Tea leaf processing parameter adaptive optimization control method and system thereof

Abstract

This invention relates to the field of intelligent control technology for tea processing, specifically to an adaptive optimization control method and system for tea processing parameters. The method involves performing near-infrared spectral scanning and hyperspectral imaging on incoming fresh leaves to obtain feature vectors; using a reaction kinetic model combining catechin enzymatic oxidation, chlorophyll demagnesiation, and amino acid thermal degradation with a data-driven residual correction network to calculate the state vectors of internal components; and having a near-end strategy optimization agent output the combined adjustment values ​​for the withering hot air temperature, rolling speed, and drying air inlet temperature, which are then sent to the processing equipment controller for execution. Through cross-process quality state transmission, the withering, rolling, and drying processes are linked for coordinated control. This invention dynamically tracks the state of internal components at the chemical reaction level, solving the technical problem of large quality fluctuations under heterogeneous fresh leaf conditions in existing lookup-based process parameter schemes.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology for tea processing, specifically to an adaptive optimization control method and system for tea processing parameters. Background Technology

[0002] As a traditional Chinese agricultural product and a major international beverage, tea's quality formation mechanism involves the synergistic effects of multiple processes, including fresh leaf picking, withering, fixation, rolling, and drying. The temperature, time, and mechanical intensity of each process decisively influence the transformation pathways of the finished tea's components. The degree of enzymatic oxidation of catechin polyphenols determines the color and flavor concentration of the tea liquor; the degree of chlorophyll demagnesiation determines the vibrancy of the dry tea's color; and the degree of amino acid thermal degradation determines the freshness and aftertaste of the finished tea. Due to differences in fresh leaf raw materials caused by varying altitudes, tea varieties, and harvesting seasons in different producing areas, the sensory quality of the finished tea obtained when the same set of processing parameters is applied to heterogeneous fresh leaf raw materials can fluctuate by 15% to 20%. The batch-to-batch quality stability in the large-scale production of premium teas has long been a core bottleneck restricting industrial upgrading.

[0003] With the penetration of computer vision, near-infrared spectroscopy, and artificial intelligence technologies into the food processing field, intelligent control of tea processing has made some progress in recent years. Chinese patent application CN118077790A discloses an intelligent tea kneading device and control method based on multimodal information. This method uses a combination of industrial cameras, electronic noses, and hyperspectral cameras to collect images, aromas, and spectral data during the tea kneading process. After extracting the tenderness characteristics of fresh leaves, the data is input into a tenderness grade prediction model. The model-predicted tenderness grade is used as a query condition to retrieve a matching combination of fixed process parameters from a pre-established process parameter database. Finally, based on a comprehensive score threshold, a hard-coded control strategy is triggered to adjust the kneading machine speed and kneading time in a stepwise manner.

[0004] However, the aforementioned approach essentially relies on a static mapping and lookup table method to process parameters based on the physical appearance characteristics of fresh leaves. It fails to dynamically perceive and track the state of internal components—such as catechin retention, chlorophyll conversion, and amino acid degradation—that truly determine the quality of the finished product during processing. The parameter combinations fixed in the database become ineffective under heterogeneous fresh leaf conditions due to drift in reaction kinetics. When the moisture content, phenol-to-amino acid ratio, or polyphenol oxidase activity of the fresh leaves deviate significantly from the established database, the process parameters obtained from looking up data based on a fixed tenderness grade cannot guarantee the stability of the retention rate of internal components and sensory quality scores in the finished tea. This still necessitates manual intervention by experienced tea masters on the production line. In other words, the current technology has failed to break through the cognitive paradigm of mapping finished product quality to physical variables and has failed to elevate the chemical state of internal components to a directly controllable object in the processing control loop. This is the fundamental reason for the long-standing bottleneck in the quality fluctuations of large-scale tea processing.

[0005] At the academic literature level, Das et al. proposed a method in their 2021 paper "Improving black tea quality through optimization of withering conditions using artificial neural network and genetic algorithm" published in J. of Food Proc. Preservation, Vol. 45, No. 15273. This method uses artificial neural networks to predict the quality indicators of tea leaves after withering and genetic algorithms to optimize process parameters such as withering temperature and time. However, this method is also based on offline optimization to obtain a fixed combination of parameters. It cannot perform online re-optimization of parameters according to the dynamic process of chemical reactions of internal components during processing. Furthermore, it does not involve the coupling and transmission of quality status between multiple processes and cannot solve the problem of quality fluctuation of heterogeneous raw materials caused by the bottleneck of the above-mentioned cognitive paradigm.

[0006] Therefore, it is necessary to provide a tea processing parameter optimization control method that can dynamically estimate the state variables of internal components online using reaction kinetics, drive process parameters to adaptively optimize under heterogeneous fresh leaf conditions using reinforcement learning agents, and achieve coupled control by transferring quality states between multiple processes. This method fundamentally breaks through the traditional paradigm of statically mapping quality using physical variables and achieves a leap in batch-to-batch stability of the quality of high-quality teas in large-scale processing. Summary of the Invention

[0007] To address the core bottleneck in existing technologies for large-scale tea processing under heterogeneous fresh leaf raw materials, which leads to significant batch-to-batch fluctuations in the quality of finished tea, this invention provides an adaptive optimization control method and system for tea processing parameters. This method integrates an online state estimator with a kinetic model of the internal components' reaction and a data-driven residual correction network, coupled with a proximal strategy optimization agent to jointly optimize parameters for the three processes of fixing, rolling, and drying. While ensuring the hardware structure of the processing equipment remains unchanged, it dynamically tracks and controls catechin retention rate, chlorophyll conversion rate, and amino acid degradation rate at the chemical reaction level, achieving batch-to-batch stability of the internal components and sensory quality of finished tea under heterogeneous fresh leaf raw materials.

[0008] The technical solution of this invention is: an adaptive optimization control method for tea processing parameters, comprising the following steps: performing near-infrared spectral scanning and hyperspectral imaging on incoming fresh leaves to obtain a feature vector of the fresh leaf raw material, wherein the feature vector of the fresh leaf raw material includes moisture content, amino acid content, phenol-amino acid ratio, and color and tenderness grading value; using a reaction kinetic model of catechin enzymatic oxidation, chlorophyll demagnesiation reaction, and amino acid thermal degradation during tea processing, combined with a data-driven residual correction network, and based on the feature vector of the fresh leaf raw material and the process parameters collected by the processing equipment, calculating the state vector of internal components every 30 seconds, wherein the state vector of internal components includes catechin retention rate and leaf... Chlorophyll conversion degree; The feature vector of the fresh leaf raw material and the state vector of the internal components are used as the state input of the reinforcement learning agent. The proximal strategy optimization agent outputs process parameter decision actions, which include the joint adjustment of the hot air temperature of blanching, the rolling speed and the drying air temperature, and sends them to the processing equipment controller for execution; The state vector of the internal components output from the blanching process is used as the cross-process quality state of the rolling process and transmitted to the proximal strategy optimization agent of the rolling process. The cross-process quality state output from the rolling process is transmitted to the proximal strategy optimization agent of the drying process to execute multi-process linkage control.

[0009] The present invention further provides an adaptive optimization control system for tea processing parameters, used to implement the above method. The system includes processing equipment for blanching, rolling, and drying processes connected in series, as well as edge computing nodes communicatively connected to each of the processing equipment. Each edge computing node internally deploys a fresh leaf quality online characterization module, a dynamic estimation module for internal components, a reinforcement learning optimization module, and a multi-process linkage control module. The fresh leaf quality online characterization module connects to a near-infrared spectrometer and a hyperspectral camera, generating a feature vector of fresh leaf raw materials based on the outputs of the near-infrared spectrometer and the hyperspectral camera. The dynamic estimation module for internal components uses a reaction kinetic model coupled with data to drive a residual correction network, calculating a corrected internal component state vector every 30 seconds based on the fresh leaf raw material feature vector and the process parameters collected by each processing equipment. The reinforcement learning optimization module deploys a proximal policy optimization agent, using the fresh leaf raw material feature vector and the corrected internal component state vector as state inputs, and outputs process parameter decision actions for the blanching, rolling, and drying processing equipment. The multi-process linkage control module maintains a cross-process quality state register chain, enabling the orderly transfer of quality states between processes.

[0010] Compared with the prior art, the present invention has the following significant advantages: Firstly, the overall sensory evaluation score of the finished tea can be stably improved by 8 to 12 points compared to the lookup table-based static parameter scheme represented by D1. The mechanism lies in the fact that this invention expands the state space of the near-end strategy optimization agent from the apparent physical characteristics of fresh leaves to a modified internal component state vector that includes catechin retention rate and chlorophyll conversion degree. This allows the agent to adjust the withering hot air temperature, rolling speed, and drying air temperature based on real-time feedback of the chemical state of internal components during strategy iteration. This ensures that the optimization process of process parameters directly targets the chemical determinants of finished product quality. In contrast, the lookup table scheme cannot overcome the enzymatic oxidation kinetic drift due to parameter solidification under heterogeneous fresh leaf conditions. This invention fundamentally eliminates the source of parameter mismatch through closed-loop feedback of chemical state.

[0011] Secondly, the amino acid retention rate of the finished tea can be improved by about 9% compared with the existing technology. The mechanism is that the present invention embeds the amino acid Maillard reaction rate equation into the internal component dynamic estimation module, and explicitly optimizes the amino acid retention rate as an independent reward item in the composite reward function. This enables the near-end strategy optimization agent to actively suppress the amino acid Maillard reaction rate in the drying process parameter decision rather than just controlling the macroscopic shape of the drying time and temperature curve. This mechanism-level optimization path cannot be reached by the lookup table scheme and the offline optimization scheme of the genetic algorithm.

[0012] Third, the coefficient of variation of the internal components and sensory scores of the finished tea under the condition of multiple batches of heterogeneous fresh leaves decreased significantly. The mechanism is that the present invention uses the catechin retention rate output from the fixation process as an additional state input of the rolling process intelligent agent through a cross-process quality state transfer protocol. The estimated value of cell breakage rate output from the rolling process and the corrected internal component state vector are spliced ​​and transferred to the drying process intelligent agent, so that the three processes form a cascade closed loop with the internal component state vector as the connecting variable. The parameter decision deviation of the previous process can be adaptively compensated by the subsequent process in the state feedback, thereby suppressing the accumulation of quality fluctuations at the system level.

[0013] Fourth, the above three types of effects exhibit a significant synergistic effect at the mechanistic level: the fusion estimator of reaction kinetics and data-driven residual correction provides the agent with accurate observation of chemical state; the adaptive optimization capability of reinforcement learning transforms this observation into closed-loop decision-making; and cross-process state transfer extends the closed-loop decision-making from single-process to multi-process coupled control. If any one of the three is missing, the overall solution cannot achieve the above-mentioned quality stability effect, demonstrating a typical 1+1>2 synergy. This is a non-obvious technological advancement that cannot be obtained by simply piling up any element in the existing technology. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the overall process of the adaptive optimization control method for tea processing parameters of the present invention.

[0015] Figure 2 This is a schematic diagram of the overall architecture of the adaptive optimization control system for tea processing parameters of the present invention.

[0016] Figure 3 This is a schematic diagram of the data flow of the reaction kinetic model coupled with the data-driven residual correction network in the component dynamic estimation module.

[0017] Figure 4 A schematic diagram of the state space, action space, and composite reward function structure of the proximal policy optimization agent in the reinforcement learning optimization module.

[0018] Figure 5 This is a timing diagram illustrating the cross-process quality status transmission protocol between the three processes of blanching, rolling, and drying in the multi-process linkage control module. Detailed Implementation

[0019] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only for explaining the present invention and are not intended to limit the scope of protection of the present invention.

[0020] Reference Figure 2In this embodiment, the adaptive optimization control system for tea processing parameters is deployed on a high-quality green tea processing production line with a daily processing capacity of 2 tons of fresh leaves. The production line consists of four sections arranged sequentially from the arrival of the fresh leaves: a fresh leaf spreading area, a fixation process equipment, a rolling process equipment, and a drying process equipment. The fixation process equipment uses a drum-type hot air fixation machine, the rolling process equipment uses a hydraulic pressure rolling machine, and the drying process equipment uses a chain-plate type hot air aroma-enhancing machine. A near-infrared spectrometer and a hyperspectral camera are installed above the fresh leaf conveyor belt at the end of the fresh leaf spreading area to perform non-contact spectral scanning and hyperspectral imaging of the fresh leaf raw materials before entering the fixation process.

[0021] The edge computing node is an industrial-grade computing unit equipped with an Intel Xeon E-series processor, 32GB of memory, and an NVIDIA Jetson AGX Orin inference acceleration module, and is deployed inside the main control cabinet of the production line. The edge computing node communicates with the programmable logic controllers (PLCs) of the blanching, rolling, and drying processing equipment via an industrial Ethernet bus, and with the near-infrared spectrometer, hyperspectral camera, and temperature sensors, wind speed sensors, and leaf temperature infrared thermometers installed inside the processing equipment via an RS-485 serial bus. The edge computing node runs on a Linux real-time operating system, and at the software level, it concurrently runs four software modules in a time-slice polling manner: the online characterization module for fresh leaf quality, the dynamic estimation module for internal components, the reinforcement learning optimization module, and the multi-process linkage control module.

[0022] Reference Figure 1 The operation flow of the method described in this embodiment is as follows: after the fresh leaves enter the factory, the fresh leaf quality online characterization module obtains the feature vector of the fresh leaf raw material; after the fresh leaves enter the blanching process, the internal component dynamic estimation module calculates the corrected internal component state vector every 30 seconds and outputs it to the reinforcement learning optimization module; the proximal strategy optimization agent of the reinforcement learning optimization module outputs the process parameter decision action and sends it to the blanching process processing equipment for execution via the processing equipment controller; at the end of the blanching process, the multi-process linkage control module encapsulates the corrected internal component state vector at the end of this process into the cross-process quality state according to the cross-process quality state transmission protocol, and sequentially transmits it to the proximal strategy optimization agents corresponding to the rolling process and the drying process, forming a closed-loop adaptive optimization control link of three processes.

[0023] The online characterization module for fresh leaf quality is responsible for non-contact, multi-dimensional characterization of the processing suitability of incoming fresh leaves, and its output is a feature vector of the fresh leaf raw material. In this embodiment, this module is implemented in the following specific manner.

[0024] The near-infrared spectrometer used is a short-wave infrared diffuse reflectance spectrometer with a wavelength range of 900nm to 1700nm and a spectral resolution of 3.5nm. It is fixed 150mm above the fresh leaf conveyor belt using a silica fiber optic probe and performs continuous non-contact scanning of the moving leaf layer at a sampling frequency of 50Hz. The near-infrared spectrometer outputs a near-infrared diffuse reflectance spectrum frame sequence of 50 frames per second. After standard normal variable transformation and SG filtering preprocessing, this sequence is input into a partial least squares model pre-trained with calibration set samples. In this embodiment, the partial least squares model uses 2000 fresh leaf samples measured by standard laboratory methods as the calibration set. The moisture content was measured according to GB 5009.3-2016, the amino acid content was measured according to GB / T 8314-2013, and the phenol-amino acid ratio was measured according to GB / T 8313-2018 (tea polyphenols) and GB / T 8314-2013 (amino acids) and then the ratio was calculated. The number of principal components of the partial least squares model was selected through cross-validation as 9 principal components for the moisture content channel, 11 principal components for the amino acid content channel, and 10 principal components for the phenol-amino acid ratio channel. The three channels output in parallel and independently, and the relative root mean square errors of prediction are no higher than 2.8%, 7.5%, and 6.2%, respectively.

[0025] The hyperspectral camera uses a pushbroom hyperspectral imaging system with a spectral range of 400nm to 1000nm, a spectral resolution of 2.5nm, and a spatial resolution of 1280×960 pixels. It is installed 200mm behind the near-infrared spectrometer, directly above the fresh leaf conveyor belt. After spectral cube acquisition, reflectance cubes are obtained through dark field and white board correction. Principal component analysis is then used to reduce the dimensionality to the first five principal components. Hue, saturation, and leaf uniformity are quantified for each frame of the fresh leaf hyperspectral image. The specific rules for the color tenderness grading value are as follows: when the hue value of the fresh leaf is greater than 120HSV, the saturation is greater than 45%, and the leaf uniformity index is greater than 0.85, a value of 5 represents the initial unfolding level of one bud and one leaf; the value is progressively lowered to 1 to represent the paired leaf level. In addition, the online fresh leaf quality characterization module also records the altitude of the production line as an auxiliary raw material characteristic through the digital twin system of the processing production line.

[0026] The online characterization module for fresh leaf quality concatenates the predicted values ​​of the three channels—moisture content, amino acid content, and phenol-amino acid ratio—with the color and tenderness grading value and the altitude of the place of origin in a fixed order into a 5-dimensional feature vector of the fresh leaf raw material. This vector is then written into the shared memory of the edge computing node and serves as the input source for the subsequent dynamic estimation module of internal components and the reinforcement learning optimization module.

[0027] The first component of the reaction kinetic model in the component dynamic estimation module is the catechin enzymatic oxidation rate equation, used to quantify the chemical state of the enzymatic oxidation and inactivation process of catechins under the action of polyphenol oxidase during the fixation process. In this embodiment, the catechin enzymatic oxidation rate equation adopts the following first-order kinetic expression in the form of the Arrhenius equation: The physical meaning, type, value range, unit, acquisition method, and technical effect of each symbol are defined as follows: Let be the instantaneous rate of change of catechin concentration over time, and be a scalar quantity with a range of . Up to 0, unit is The instantaneous progress of the catechin enzymatic oxidation process is calculated from this rate equation. is the pre-exponential factor, and is a scalar constant with a range of values ​​of . to The unit is The upper limit of the reaction frequency is obtained by linear fitting of the calibration experiment to the double logarithmic form of the Arrhenius equation. is the apparent activation energy of the enzymatic oxidation of catechins, and is a scalar constant ranging from 45 to 65, with units of . The value is obtained by kinetic determination at at least 4 temperature points in the range of 150℃ to 280℃ and fitting the slope of the Arrhenius plot. The value directly determines the sensitivity of the reaction rate to temperature. If the value is too small, the temperature feedback sensitivity will be insufficient. If the value is too large, the rate equation will cause numerical overflow in the low temperature region. Let be the universal gas constant, and be a scalar with a value of . The unit is It is determined by physical constants; The absolute temperature at which the reaction occurs is , which is a scalar quantity ranging from 423 to 553, and its unit is . It is obtained by weighted fusion of the current blanching hot air temperature and leaf temperature collected by the processing equipment, and is used to characterize the thermodynamic input in the driving variables; The current catechin concentration is a scalar quantity, ranging from 0 to 1, and the unit is dimensionless retention rate. The initial value is derived from the phenol-amino acid ratio in the characteristic vector of the fresh leaf raw material, and it characterizes the instantaneous content of the reactant. The moisture content influencing factor is specifically defined as follows: ,in The current moisture content is carried by the feature vector of the fresh leaf raw material. Let be the semi-saturated moisture constant, and take the value of . Dimensionless, characterizing the water dependence of enzyme-catalyzed reactions as exhibiting Michaelis saturation characteristics; It is a polyphenol oxidase activity factor, specifically defined as ,in The initial polyphenol oxidase activity ranges from 0.8 to 1.2, with units of U / g. This is the enzyme inactivation constant, with values ​​ranging from [value missing]. The unit is ; This is the enzyme inactivation threshold temperature, taken as 453, in units of... ; To obtain the non-negative component operator, it is used to characterize the exponential decay of polyphenol oxidase activity when the blanching hot air temperature exceeds the enzyme inactivation threshold temperature; This is the operator for the natural exponential function.

[0028] The catechin enzymatic oxidation rate equation is discretely accumulated with a sampling period of 30 seconds to obtain the increment of catechin retention rate at the current moment. ,in The sampling period is 30 seconds, and the cumulative value yields the catechin retention rate at the current moment. (Left side of the equation) Units are ; right side of the equation Units are , middle The unit is That is, dimensionless, therefore The term is dimensionless; Dimensionless retention rate; It is a dimensionless Michaelis saturation factor; unit After normalization, the activity was characterized as a dimensionless relative activity; therefore, the final unit on the right side of the equation is... This is strictly consistent with the left side of the equation.

[0029] During implementation, the dynamic estimation module for internal components integrates the aforementioned rate equation using the Runge-Kutta fourth-order numerical integration method within each 30-second sampling period to obtain the increment of catechin retention rate within that period. This increment is then summed with the catechin retention rate from the previous period to update the current catechin retention rate state component. For the apparent activation energy in heterogeneous fresh leaf scenarios... With pre-exponential factors The drift problem caused by variety differences is dynamically compensated in this embodiment by time-series modeling of the deviation sequence in the reaction kinetics and data-driven residual correction network fusion estimator described in subsequent chapters. Therefore, the rate equation itself only needs to retain the calibration value.

[0030] The second component of the reaction kinetic model is the chlorophyll demagnesification reaction rate equation, used to quantitatively characterize the chemical process of chlorophyll a converting to demagnesified chlorophyll a under the synergistic effects of heat and acidity during the fixation and drying processes. The progress of this reaction directly affects the color and vibrancy of the dried tea and tea liquor. In this embodiment, the chlorophyll demagnesification reaction rate equation adopts the following first-order kinetic expression including an acidity coupling term: The symbols are defined as follows: Let be the instantaneous rate of change of chlorophyll conversion over time, and let be a scalar value ranging from 0 to 1. The unit is Calculated from this rate equation, it characterizes the instantaneous progress of the chlorophyll demagnesiation reaction and is the source of the time derivative of the chlorophyll conversion component in the state vector of the contained components. is the pre-exponential factor for the chlorophyll demagnesiation reaction, and is a scalar constant with a range of values ​​of . to The unit is The upper limit of the reaction frequency was obtained by fitting the offline blanching kinetic calibration experiment of fresh leaves of different altitude varieties. is the apparent activation energy of the chlorophyll demagnesiation reaction, and is a scalar constant ranging from 55 to 75, with units of . If the value is too low, the magnesium removal reaction will quickly saturate and lose its controllable range in the early stage of fixation; if the value is too high, the chlorophyll conversion will be difficult to trigger. In this embodiment, the value is fitted according to the typical green tea fresh leaves. ; This is a universal gas constant, and its definition is the same as that of the aforementioned catechin enzymatic oxidation rate equation, so it will not be repeated here. The current reaction temperature is defined as the same as the aforementioned catechin enzymatic oxidation rate equation, and will not be repeated here; The equilibrium degree of chlorophyll demagnesiation under the current temperature and acidity conditions is represented by a scalar value ranging from 0.85 to 0.98, in dimensionless units, derived from an empirical formula. The calculation yielded, where The temperature balance coefficient is given by a value of [value missing]. ; Let be the chlorophyll conversion degree at the current moment, a scalar value ranging from 0 to 1. The unit is dimensionless and is initialized by the chlorophyll conversion state component of the previous sampling period, representing the instantaneous progress of the reaction object. The acidity coupling factor is specifically defined as follows: ,in The neutral reference acidity is set at 6.0. The apparent acidity of the leaf juice at the current moment is obtained by empirical conversion from the phenol-to-amino acid ratio in the feature vector of the fresh leaf raw material. The acidity sensitivity coefficient, with a value of 0.8 and a dimensionless unit, characterizes the increase in the rate of magnesium removal reaction for every 1 pH unit decrease in the apparent acidity of the leaf sap. The physical equivalent of this is the demagnesiation reaction. The chemical mechanism of catalysis; and The definition is the same as before.

[0031] The chlorophyll demagnesiation reaction rate equation is also discretely accumulated with a 30-second sampling period to obtain the chlorophyll conversion increment within the current sampling period. The chlorophyll conversion degree is summed with the chlorophyll conversion degree of the previous period to update the chlorophyll conversion degree state component at the current moment, and output as the second component of the internal component state vector. The unit on the left side of the equation is... ; right side of the equation for Dimensionless Dimensionless Dimensionless, total unit , consistent with the left side.

[0032] In this embodiment, the chlorophyll magnesium removal reaction rate equation and the catechin enzymatic oxidation rate equation share the same temperature variable. With the same gas constant The integrals of the two equations are synchronously executed by the edge computing node in parallel numerical integration within the same 30s sampling period. After completion, the increments of the two equations are accumulated into the corresponding components in the state vector of the internal components, ensuring that the chemical states of the two types of internal components advance synchronously with the same time base.

[0033] The third component of the reaction kinetic model is the Maillard reaction rate equation for amino acids, used to quantitatively characterize the Maillard reaction between amino acids and reducing sugars under thermal action during the drying process. The Maillard reaction is the main chemical pathway for amino acid loss during the drying stage, and its rate directly determines the freshness and amino acid retention rate of the finished tea, serving as the physical source of the amino acid retention rate reward term in the composite reward function. In this embodiment, the amino acid Maillard reaction rate equation adopts the following first-order kinetic expression, which includes the product of sugar and amino acid concentrations and a moisture constraint term: , The symbols are defined as follows: Let be the instantaneous rate of change of amino acid concentration over time, and let be a scalar with a range of values. Up to 0, unit is The rate is calculated from this rate equation and characterizes the instantaneous consumption rate of amino acids in the Maillard reaction. Let be the pre-exponential factor of the Maillard reaction, and be a scalar constant with a range of values ​​of . to The unit is The amino acid loss kinetics data from historical drying batches were obtained by fitting the slope intercept of the Arrhenius plot. is the apparent activation energy of the Maillard reaction, and is a scalar constant ranging from 70 to 95, with units of . The value of this value determines the sensitivity of amino acid loss to the drying air inlet temperature. and The definition is the same as the aforementioned rate equation, and will not be repeated here; The relative amino acid content at the current moment is a scalar quantity with a value range of 0 to 1, and the unit is a dimensionless relative retention rate. The initial value is obtained by normalizing the amino acid content in the feature vector of the fresh leaf raw material. The relative content of reducing sugars at the current moment is a scalar value, ranging from 0 to 1, with the unit being dimensionless relative content. The initial value is obtained from the offline calibration of the reducing sugar content of fresh leaves, and is determined during the drying process according to... Approximately coupled to chlorophyll conversion; The moisture content of the tea leaves at the current moment is a scalar quantity, ranging from 0.04 to 0.75, and is expressed as a dimensionless mass fraction. It is obtained by inversely calculating the humidity from the humidity sensor inside the drying oven collected by the processing equipment, and characterizes the dependence of the Maillard reaction on moisture. These represent the reaction orders for amino acids, reducing sugars, and moisture content, respectively, and are all scalar constants. In this example, the values ​​are taken according to the typical Maillard reaction mechanism in tea. ,in This reflects the half-power characteristic that the Maillard reaction is suppressed under extremely low moisture conditions due to restricted molecular migration; This is the natural exponential function operator, defined as before.

[0034] The Maillard reaction rate equation for amino acids is activated during the drying process, and reset to zero during the fixation and rolling processes because the temperature or moisture conditions do not meet the reaction initiation threshold. Within a 30-second drying sampling cycle, the incremental amino acid loss in the current cycle is calculated using Runge-Kutta fourth-order numerical integration. The amino acid retention rate is updated to [value] at the current moment. The dynamic estimated value of the amino acid retention rate in the finished tea is output to the reinforcement learning optimization module. The unit on the left side of the equation is... ; right side of the equation for , Dimensionless All are dimensionless, total units , consistent with the left side.

[0035] Compared to the aforementioned rate equations for catechin enzymatic oxidation and chlorophyll demagnesiation, the unique feature of the Maillard reaction rate equation for amino acids lies in its reactants being a two-component product of amino acids and reducing sugars, resulting in a combination of reaction orders. This results in significant temperature-moisture synergistic control characteristics. The proximal strategy optimization agent, in the decision-making process for drying parameters, jointly suppresses the Maillard reaction rate by adjusting the drying inlet air temperature and outlet air humidity. This maximizes amino acid retention while ensuring the tea leaves are dried to a safe moisture level. This synergistic control path at the mechanistic level constitutes one of the core advantages of this invention compared to existing lookup-based process parameter schemes.

[0036] Reference Figure 3 The internal component dynamic estimation module fuses the outputs of the above three types of reaction kinetic rate equations with the output of the data-driven residual correction network to obtain the corrected internal component state vector. The core idea of ​​the fusion structure is that the pure kinetic calculation value has parameter drift problems caused by the differences in polyphenol oxidase activity of different varieties and the differences in leaf structure at different altitudes, while the pure data-driven method faces the problems of limited historical processing batch samples and insufficient interpretability. The fusion of the two in a causal complementary manner can achieve both the physical interpretability of the reaction kinetic model and the empirical correction capability of data-driven methods.

[0037] In this embodiment, the mathematical expression of the fusion estimator is: , The symbols are defined as follows: The current modified internal component state vector is a 3-dimensional vector with each component ranging from 0 to 1, and the unit is dimensionless retention rate or conversion degree. It is calculated by this fusion formula and includes three components: catechin retention rate, chlorophyll conversion degree and amino acid retention rate. It is the main source of state input for the reinforcement learning optimization module. This is a vector of reaction kinetic values ​​calculated by integrating the rate equations of the three types of reaction kinetics mentioned above at the current moment. It is a 3-dimensional vector with the same range and unit as the above. It is obtained from the integral output of the reaction kinetic model in the current sampling period; This is the correction increment vector output by the data-driven residual correction network at the current moment. It is a 3-dimensional vector with values ​​ranging from [value range missing]. to The unit is dimensionless and represents the correction magnitude of the data-driven method to the purely dynamic calculation value. Let be the vector of deviations between the calculated reaction kinetics value and the actual laboratory measurement value at the current moment, defined as ,in The vector of actual values ​​of the contents of the components measured by standard laboratory methods in historical processing batches serves as a supervision signal during the offline training phase. is the time window length of the deviation sequence, and is a scalar constant with a value of 16, which means that the deviation sequence of 16 30-second sampling periods in the past 8 minutes is used as the input history length of the residual correction network; For parameters The bidirectional long short-term memory network (LSTM) temporal mapping operator is used in this embodiment. The LSM is configured as a two-layer stack, with 64 hidden units in each layer for both forward and backward directions. The hidden states of the last layer are mapped to a 3D output vector through a fully connected layer. .

[0038] The offline training process of the data-driven residual correction network is as follows: Data from 500 batches of tea processing covering multiple production areas and varieties is collected. For each batch, the feature vector of the fresh leaf raw material, the reaction kinetics estimation vector for each 30-second sampling period, and the measured values ​​of catechin retention rate, chlorophyll conversion rate, and amino acid retention rate at key time points of each batch, determined using standard laboratory methods, are recorded. The bias sequence is then used for training. For input, with actual deviation The labels are defined as follows: mean squared error is used as the loss function; and the Adam optimizer is used with a batch size of 32 and an initial learning rate of 100%. Parameters of bidirectional long short-term memory networks Offline training was performed, and convergence occurred after 200 training epochs; after convergence... The deployment is permanently installed on the edge computing node. Equation middle and All are dimensionless retention rate or transformation vectors, which, when added together... Units of the same type, consistent on both sides; equation The input to the bidirectional long short-term memory network is a bias sequence vector, and the output is generated as a dimensionless correction increment vector through a fully connected layer, with consistent units on both sides.

[0039] The execution timing of the fusion estimator in each 30s sampling period is as follows: the reaction kinetic model is output first. The processing equipment controller bypasses sampling at a specified time point and sends it to the laboratory for verification to obtain several sparse samples. The data-driven residual correction network performs inference on the most recent time window deviation sequence to obtain the sample, which is used to calculate the deviation and slide to update the deviation sequence buffer. The two are added together to obtain the corrected state vector of the internal components. This fusion structure enables the inherent parameter drift of the pure reaction kinetic model to be dynamically compensated in a time-series manner by a data-driven residual correction network when the polyphenol oxidase activity, carotenoid ratio of the variety, or physiological age at the harvest time of the fresh leaf raw material changes. This achieves high-precision dynamic tracking of the corrected state vector of the internal components, and the fusion estimator controls the calculation deviation of catechin retention rate, chlorophyll conversion degree, and amino acid retention rate to within 5%.

[0040] Reference Figure 4 The reinforcement learning optimization module deploys the proximal policy optimization agent to adaptively optimize the decision-making actions for the process parameters in the three processes of fixing, rolling, and drying. In this embodiment, the proximal policy optimization agent is deployed according to the Actor-Critic architecture, and the state space, action space, and policy optimization objective function are customized for the tea processing scenario.

[0041] The state space of the near-end policy optimization agent is defined as follows: ,in The feature vector of the fresh leaf raw material is a 5-dimensional vector. The modified internal component state vector is a 3-dimensional vector. The process parameter decision action from the previous moment is represented by a 3D vector. The normalized value of the current process's running time is a 1-dimensional scalar with a value range of 0 to 1. The quality state across processes, inherited from the previous process, is represented by a 3-dimensional vector (this component is set to zero in the final drying process). The total dimension of the state vector is 15. After being normalized to zero mean and unit variance, it is input into the Actor network and the Critic network. The action space is defined as a continuous action space. The three components are the adjustment amount of the hot air temperature for blanching. (Corresponding to the step-by-step adjustment of the blanching hot air temperature within the range of 150℃ to 280℃), and the adjustment amount of the kneading speed. (Corresponding to the step adjustment of the kneading speed within the range of 40 r / min to 80 r / min), and the adjustment amount of the drying air inlet temperature. (Corresponding to the step adjustment of the drying air inlet temperature within the range of 80℃ to 130℃).

[0042] The Actor network of the near-end policy optimization agent The Critic network employs a 3-layer fully connected network structure (15-dimensional input layer, 256-dimensional hidden layer, 128-dimensional hidden layer, and an output layer with 3-dimensional Gaussian strategy, representing the mean and log-variance, respectively). A symmetrical 3-layer fully connected network (with the output layer being a 1-dimensional state value scalar) is employed, and both networks use Tanh activation functions to ensure bounded outputs. The policy optimization objective function of the proximal policy optimization agent adopts the PPO-Clip objective function, constructed as follows: , The symbols are defined as follows: For strategy parameters The corresponding PPO-Clip objective function value is a scalar, and its range is not hard-constrained (typical value is within...). to (between), the unit is a dimensionless reward scale, calculated by this formula, representing the expected improvement of the current strategy relative to the old strategy, and is the optimization objective of gradient ascent of the Actor network parameters; The probability ratio of the current policy to the old policy is defined as follows: , is a scalar, typically ranging from 0.5 to 2.0, dimensionless, and represents the relative sampling probability of the same state-action pair under the old and new policies; and The parameters are respectively The new strategy and parameters are The old strategy in the state Down Output Action The probability density is obtained by calculating the three-dimensional Gaussian probability density function output by the Actor network, and both are scalar dimensionless. For the current time step, the advantage estimate is calculated using the generalized advantage estimation method. The calculation yielded, where To assign a value to the discount factor Dimensionless To determine the smoothing factor value for generalized dominance estimation Dimensionless The unit of the single-step reward for the current time step is a dimensionless reward scale. The output of the Critic network for estimating state value is a dimensionless reward scale. The scalar dimensionless reward scale represents the excess return of the current action relative to the baseline value; The PPO policy ratio pruning threshold is a scalar constant, which is set to 0.2 in this embodiment. It is dimensionless. If the value is too large, it will weaken the policy constraint of PPO and cause training divergence. If the value is too small, it will hinder the policy improvement speed. 0.2 is the empirical value recommended by Schulman et al. in their 2017 paper. To select the smaller of the two input values; The clipping operator is defined as follows: Limit the probability ratio to Within the range; The time step expectation operator is obtained based on empirical samples.

[0043] The proximal policy optimization agent performs policy parameter updates according to the proximal policy optimization algorithm proposed by Schulman in 2017: 32 trajectories are collected in parallel within each policy collection cycle, each trajectory containing 128 time steps; based on the collected trajectories, the policy parameters are updated using the aforementioned PPO-Clip objective function with the Adam optimizer and an initial learning rate. For Actor network parameters Perform mini-batch gradient ascent for 10 epochs; simultaneously, adjust the Critic network parameters using mean squared TD error. The training process involves pre-training for 10,000 policy update cycles in a digital twin simulation environment, followed by online fine-tuning on a real production line. (Left side of the equation) For dimensionless reward scales; right side of the equation Dimensionless The product of the two is a dimensionless reward scale; The operator remains dimensionless after it is applied. The operator does not change the unit; the expectation operator does not change the unit, and the right side of the equation is ultimately a dimensionless reward scale, consistent with the left side.

[0044] The composite reward function in the reinforcement learning optimization module is the value feedback source that drives the proximal policy optimization agent to learn the decision-making strategy for the optimal process parameters of tea processing. In this embodiment, the composite reward function is specifically constructed in the following form: , The symbols are defined as follows: For a single batch of compound rewards, is a scalar quantity, and its value range is [value range missing]. to The unit is a dimensionless reward scale, which is calculated by this formula and represents the comprehensive quality score of a single processing batch. It is the direct value signal source for the policy gradient update of the near-end policy optimization agent. The total sensory evaluation score of the finished tea is a scalar value ranging from 0 to 100, expressed in points. It is obtained by a tea evaluation group with senior tea taster qualifications, which scores each batch of finished tea for five factors: appearance, aroma, liquor color, taste, and infused leaf color according to the sensory evaluation method for tea specified in GB / T 23776-2018. The scores are then weighted and summed to represent the overall sensory quality of the finished tea. The amino acid retention rate is a scalar value ranging from 0 to 1, and the unit is a dimensionless relative retention rate. It is obtained by normalizing the amino acid component in the corrected internal component state vector output by the internal component dynamic estimation module of the present invention at the end of the drying process, and characterizes the degree of retention of amino acids as the material basis of freshness during the processing. The deviation of the current batch's quality index from the historical average is given by , where is a scalar quantity and its value range is . to The unit is cents, defined as ,in The moving average of the total sensory evaluation scores of the same type of finished tea from the same historical processing batches is used, with a time window of 30 batches, representing the deviation of the current batch from the historical stable level. It is an absolute value operator; These are preset weighting coefficients, all of which are scalar constants. In this embodiment, they are set to values ​​of [value missing]. ,in The weighting of the sensory evaluation score reflects the degree of importance attached to absolute quality. The amino acid retention rate weight was scaled 100 times to match the sensory score magnitude, representing the degree of importance attached to the material basis of freshness. The batch deviation penalty weight, scaled by a factor of 5, represents the degree of importance attached to quality stability. All three conditions must be met. The relative weight.

[0045] In the step-by-step reward setting of this embodiment Item and The penalty is a batch-level sparse reward injected after the entire batch processing is completed. The dense reward is injected every 30 seconds during the drying process. The three types of reward signals, after time discounting, constitute the generalized dominance estimation. Time-series reward sequence. Left side of the equation. For dimensionless reward scales; right side of the equation for The weights are normalized and standardized to a single reward scale. middle Dimensionless Absorb the scaling transformation from dimensionless to reward scale; Similarly, the scale conversion from absorption score to reward scale, the sum of the three terms is still a dimensionless reward scale, consistent with the left side.

[0046] The three-item combination design of the composite reward function is an original design of this invention for the special engineering context of tea processing. The sensory evaluation total score directly targets the final evaluation standard of finished product quality, the amino acid retention rate directly targets the chemical basis of freshness, and the batch deviation penalty directly targets the quality stability requirements of large-scale processing. The three reward signals work together to drive the near-end strategy optimization agent to learn and accurately adaptively optimize the process parameters under heterogeneous fresh leaf conditions.

[0047] Reference Figure 5The multi-process linkage control module maintains the cross-process quality status register chain, transferring the chemical state of the internal components between the three processes of blanching, rolling, and drying in a continuous manner, so that the three processes form a cascaded closed loop with the modified internal component state vector as the link. In this embodiment, the cross-process quality status transfer protocol is constructed according to the following mathematical form: , The symbols are defined as follows: For the first The initial state vector of the proximal policy optimization agent at the start of the process. Corresponding to the kneading process, For the corresponding drying process, each is a 15-dimensional vector, with the value range and unit being the same as the aforementioned state space definition; For the first The end time of the process is the duration of the process. The final state vector at time , Corresponding to the blanching process, The corresponding kneading process is represented by a 15-dimensional vector. The catechin retention rate is a scalar value at the end of the fixation process, ranging from 0 to 1, and the unit is dimensionless retention rate. It is extracted from the catechin component in the corrected internal component state vector output by the internal component dynamic estimation module at the end of the fixation process. The output state vector of the kneading process at the end of the kneading process is a 4-dimensional vector. The first 3 dimensions are the corrected internal component state vector at the end of the kneading process, namely the three components of catechin retention rate, chlorophyll conversion degree, and amino acid retention rate. The 4th dimension is the estimated value of cell breakage rate output by the kneading process. The value range of each component is 0 to 1. For the state coding operator of the greening to rolling process, the scalar quantity Position-based encoding is expanded into a 3D vector. ,in The phase constant value is embedded in the position of the blanching process. Dimensionless, the encoded vector serves as an additional component of the state during the kneading process; For the state encoding operator of the kneading to drying process, a 4-dimensional vector is used. via fully connected layer The mapping is performed as a 3D vector as an additional component representing the state of the drying process, where... A 3x4 parameter matrix The 3D bias vectors are all pre-trained in a digital twin simulation environment using the sensory evaluation total score of the downstream drying process as a supervisory signal. This is the vector vertical concatenation operator.

[0048] The cross-process quality status transmission protocol is executed in the actual production line according to the following timing sequence: When the processing equipment controller detects the completion signal of the blanching process, the multi-process linkage control module reads the catechin retention rate component from the dynamic estimation module of the internal components of the blanching process at the last moment in the corrected internal component status vector. ,through The encoded state is concatenated with the initial state vector of the kneading process and written into the observation buffer of the proximal policy optimization agent of the kneading process; similarly, at the end of the kneading process, a second-level transfer is performed, passing the output state of the kneading process through... The encoding is injected into the observation buffer of the agent in the drying process. (Left side of the equation) The state vector is a 15-dimensional composite dimension (each component is dimensionless after normalization); the right side of the equation It is also a 15-dimensional state vector with the same dimensions. and The input and output of the operator are both normalized into dimensionless vectors, and after concatenation, their dimensions are strictly consistent with those on the left.

[0049] The technical effect of the cross-process quality status transmission protocol is that, after the catechin retention rate at the end of the fixation process is used as an additional status input for the rolling process intelligent agent, the rolling process intelligent agent can dynamically adapt the rolling mechanical action intensity by referring to the catechin conversion status of the upstream fixation process when selecting rolling pressure and speed; after the cell breakage rate and internal component status at the end of the rolling process are used as the status input for the drying process intelligent agent, the drying intelligent agent can perform feedforward adjustment on the drying air inlet temperature curve to avoid the aggravation of drying amino acid loss due to excessive breakage in the rolling process. The three processes form a cascade closed-loop control link connected by internal components and cell physical status.

[0050] At the end of the kneading process, the multi-process linkage control module needs to transmit the estimated cell breakage rate of the kneaded leaves as a component of the output state of the kneading process to the drying process. Cell breakage rate is a physical indicator of kneading adequacy and is coupled with the moisture evaporation rate and amino acid Maillard reaction kinetics in the subsequent drying process. In this embodiment, the estimated cell breakage rate is calculated based on an empirical kinetic model of the effect of kneading speed and kneading time on the pressing effect. This empirical kinetic model is specifically constructed as follows: , The symbols are defined as follows: The current kneading time The estimated cell breakage rate is a scalar, ranging from 0 to 1, and the unit is dimensionless breakage rate. It is calculated by this formula and represents the proportion of cell walls in the kneaded leaf that are mechanically damaged. It serves as the fourth dimension of the output state of the kneading process and participates in the cross-process state transmission. The initial cell breakage rate before the start of rolling is a scalar constant with a value ranging from 0.02 to 0.08. The unit is dimensionless and is estimated offline by a pressing damage detection algorithm from hyperspectral camera images of fresh leaves upon arrival at the factory. It characterizes the initial cell damage caused during the picking and transportation of fresh leaves. Let be the kneading and crushing kinetic constant, and be a scalar constant with a range of values. to The unit is derived from the dimension consistency (see the dimension verification section below for details), and is obtained by logarithmic linear fitting of empirical data on cell breakage rate, rolling speed and rolling time of historical rolling batches, characterizing the efficiency of the mechanical action intensity of rolling on cell wall destruction. The kneading speed is a scalar value, ranging from 40 to 80, and the unit is _____. The speed signal of the kneading machine, collected by the processing equipment, is used as the driving variable for the intensity of the kneading mechanical action. This is the cumulative kneading time since the start of kneading; it is a scalar quantity, ranging from 0 to 1800, and the unit is 1800. This is provided by the timer of the processing equipment controller; is the power exponent of the kneading speed, and is a scalar constant, which is taken as 1.2 in this embodiment. It is dimensionless, and its value is greater than 1, which reflects the superlinear dependence of cell disruption on the kneading speed. is the exponent of kneading time, and is a scalar constant, which is taken as 0.85 in this embodiment. It is dimensionless, and its value is less than 1, which reflects the gradual saturation of cell breakage as kneading time progresses. This is the operator for the natural exponential function.

[0051] The empirical dynamics model takes the instant of feeding into the kneading machine as the zero point of time. The processing equipment controller samples the current kneading speed and cumulative kneading time every 10 seconds, substitutes these data into the above formula to calculate the estimated cell breakage rate, and writes it into the shared register of the multi-process linkage control module. When the kneading process ends, the estimated cell breakage rate at the final moment is... The modified state vector containing the internal components is incorporated into the output state of the kneading process and transmitted to the intelligent agent in the drying process. (Left side of equation) The value is the dimensionless breakage rate; the parentheses on the right side of the equation... The term requires that the exponential parameter be dimensionless, i.e. It must be dimensionless. Units are Units are ,therefore Units are After substitution, the exponential parameter becomes dimensionless, and the overall expression becomes dimensionless, consistent with the left side.

[0052] When the method described in this invention is applied to black tea processing, in addition to the aforementioned three processes of fixing, rolling, and drying, a fermentation process is added between the rolling and drying processes to promote the oxidation and polymerization of catechins to generate theaflavins and thearubigins. In this embodiment, a colorimeter is installed 150 mm above the surface of the fermenting leaves in the fermentation chamber. The colorimeter is a portable spectrophotometer with CIE Lab color space output, which performs online monitoring of the rate of color change of the fermenting leaves at a sampling frequency of 1 Hz. The colorimeter outputs the L, a, and b color component values ​​of the fermenting leaves at the current moment, and the color difference change rate is calculated in real time by the edge computing node.

[0053] Specifically, the color difference rate is obtained by calculating the color difference rate from the fermented leaf color signal output by the colorimeter using time difference calculation. The color difference change rate is defined as the total color difference. The rate of change of total color difference over time relative to the previous sampling time, Total color difference These are the CIE Lab three-component baseline values ​​of the rolled leaves before fermentation begins. When the color difference change rate drops below a preset inflection point threshold, the near-end strategy optimization agent determines that the fermentation process has reached its optimal endpoint, triggers a discharge command, and updates the fermentation endpoint status accordingly. The cross-process quality status is written in the form of [data] and transmitted to the near-end strategy optimization agent in the drying process. The total color difference accumulated at the end of fermentation. The rate of color difference change at the end of fermentation, This is the cumulative fermentation time. The preset inflection point threshold is obtained by calibrating the color difference change rate trajectory of historical high-quality black tea batches; in this embodiment, the value is taken as follows: .

[0054] The online color difference monitoring function of the black tea fermentation process shares the edge computing node and the near-end policy optimization agent with the three-process linkage control framework of the aforementioned green tea processing. When running, the two switch the corresponding process link through the tea type flag bit to achieve compatibility of the same production line for processing multiple tea types.

[0055] To address the long-term drift of fresh leaf raw material characteristics with seasons and years, and the gradual mismatch of model parameters due to the aging of production line equipment, the dynamic estimation module for internal components and the reinforcement learning optimization module still need to perform online incremental updates after deployment on the production line. In this embodiment, after each batch of processing is completed, the edge computing node writes a quadruple sample consisting of the feature vector of the fresh leaf raw material, the time series of the corrected internal component state vector, the time series of the process parameter decision actions, and the total sensory evaluation score of the finished tea obtained through sensory evaluation into the historical processing batch database.

[0056] The edge computing nodes perform a round of supervised fine-tuning on the data-driven residual correction network and a round of policy parameter fine-tuning on the near-end policy optimization agent at a preset period (in this embodiment, this is taken as every 50 newly added batches of samples). During the incremental update phase, the most recent 500 batches of samples in the current database are used as the training set, with newly added batches participating in the dual network parameter updates at a 7:3 ratio. To ensure that model performance does not deteriorate due to data contamination by individual abnormal batches during online updates, this embodiment also maintains the policy parameters of the near-end policy optimization agent. moving average ,in The moving average coefficient is set to 0.95. To update the cycle number, when it is detected that the average return of the new strategy in the digital twin simulation environment has decreased by more than 5% compared to the moving average strategy, the strategy is rolled back to the moving average strategy. This ensures that the monotonic control performance of the production line does not decline.

[0057] The fresh leaf raw material feature vector, the internal component state vector, and the process parameter decision-making action involve real-time data interaction between the processing equipment controller and the edge computing node. This includes raw material quality characteristics and process parameter secrets, requiring information security measures in the transmission link. In this embodiment, the aforementioned data stream is encrypted and transmitted between the edge computing node and the processing equipment controller using the national cryptographic SM4 algorithm in a block ciphertext format, with a block length of 128 bits and a key length of 128 bits. Simultaneously, a 256-bit integrity check value is calculated for each data packet using the national cryptographic SM3 algorithm and appended to the end of the data packet. After unpacking, the receiving end recalculates the SM3 check value and compares it with the packet end value; if they do not match, the data is deemed tampered with and the packet is discarded. The SM4 and SM3 algorithms are implemented according to the national cryptographic industry standards GM / T 0002-2012 and GM / T 0004-2012.

[0058] To verify the technical effect of the present invention compared with the prior art, the applicant selected fresh leaves of Longjing No. 43 from three different altitude production areas in the spring tea season of 2025: Xihu District of Zhejiang Province, Huangshan District of Anhui Province, and Wuyuan County of Jiangxi Province, as heterogeneous raw material samples. A total of 180 batches of fresh leaves were collected from each production area, with 60 batches collected from each area. The processing control experiment was carried out according to the method described in the present invention and the table-based process parameter scheme represented by publication number CN118077790A. The finished tea was subjected to sensory evaluation by 10 tea tasters with senior tea taster qualifications according to GB / T 23776-2018, and the amino acid content was determined according to GB / T 8314-2013.

[0059] Experimental results show that the sensory evaluation scores of the finished tea processed by the method of this invention reached 89.4, 88.7, and 87.9 points in the three production areas, respectively, representing increases of 10.8, 9.6, and 8.2 points compared to the control scheme, with an average increase of 9.5 points across the three production areas. The amino acid retention rates of the finished tea reached 84.6%, 82.9%, and 81.3% in the three production areas, respectively, representing increases of 9.8%, 9.1%, and 8.4% compared to the control scheme, with an average increase of 9.1% across the three production areas. The coefficient of variation of the sensory evaluation scores of 180 batches of finished tea decreased from 7.8% in the control scheme to 3.2% in the method of this invention, indicating a significant improvement in quality stability under heterogeneous fresh leaf conditions. The experimental data strongly support the assertions regarding the beneficial effects stated in the invention.

[0060] This invention is not only applicable to the processing scenario of premium green tea described in this embodiment, but can also be applied to black tea processing by switching the tea category flag of the proximal strategy optimization agent and enabling the color difference monitoring module of the fermentation process in the black tea processing link. By recalibrating the activation energy and pre-exponential factor parameters of each rate equation in one go, it can also be extended to adaptive control of processing parameters of other tea categories such as yellow tea, black tea, white tea, and oolong tea.

[0061] The embodiments of the present invention are not limited to the specific embodiments described above. Those skilled in the art can make various equivalent changes or substitutions based on the technical solutions of the present invention, and all such changes or substitutions should be included within the protection scope of the present invention.

Claims

1. An adaptive optimization control method for tea processing parameters, characterized in that, The method includes the following steps: performing near-infrared spectroscopy scanning and hyperspectral imaging on the incoming fresh leaves to obtain a feature vector of the fresh leaf raw materials, the feature vector of the fresh leaf raw materials including moisture content, amino acid content, phenol-amino acid ratio, and color and tenderness grading value; using a reaction kinetic model of catechin enzymatic oxidation, chlorophyll demagnesiation reaction, and amino acid thermal degradation during tea processing, combined with a data-driven residual correction network, calculating the internal component state vector every 30 seconds based on the feature vector of the fresh leaf raw materials and the process parameters collected by the processing equipment, the internal component state vector including catechin retention rate and chlorophyll conversion degree; and then processing the fresh leaves... The raw material feature vector and the state vector of the contained components are used as the state input of the reinforcement learning agent. The proximal policy optimization agent outputs process parameter decision actions, which include the joint adjustment of the blanching hot air temperature, the rolling speed and the drying air inlet temperature, and sends them to the processing equipment controller for execution. The state vector of the contained components output from the blanching process is used as the cross-process quality state of the rolling process and is transmitted to the proximal policy optimization agent of the rolling process. The cross-process quality state output from the rolling process is transmitted to the proximal policy optimization agent of the drying process to execute multi-process linkage control.

2. The adaptive optimization control method for tea processing parameters according to claim 1, characterized in that, The process of obtaining the feature vector of the fresh leaf raw material includes: performing non-contact diffuse reflectance scanning on the incoming fresh leaves using a near-infrared spectrometer with a wavelength range of 900nm to 1700nm; simultaneously predicting the three processing suitability indicators of moisture content, amino acid content, and phenol-amino acid ratio using a partial least squares model; performing visual quantification on the color uniformity and tenderness grading of the fresh leaves using hyperspectral imaging to obtain the color and tenderness grading value; and concatenating the moisture content, amino acid content, phenol-amino acid ratio, color and tenderness grading value with the altitude of the place of origin to form the feature vector of the fresh leaf raw material.

3. The adaptive optimization control method for tea processing parameters according to claim 2, characterized in that, The reaction kinetic model adopts a three-parameter first-order kinetic framework, which includes three types of rate equations: the catechin enzymatic oxidation rate equation, the chlorophyll demagnesiation rate equation, and the amino acid Maillard reaction rate equation. Each rate equation uses the current hot air temperature, hot air velocity, and leaf temperature collected by the processing equipment as driving variables, and the moisture content and phenol-to-amino acid ratio in the fresh leaf raw material feature vector as initial conditions. The influence of temperature on the reaction rate constant is characterized in the form of the Arrhenius equation. The model outputs the increment of catechin retention rate, chlorophyll conversion rate, and amino acid degradation rate in each sampling period, and accumulates them over time steps to obtain the state vector of the internal components.

4. The adaptive optimization control method for tea processing parameters according to claim 3, characterized in that, The increments in catechin retention, chlorophyll conversion, and amino acid degradation output by the reaction kinetic model are corrected by a data-driven residual correction network to obtain the corrected internal component state vector. The data-driven residual correction network uses the deviation sequence between the reaction kinetic estimated values ​​and the actual laboratory measured values ​​in historical processing batches as training samples, and performs time-series modeling of the deviation sequence using a bidirectional long short-term memory network structure to output the correction increment of the reaction kinetic estimated values. The correction increment is added to the reaction kinetic estimated values ​​to obtain the corrected internal component state vector. The corrected internal component state vector is used as the state input of the proximal policy optimization agent.

5. The adaptive optimization control method for tea processing parameters according to claim 4, characterized in that, The proximal policy optimization agent is trained using an Actor-Critic architecture. Its state space includes the feature vector of the fresh leaf raw material and the state vector of the corrected intrinsic components. Its action space includes the adjustment amount of the hot air temperature for blanching within the range of 150℃ to 280℃, the adjustment amount of the rolling speed within the range of 40 r / min to 80 r / min, and the adjustment amount of the drying air inlet temperature within the range of 80℃ to 130℃. The composite reward function is constructed as follows: Where R represents a single-batch compound reward, This indicates the total score of the sensory evaluation of the finished tea. Indicates the amino acid retention rate. The deviation of the current batch quality index from the historical average is represented by α, β, and γ, which represent preset weighting coefficients. The near-end strategy optimization agent executes strategy parameter iteration according to the principle of maximizing the composite reward function.

6. The adaptive optimization control method for tea processing parameters according to claim 5, characterized in that, The multi-process linkage control includes cross-process quality state transfer between the fixation process, the rolling process, and the drying process. The catechin retention rate in the modified internal component state vector at the end of the fixation process is used as an additional state input to the proximal strategy optimization agent of the rolling process, driving the adjustment of rolling pressure and rolling time parameters. At the end of the rolling process, the estimated cell breakage rate of the rolled leaves and the modified internal component state vector are concatenated to form the rolling process output state. The estimated cell breakage rate is calculated based on the empirical dynamic model of the effect of rolling speed and rolling time on the pressing effect. The rolling process output state is used as an additional state input to the proximal strategy optimization agent of the drying process, driving the feedforward adjustment of the drying air inlet temperature curve.

7. The adaptive optimization control method for tea processing parameters according to claim 1, characterized in that, When the method is applied to the black tea processing scenario, a colorimeter is set up in the fermentation process to monitor the rate of color change of the fermented leaves online. When the rate of color change drops to a preset inflection point, the near-end strategy optimization agent triggers the discharge command and writes the fermentation endpoint state into the cross-process quality state, which is then transmitted to the drying process to execute subsequent parameter decisions.

8. The adaptive optimization control method for tea processing parameters according to claim 4, characterized in that, After each batch of processing is completed, the feature vector of the fresh leaf raw material, the time series of the corrected internal component state vector, the time series of the process parameter decision action, and the total sensory evaluation score of the finished tea are combined into a quadruple sample and written into the historical processing batch database. According to a preset period, the data-driven residual correction network and the proximal strategy optimization agent are updated online with new samples. During the update process, the moving average of the historical strategy parameters is retained as a fallback guarantee.

9. The adaptive optimization control method for tea processing parameters according to claim 1, characterized in that, The data streams of the fresh leaf raw material feature vector, the internal component state vector, and the process parameter decision action are transmitted between the processing equipment controller and the edge computing node using the national cryptographic SM4 algorithm in a packet-encrypted manner, and a data integrity verification value is generated using the SM3 algorithm and appended to the end of the data packet.

10. An adaptive optimization control system for tea processing parameters, used to implement the adaptive optimization control method for tea processing parameters according to any one of claims 1 to 9, characterized in that, The system includes processing equipment for blanching, rolling, and drying processes connected in series, as well as edge computing nodes communicatively connected to each of the processing equipment. Each edge computing node internally deploys a fresh leaf quality online characterization module, a dynamic estimation module for internal components, a reinforcement learning optimization module, and a multi-process linkage control module. The fresh leaf quality online characterization module connects to a near-infrared spectrometer and a hyperspectral camera, generating a feature vector of the fresh leaf raw material based on the outputs of the near-infrared spectrometer and the hyperspectral camera. The dynamic estimation module for internal components uses a reaction kinetic model coupled with data to drive a residual correction network, based on the feature vector of the fresh leaf raw material and the characteristics of each processing equipment. The process parameters collected are calculated every 30 seconds to deduce the corrected internal component state vector; the reinforcement learning optimization module deploys a proximal policy optimization agent, which takes the fresh leaf raw material feature vector and the corrected internal component state vector as state input, and outputs the process parameter decision actions of the blanching process equipment, the rolling process equipment, and the drying process equipment; the multi-process linkage control module maintains a cross-process quality state register chain, transmits the corrected internal component state vector output by the blanching process to the proximal policy optimization agent of the rolling process, and transmits the rolling process output state to the proximal policy optimization agent of the drying process.

Images