production process, and the quality-related feature set of the production process was collected. The quality assessment and fusion analysis were performed through a multimodal feature fusion network to obtain a comprehensive quality assessment result, including:

[0082] The temperature and stirring speed control strategies are applied to the tagatose sugar-free tea production process, and the quality-related feature set of the production process is collected and input into the multimodal feature fusion network, which includes an attention-guided multi-scale feature fusion coordination module, an indicator calibration module, and an adaptive deep fusion module.

[0083] The quality-related feature set is input into the attention-guided multi-scale feature fusion coordination module, and the quality-related feature set is weightedly fused to obtain the multi-scale fusion feature;

[0084] Input the multi-scale fusion features into the index calibration module for feature calibration to obtain the calibrated fusion features;

[0085] The calibrated fusion features are input into the adaptive deep fusion module. The adaptive deep fusion module contains multiple residual blocks and an adaptive weight layer. Each residual block consists of two 3x3 convolutional layers and a short-circuit connection. The adaptive weight layer uses the softmax function to dynamically adjust the importance of different deep features to obtain the global fusion features.

[0086] The global fusion features are input into the regression head network for quality assessment to generate a comprehensive quality assessment result. The regression head network includes two fully connected layers. The first fully connected layer in the regression head network uses 512 neurons and ReLU activation function, and the second fully connected layer in the regression head network uses a linear activation function to output the quality assessment score.

[0087] In this example, the optimized temperature and stirring speed control strategies are applied to the actual production process. At the same time, the quality-related feature set of the production process is collected in real time, including parameters such as solubility, dissolution uniformity, pH value, and tagatose concentration. The quality-related feature set is input into the multimodal feature fusion network, which generates a comprehensive evaluation of product quality by comprehensive processing of multi-source features. The multimodal feature fusion network consists of three main modules: an attention-guided multiscale feature fusion coordination module, an indicator calibration module, and an adaptive deep fusion module. Each module processes and optimizes features at different stages to ensure the accuracy and stability of the final quality assessment results. The quality-related feature set is input into the attention-guided multiscale feature fusion coordination module. This module performs weighted fusion of features of different scales through the attention mechanism to capture feature information at different levels in the production process. The process of multiscale feature fusion is implemented using convolution operations of different scales. Each convolution branch performs convolution operations of different sizes on the input feature matrix to extract point features, local correlation features, and target range features. For example, let the feature matrix be ,in represents the time step, Indicates the number of features. Used in a convolution branch The convolution kernel extracts point features, and the result is expressed as , the calculation formula is:

[0088] ;

[0089] in, for The weight matrix of the convolution kernel, * represents the convolution operation, is the bias term. Used in other convolution branches and The convolution kernel extracts local correlation features and target range features, and the results are and By splicing feature maps of different scales together, a multi-scale fusion feature matrix is ​​obtained. , which fully reflects the multi-level associations between different features in the production process. The multi-scale fusion feature matrix is ​​input into the index calibration module for feature calibration. The index calibration module performs linear mapping and standardization on the features through the full connection layer and normalization operation to eliminate the scale differences between the features and ensure the uniformity of the features. The dimension is ,in Represents the number of features after concatenation. Input to a fully connected layer, which maps the input features to a low-dimensional space, and the output is:

[0090] ;

[0091] in, is the weight matrix of the fully connected layer, represents matrix multiplication, is the bias term. The mapped features are batch normalized to adjust the mean of each feature to zero and the variance to 1 to obtain a standardized feature matrix The calibrated fusion features are input into the adaptive deep fusion module. The adaptive deep fusion module includes multiple residual blocks and an adaptive weight layer. The residual block retains the information of the input features through short-circuit connections, thereby effectively solving the gradient vanishing problem in the deep network and improving the training effect. Each residual block consists of two The convolution layer is composed of a short-circuit connection, and the input feature is , the intermediate features obtained after the first convolutional layer are:

[0092] ;

[0093] in, is the weight of the convolution kernel, is the bias term, and ReLU is the activation function. After the second convolutional layer, the output is , and then input the feature Add to the output, that is:

[0094] ;

[0095] The input features are combined with the output features to enhance the feature expression capability. The features processed by the residual block are input into the adaptive weight layer. The adaptive weight layer uses the softmax function to dynamically adjust the importance of features at different depths. The output of softmax is the weight of each feature. These weights are used to perform weighted summation of features at different depths to obtain global fusion features. The global fusion features are input into the regression head network for quality assessment. The regression head network consists of two fully connected layers. The first fully connected layer contains 512 neurons and uses the ReLU activation function to perform nonlinear mapping on the input features to enhance the expressiveness of the features. Assume that the input is , the output of the first fully connected layer is:

[0096] ;

[0097] in, is the weight matrix of the fully connected layer, is the bias term. The output The input is fed into the second fully connected layer, which uses a linear activation function to output a comprehensive quality assessment score, expressed as:

[0098] ;

[0099] in, is the weight matrix of the second fully connected layer, is the bias term, The final quality assessment score.

[0100] In one example, the quality-related feature set is input into the attention-guided multi-scale feature fusion coordination module, and the quality-related feature set is weightedly fused to obtain a multi-scale fusion feature, including:

[0101] Perform data preprocessing on the quality-related feature set and adjust all features to the same dimension to obtain a preprocessed feature matrix;

[0102] The preprocessed feature matrix is ​​input into the first convolution branch in the attention-guided multi-scale feature fusion coordination module, and 64 1x1 convolution kernels are used for convolution operation to obtain the first-scale feature map, which mainly captures the point features of various indicators of tagatose sugar-free tea;

[0103] The preprocessed feature matrix is ​​input into the second convolution branch in the attention-guided multi-scale feature fusion coordination module. First, 32 1x1 convolution kernels are used for dimensionality reduction, and then 64 3x3 convolution kernels are used for convolution operation to obtain the second-scale feature map, which mainly captures the local correlation characteristics of various indicators of tagatose sugar-free tea.

[0104] The preprocessed feature matrix is ​​input into the third convolution branch in the attention-guided multi-scale feature fusion coordination module. First, 16 1x1 convolution kernels are used for dimensionality reduction, and then 64 5x5 convolution kernels are used for convolution operation to obtain the third-scale feature map, which mainly captures the target range correlation characteristics of various indicators of tagatose sugar-free tea.

[0105] The first scale feature map, the second scale feature map and the third scale feature map are spliced ​​in the channel dimension to obtain a multi-scale feature map, and the channel attention weight is calculated based on the multi-scale feature map. First, global average pooling and global maximum pooling operations are performed to obtain an average pooling channel descriptor and a maximum pooling channel descriptor respectively, and then the average pooling channel descriptor and the maximum pooling channel descriptor are added element-wise to obtain a fused channel descriptor;

[0106] The fused channel descriptor is multiplied channel by channel with the multi-scale feature map and summed along the channel dimension to obtain the multi-scale fused feature.

[0107] In this example, data preprocessing is performed on a set of quality-related features. The steps of data preprocessing include outlier processing, missing value filling, and standardization, which aim to eliminate noise in the data and ensure the stability of the features. After preprocessing, all features are adjusted to the same dimension to obtain the preprocessed feature matrix. Let the feature matrix be ,in represents the number of time steps, The preprocessed feature matrix is ​​input into the first convolution branch in the attention-guided multi-scale feature fusion coordination module. In the first convolution branch, 64 The convolution kernel of performs convolution operation on the input feature matrix to capture the point features of various indicators in the production process of tagatose sugar-free tea. Convolution realizes the linear combination of each feature at each position, highlighting the individual contribution of each feature. Assume that the weight matrix of the convolution kernel is , then the calculation formula of the first scale feature map is:

[0108] ;

[0109] Among them, * represents the convolution operation, is the first scale feature map, The result of this operation is a feature map with 64 channels, each of which captures the point information of each feature in the production process. The preprocessed feature matrix is ​​input to the second convolution branch in the attention-guided multi-scale feature fusion coordination module. In this branch, 32 The convolution kernel performs dimensionality reduction on the features, reducing the computational complexity while retaining the main feature information. Assume that the weight matrix of the convolution kernel is , the feature matrix after dimension reduction is:

[0110] ;

[0111] in, is the feature after dimensionality reduction, is the bias term. 64 are used for the feature matrix after dimensionality reduction The convolution kernel is used to perform convolution operation to extract local correlation features. The weight matrix of the convolution kernel is , the calculation of the second scale feature map is:

[0112] ;

[0113] This feature map mainly captures the local correlation between various indicators of tagatose sugar-free tea, such as the local effect of temperature and stirring speed on solubility. At the same time, the preprocessed feature matrix is ​​input into the third convolution branch, using 16 The convolution kernel is used to reduce the dimension. Assume that the feature matrix after dimension reduction is , then use 64 The convolution kernel Convolution operation is performed to extract target correlation features in a larger range. The result of the convolution operation is the third scale feature map These feature maps are used to describe the larger range of feature correlations in the production process, such as the macroscopic effect of overall temperature and stirring speed on the uniformity of product dissolution. , the second scale feature map And the third scale feature map Splicing is performed on the channel dimension to obtain a multi-scale feature map . Assume that the size of each feature map is , then the dimension of the concatenated feature map is , where 192=64+64+64. The multi-scale feature map contains rich information from point features, local correlation features to target range features. Calculate the channel attention weight. Perform global average pooling and global maximum pooling operations on the multi-scale feature map to obtain the average pooled channel descriptors respectively. and max-pooled channel descriptor . Global average pooling calculates the average value of each channel, expressed as:

[0114] ;

[0115] in, Indicates The average value of the channels, and Represent the height and width of the feature map respectively. Similarly, the global maximum pooling calculates the maximum value of each channel, expressed as:

[0116] ;

[0117] Add the average pooled channel descriptor and the maximum pooled channel descriptor element-wise to get the fused channel descriptor , which combines the global average information and the global maximum information to more comprehensively reflect the importance of each channel. The fused channel descriptor is multiplied by the multi-scale feature map channel by channel to dynamically adjust the feature importance of each channel and achieve weighted fusion. Suppose the weighted channel feature is , then its calculation is:

[0118] ;

[0119] Through the channel-by-channel multiplication operation, the more important features in the production process are highlighted, and unimportant or redundant features are suppressed. The weighted features are summed along the channel dimension to obtain multi-scale fusion features. This feature integrates important information at different scales and is a highly concentrated expression of various quality characteristics in the production process.

[0120] In one example, the multi-scale fusion feature is input into the index calibration module for feature calibration to obtain the calibrated fusion feature, including:

[0121] Perform dimensional analysis on the multi-scale fusion features to determine the dimensional size of the input features, obtain the feature dimension parameters, and build a fully connected layer based on the feature dimension parameters. Set the number of input neurons of the fully connected layer to the value of the feature dimension parameters, and set the number of output neurons to 512 to obtain the indicator mapping fully connected layer.

[0122] The multi-scale fusion feature is input into the index mapping fully connected layer, and the ReLU activation function is used for nonlinear transformation to obtain the preliminary mapping features. The preliminary mapping features are statistically analyzed to calculate the mean and variance of each feature dimension to obtain the feature statistical parameters.

[0123] A BatchNormalization layer is constructed based on the feature statistical parameters to obtain a feature normalization layer, and the preliminary mapping features are input into the feature normalization layer for normalization processing to obtain standardized features;

[0124] The standardized features are scaled and translated to obtain calibrated fused features, and the calibrated fused features are residually connected with the multi-scale fused features to obtain the calibrated fused features.

[0125] In this example, the multi-scale fusion feature is dimensionally analyzed to determine the dimensionality of the input feature. The dimensionality of the multi-scale fusion feature includes information such as the length of the time series, the number of different features, and the number of channels obtained after the multi-scale fusion operation. Let the multi-scale fusion feature matrix be ,in is the number of time steps, is the number of features, is the number of channels. By analyzing these dimensions, the total dimension size of the input features is . A fully connected layer is constructed based on the feature dimension parameter, and the number of input neurons of the fully connected layer is set to the value of the feature dimension parameter, that is, , and the number of output neurons is set to 512, and the index mapping fully connected layer is obtained. The function of this fully connected layer is to map the high-dimensional multi-scale fusion features into a more compact feature space and extract more abstract feature expressions. is the weight matrix of the fully connected layer, is the bias term, then the initial features after mapping It is expressed as:

[0126] ;

[0127] where · represents matrix multiplication, is the multi-scale fusion feature of the input, is the output preliminary mapping feature. In order to increase the nonlinear expression capability, the preliminary mapping feature is input into the ReLU activation function for nonlinear transformation. The ReLU activation function is defined as:

[0128] ;

[0129] The ReLU function effectively suppresses the negative value part of the feature, making the model more capable of nonlinear fitting. Perform statistical analysis, calculate the mean and variance of each feature dimension, and obtain feature statistical parameters. These statistical parameters are used to measure the central tendency and distribution degree of the feature. Based on the feature statistical parameters, construct the BatchNormalization layer, which is used to standardize the features to eliminate the scale differences between different features and improve the training stability and convergence speed of the model. , the output of batch normalization Calculated as:

[0130] ;

[0131] in, is a small constant used to avoid the denominator being zero. is the normalized feature. After completing the standardization of the features, the standardized features are scaled and translated to enhance the model's adaptability to different data distributions. Suppose the scaling factor is and the offset coefficient is , then the scaled and translated features Calculated as:

[0132] ;

[0133] in, and It is a parameter learned through training, which is used to control the distribution of features after standardization. The calibrated fusion features are combined with the original multi-scale fusion features. Perform residual connection to obtain the final calibration fusion feature The residual connection operation is expressed as:

[0134] ;

[0135] Through the residual connection method, the information of the original multi-scale fusion features is effectively retained, ensuring that key feature information is not lost during the feature calibration and mapping process, while enhancing the learning ability of the network.

[0136] Reference Figure 2 This embodiment provides a production monitoring and management system for tagatose sugar-free tea, including:

[0137] Analysis module 1 is used to perform correlation analysis on key parameters in the production process of tagatose sugar-free tea, and obtain a set of correlation production parameters including temperature, stirring speed, pH value, solubility and tagatose addition amount, tea extraction temperature and tea extraction time;

[0138] Prediction module 2, used for organizing the relevant production parameter set into a two-dimensional data matrix according to the time series, and inputting the two-dimensional data matrix into the quality index prediction model to predict the quality index, and obtain the quality index prediction value;

[0139] Create Module 3 to use the quality indicator prediction values ​​and the temperature-stirring speed data subset as inputs to a two-stage learner-evaluator multi-task sequential neural model to predict the tagatose dissolution uniformity and create temperature and stirring speed control strategies;

[0140] Evaluation module 4 is used to apply the temperature and stirring speed control strategies to the tagatose sugar-free tea production process, collect the quality-related feature set of the production process, and perform quality evaluation and fusion analysis through a multimodal feature fusion network to obtain a comprehensive quality evaluation result.

[0141] In this embodiment, for the specific implementation of each unit in the above system embodiment, please refer to the above method embodiment, which will not be repeated here.

[0142] Reference Figure 3 In an embodiment of the present invention, a computer device is also provided. The computer device may be a server, and its internal structure may be as follows: Figure 3 As shown. The computer device includes a processor, a memory, a display screen, an input device, a network interface and a database connected through a system bus. Among them, the processor designed by the computer is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store the corresponding data in this embodiment. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, the above method is implemented.

[0143] Those skilled in the art will understand that Figure 3 The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present invention, and does not constitute a limitation on the computer device to which the solution of the present invention is applied.

[0144] An embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the above method is implemented. It can be understood that the computer-readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0145] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media provided by the present invention and used in the embodiments may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double-speed data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM.

[0146] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, system, article or method including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, system, article or method. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the presence of other identical elements in the process, system, article or method including the element.

[0147] The above description is only a preferred embodiment of the present invention, and does not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A production monitoring and management method for tagatose sugar-free tea, characterized in that: The following steps are involved: The key parameters in the production process of tagatose sugar-free tea were analyzed for correlation, and a set of correlation production parameters including temperature, stirring speed, pH value, solubility and tagatose addition amount, tea extraction temperature and tea extraction time were obtained; Organizing the correlation production parameter set into a two-dimensional data matrix according to the time series, and inputting the two-dimensional data matrix into a quality indicator prediction model to predict the quality indicator to obtain a quality indicator prediction value, wherein the quality indicator prediction model includes a first input layer, three convolution blocks, two fully connected layers and a first output layer, wherein each convolution block includes a convolution layer, a batch normalization layer, an activation function layer and a pooling layer; The quality index prediction values ​​and the temperature-stirring speed data subset are used as inputs of a two-stage learner-evaluator multi-task sequential neural model to predict the dissolution uniformity of tagatose and create temperature and stirring speed control strategies; The temperature and stirring speed control strategy is applied to the tagatose sugar-free tea production process, the quality-related feature set of the production process is collected, and quality evaluation and fusion analysis are performed through a multimodal feature fusion network to obtain a comprehensive quality evaluation result; specifically comprising: applying the temperature and stirring speed control strategy to the tagatose sugar-free tea production process, collecting the quality-related feature set of the production process, and inputting the quality-related feature set into a multimodal feature fusion network, wherein the multimodal feature fusion network includes an attention-guided multi-scale feature fusion coordination module, an indicator calibration module, and an adaptive deep fusion module; inputting the quality-related feature set into the attention-guided multi-scale feature fusion coordination module, performing weighted fusion on the quality-related feature set, and obtaining a multi-scale fusion feature; The scale fusion feature is input into the index calibration module for feature calibration to obtain the calibrated fusion feature; the calibrated fusion feature is input into the adaptive deep fusion module, the adaptive deep fusion module includes multiple residual blocks and an adaptive weight layer, each residual block is composed of two 3x3 convolutional layers and a short-circuit connection, the adaptive weight layer uses the softmax function to dynamically adjust the importance of different depth features to obtain the global fusion feature; the global fusion feature is input into the regression head network for quality assessment to generate a comprehensive quality assessment result, the regression head network includes two fully connected layers, the first fully connected layer in the regression head network uses 512 neurons and ReLU activation function, and the second fully connected layer in the regression head network uses a linear activation function to output a quality assessment score.

2. The production monitoring and management method of tagatose sugar-free tea according to claim 1, characterized in that: The correlation analysis of the key parameters in the tagatose sugar-free tea production process is performed to obtain a set of correlation production parameters including temperature, stirring speed, pH value, solubility, tagatose addition amount, tea extraction temperature and tea extraction time, including: High-frequency data collection is performed on key parameters in the production process of tagatose sugar-free tea to obtain a multi-dimensional original parameter data set, and outlier detection, missing value interpolation and data standardization are performed on the multi-dimensional original parameter data set to obtain a cleaned standardized parameter data set; Performing maximum information coefficient calculation on each pair of parameters in the cleaned standardized parameter data set to obtain a MIC correlation matrix, and calculating the linear correlation degree of each pair of parameters based on the cleaned standardized parameter data set using a Pearson correlation coefficient calculation formula to obtain a Pearson correlation matrix; The MIC correlation matrix and the Pearson correlation matrix are weightedly fused to obtain a fused correlation matrix, and key parameters are grouped according to the fused correlation matrix to obtain a parameter clustering result, wherein the parameter clustering result includes a plurality of clusters; Based on the parameter clustering results, combined with the tagatose solubility characteristics influencing factor weight table predefined in the expert knowledge base, the weighted score is calculated for the parameters in each cluster, and the parameter with the highest score in each cluster is selected to obtain a set of relevant production parameters, wherein the set of relevant production parameters includes temperature, stirring speed, pH value, solubility and tagatose addition amount, tea extraction temperature and tea extraction time.

3. The production monitoring and management method of tagatose sugar-free tea according to claim 2, characterized in that: The step of organizing the set of related production parameters into a two-dimensional data matrix according to a time series, and inputting the two-dimensional data matrix into a quality indicator prediction model to perform quality indicator prediction to obtain a quality indicator prediction value includes: Sorting the correlation production parameter set according to timestamps to obtain a time-ordered parameter sequence, and sampling the time-ordered parameter sequence at equal intervals to obtain parameter samples at fixed time intervals; Organizing the parameter samples at the fixed time interval into a two-dimensional data matrix, where the rows of the matrix represent different time points and the columns represent different parameters, to obtain an original time series feature matrix, and dividing the original time series feature matrix into time windows to obtain a first time series feature sub-matrix; Normalizing the first time series feature submatrix to obtain a second time series feature submatrix, and inputting the second time series feature submatrix into a quality indicator prediction model; In the first convolution block of the quality indicator prediction model, a convolution operation is performed on the input second time series feature submatrix using 64 3x3 convolution kernels with a step size of 1 and the same padding to obtain a first convolution feature vector, and then batch normalization is performed, a nonlinear transformation is performed using a ReLU activation function, and finally a 2x2 maximum pooling is used for downsampling to obtain a first transformed feature vector; In the second convolution block of the quality index prediction model, the first transformed feature vector is convolved using 128 3x3 convolution kernels with a step size of 1 and the same padding to obtain a second convolution feature vector, and then batch normalization is performed, a ReLU activation function is used for nonlinear transformation, and finally 2x2 maximum pooling is used for downsampling to obtain a second transformed feature vector; In the third convolution block of the quality indicator prediction model, the second convolution feature vector is convolved using 256 3x3 convolution kernels with a step size of 1 and padding of the same to obtain a feature map, and then batch normalization is performed, a ReLU activation function is used for nonlinear transformation, and finally 2x2 maximum pooling is used for downsampling to obtain a third transformed feature vector; Flatten the third transformed feature vector, connect it to a fully connected layer with 1024 neurons, use the ReLU activation function, and apply regularization with a dropout rate of 0.5 to obtain a high-level feature representation, and input the high-level feature representation to a second fully connected layer with 512 neurons, use the ReLU activation function, and apply regularization with a dropout rate of 0.5 to obtain a target feature representation; The target feature representation is input to the output layer, the number of neurons in the output layer is the same as the number of predicted quality indicators, and a linear activation function is used to obtain the predicted value of the quality indicator.

4. The production monitoring and management method of tagatose sugar-free tea according to claim 3, characterized in that: The quality indicator prediction value and the temperature-stirring speed data subset are used as inputs of a two-stage learner-evaluator multi-task sequential neural model to predict the dissolution uniformity of tagatose and create a temperature and stirring speed control strategy, including: Dividing the cleaned standardized parameter data set into a temperature-stirring speed data subset and a dissolution uniformity data subset; Construct an evaluator network, including a second input layer, three first hidden layers and a second output layer, wherein the second input layer receives a subset of temperature-stirring speed data, the three first hidden layers contain 64, 128 and 256 neurons respectively, each first hidden layer uses a ReLU activation function and batch normalization, and the second output layer uses a Sigmoid activation function to output a predicted value of dissolution uniformity; Inputting the quality index prediction value and the temperature-stirring speed data subset into the evaluator network for training to obtain a dissolution uniformity prediction model; Construct an initial learner network, including a third input layer, four second hidden layers and a third output layer, wherein the third input layer receives the target dissolution uniformity, the four second hidden layers contain 128, 256, 128 and 64 neurons respectively, each second hidden layer uses a Leaky ReLU activation function and dropout regularization, and the third output layer uses a Tanh activation function to output temperature and stirring speed control parameters; Inputting the target dissolution uniformity into the initial learner network to obtain initial temperature and stirring speed control parameters, and inputting the initial temperature and stirring speed control parameters into the evaluator network to obtain predicted dissolution uniformity; Calculate the mean square error between the predicted dissolution uniformity and the target dissolution uniformity, and construct a multi-task loss function in combination with the change amplitude of temperature and stirring speed, wherein the weight of the dissolution uniformity error is 0.7, the weight of the temperature change amplitude is 0.15, and the weight of the stirring speed change amplitude is 0.15; The initial learner network is back-propagated and trained using the multi-task loss function to update network parameters to obtain a target learner network. The target dissolution uniformity is input into the target learner network for prediction to obtain a temperature and stirring speed control strategy.

5. The production monitoring and management method of tagatose sugar-free tea according to claim 1, characterized in that: The step of inputting the quality-related feature set into the attention-guided multi-scale feature fusion coordination module, performing weighted fusion on the quality-related feature set, and obtaining a multi-scale fusion feature comprises: Performing data preprocessing on the quality-related feature set and adjusting all features to the same dimension to obtain a preprocessed feature matrix; Input the preprocessed feature matrix into the first convolution branch in the attention-guided multi-scale feature fusion coordination module, and perform a convolution operation using 64 1x1 convolution kernels to obtain a first-scale feature map, which mainly captures the point features of various indicators of tagatose sugar-free tea; The preprocessed feature matrix is ​​input into the second convolution branch in the attention-guided multi-scale feature fusion coordination module, 32 1x1 convolution kernels are first used for dimensionality reduction, and then 64 3x3 convolution kernels are used for convolution operation to obtain a second-scale feature map, which mainly captures the local correlation characteristics of various indicators of tagatose sugar-free tea; The preprocessed feature matrix is ​​input into the third convolution branch in the attention-guided multi-scale feature fusion coordination module, and 16 1x1 convolution kernels are first used for dimensionality reduction, and then 64 5x5 convolution kernels are used for convolution operation to obtain a third-scale feature map, which mainly captures the target range correlation characteristics of various indicators of tagatose sugar-free tea; The first scale feature map, the second scale feature map and the third scale feature map are spliced ​​in channel dimension to obtain a multi-scale feature map, and the channel attention weight is calculated based on the multi-scale feature map, and firstly global average pooling and global maximum pooling operations are performed to obtain an average pooling channel descriptor and a maximum pooling channel descriptor respectively, and then the average pooling channel descriptor and the maximum pooling channel descriptor are added element-wise to obtain a fused channel descriptor; The fused channel descriptor is multiplied by the multi-scale feature map channel by channel, and the sum is taken along the channel dimension to obtain a multi-scale fused feature.

6. The production monitoring and management method of tagatose sugar-free tea according to claim 1, characterized in that: Inputting the multi-scale fusion feature into the index calibration module for feature calibration to obtain the calibrated fusion feature includes: Performing dimensional analysis on the multi-scale fusion feature to determine the dimensional size of the input feature, obtaining a feature dimension parameter, and constructing a fully connected layer based on the feature dimension parameter, setting the number of input neurons of the fully connected layer to the value of the feature dimension parameter, and setting the number of output neurons to 512, to obtain an indicator mapping fully connected layer; Input the multi-scale fusion features into the index mapping fully connected layer, use the ReLU activation function to perform nonlinear transformation to obtain preliminary mapping features, and perform statistical analysis on the preliminary mapping features to calculate the mean and variance of each feature dimension to obtain feature statistical parameters; Constructing a BatchNormalization layer based on the feature statistical parameters to obtain a feature normalization layer, and inputting the preliminary mapping features into the feature normalization layer to perform a normalization process to obtain a standardized feature; The standardized features are scaled and translated to obtain calibrated fused features, and the calibrated fused features are residually connected with the multi-scale fused features to obtain calibrated fused features.

7. A production monitoring and management system for tagatose sugar-free tea, characterized in that: For implementing the steps of the method according to any one of claims 1 to 6, the system comprises: An analysis module is used to perform correlation analysis on key parameters in the production process of tagatose sugar-free tea, and obtain a set of correlation production parameters including temperature, stirring speed, pH value, solubility and tagatose addition amount, tea extraction temperature and tea extraction time; A prediction module, used for organizing the relevant production parameter set into a two-dimensional data matrix according to a time series, and inputting the two-dimensional data matrix into a quality index prediction model to perform quality index prediction to obtain a quality index prediction value; Creating a module for using the quality indicator prediction value and the temperature-stirring speed data subset as inputs to a two-stage learner-evaluator multi-task sequential neural model to predict the dissolution uniformity of tagatose and create a temperature and stirring speed control strategy; The evaluation module is used to apply the temperature and stirring speed control strategy to the tagatose sugar-free tea production process, collect the quality-related feature set of the production process, and perform quality evaluation and fusion analysis through a multimodal feature fusion network to obtain a comprehensive quality evaluation result.

8. A computer device comprising a memory and a processor, wherein a computer program is stored in the memory, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images