Tea quality prediction method and system based on processing process digital twinning construction

Abstract

This invention discloses a method and system for predicting tea quality based on a digital twin of the tea processing process. First, time-series process parameters and sample physicochemical composition data of the entire tea processing process are acquired to construct a basic digital twin dataset. Next, feature correlation analysis is performed based on the mutual information entropy algorithm to identify key processing steps and extract their multi-source heterogeneous feature vectors, thereby constructing a material migration and transformation map to clarify the evolution trajectory of biochemical components. Then, a digital twin model of the processing process is constructed by combining the basic data and the evolution trajectory, and a quality prediction model is obtained by training it using a long short-term memory network. Finally, the real-time processing data stream of the batch to be tested at key stages is input into the model to output the tea quality prediction result. This invention achieves dynamic monitoring and accurate prediction of quality changes during tea processing, which helps to optimize processes and stabilize and improve tea quality.

Description

Technical Field

[0001] This invention relates to the field of tea processing and quality control technology, and in particular to a method and system for predicting tea quality based on digital twins of the processing process. Background Technology

[0002] The formation of tea quality is highly dependent on its complex processing, which involves multiple stages such as withering, fixation, rolling, fermentation, and drying. The dynamic changes in process parameters at each stage profoundly affect the migration and transformation of substances within the tea leaves, ultimately determining its quality characteristics such as color, aroma, taste, and shape. Traditional tea quality control mainly relies on the experience and judgment of processing personnel and post-processing physicochemical testing. This method has significant drawbacks, including strong lag, difficulty in real-time control, and insufficient stability, and cannot meet the needs of modern precision manufacturing and quality traceability.

[0003] In recent years, the integration of IoT, big data, and AI technologies has provided new insights into optimizing industrial production processes. In the field of tea processing, existing research largely focuses on optimizing process parameters in single stages or static analysis of finished product quality, failing to construct a digital mapping of the co-evolution of process flow and material flow from a dynamic, interconnected perspective. This results in insufficient understanding of the quality formation mechanisms during processing, making it difficult to achieve accurate prediction and online control of the final tea quality. Therefore, there is an urgent need for an intelligent method that can deeply integrate multi-source data from the entire processing process, reveal the inherent biochemical evolution laws, and achieve advanced quality prediction, thereby promoting the digital transformation and precise quality control of tea processing. Summary of the Invention

[0004] To address at least one of the aforementioned technical problems, this invention proposes a method and system for predicting tea quality based on digital twins of the processing process.

[0005] The first aspect of this invention provides a method for predicting tea quality based on digital twin construction of the processing procedure, comprising: Acquire the time-series process parameter data and corresponding physicochemical component data of tea samples throughout the entire tea processing process, and construct a digital twin basic dataset of the tea processing process based on the time-series process parameter data and the physicochemical component data of tea samples. Based on the mutual information entropy algorithm, feature correlation analysis is performed on the digital twin basic dataset to identify the processing links that affect tea quality, and multi-source heterogeneous feature vectors of the processing links that affect tea quality are extracted. Based on the multi-source heterogeneous feature vectors, a material migration and transformation map of tea is constructed, and the evolution trajectory of biochemical components affecting tea quality in the processing stage is determined based on the material migration and transformation map. A digital twin model of tea processing is constructed based on the aforementioned digital twin dataset and the evolution trajectory of the biochemical components. The digital twin model is then trained using a long short-term memory network to obtain a tea quality prediction model. The real-time processing data stream of the tea batch to be predicted is obtained at the processing stage that affects the quality of the tea. The real-time processing data stream is then input into the tea quality prediction model to predict the quality of the tea batch and obtain the tea quality prediction result.

[0006] In this solution, the acquisition of time-series process parameter data and corresponding physicochemical component data of tea samples throughout the entire tea processing process, and the construction of a digital twin dataset of the tea processing process based on the time-series process parameter data and the physicochemical component data of tea samples, specifically involves: The time-series process parameter data of each processing stage of the tea processing production line is obtained based on IoT sensors. The time-series process parameter data includes the time-series changes in leaf processing temperature, humidity, fixation temperature, rolling pressure and duration, fermentation environment temperature, humidity and oxygen concentration, and drying temperature. At the synchronous time node of the acquisition of the time sequence process parameter data, the produced tea sample is subjected to spectral analysis to obtain the physicochemical composition data of the tea sample at the corresponding time. The physicochemical composition data includes the content data of tea polyphenols, amino acids, caffeine, and aroma components. A dynamic time warping algorithm is introduced to align the time series of the time-series process parameter data with the sampling time series corresponding to the physicochemical composition data, thereby obtaining time-aligned process-composition data pairs. Based on the sequence of tea processing techniques, the process-component data is segmented and reorganized according to the processing steps of withering, fixation, rolling, fermentation, and drying. A structured data set is constructed with the processing steps as the index, and each processing step contains time-series process parameter data and physicochemical component vectors of tea samples, thus obtaining a digital twin basic dataset of the tea processing process.

[0007] In this solution, the feature correlation analysis of the digital twin dataset based on the mutual information entropy algorithm is performed to identify the processing links that affect tea quality, and multi-source heterogeneous feature vectors of the processing links that affect tea quality are extracted, specifically as follows: Joint probability density estimation is performed on the time-series process parameter data and tea sample physicochemical component data of each processing step in the digital twin basic dataset. The mutual information entropy value between each process parameter and each physicochemical component in each processing step is calculated, and the mutual information entropy matrix between the process parameters and physicochemical components in each processing step is constructed based on the mutual information entropy value. Set a mutual information entropy threshold, mark process parameters in the mutual information entropy matrix that are greater than the mutual information entropy threshold as key process parameters, and mark physicochemical components that are related to the key process parameters as key quality indicators; The processing steps are ranked from high to low based on the correlation strength between key process parameters and key quality indicators, and the processing steps with the highest correlation strength are selected as the processing steps that affect the quality of tea. The time series data of key process parameters affecting the quality of tea in the processing stage are decomposed into energy feature vectors of different frequency bands by wavelet packet transform. The change trend feature vector of key quality indicators is constructed by calculating the first-order and second-order difference features of the change data. The energy feature vector and the trend feature vector are spliced ​​and standardized to form a multi-source heterogeneous feature vector that influences the processing of tea quality.

[0008] In this scheme, the step of constructing a material migration and transformation map of tea based on the multi-source heterogeneous feature vector, and determining the evolution trajectory of biochemical components affecting tea quality in the processing stage based on the material migration and transformation map, specifically involves: Based on graph neural networks, material nodes and reaction edges are constructed, and the energy feature vector representing process parameters and the change trend feature vector representing quality indicators in the multi-source heterogeneous feature vector are respectively mapped to process condition nodes and material state nodes in the graph. Define the reaction types between nodes, including enzyme-catalyzed conversion, pyrolysis reaction, and oxidative polymerization. Connect the process condition nodes and material state nodes according to the reaction types and initialize the graph adjacency matrix. An attention mechanism is introduced, using the time series features of process parameters as query vectors and the sequence features of material states as key vectors. The influence weights of process condition nodes on material state nodes are calculated, and the adjacency matrix is ​​weighted according to the influence weights to construct an initial material migration and transformation map. Based on the evolution sequence of process-component data pairs in the digital twin basic dataset, a graph comparison learning method is used to calculate the difference between the predicted material state sequence generated by the initial material migration and transformation graph and the actual physicochemical component change sequence. The influence weights and node feature representations are then optimized through gradient backpropagation. When the difference is less than the preset loss threshold, the optimized map is used as the final map of material migration and transformation in tea. The feature vectors of the material state nodes corresponding to the key quality indicators in the map are extracted along the time dimension at each tea quality-affecting processing stage. The feature vectors are connected according to the processing order to form the evolution trajectory of biochemical components.

[0009] In this solution, the step of constructing a digital twin model of the tea processing process based on the digital twin basic dataset and the evolution trajectory of the biochemical components, and training the digital twin model based on a long short-term memory network to obtain a tea quality prediction model, specifically involves: Define the state space of the digital twin model of tea processing, which includes the state variables of the time-series process parameters of the processing stage and the material state variables in the evolution trajectory of the biochemical components; Using the time-series process parameter data in the digital twin basic dataset as model input, and the evolution trajectory of the biochemical components as the supervision target of the model state evolution, a state transition sequence with the processing steps as time steps is constructed. A dynamic state transition unit based on a long short-term memory network is constructed. The dynamic state transition unit takes the time-series process parameter state variables and material state variables of the previous time step as inputs, and controls the updating and transmission of state information through forget gate, input gate and output gate, and outputs the predicted value of the material state variables of the current time step. The actual material state vector at each time step in the evolution trajectory of the biochemical components is compared with the predicted value of the material state variable output by the dynamic state transition unit, and the state prediction loss is calculated. The backpropagation algorithm is used to iteratively optimize the network weight parameters in the dynamic state transition unit based on the state prediction loss until the state prediction loss converges to below a preset threshold. The optimized dynamic state transition unit is then integrated with the state space to obtain the trained tea quality prediction model.

[0010] In this solution, the real-time processing data stream of the tea batch to be predicted at the processing stage affecting tea quality is obtained, and the real-time processing data stream is input into the tea quality prediction model to predict the tea quality of the tea batch to be predicted, thereby obtaining the tea quality prediction result. Specifically: The real-time processing data stream of the tea batch to be predicted in the processing stage affecting the tea quality is obtained, and the real-time processing data stream is input into the tea quality prediction model. The tea quality prediction model, based on the real-time processing data stream and the evolution trajectory of biochemical components stored internally, extrapolates the predicted value of the material state variables of the current processing stage step by step through the dynamic state transition unit. The predicted value of the material state variable at the last time step output by the dynamic state transition unit is decoded, and the predicted value of the material state variable is mapped into a physicochemical component content prediction vector. The quality of tea is evaluated based on the predicted vector of physicochemical component content, and the tea quality prediction result is obtained.

[0011] A second aspect of the present invention also provides a tea quality prediction system based on a digital twin of the processing process. The system includes a memory and a processor. The memory includes a tea quality prediction method program based on a digital twin of the processing process. When the processor executes the tea quality prediction method program based on a digital twin of the processing process, it performs the following steps: Acquire the time-series process parameter data and corresponding physicochemical component data of tea samples throughout the entire tea processing process, and construct a digital twin basic dataset of the tea processing process based on the time-series process parameter data and the physicochemical component data of tea samples. Based on the mutual information entropy algorithm, feature correlation analysis is performed on the digital twin basic dataset to identify the processing links that affect tea quality, and multi-source heterogeneous feature vectors of the processing links that affect tea quality are extracted. Based on the multi-source heterogeneous feature vectors, a material migration and transformation map of tea is constructed, and the evolution trajectory of biochemical components affecting tea quality in the processing stage is determined based on the material migration and transformation map. A digital twin model of tea processing is constructed based on the aforementioned digital twin dataset and the evolution trajectory of the biochemical components. The digital twin model is then trained using a long short-term memory network to obtain a tea quality prediction model. The real-time processing data stream of the tea batch to be predicted is obtained at the processing stage that affects the quality of the tea. The real-time processing data stream is then input into the tea quality prediction model to predict the quality of the tea batch and obtain the tea quality prediction result.

[0012] This invention discloses a method and system for predicting tea quality based on a digital twin of the tea processing process. First, time-series process parameters and sample physicochemical composition data of the entire tea processing process are acquired to construct a basic digital twin dataset. Next, feature correlation analysis is performed based on the mutual information entropy algorithm to identify key processing steps and extract their multi-source heterogeneous feature vectors, thereby constructing a material migration and transformation map to clarify the evolution trajectory of biochemical components. Then, a digital twin model of the processing process is constructed by combining the basic data and the evolution trajectory, and a quality prediction model is obtained by training it using a long short-term memory network. Finally, the real-time processing data stream of the batch to be tested at key stages is input into the model to output the tea quality prediction result. This invention achieves dynamic monitoring and accurate prediction of quality changes during tea processing, which helps to optimize processes and stabilize and improve tea quality. Attached Figure Description

[0013] Figure 1 A flowchart of a tea quality prediction method based on digital twin construction of the processing process is shown in this invention. Figure 2 This invention illustrates a flowchart of the evolution trajectory of biochemical components affecting tea quality during the processing stage. Figure 3 The flowchart of the tea quality prediction model obtained by the present invention is shown; Figure 4 The diagram shows a block diagram of a tea quality prediction system based on a digital twin of the processing procedure, according to the present invention. Detailed Implementation

[0014] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0015] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0016] Figure 1 The flowchart of a tea quality prediction method based on digital twin construction of the processing process is shown.

[0017] like Figure 1 As shown, the first aspect of the present invention provides a method for predicting tea quality based on digital twin construction of the processing process, comprising: S102, acquire the time-series process parameter data and corresponding physicochemical component data of tea samples for the entire tea processing process, and construct a digital twin basic dataset of the tea processing process based on the time-series process parameter data and the physicochemical component data of tea samples. S104, Based on the mutual information entropy algorithm, perform feature correlation analysis on the digital twin basic dataset, identify the tea quality affecting the processing links, and extract the multi-source heterogeneous feature vectors of the tea quality affecting the processing links. S106, construct a material migration and transformation map of tea based on the multi-source heterogeneous feature vector, and determine the evolution trajectory of biochemical components that affect the quality of tea in the processing stage based on the material migration and transformation map. S108. Construct a digital twin model of tea processing based on the digital twin basic dataset and the evolution trajectory of biochemical components. Train the digital twin model based on a long short-term memory network to obtain a tea quality prediction model. S110, acquire the real-time processing data stream of the tea batch to be predicted in the processing stage that affects the quality of tea, input the real-time processing data stream into the tea quality prediction model, predict the quality of the tea batch to be predicted, and obtain the tea quality prediction result.

[0018] It should be noted that by collecting and fusing time-series process parameters and synchronous physicochemical component data throughout the entire processing, a spatiotemporally aligned digital twin dataset is constructed. Then, a mutual information entropy algorithm is used for feature correlation analysis to assess the nonlinear correlation between various process parameters and quality indicators, identifying key processing steps that decisively influence tea quality formation. Multi-source heterogeneous feature vectors containing time-frequency domain energy and gradient changes are extracted, achieving effective dimensionality reduction and focusing of key information. Subsequently, a material migration and transformation map is constructed based on graph neural networks and attention mechanisms. This map represents the dynamic interaction and reaction paths between key process conditions and the states of various biochemical components, thereby revealing a clear evolution trajectory of biochemical components. A dynamic state transition model is constructed with a long short-term memory network as its core. Through training, it learns the temporal dependencies of the processing process, simulating the continuous evolution of internal material states under a given process input. Finally, by inputting real-time process data streams into this model, the physicochemical component content prediction vector for the current batch can be derived online, step-by-step, achieving quantitative and forward-looking prediction of tea quality.

[0019] According to an embodiment of the present invention, the step of obtaining time-series process parameter data and corresponding physicochemical component data of tea samples throughout the entire tea processing process, and constructing a digital twin basic dataset of the tea processing process based on the time-series process parameter data and the physicochemical component data of tea samples, specifically involves: The time-series process parameter data of each processing stage of the tea processing production line is obtained based on IoT sensors. The time-series process parameter data includes the time-series changes in leaf processing temperature, humidity, fixation temperature, rolling pressure and duration, fermentation environment temperature, humidity and oxygen concentration, and drying temperature. At the synchronous time node of the acquisition of the time sequence process parameter data, the produced tea sample is subjected to spectral analysis to obtain the physicochemical composition data of the tea sample at the corresponding time. The physicochemical composition data includes the content data of tea polyphenols, amino acids, caffeine, and aroma components. A dynamic time warping algorithm is introduced to align the time series of the time-series process parameter data with the sampling time series corresponding to the physicochemical composition data, thereby obtaining time-aligned process-composition data pairs. Based on the sequence of tea processing techniques, the process-component data is segmented and reorganized according to the processing steps of withering, fixation, rolling, fermentation, and drying. A structured data set is constructed with the processing steps as the index, and each processing step contains time-series process parameter data and physicochemical component vectors of tea samples, thus obtaining a digital twin basic dataset of the tea processing process.

[0020] It should be noted that by deploying IoT sensors and spectral analysis technology, continuous monitoring of process parameters and synchronous sampling of corresponding physicochemical components throughout the entire tea processing stages—withering, fixation, rolling, fermentation, and drying—was achieved, ensuring the comprehensiveness and spatiotemporal synchronization of data acquisition. To address the potential non-uniform and nonlinear offsets between the process flow and component flow in time series, a dynamic time warping algorithm was introduced for time series alignment, accurately matching the causal relationship between process stimuli and material responses, thereby constructing high-fidelity process-component data pairs. Based on this, according to the inherent sequence of tea processing, these aligned data pairs were structurally reorganized according to processing stages, ultimately forming a structured data set indexed by processing stages and embedding time-series process parameters and multi-dimensional physicochemical component vectors.

[0021] According to an embodiment of the present invention, the step of performing feature correlation analysis on the digital twin basic dataset based on the mutual information entropy algorithm to identify the processing links that affect tea quality, and extracting the multi-source heterogeneous feature vectors of the processing links that affect tea quality, specifically involves: Joint probability density estimation is performed on the time-series process parameter data and tea sample physicochemical component data of each processing step in the digital twin basic dataset. The mutual information entropy value between each process parameter and each physicochemical component in each processing step is calculated, and the mutual information entropy matrix between the process parameters and physicochemical components in each processing step is constructed based on the mutual information entropy value. Set a mutual information entropy threshold, mark process parameters in the mutual information entropy matrix that are greater than the mutual information entropy threshold as key process parameters, and mark physicochemical components that are related to the key process parameters as key quality indicators; The processing steps are ranked from high to low based on the correlation strength between key process parameters and key quality indicators, and the processing steps with the highest correlation strength are selected as the processing steps that affect the quality of tea. The time series data of key process parameters affecting the quality of tea in the processing stage are decomposed into energy feature vectors of different frequency bands by wavelet packet transform. The change trend feature vector of key quality indicators is constructed by calculating the first-order and second-order difference features of the change data. The energy feature vector and the trend feature vector are spliced ​​and standardized to form a multi-source heterogeneous feature vector that influences the processing of tea quality.

[0022] It should be noted that by using the mutual information entropy algorithm to evaluate the nonlinear statistical dependence between process parameters and physicochemical components in each processing stage, the process variables with strong driving effects on material transformation and quality formation can be effectively revealed. By setting thresholds and ranking, the key processing stages with the highest correlation strength can be screened from numerous stages, achieving precise positioning of the analysis focus and significantly improving the efficiency and relevance of subsequent model construction. In the feature extraction stage, for the time-series data of key process parameters, wavelet packet transform is used to decompose them into signals of different frequency bands, and energy feature vectors of each frequency band are extracted to characterize the fluctuation patterns and energy distribution at different time scales contained in the dynamic changes of process parameters. For key quality index data, their first-order and second-order difference features are calculated, where the first-order difference reflects the rate of change between adjacent sampling points, and the second-order difference reflects the acceleration of the rate of change. Together, they constitute the feature vector describing the trend and law of quality index change. Finally, these heterogeneous features extracted from the time-frequency domain and the gradient domain are fused and standardized to form a multi-source heterogeneous feature vector that can comprehensively and deeply characterize the dynamic process and state evolution of key processing stages.

[0023] Figure 2 The flowchart illustrates the evolution trajectory of biochemical components that influence tea quality during the processing stage, as determined by this invention.

[0024] According to an embodiment of the present invention, the step of constructing a material migration and transformation map of tea based on the multi-source heterogeneous feature vector, and determining the evolution trajectory of biochemical components affecting tea quality during processing based on the material migration and transformation map, specifically includes: Based on graph neural networks, material nodes and reaction edges are constructed, and the energy feature vector representing process parameters and the change trend feature vector representing quality indicators in the multi-source heterogeneous feature vector are respectively mapped to process condition nodes and material state nodes in the graph. Define the reaction types between nodes, including enzyme-catalyzed conversion, pyrolysis reaction, and oxidative polymerization. Connect the process condition nodes and material state nodes according to the reaction types and initialize the graph adjacency matrix. An attention mechanism is introduced, using the time series features of process parameters as query vectors and the sequence features of material states as key vectors. The influence weights of process condition nodes on material state nodes are calculated, and the adjacency matrix is ​​weighted according to the influence weights to construct an initial material migration and transformation map. Based on the evolution sequence of process-component data pairs in the digital twin basic dataset, a graph comparison learning method is used to calculate the difference between the predicted material state sequence generated by the initial material migration and transformation graph and the actual physicochemical component change sequence. The influence weights and node feature representations are then optimized through gradient backpropagation. When the difference is less than the preset loss threshold, the optimized map is used as the final map of material migration and transformation in tea. The feature vectors of the material state nodes corresponding to the key quality indicators in the map are extracted along the time dimension at each tea quality-affecting processing stage. The feature vectors are connected according to the processing order to form the evolution trajectory of biochemical components.

[0025] It should be noted that by constructing a material migration and transformation map within tea leaves and extracting the evolution trajectory of biochemical components, the complex processing biochemical process is transformed into a visualized dynamic network, allowing for an intuitive understanding of the transformation paths between substances under specific processing conditions. The introduced attention mechanism can quantify the dynamic weights of the influence of different process parameters on the state of matter, such as accurately characterizing the temporal influence of temperature on the formation of theaflavins during fermentation, thus providing a direct basis for precise process control. The map models the synergistic and chain reactions between multiple material nodes through graph neural networks, and the extracted evolution trajectory is a high-fidelity continuous state sequence containing the reaction mechanism. Ultimately, the optimized map itself constitutes a computable digital experimental field, supporting the simulation and deduction of quality changes by intervening in process nodes, realizing virtual debugging and optimization of the process. Among them, the state of matter is a multi-dimensional dynamic feature vector, which not only includes the real-time absolute content of core physicochemical components such as tea polyphenols, amino acids, caffeine, and key aroma components, but also integrates the changing trend characteristics describing the rate and acceleration of these components' changes, together forming a complete dynamic snapshot characterizing the instantaneous dynamic properties of processing. The state of matter includes the real-time content of physicochemical components such as tea polyphenols, amino acids, and caffeine.

[0026] Figure 3 The flowchart of the tea quality prediction model obtained by the present invention is shown.

[0027] According to an embodiment of the present invention, the step of constructing a digital twin model of the tea processing process based on the digital twin basic dataset and the evolution trajectory of the biochemical components, and training the digital twin model based on a long short-term memory network to obtain a tea quality prediction model, specifically includes: Define the state space of the digital twin model of tea processing, which includes the state variables of the time-series process parameters of the processing stage and the material state variables in the evolution trajectory of the biochemical components; Using the time-series process parameter data in the digital twin basic dataset as model input, and the evolution trajectory of the biochemical components as the supervision target of the model state evolution, a state transition sequence with the processing steps as time steps is constructed. A dynamic state transition unit based on a long short-term memory network is constructed. The dynamic state transition unit takes the time-series process parameter state variables and material state variables of the previous time step as inputs, and controls the updating and transmission of state information through forget gate, input gate and output gate, and outputs the predicted value of the material state variables of the current time step. The actual material state vector at each time step in the evolution trajectory of the biochemical components is compared with the predicted value of the material state variable output by the dynamic state transition unit, and the state prediction loss is calculated. The backpropagation algorithm is used to iteratively optimize the network weight parameters in the dynamic state transition unit based on the state prediction loss until the state prediction loss converges to below a preset threshold. The optimized dynamic state transition unit is then integrated with the state space to obtain the trained tea quality prediction model.

[0028] It should be noted that by defining the state space of the digital twin model of the tea processing process, the temporal process parameter state variables of the processing steps and the material state variables in the evolution trajectory of biochemical components are incorporated into a unified framework. Next, the temporal process parameter data in the digital twin basic dataset is used as the model input, while the evolution trajectory of biochemical components is used as the supervised target for model state evolution. A state transition sequence with the processing steps as time steps is constructed, thus establishing a temporal correspondence between process stimuli and material state responses at the data level. Then, a dynamic state transition unit based on a long short-term memory network is constructed. This unit uses the temporal process parameter state variables and material state variables of the previous time step as input. Utilizing forget gates, input gates, and output gates, it can effectively control the selection and discarding of historical information, the integration of new information, and the output of the final state, thereby simulating the complex temporal dependence and continuous evolution of material states during processing and outputting the predicted value of the material state variable at the current time step. Finally, the actual material state vector at each time step in the evolution trajectory of biochemical components is compared with the predicted value output by the dynamic state transition unit, and the state prediction loss is calculated. Finally, the backpropagation algorithm is employed to iteratively optimize the network weight parameters in the dynamic state transition unit based on the state prediction loss. This allows the model to minimize the prediction error by continuously adjusting its internal parameters, thereby gradually approximating the actual processing dynamics. When the state prediction loss converges to below a preset threshold, it indicates that the model has learned the inherent laws of material state transition driven by process parameters. At this point, the optimized dynamic state transition unit is integrated with the state space, resulting in the trained tea quality prediction model. This model simulates the continuous changes in the internal components of tea during processing, ultimately achieving time-step online extrapolation and forward-looking prediction of the content of key physicochemical components in tea.

[0029] According to an embodiment of the present invention, the step of acquiring the real-time processing data stream of the tea batch to be predicted in the processing stage affecting tea quality, inputting the real-time processing data stream into the tea quality prediction model, and predicting the tea quality of the tea batch to be predicted to obtain the tea quality prediction result is as follows: The real-time processing data stream of the tea batch to be predicted in the processing stage affecting the tea quality is obtained, and the real-time processing data stream is input into the tea quality prediction model. The tea quality prediction model, based on the real-time processing data stream and the evolution trajectory of biochemical components stored internally, extrapolates the predicted value of the material state variables of the current processing stage step by step through the dynamic state transition unit. The predicted value of the material state variable at the last time step output by the dynamic state transition unit is decoded, and the predicted value of the material state variable is mapped into a physicochemical component content prediction vector. The quality of tea is evaluated based on the predicted vector of physicochemical component content, and the tea quality prediction result is obtained.

[0030] It should be noted that by acquiring real-time processing data streams of the batch to be predicted at key processing stages and inputting them into a pre-trained tea quality prediction model, the model can utilize its internally learned state transition patterns, shaped by historical data and biochemical mechanisms, combined with prior knowledge provided by the evolution trajectories of internally stored biochemical components, to drive the dynamic state transition unit to extrapolate the material state of the current batch step by step. This extrapolation is not a simple endpoint estimation, but rather a simulation of the continuous dynamic process of tea's internal quality changing with processing conditions, thus allowing for a forward-looking insight into the quality formation path before processing is complete. Subsequently, the predicted values ​​of the material state variables at the final time step output by the dynamic state transition unit are decoded and mapped into quantifiable physicochemical component content prediction vectors, such as the content of specific components like tea polyphenols and amino acids. Finally, the tea quality is evaluated based on these physicochemical component content prediction vectors, transforming quality judgments that previously relied on post-processing sensory evaluation and offline testing into objective, quantifiable prediction results obtainable in real time during processing. The real-time processing data stream includes time-series data on leaf processing temperature, humidity, fixation temperature, rolling pressure and duration, and fermentation environment temperature, humidity, and oxygen concentration. The physicochemical component content prediction vector includes the predicted contents of tea polyphenols, amino acids, caffeine, and aroma components.

[0031] According to an embodiment of the present invention, it further includes: Obtain the original sensor real-time data stream corresponding to the key process parameters that affect the quality of tea in the processing stage, and simultaneously collect the standard reference measurement data at the time point corresponding to the original sensor real-time data stream. Construct the sensor drift detection dataset for the current processing batch based on the original sensor real-time data stream and the standard reference measurement data. Based on the sensor drift detection dataset, the original sensor real-time data stream is divided into sliding windows. Within each sliding window, the dynamic deviation sequence between the original sensor real-time data stream and the measurement data of the standard reference is calculated. The dynamic deviation sequence is then input into a pre-trained sensor health status assessment model. The sensor health status assessment model outputs a real-time drift factor that characterizes the degree of sensor contamination or performance degradation. The real-time drift factor is fused with the original sensor real-time data stream to construct a sensor drift correction state vector. The sensor drift correction state vector is then input into a sensor data correction twin network. The sensor data correction twin network dynamically allocates the weights of historical normal data and current deviation data based on an attention mechanism to reconstruct the corrected time-series process parameter data. The corrected time-series process parameter data is fed back and updated to the input of the digital twin model of the tea processing process, replacing the real-time processing data stream obtained in the original invention steps. This allows the tea quality prediction model to extrapolate and predict tea quality based on the corrected process parameters, resulting in tea quality prediction results that are resistant to data distortion.

[0032] According to an embodiment of the present invention, based on the sensor drift detection dataset, the original sensor real-time data stream is divided into sliding window segments. Within each sliding window, a dynamic deviation sequence between the original sensor real-time data stream and the measurement data of the standard reference is calculated. The dynamic deviation sequence is then input into a pre-trained sensor health status assessment model. The sensor health status assessment model outputs a real-time drift factor characterizing the degree of sensor contamination or performance degradation. Specifically: The original real-time sensor data stream in the sensor drift detection dataset is divided into time windows to generate a series of continuous time-series data segments. For each time-series data segment, the measurement data of the standard reference object aligned with its time is extracted synchronously. The instantaneous deviation between the original real-time sensor data and the measurement data of the standard reference object at each time point is calculated to form the dynamic deviation sequence. Feature extraction is performed on the dynamic deviation sequence. The extracted features include the changing trends of the mean, variance, and moving average of the deviation, as well as the autocorrelation function features of the deviation. The extracted features are combined into a sensor anomaly feature vector. The abnormal feature vector of the sensor is input into the sensor health status assessment model, which is a classification and regression model based on gradient boosting tree, and is trained using drift state data of the sensor from clean to contaminated in the entire process in historical batches. After receiving the sensor's abnormal feature vector, the model outputs a continuously changing real-time drift factor. The real-time drift factor is used to quantify the degree of distortion of the current sensor data relative to the reference state, and the real-time drift factor is passed to the sensor data correction twin network.

[0033] It is important to note that under the harsh conditions of high temperature and humidity during tea processing, sensors are highly susceptible to gradual or abrupt drift in their monitoring data due to tea hairs, moisture condensation, or oil contamination. This distortion at the data source directly contaminates the input of the digital twin model, causing quality predictions and extrapolations based on distorted process parameters to deviate completely from the actual biochemical process, ultimately leading to prediction failure. Therefore, by introducing standard reference measurement data as a baseline truth, a dynamic sequence describing the real-time bias of the sensors can be constructed. Using a sliding window and feature engineering, the original bias sequence is transformed into a sensor anomaly feature vector containing statistical characteristics and time dependencies. The sensor health status assessment model trained on a gradient boosting tree can accurately analyze this feature vector and output a continuously changing real-time drift factor. This factor not only indicates the presence of contamination but also precisely quantifies the degree of distortion. This drift factor, as a key input to the subsequent correction network, provides the sensor data correction twin network with quantitative prior knowledge about the degree of data reliability decay. This enables the network to dynamically weigh the contribution weights of historical reliable data and current distorted data streams when reconstructing process parameters based on an attention mechanism. This results in the generation of corrected time-series process parameter data that is closer to physical reality, ultimately ensuring the quality of the input data of the digital twin model and improving its robustness and reliability in quality prediction.

[0034] Figure 4 The diagram shows a block diagram of a tea quality prediction system based on a digital twin of the processing procedure, according to the present invention.

[0035] A second aspect of the present invention also provides a tea quality prediction system based on a digital twin of the processing process. The system includes a memory 401, a processor 402, and a communication interface 403. The memory includes a tea quality prediction method program based on a digital twin of the processing process. The communication interface is used for data connection and communication between the memory and the processor. When the processor executes the tea quality prediction method program based on the digital twin of the processing process, it performs the following steps: Acquire the time-series process parameter data and corresponding physicochemical component data of tea samples throughout the entire tea processing process, and construct a digital twin basic dataset of the tea processing process based on the time-series process parameter data and the physicochemical component data of tea samples. Based on the mutual information entropy algorithm, feature correlation analysis is performed on the digital twin basic dataset to identify the processing links that affect tea quality, and multi-source heterogeneous feature vectors of the processing links that affect tea quality are extracted. Based on the multi-source heterogeneous feature vectors, a material migration and transformation map of tea is constructed, and the evolution trajectory of biochemical components affecting tea quality in the processing stage is determined based on the material migration and transformation map. A digital twin model of tea processing is constructed based on the aforementioned digital twin dataset and the evolution trajectory of the biochemical components. The digital twin model is then trained using a long short-term memory network to obtain a tea quality prediction model. The real-time processing data stream of the tea batch to be predicted is obtained at the processing stage that affects the quality of the tea. The real-time processing data stream is then input into the tea quality prediction model to predict the quality of the tea batch and obtain the tea quality prediction result.

[0036] This invention discloses a method and system for predicting tea quality based on a digital twin of the tea processing process. First, time-series process parameters and sample physicochemical composition data of the entire tea processing process are acquired to construct a basic digital twin dataset. Next, feature correlation analysis is performed based on the mutual information entropy algorithm to identify key processing steps and extract their multi-source heterogeneous feature vectors, thereby constructing a material migration and transformation map to clarify the evolution trajectory of biochemical components. Then, a digital twin model of the processing process is constructed by combining the basic data and the evolution trajectory, and a quality prediction model is obtained by training it using a long short-term memory network. Finally, the real-time processing data stream of the batch to be tested at key stages is input into the model to output the tea quality prediction result. This invention achieves dynamic monitoring and accurate prediction of quality changes during tea processing, which helps to optimize processes and stabilize and improve tea quality.

[0037] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0038] Alternatively, if the integrated units of this invention are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROM, RAM, magnetic disks, or optical disks.

[0039] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A tea quality prediction method based on a process digital twin construction, characterized by, Includes the following steps: Acquire the time-series process parameter data and corresponding physicochemical component data of tea samples throughout the entire tea processing process, and construct a digital twin basic dataset of the tea processing process based on the time-series process parameter data and the physicochemical component data of tea samples. Based on the mutual information entropy algorithm, feature correlation analysis is performed on the digital twin basic dataset to identify the processing links that affect tea quality, and multi-source heterogeneous feature vectors of the processing links that affect tea quality are extracted. Based on the multi-source heterogeneous feature vectors, a material migration and transformation map of tea is constructed, and the evolution trajectory of biochemical components affecting tea quality in the processing stage is determined based on the material migration and transformation map. A digital twin model of tea processing is constructed based on the aforementioned digital twin dataset and the evolution trajectory of the biochemical components. The digital twin model is then trained using a long short-term memory network to obtain a tea quality prediction model. The real-time processing data stream of the tea batch to be predicted is obtained at the processing stage that affects the quality of the tea. The real-time processing data stream is then input into the tea quality prediction model to predict the quality of the tea batch and obtain the tea quality prediction result.

2. The tea quality prediction method based on the process digital twin construction according to claim 1, characterized in that, The process involves acquiring time-series process parameter data and corresponding physicochemical component data of tea samples throughout the entire tea processing process, and constructing a digital twin dataset of the tea processing process based on the time-series process parameter data and the physicochemical component data of tea samples. Specifically, this includes: The time-series process parameter data of each processing stage of the tea processing production line is obtained based on IoT sensors. The time-series process parameter data includes the time-series changes in leaf processing temperature, humidity, fixation temperature, rolling pressure and duration, fermentation environment temperature, humidity and oxygen concentration, and drying temperature. At the synchronous time node of the acquisition of the time sequence process parameter data, the produced tea sample is subjected to spectral analysis to obtain the physicochemical composition data of the tea sample at the corresponding time. The physicochemical composition data includes the content data of tea polyphenols, amino acids, caffeine, and aroma components. A dynamic time warping algorithm is introduced to align the time series of the time-series process parameter data with the sampling time series corresponding to the physicochemical composition data, thereby obtaining time-aligned process-composition data pairs. Based on the sequence of tea processing techniques, the process-component data is segmented and reorganized according to the processing steps of withering, fixation, rolling, fermentation, and drying. A structured data set is constructed with the processing steps as the index, and each processing step contains time-series process parameter data and physicochemical component vectors of tea samples, thus obtaining a digital twin basic dataset of the tea processing process.

3. The tea quality prediction method based on digital twin construction of the processing process according to claim 1, characterized in that, The method involves performing feature correlation analysis on the digital twin dataset based on the mutual information entropy algorithm to identify the processing steps that affect tea quality, and extracting multi-source heterogeneous feature vectors for these processing steps. Specifically: Joint probability density estimation is performed on the time-series process parameter data and tea sample physicochemical component data of each processing step in the digital twin basic dataset. The mutual information entropy value between each process parameter and each physicochemical component in each processing step is calculated, and the mutual information entropy matrix between the process parameters and physicochemical components in each processing step is constructed based on the mutual information entropy value. Set a mutual information entropy threshold, mark process parameters in the mutual information entropy matrix that are greater than the mutual information entropy threshold as key process parameters, and mark physicochemical components that are related to the key process parameters as key quality indicators; The processing steps are ranked from high to low based on the correlation strength between key process parameters and key quality indicators, and the processing steps with the highest correlation strength are selected as the processing steps that affect the quality of tea. The time series data of key process parameters affecting the quality of tea in the processing stage are decomposed into energy feature vectors of different frequency bands by wavelet packet transform. The change trend feature vector of key quality indicators is constructed by calculating the first-order and second-order difference features of the change data. The energy feature vector and the trend feature vector are spliced ​​and standardized to form a multi-source heterogeneous feature vector that influences the processing of tea quality.

4. The tea quality prediction method based on digital twin construction of the processing process according to claim 1, characterized in that, The process of constructing a substance migration and transformation map within tea leaves based on the multi-source heterogeneous feature vectors, and determining the evolution trajectory of biochemical components affecting tea quality during processing based on the substance migration and transformation map, specifically involves: Based on graph neural networks, material nodes and reaction edges are constructed, and the energy feature vector representing process parameters and the change trend feature vector representing quality indicators in the multi-source heterogeneous feature vector are respectively mapped to process condition nodes and material state nodes in the graph. Define the reaction types between nodes, including enzyme-catalyzed conversion, pyrolysis reaction, and oxidative polymerization. Connect the process condition nodes and material state nodes according to the reaction types and initialize the graph adjacency matrix. An attention mechanism is introduced, using the time series features of process parameters as query vectors and the sequence features of material states as key vectors. The influence weights of process condition nodes on material state nodes are calculated, and the adjacency matrix is ​​weighted according to the influence weights to construct an initial material migration and transformation map. Based on the evolution sequence of process-component data pairs in the digital twin basic dataset, a graph comparison learning method is used to calculate the difference between the predicted material state sequence generated by the initial material migration and transformation graph and the actual physicochemical component change sequence. The influence weights and node feature representations are then optimized through gradient backpropagation. When the difference is less than the preset loss threshold, the optimized map is used as the final map of material migration and transformation in tea. The feature vectors of the material state nodes corresponding to the key quality indicators in the map are extracted along the time dimension at each tea quality-affecting processing stage. The feature vectors are connected according to the processing order to form the evolution trajectory of biochemical components.

5. The tea quality prediction method based on digital twin construction of the processing process according to claim 1, characterized in that, The process involves constructing a digital twin model of the tea processing procedure based on the digital twin dataset and the evolutionary trajectory of the biochemical components. This digital twin model is then trained using a long short-term memory network to obtain a tea quality prediction model. Specifically: Define the state space of the digital twin model of tea processing, which includes the state variables of the time-series process parameters of the processing stage and the material state variables in the evolution trajectory of the biochemical components; Using the time-series process parameter data in the digital twin basic dataset as model input, and the evolution trajectory of the biochemical components as the supervision target of the model state evolution, a state transition sequence with the processing steps as time steps is constructed. A dynamic state transition unit based on a long short-term memory network is constructed. The dynamic state transition unit takes the time-series process parameter state variables and material state variables of the previous time step as inputs, and controls the updating and transmission of state information through forget gate, input gate and output gate, and outputs the predicted value of the material state variables of the current time step. The actual material state vector at each time step in the evolution trajectory of the biochemical components is compared with the predicted value of the material state variable output by the dynamic state transition unit, and the state prediction loss is calculated. The backpropagation algorithm is used to iteratively optimize the network weight parameters in the dynamic state transition unit based on the state prediction loss until the state prediction loss converges to below a preset threshold. The optimized dynamic state transition unit is then integrated with the state space to obtain the trained tea quality prediction model.

6. The tea quality prediction method based on digital twin construction of the processing process according to claim 1, characterized in that, The process involves acquiring real-time processing data streams of the tea batch to be predicted at the stages affecting tea quality, inputting these data streams into the tea quality prediction model, and predicting the tea quality of the tea batch to obtain the tea quality prediction result. Specifically: The real-time processing data stream of the tea batch to be predicted in the processing stage affecting the tea quality is obtained, and the real-time processing data stream is input into the tea quality prediction model. The tea quality prediction model, based on the real-time processing data stream and the evolution trajectory of biochemical components stored internally, extrapolates the predicted value of the material state variables of the current processing stage step by step through the dynamic state transition unit. The predicted value of the material state variable at the last time step output by the dynamic state transition unit is decoded, and the predicted value of the material state variable is mapped into a physicochemical component content prediction vector. The quality of tea is evaluated based on the predicted vector of physicochemical component content, and the tea quality prediction result is obtained.

7. A tea quality prediction system based on digital twin construction of the processing process, characterized in that, The tea quality prediction system based on a digital twin of the processing process includes a storage device and a processor. The storage device includes a tea quality prediction method program based on a digital twin of the processing process. When the processor executes the tea quality prediction method program based on a digital twin of the processing process, it performs the following steps: Acquire the time-series process parameter data and corresponding physicochemical component data of tea samples throughout the entire tea processing process, and construct a digital twin basic dataset of the tea processing process based on the time-series process parameter data and the physicochemical component data of tea samples. Based on the mutual information entropy algorithm, feature correlation analysis is performed on the digital twin basic dataset to identify the processing links that affect tea quality, and multi-source heterogeneous feature vectors of the processing links that affect tea quality are extracted. Based on the multi-source heterogeneous feature vectors, a material migration and transformation map of tea is constructed, and the evolution trajectory of biochemical components affecting tea quality in the processing stage is determined based on the material migration and transformation map. A digital twin model of tea processing is constructed based on the aforementioned digital twin dataset and the evolution trajectory of the biochemical components. The digital twin model is then trained using a long short-term memory network to obtain a tea quality prediction model. The real-time processing data stream of the tea batch to be predicted is obtained at the processing stage that affects the quality of the tea. The real-time processing data stream is then input into the tea quality prediction model to predict the quality of the tea batch and obtain the tea quality prediction result.

Images