A method, system and medium for adjusting a tea leaf fixation process

Abstract

The application discloses a tea leaf fixation process adjusting method and system and medium, aiming at realizing the synergistic regulation of enzyme activity and aroma substances in the fixation process. The method comprises the following steps: collecting multi-modal sensing information in the fixation process, extracting features from the multi-modal sensing information to obtain multi-dimensional features; dynamically constructing an LSTM prediction model based on the multi-dimensional features, and outputting an enzyme activity prediction value sequence; inputting the enzyme activity prediction value sequence into a fuzzy PID controller optimized by reinforcement learning to generate temperature-rotation speed adjusting instructions, and dynamically updating the adjusting parameters of the controller through a reward function of chlorophyll retention rate and energy consumption ratio; adjusting the hot air temperature of each temperature zone of the fixation chamber, the rotation speed of the fixation cylinder and the axial segmented wind speed according to the temperature-rotation speed adjusting instructions; and collecting fixation end judgment information, and terminating the fixation when the preset conditions are met. The application solves the problems of unstable quality and excessive energy consumption caused by parameter fixation in the traditional fixation process, and realizes dynamic optimization and precise control of the fixation process.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of tea processing, and in particular to a tea fixation process adjustment method, system and medium. BACKGROUND

[0002] Tea fixation is a key process in tea processing, and its core is to quickly inactivate enzyme activity (such as polyphenol oxidase) through high temperature, while retaining quality ingredients such as chlorophyll, and promoting the formation of aroma substances. Traditional fixation processes mostly rely on fixed temperature curves or manual experience to adjust hot air temperature, drum speed and other parameters. However, only monitoring temperature through a thermocouple cannot capture the relationship between enzyme activity dynamic changes and aroma substance volatilization in real time, which can easily lead to a lag in fixation degree judgment, and traditional PID control cannot adapt to tea moisture content, variety differences and other variables, which can easily cause local over-fixation or fixation deficiency. In addition, only using time or apparent temperature as a standard, ignoring the synergistic indicators of chlorophyll retention rate and key aroma components, can lead to large fixation quality fluctuations and high energy consumption, and long downtime is required to adjust parameters when switching tea varieties. SUMMARY

[0003] Therefore, the purpose of the present application is to provide a tea fixation process adjustment method, system and medium, which can realize synergistic and accurate regulation of enzyme activity and aroma substances in the fixation process, and solve the problems of unstable quality and high energy consumption in traditional fixation processes.

[0004] In order to achieve the above technical purposes, in a first aspect, the present application provides a tea fixation process adjustment method, comprising:

[0005] Collecting multi-modal sensing information in the fixation process, the multi-modal sensing information including infrared temperature time series information, chlorophyll fluorescence intensity and volatile organic compound concentration;

[0006] Performing feature extraction on the multi-modal sensing information to obtain multi-dimensional features, the multi-dimensional features including enzyme activity decay curve, aroma volatilization characteristic spectrum and temperature fluctuation characteristics;

[0007] Dynamically constructing an LSTM prediction model according to the multi-dimensional features, and outputting an enzyme activity prediction value sequence of a prediction time segment;

[0008] Inputting the enzyme activity prediction value sequence into a fuzzy PID controller optimized by reinforcement learning to generate temperature-rotation speed adjustment instructions, and the adjustment parameters of the fuzzy PID controller are configured to be dynamically updated by a reward function of chlorophyll retention rate and energy consumption ratio;

[0009] Controlling edge execution nodes in each temperature zone in the fixation chamber according to the temperature-rotation speed adjustment instructions, and the edge execution nodes adjust hot air temperature, fixation drum speed and axial segmented wind speed;

[0010] Collecting the fixation endpoint determination information, the fixation endpoint determination information includes chlorophyll fluorescence threshold, volatile organic matter ratio and leaf surface temperature, and the fixation is terminated when the fixation endpoint determination information meets the preset determination condition.

[0011] In some embodiments, feature extraction is performed on the multi-modal sensing information to obtain multi-dimensional features, including:

[0012] The infrared temperature time series information is segmented by a sliding window to obtain a plurality of temperature sequence windows;

[0013] The temperature data in each temperature sequence window is subjected to frequency spectrum analysis to extract the main frequency component and the energy distribution feature corresponding to each main frequency component, and the temperature fluctuation feature is obtained;

[0014] And the chlorophyll fluorescence intensity is subjected to digital filtering processing to remove high-frequency noise and obtain smooth fluorescence intensity information;

[0015] The smooth fluorescence intensity information is subjected to difference calculation to obtain the instantaneous change rate of enzyme activity;

[0016] The instantaneous change rate of enzyme activity is subjected to time fitting by a nonlinear least squares method to obtain an enzyme activity decay curve of the fluorescence intensity with time;

[0017] The concentrations of each component in the volatile organic matter are monitored in real time, and the dynamic concentration ratio of the preset aroma substance is calculated;

[0018] Based on the dynamic concentration ratio, a conversion relationship model between each component of the volatile organic matter is established by time series regression analysis;

[0019] The dynamic change characteristics of the aroma substance are extracted according to the conversion relationship model to construct an aroma volatilization characteristic spectrum;

[0020] The temperature fluctuation feature, the enzyme activity decay curve and the aroma volatilization characteristic spectrum are subjected to normalization processing to generate multi-dimensional features.

[0021] In some embodiments, an LSTM prediction model is dynamically constructed according to the multi-dimensional features, and an enzyme activity prediction value sequence of a prediction time section is output, including:

[0022] The temperature fluctuation feature, the enzyme activity decay curve and the aroma volatilization characteristic spectrum are subjected to time sequence alignment processing to construct a multi-dimensional feature input sequence;

[0023] The multi-dimensional feature data collected in real time is input into the LSTM network model trained;

[0024] The original enzyme activity prediction value sequence of the prediction time section is output by forward propagation calculation of the LSTM network;

[0025] Smooth the original enzyme activity prediction value sequence to obtain an enzyme activity prediction value sequence;

[0026] The LSTM prediction model is constructed by the following steps:

[0027] Initialize the infrastructure of the LSTM prediction model, and configure the number of neurons of the input layer, the hidden layer and the output layer;

[0028] The sample training features of the same tea category are divided into a training sample set by using a sliding time window method;

[0029] The LSTM prediction model is iteratively trained using the training sample set, and the network weight parameters are optimized until the LSTM prediction model is trained.

[0030] In some embodiments, the enzyme activity prediction value sequence is input into a fuzzy PID controller optimized by reinforcement learning to generate temperature-rotation speed adjustment instructions, including:

[0031] The membership functions of the input variables and the membership functions of the output variables of the fuzzy PID controller are established, the input variables include the deviation and the rate of change of the deviation of the enzyme activity prediction value sequence, and the output variables include the temperature adjustment amount and the rotation speed adjustment amount;

[0032] A reinforcement learning agent is constructed, and the state space, action space and reward function of the reinforcement learning agent are configured, the state space includes the current PID parameters and the control effect evaluation index, the action space is the adjustment amplitude of the PID parameters, and the reward function is calculated based on the chlorophyll retention rate and the energy consumption ratio;

[0033] The reinforcement learning agent interacts with the fixation device to generate interaction information, and the parameter correction strategy of the fuzzy PID controller is dynamically optimized according to the interaction information;

[0034] The enzyme activity prediction value sequence is input into the optimized fuzzy PID controller, and the fuzzy control output information of the temperature adjustment amount and the rotation speed adjustment amount is calculated;

[0035] The fuzzy control output information is de-fuzzified to generate accurate initial temperature-rotation speed adjustment instructions;

[0036] The initial temperature-rotation speed adjustment instructions are checked for feasibility according to the current working state of the fixation chamber;

[0037] The initial temperature-rotation speed adjustment instructions that pass the check are recorded as temperature-rotation speed adjustment instructions and output.

[0038] In some embodiments, a reinforcement learning agent is constructed, a state space, an action space and a reward function of the reinforcement learning agent are configured, the state space includes current PID parameters and control effect evaluation indexes, the action space is an adjustment amplitude of the PID parameters, and the reward function is calculated based on chlorophyll retention rate and energy consumption ratio, including:

[0039] The state space is represented by formula (1), and formula (1) is as follows:

[0040] ;

[0041] In formula (1), is the state space, is a parameter combination of the fuzzy PID controller, is a proportional coefficient of the fuzzy PID controller, is an integral coefficient of the fuzzy PID controller, is a differential coefficient of the fuzzy PID controller, is a control effect evaluation index, is a chlorophyll retention rate deviation, is an energy consumption ratio deviation;

[0042] The action space is represented by formula (2), and formula (2) is as follows:

[0043] ;

[0044] In formula (2), is the action space, is an adjustment amount of the proportional coefficient, , is a limit adjustment amplitude, is an adjustment amount of the integral coefficient, , is a limit adjustment amplitude, is an adjustment amount of the differential coefficient, , is a limit adjustment amplitude;

[0045] The reward function is represented by formula (3), and formula (3) is as follows:

[0046] ;

[0047] In formula (3), is the reward function, is a weight coefficient of the chlorophyll retention rate, is a weight coefficient of the energy consumption efficiency, is an absolute value deviation of the chlorophyll retention rate, to take the upper limit of the energy consumption ratio deviation,

[0048] Initialize the policy network and value network of the reinforcement learning agent, and configure the training hyperparameters, including the learning rate and the discount factor;

[0049] The interaction information generated by the interaction between the reinforcement learning agent and the kill device, and the parameter correction strategy of the fuzzy PID controller is dynamically optimized according to the interaction information, including:

[0050] Collect the interaction information generated by the real-time interaction process between the reinforcement learning agent and the kill device, including control state, execution action, reward obtained and new control state;

[0051] Based on the interaction information, an experience sample library is constructed, and training data is extracted from the experience sample library using random sampling;

[0052] The training data is used to update the parameter weights of the policy network and the value network, and a new parameter adjustment strategy is generated;

[0053] The new parameter adjustment strategy is applied to the parameter update of the fuzzy PID controller, including the update of the proportional coefficient, the integral coefficient and the differential coefficient;

[0054] The updated fuzzy PID controller is evaluated in terms of value, and the fuzzy PID controller that passes the value evaluation is verified for stability to keep the control parameters within a predetermined safe range;

[0055] The new parameter adjustment strategy that passes the stability verification is output and executed.

[0056] In some embodiments, the enzyme activity prediction value sequence is input into the optimized fuzzy PID controller to calculate the fuzzy control output information of the temperature adjustment amount and the speed adjustment amount, including:

[0057] The enzyme activity prediction value sequence is subjected to deviation calculation to obtain the enzyme activity prediction value deviation and the deviation change rate;

[0058] The enzyme activity prediction value deviation and the deviation change rate are input into the fuzzification interface to be converted into corresponding fuzzy language variables;

[0059] The fuzzy PID controller matches the applicable temperature adjustment fuzzy rule and the speed adjustment fuzzy rule;

[0060] The matched temperature adjustment fuzzy rule and speed adjustment fuzzy rule are subjected to inference operation to generate fuzzy control output information containing temperature adjustment amount and speed adjustment amount;

[0061] ​The fuzzy control output information is de-fuzzified to generate accurate initial temperature-rotation speed adjustment instructions, including:

[0062] The temperature adjustment amount in the fuzzy control output information is calculated by the barycenter method to obtain an initial temperature adjustment amount;

[0063] The rotation speed adjustment amount in the fuzzy control output information is calculated by the maximum membership degree method to obtain an initial rotation speed adjustment amount;

[0064] The initial temperature adjustment amount and the initial rotation speed adjustment amount are combined in time sequence to generate initial temperature-rotation speed adjustment instructions.

[0065] In some embodiments, the edge execution nodes of each temperature zone in the fixation chamber are controlled according to the temperature-rotation speed adjustment instructions, and the edge execution nodes adjust the hot air temperature, the fixation drum rotation speed, and the axial segmented air speed, including:

[0066] The temperature-rotation speed adjustment instructions are decomposed into independent control instructions for each temperature zone, and the independent control instructions for each temperature zone include preheating zone temperature instructions, main fixation zone temperature instructions, and stabilization zone temperature instructions;

[0067] The preheating zone temperature instructions are signal-converted to generate a gas valve opening degree control signal to adjust the gas combustion intensity of the preheating zone;

[0068] The main fixation zone temperature instructions are power-modulated and converted to obtain the output power of the heating element to adjust the hot air temperature of the main fixation zone;

[0069] The stabilization zone temperature instructions are subjected to PID operation to obtain a fan rotation speed adjustment amount to adjust the temperature interval of the stabilization zone;

[0070] And the temperature-rotation speed adjustment instructions are input into a frequency conversion driver to adjust the operation frequency of the drive unit of the fixation drum.

[0071] In some embodiments, fixation endpoint determination information is collected, the fixation endpoint determination information includes a chlorophyll fluorescence threshold value, a volatile organic matter ratio, and a leaf surface temperature, and the fixation is terminated when the fixation endpoint determination information meets a preset determination condition, including:

[0072] The chlorophyll fluorescence threshold value is sampled in real time to obtain a current fluorescence intensity value;

[0073] The current fluorescence intensity value is compared with a preset fluorescence threshold value to calculate a fluorescence intensity decay rate;

[0074] When the fluorescence intensity decay rate reaches a preset enzyme inactivation threshold value, a first termination signal is obtained;

[0075] and, real-time monitoring of the concentration of each component in the volatile organic compounds, and calculating a volatile organic compound ratio, the volatile organic compound ratio including a real-time concentration ratio of linalool to coniferyl aldehyde;

[0076] When the volatile organic compound ratio reaches a preset aroma balance threshold, a second termination signal is obtained;

[0077] and, real-time acquisition of the leaf surface temperature, and calculating an average value of the leaf surface temperature to obtain a leaf surface temperature average value;

[0078] When the leaf surface temperature average value is lower than a preset surface temperature threshold, a third termination signal is obtained;

[0079] The method further comprises:

[0080] When the first termination signal and the second termination signal are simultaneously received, a fixation termination instruction is immediately triggered;

[0081] When only the first termination signal or the second termination signal is received, a delay determination procedure is started, and the third termination signal is detected within a preset time.

[0082] In a second aspect, the present application further provides a tea leaf fixation process adjustment system, which is suitable for the adjustment method of the first aspect, and comprises a fixation device, a driving unit, a hot air assembly, a temperature sensor, an olfactory sensor, an image acquisition unit and a control unit. The fixation device comprises a fixation cylinder, and the fixation cylinder has a fixation chamber in which the tea leaves to be fixed are contained. The driving unit is in transmission connection with the fixation cylinder. The hot air assembly comprises at least two fans, and the at least two fans are distributed in the fixation cylinder in a preset manner. The temperature sensor is arranged in the fixation cylinder. The olfactory sensor is arranged in the fixation cylinder. The image acquisition unit is arranged in the fixation cylinder. The control unit is electrically connected with the driving unit, the hot air assembly, the temperature sensor, the olfactory sensor and the image acquisition unit respectively, and is used for executing the method of the first aspect.

[0083] In a third aspect, the present application further provides a computer readable storage medium, which stores computer program instructions, and the computer program instructions are used for implementing the method of the first aspect when executed by a processor.

[0084] By using the above technical solution, the present application has the following beneficial effects compared with the prior art:

[0085] The application provides a tea leaf fixation process adjusting method, system and medium, aiming to realize the synergistic and accurate regulation of enzyme activity and aroma substances in the fixation process. The method comprises the following steps: collecting multi-modal sensing information in the fixation process, including infrared temperature time sequence information, chlorophyll fluorescence intensity and volatile organic matter concentration; performing feature extraction on the multi-modal sensing information to obtain an enzyme activity decay curve, an aroma volatile feature spectrum and a temperature fluctuation feature; dynamically constructing an LSTM prediction model based on the multi-dimensional features to output an enzyme activity prediction value sequence; inputting the enzyme activity prediction value sequence into a fuzzy PID controller optimized by reinforcement learning to generate a temperature-rotation speed adjusting instruction, and dynamically updating the adjusting parameters of the PID controller through a reward function of chlorophyll retention rate and energy consumption ratio; adjusting the edge execution nodes of each temperature zone of the fixation chamber according to the temperature-rotation speed adjusting instruction to adjust the hot air temperature, the fixation cylinder rotation speed and the axial segmented air speed; and collecting fixation end point judgment information and terminating the fixation when the preset condition is met. The application solves the problems of unstable quality and excessive energy consumption caused by the parameter fixation of the traditional fixation process, and realizes the dynamic optimization and accurate control of the fixation process. BRIEF DESCRIPTION OF DRAWINGS

[0086] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or the prior art description. Obviously, the drawings in the following description only constitute some embodiments of the present application, and all other drawings obtained by those skilled in the art without creative labor on the basis of these drawings are also within the scope of the present application.

[0087] Figure 1 is a method step diagram of steps S101 to S106 of the adjusting method described in the specific embodiment;

[0088] Figure 2 is a method step diagram of steps S201 to S208 of the adjusting method described in the specific embodiment. DETAILED DESCRIPTION

[0089] The application will be further described in detail below in combination with the drawings and embodiments. It is particularly pointed out that the following embodiments are only used to illustrate the present application, but do not limit the scope of the present application. Similarly, the following embodiments are only some embodiments of the present application, not all embodiments, and all other embodiments obtained by those skilled in the art without creative labor are within the scope of the present application.

[0090] Please refer to Figure 1 In the first aspect, the present embodiment provides a tea leaf fixation process adjusting method, comprising:

[0091] S101, collect multi-modal sensing information in the process of fixation, the multi-modal sensing information includes infrared temperature time series information, chlorophyll fluorescence intensity and volatile organic compound concentration;

[0092] S102, feature extraction is performed on the multi-modal sensing information to obtain multi-dimensional features, the multi-dimensional features include enzyme activity decay curve, aroma volatile characteristic spectrum and temperature fluctuation characteristics;

[0093] S103, dynamically constructing an LSTM prediction model according to the multi-dimensional features, and outputting an enzyme activity prediction value sequence of a prediction time section;

[0094] S104, inputting the enzyme activity prediction value sequence into a fuzzy PID controller optimized by reinforcement learning to generate temperature-rotational speed adjustment instructions, and the adjustment parameters of the fuzzy PID controller are configured to be dynamically updated by a reward function of chlorophyll retention rate and energy consumption ratio;

[0095] S105, controlling edge execution nodes of each temperature zone in the fixation chamber according to the temperature-rotational speed adjustment instructions, and the edge execution nodes adjust hot air temperature, fixation cylinder rotational speed and axial segmented air speed;

[0096] S106, collecting fixation endpoint determination information, the fixation endpoint determination information includes chlorophyll fluorescence threshold, volatile organic compound ratio and leaf surface temperature, and the fixation is terminated when the fixation endpoint determination information meets the preset determination condition; further, the fixation process parameters, adjustment instructions and determination results are stored in association to a block chain to generate a traceable process chain including enzyme activity dynamic curve, aroma quality score and energy consumption distribution.

[0097] In step S101, the collection process of multi-modal sensing information adopts an adaptive sampling strategy, preferably, the spatial resolution of infrared temperature time series information is dynamically adjusted according to the rotational speed of the fixation cylinder, and when the rotational speed is increased, the density of transverse temperature measurement points is increased to compensate for the monitoring blind area caused by leaf turning; the detection wavelength of chlorophyll fluorescence intensity is adaptively selected according to the spectral characteristics of tea varieties to ensure the sensitivity to different chlorophyll isomers; the monitoring of volatile organic compound concentration focuses on establishing a dynamic targeted detection list for characteristic aroma components of varieties, and the monitoring intensity of terpenes is automatically enhanced in the later stage of fixation. Through intelligent adjustment of the sensing strategy, the problem of insufficient adaptability of traditional fixed parameter sensing system to process changes is effectively solved.

[0098] In step S102, preferably, feature extraction introduces an attention mechanism to enhance the recognition of key features. When extracting the enzyme activity decay curve, time-domain attention weights are used to highlight the feature points in the temperature-sensitive stage. For the aroma volatile feature spectrum, frequency-domain attention is used to focus on the resonance peaks of the characteristic aroma components. The temperature fluctuation feature is strengthened by a spatial attention mechanism to represent the high-variation region. This step significantly improves the ability of the subsequent model to capture weak quality signals, especially the early identification of the fixation turning point.

[0099] In step S103, the LSTM prediction model uses a multi-task learning architecture. The main task is to predict enzyme activity, while the auxiliary task is to simultaneously predict chlorophyll degradation trends and aroma generation potential. The hidden layers of the three tasks share feature representations, but the output layers are independent. Gradient masking mechanism is used to achieve collaborative optimization of different prediction targets, so that the model can capture the deep relationship between enzyme activity changes and quality indicators, and the prediction results are more consistent with the coupling characteristics of the actual process.

[0100] In step S104, preferably, the parameter optimization of the fuzzy PID controller uses a hierarchical reinforcement learning strategy. The high-level strategy quickly initializes control parameters during variety switching through a meta-learning framework, and the low-level strategy adjusts the proportional, integral, and derivative coefficients in real time through online learning. The reward function introduces a dynamic weight mechanism, focusing on chlorophyll retention rate at the early stage of fixation, and gradually increasing the energy consumption weight at the later stage. This time-varying optimization objective is more consistent with the phased characteristics of tea fixation.

[0101] In step S105, the regulation of the edge execution node can introduce a feedforward-feedback composite control. The feedforward channel adjusts the hot air temperature in advance based on the tea moisture content prediction model, and the feedback channel compensates for the regulation deviation based on real-time sensor data. Axial segmented wind speed control uses computational fluid dynamics simulation assisted inverse model to calculate the optimal opening combination of the guide vane, achieving precise spatial regulation of the heat field in the fixation cylinder.

[0102] In step S106, the fixation endpoint determination condition uses a fuzzy logic dynamic adjustment strategy. The chlorophyll fluorescence threshold determination interval is automatically scaled based on the initial chlorophyll content. Preferably, the volatile organic compound ratio introduces a moving time window statistical verification mechanism, requiring three consecutive sampling periods to meet the standard. The gradient condition of the leaf surface temperature is gradually tightened as the fixation stage progresses. Further, the blockchain storage can use a lightweight Merkle tree structure to ensure traceability while adapting to the data throughput requirements of the industrial site. The quality scoring module in the process chain integrates transfer learning capabilities, which can quickly adapt to the evaluation standards of new varieties.

[0103] The embodiment realizes the synergistic regulation of enzyme activity and aroma generation by real-time capturing the temperature distribution, chlorophyll stability and dynamic changes of aroma substances in the fixation process through multi-modal sensor information, and based on the LSTM prediction model and the fuzzy PID controller optimized by reinforcement learning. Through dynamic extraction and modeling of multi-dimensional features, the quality fluctuation problem caused by the solidification of traditional fixation process parameters is solved; the distributed control architecture of edge execution node is adopted to realize the precise synergistic adjustment of hot air temperature, fixation cylinder speed and axial segmented wind speed; combined with the blockchain traceable process chain, the complete association of process parameters and quality indicators is ensured, which improves the fixation uniformity while reducing energy consumption, providing a standardized and adaptive intelligent control paradigm for tea processing.

[0104] Referring to Figure 2 In some embodiments, multi-modal sensor information is feature extracted to obtain multi-dimensional features, including:

[0105] S201, the infrared temperature time series information is segmented by a sliding window to obtain a plurality of temperature sequence windows;

[0106] S202, the temperature data in each temperature sequence window is subjected to frequency spectrum analysis to extract the main frequency component and the energy distribution feature corresponding to each main frequency component, and the temperature fluctuation feature is obtained;

[0107] And, the chlorophyll fluorescence intensity is subjected to digital filtering processing to remove high-frequency noise and obtain smooth fluorescence intensity information;

[0108] S203, the smooth fluorescence intensity information is subjected to difference calculation to obtain the instantaneous change rate of enzyme activity;

[0109] S204, the instantaneous change rate of enzyme activity is subjected to time fitting by a nonlinear least squares method to obtain an enzyme activity decay curve of the fluorescence intensity with time;

[0110] S205, the concentration of each component in volatile organic compounds is monitored in real time, and the dynamic concentration ratio of the preset aroma substance is calculated;

[0111] S206, based on the dynamic concentration ratio, a conversion relationship model between each component of volatile organic compounds is established by time series regression analysis;

[0112] S207, the dynamic change feature of aroma substances is extracted according to the conversion relationship model, and an aroma volatilization feature spectrum is constructed;

[0113] S208, the temperature fluctuation feature, the enzyme activity decay curve and the aroma volatilization feature spectrum are subjected to normalization processing to generate multi-dimensional features.

[0114] In step S201, the sliding window segmentation adopts an adaptive window length strategy. Preferably, the window length is dynamically adjusted according to the speed of the fixation cylinder, ensuring that each window contains a complete temperature fluctuation period. The division of the temperature sequence window is achieved by timestamp alignment, ensuring the time sequence consistency with the chlorophyll fluorescence intensity data. The spectral analysis adopts the fast Fourier transform algorithm, and the extraction of the main frequency component is completed by peak detection of the power spectral density. The energy distribution characteristics are obtained by calculating the energy proportion of each frequency band. These characteristics together constitute the temperature fluctuation characteristics reflecting the heat conduction characteristics in the fixation chamber.

[0115] In step S202, the chlorophyll fluorescence intensity is digitally filtered using a Butterworth low-pass filter. The cutoff frequency is set according to the typical fluctuation range of the chlorophyll fluorescence signal, which can effectively suppress high-frequency noise while retaining the true fluorescence intensity trend. Further, the smoothed fluorescence intensity information is obtained through recursive filtering, which minimizes phase distortion while ensuring data smoothness.

[0116] In step S203, the differential calculation adopts the central difference method, and the calculation step of the enzyme activity transient change rate is synchronized with the temperature sequence window, ensuring the time alignment of the multi-modal characteristics.

[0117] In step S204, the nonlinear least squares fitting selects an exponential decay model, and the initial parameters are initialized by prior knowledge of enzyme activity changes. Preferably, the fitting process is optimized using the Levenberg-Marquardt algorithm, and the obtained enzyme activity decay curve contains characteristic parameters such as half-life and decay rate constant, which characterize the enzyme inactivation kinetics. The enzyme activity decay curve is verified by residual analysis to ensure that it accurately reflects the activity change of the key enzyme during the fixation process.

[0118] In step S205, the preset aroma substance refers to the characteristic aromatic components produced during the fixation process of tea leaves, such as terpenes, alcohols, and other key volatile organic compounds that determine the tea aroma type. Preferably, the dynamic concentration ratio calculation of the preset aroma substance adopts the relative peak area method, and the target components are determined according to the characteristic aroma fingerprint of the tea variety. The ratio calculation introduces moving average smoothing processing to eliminate the influence of instantaneous detection fluctuations. The update frequency of the dynamic concentration ratio is matched with the fixation process adjustment period, providing stable input sequences for subsequent conversion relationship modeling.

[0119] In steps S206 and S207, the time series regression analysis adopts a vector autoregressive model, and the lag order is automatically determined by information criteria. The establishment of the conversion relationship model includes the calculation of the conversion rate between components and the significance test, and only the statistically significant conversion paths are retained. The construction of the aroma volatilization characteristic spectrum is completed by extracting the key conversion coefficients in the model, forming a characteristic matrix reflecting the dynamic balance between the generation and degradation of aroma substances.

[0120] In step S208, the normalization processing preferably adopts the dynamic extreme value method of each feature parameter, and the scaling coefficient is adaptively adjusted according to the process characteristics of the current batch of tea. The generation of multi-dimensional features is realized by feature vector splicing, and the original dimensional relationship of the temperature fluctuation feature, the enzyme activity decay curve and the aroma volatilization feature spectrum is maintained. The normalized feature vector verifies the independence between features through covariance matrix analysis, and ensures that the feature space input into the prediction model has good separability.

[0121] In this embodiment, the temperature fluctuation feature containing the main frequency component and the energy distribution characteristic is extracted by sliding window segmentation and spectrum analysis of the infrared temperature time series information; at the same time, the enzyme activity instantaneous change rate is obtained by digital filtering and difference calculation of the chlorophyll fluorescence intensity, and is fitted into an enzyme activity decay curve; combined with the dynamic concentration ratio analysis of volatile organic compounds, an aroma substance conversion relationship model is established and an aroma volatilization feature spectrum is constructed. The combination of sliding window segmentation and spectrum analysis completely retains the periodic characteristics of temperature fluctuation; through digital filtering processing and nonlinear least squares fitting, the dynamic process of enzyme activity change is accurately represented; based on the conversion relationship model established by time series regression analysis, the dynamic conversion law of aroma substances is effectively reflected; and finally, the multi-dimensional features generated through normalization processing realize the organic integration of temperature fluctuation feature, enzyme activity decay curve and aroma volatilization feature spectrum, and provide comprehensive and reliable features for process optimization.

[0122] In some embodiments, an LSTM prediction model is dynamically constructed according to the multi-dimensional features, and an enzyme activity prediction value sequence of a prediction time section is output, including:

[0123] The temperature fluctuation feature, the enzyme activity decay curve and the aroma volatilization feature spectrum are subjected to time sequence alignment processing to construct a multi-dimensional feature input sequence;

[0124] The real-time collected multi-dimensional feature data is input into the trained LSTM network model;

[0125] The original enzyme activity prediction value sequence of the prediction time section is output through forward propagation calculation of the LSTM network;

[0126] The original enzyme activity prediction value sequence is subjected to smoothing processing to obtain an enzyme activity prediction value sequence;

[0127] The LSTM prediction model is constructed through the following steps:

[0128] The basic structure of the LSTM prediction model is initialized, and the number of neurons of the input layer, the hidden layer and the output layer is configured;

[0129] The sample training features of the same tea category are divided into a training sample set in a sliding time window manner;

[0130] The LSTM prediction model is iteratively trained using the training sample set to optimize the network weight parameters until the LSTM prediction model is trained.

[0131] In this embodiment, the temperature fluctuation feature refers to a periodic feature extracted from the infrared temperature time series data through sliding window segmentation and spectral analysis, containing the main frequency component and energy distribution information, for representing the heat conduction characteristics in the fixation chamber. The aroma volatilization feature spectrum is a transformation relationship model established based on the dynamic concentration ratio analysis of volatile organic compounds, reflecting the dynamic balance characteristics of aroma substance generation and degradation.

[0132] The time series alignment processing refers to calibrating the temperature fluctuation feature, enzyme activity decay curve and aroma volatilization feature spectrum according to a unified time reference, ensuring the consistency of the multi-dimensional feature input sequence in the time dimension. In the basic structure configuration of the LSTM network model, the number of input layer neurons matches the dimension of the multi-dimensional features, the number of hidden layer neurons is determined according to the characteristics of tea categories, and the output layer corresponds to the enzyme activity prediction value sequence of the prediction time segment. The division of the sample training features of the same tea category into training sample sets by using the sliding time window method refers to dividing the training samples according to fixed time intervals, and the window length is adaptively adjusted according to the process cycle characteristics, ensuring that each window contains a complete process change cycle.

[0133] Further, the division of the training sample set needs to consider the process consistency of the same tea category, and the sample features should include the temperature fluctuation, enzyme activity change and aroma substance transformation data under typical process conditions.

[0134] Preferably, the optimization of network weight parameters is realized by the back propagation algorithm, and the adaptive learning rate strategy is used to balance the training speed and convergence stability; the smoothing processing of the original enzyme activity prediction value sequence uses the moving average algorithm, and the window size is synchronized with the process adjustment period to eliminate transient fluctuation interference.

[0135] This embodiment models the long-range dependence relationship of time series features through the LSTM network, encodes the synergistic change law of temperature fluctuation features, enzyme activity decay curves and aroma volatilization feature spectra into network weights, and realizes accurate prediction of enzyme activity change in the future time segment. Among them, the temperature fluctuation feature reflects the external thermal action condition, the enzyme activity decay curve represents the internal biochemical reaction process, and the aroma volatilization feature spectrum indicates the quality formation trend, which together constitute the multi-dimensional feature space of the prediction model. The prediction model learns the nonlinear mapping relationship between process parameters and quality indicators through iterative training, and finally outputs the smoothed enzyme activity prediction value sequence, providing decision basis for real-time process regulation.

[0136] The embodiment realizes accurate prediction of the change of enzyme activity by time sequence alignment processing of the temperature fluctuation characteristics, enzyme activity attenuation curve and aroma volatile characteristics spectrum and inputting the LSTM prediction model. The synergistic analysis of multi-dimensional characteristics fully reflects the interaction of heat conduction characteristics, enzyme inactivation kinetics and aroma conversion law in the fixation process; the long-range modeling capability of the LSTM network for time sequence characteristics ensures the reliability of the prediction results; the smoothed enzyme activity prediction value sequence provides a stable basis for process control, effectively improving the accuracy and timeliness of tea quality control.

[0137] In some embodiments, the enzyme activity prediction value sequence is input into a fuzzy PID controller optimized by reinforcement learning to generate temperature-rotational speed adjustment instructions, including:

[0138] The membership functions of the input variables and the membership functions of the output variables of the fuzzy PID controller are established, the input variables including the deviation and the rate of change of the deviation of the enzyme activity prediction value sequence, and the output variables including the temperature adjustment amount and the rotational speed adjustment amount;

[0139] The reinforcement learning agent is constructed, and the state space, action space and reward function of the reinforcement learning agent are configured, the state space containing the current PID parameters and the control effect evaluation index, the action space being the adjustment amplitude of the PID parameters, and the reward function being calculated based on the chlorophyll retention rate and the energy consumption ratio;

[0140] The interactive information generated by the interaction between the reinforcement learning agent and the fixation device is used to dynamically optimize the parameter correction strategy of the fuzzy PID controller;

[0141] The enzyme activity prediction value sequence is input into the optimized fuzzy PID controller, and the fuzzy control output information of the temperature adjustment amount and the rotational speed adjustment amount is calculated;

[0142] The fuzzy control output information is de-fuzzified to generate accurate initial temperature-rotational speed adjustment instructions;

[0143] The initial temperature-rotational speed adjustment instructions are checked for feasibility according to the current working state in the fixation chamber;

[0144] The initial temperature-rotational speed adjustment instructions that pass the check are recorded as temperature-rotational speed adjustment instructions and output.

[0145] In the embodiment, the input variable membership functions of the fuzzy PID controller are used to define the fuzzy conversion rules of the deviation and the rate of change of the deviation of the enzyme activity prediction value sequence, and the output variable membership functions correspond to the fuzzy control output range of the temperature adjustment amount and the rotational speed adjustment amount.

[0146] The state space of the reinforcement learning agent contains the current PID parameter combination and its corresponding control effect evaluation index, the action space is defined as the adjustable range of the PID parameters, the reward function is obtained by weighted calculation of the chlorophyll retention rate and the energy consumption ratio, and is used to guide the optimization direction of the agent.

[0147] The interaction information refers to the state transition data and control effect feedback generated by the reinforcement learning agent and the fixation device during the control process, which is used to dynamically optimize the parameter correction strategy of the fuzzy PID controller. Preferably, the defuzzification process converts the fuzzy control output information into accurate initial temperature-rotation speed adjustment instructions using the barycenter method. The feasibility check is based on the current working state parameters of the fixation chamber, including the real-time temperature distribution and mechanical load capacity, to ensure the executability of the adjustment instructions.

[0148] The embodiment realizes adaptive control based on the predicted value of enzyme activity through the collaborative optimization of the reinforcement learning agent and the fuzzy PID controller. The fuzzy PID controller processes the deviation relationship between the predicted value and the target value, and the reinforcement learning agent optimizes the PID parameter adjustment strategy based on the long-term control effect. The combination of the two not only retains the rapid response characteristics of fuzzy control, but also has the continuous optimization ability of reinforcement learning. The finally generated temperature-rotation speed adjustment instructions ensure the process precision while considering the equipment operation stability, forming a closed-loop control system.

[0149] The embodiment realizes precise regulation and control of the predicted value of enzyme activity through the fuzzy PID controller optimized by reinforcement learning. The membership function of the fuzzy PID controller ensures reasonable fuzzy processing of the enzyme activity deviation. The reinforcement learning agent optimizes the PID parameter adjustment strategy based on the chlorophyll retention rate and the energy consumption ratio, improving the adaptive ability of the control system. The defuzzification process and the feasibility check ensure the accuracy and executability of the temperature-rotation speed adjustment instructions, ultimately realizing intelligent optimization control of the fixation process parameters.

[0150] In some embodiments, a reinforcement learning agent is constructed, and the state space, action space, and reward function of the reinforcement learning agent are configured. The state space contains the current PID parameters and control effect evaluation indexes, the action space is the adjustment range of the PID parameters, and the reward function is calculated based on the chlorophyll retention rate and the energy consumption ratio, including:

[0151] The state space is represented by formula (1), and formula (1) is as follows:

[0152] ;

[0153] In formula (1), is the state space, is the parameter combination of the fuzzy PID controller, is the proportional coefficient of the fuzzy PID controller, which controls the reaction strength to the current error, is an integral coefficient of the fuzzy PID controller, used to eliminate steady-state error, is a differential coefficient of the fuzzy PID controller, used to suppress system oscillation, is a control effect evaluation index, is a chlorophyll retention rate deviation, is an energy consumption ratio deviation;

[0154] The action space is represented by formula (2), which is as follows:

[0155] ;

[0156] In formula (2), is an action space, is an adjustment amount of a proportional coefficient, , is a limit adjustment amplitude, is an adjustment amount of an integral coefficient, , is a limit adjustment amplitude, is an adjustment amount of a differential coefficient, , is a limit adjustment amplitude;

[0157] The reward function is represented by formula (3), which is as follows:

[0158] ;

[0159] In formula (3), is a reward function, is a weight coefficient of the chlorophyll retention rate, wherein , is a weight coefficient of the energy consumption efficiency, wherein , is an absolute value deviation of the chlorophyll retention rate, is an energy consumption ratio deviation truncated by an upper limit, ;

[0160] The policy network and the value network of the reinforcement learning agent are initialized, and training hyperparameters are configured, including a learning rate and a discount factor, wherein the learning rate controls the parameter update step, and the discount factor balances immediate rewards and long-term benefits;

[0161] The interaction information generated by the interaction between the reinforcement learning agent and the killing device is obtained, and the parameter correction strategy of the fuzzy PID controller is dynamically optimized according to the interaction information, including:

[0162] Interaction information is collected from the real-time interaction process between the reinforcement learning agent and the killing device, and the interaction information includes control state, executed action, obtained reward and new control state, and the interaction information is expressed by formula (4) , wherein is the state observation value at time , which contains , is the executed parameter adjustment action , is the immediate reward value, which is calculated according to the reward function , and is the new state after executing the adjustment action.

[0163] Based on the interaction information, an experience sample library is constructed, training data is extracted from the experience sample library by random sampling, and a time difference error is calculated, which is expressed by formula (5):

[0164] ;

[0165] , wherein is the time difference error, is a discount factor for balancing immediate reward and long-term return, is a value network for evaluating the value of the current state, is the parameter of the value network.

[0166] The parameter weights of the policy network and the value network are updated using the training data to generate a new parameter adjustment policy, and the updated parameter of the value network is expressed by formula , further, the advantage function is calculated and expressed by the following formula:

[0167] ;

[0168] , wherein is the learning rate, is the current sampling batch, is the batch sampling index, is the time difference error of the sample in the batch, is the probability of the policy network selecting action under state ; the updated parameter of the policy network is expressed by formula , wherein is the policy network, is the parameter of the policy network.

[0169] The new parameter adjustment strategy is applied to parameter updating of the fuzzy PID controller, and the parameter updating includes proportional coefficient updating, integral coefficient updating and differential coefficient updating, and is expressed by the following formula (6):

[0170] ;

[0171] wherein, is the updated proportional coefficient of the fuzzy PID controller, is the updated integral coefficient of the fuzzy PID controller, is the updated differential coefficient of the fuzzy PID controller;

[0172] The updated fuzzy PID controller is evaluated in value, and the fuzzy PID controller passing the value evaluation is verified in stability, so that the control parameters are in a preset safe range;

[0173] The new parameter adjustment strategy passing the stability verification is output and executed.

[0174] In the embodiment, the state space is constituted by parameter combination of the fuzzy PID controller and control effect evaluation indexes, and the control effect evaluation indexes characterize the deviation degree of the chlorophyll retention rate from the target value, reflect the deviation of the actual energy consumption ratio from the expected value. The action space is defined as the adjustment range of the PID parameters, and the limit adjustment range of each parameter 、 、 is determined according to process sensitivity analysis.

[0175] The reward function is constructed by weighted combination of the chlorophyll retention rate deviation and the energy consumption ratio deviation, and the weight coefficients and satisfy the normalization constraint, and are used to balance the dual objectives of quality retention and energy consumption optimization. The experience sample library stores interaction information tuples , and the time difference error is calculated by the difference between the current reward and the state value evaluated by the value network, and the discount factor adjusts the proportion of long-term income.

[0176] The strategy network outputs the probability distribution of the parameter adjustment action, and the value network evaluates the state value. The parameter updating process adopts the advantage function to calculate the strategy gradient, and the current sampling batch The training variance is reduced. The updated PID parameters are adjusted by the action selection probability of the policy network to ensure smooth transition of parameter changes. The stability verification is constrained by the preset safety range to prevent parameter updating from causing system instability. Preferably, the preset safety range is determined by the allowable fluctuation range of the process parameters, combined with the equipment operation limits and product quality requirements, and is verified by experimental tests to ensure the rationality and stable operation of the system.

[0177] The embodiment realizes adaptive adjustment of control parameters through iterative optimization of the reinforcement learning agent and the fuzzy PID controller. The state space completely represents the system control characteristics, the action space provides feasible parameter adjustment directions, and the reward function guides the optimization goal. The experience sample library and the time difference learning realize efficient use of experience, and the collaborative training of the policy network and the value network ensures the reliability of the parameter adjustment strategy. The finally formed closed-loop control system can dynamically optimize the PID parameters according to the process state, while ensuring the retention rate of chlorophyll and improving the energy utilization efficiency.

[0178] The embodiment realizes intelligent optimization of the fuzzy PID controller by constructing a state space containing PID parameters and control effect evaluation indicators, defining an action space for parameter adjustment amplitude, and defining a reward function based on chlorophyll retention rate and energy consumption ratio. The collaborative training mechanism of the policy network and the value network is used to ensure the reliability and stability of the parameter adjustment strategy. Finally, an adaptive closed-loop control system is formed to maintain the chlorophyll retention rate while effectively optimizing the energy consumption efficiency.

[0179] In some embodiments, the enzyme activity prediction value sequence is input into the optimized fuzzy PID controller to calculate the fuzzy control output information of the temperature adjustment amount and the speed adjustment amount, including:

[0180] The enzyme activity prediction value sequence is subjected to deviation calculation to obtain the enzyme activity prediction value deviation and the deviation change rate;

[0181] The enzyme activity prediction value deviation and the deviation change rate are input into the fuzzification interface to be converted into corresponding fuzzy language variables;

[0182] The fuzzy PID controller matches the applicable temperature adjustment fuzzy rules and speed adjustment fuzzy rules;

[0183] The matched temperature adjustment fuzzy rules and speed adjustment fuzzy rules are subjected to inference operation to generate fuzzy control output information containing temperature adjustment amount and speed adjustment amount;

[0184] The fuzzy control output information is subjected to defuzzification processing to generate accurate initial temperature-speed adjustment instructions, including:

[0185] The temperature adjustment amount in the fuzzy control output information is calculated by using the barycenter method to obtain an initial temperature adjustment amount.

[0186] The speed adjustment amount in the fuzzy control output information is calculated by using the maximum membership degree method to obtain an initial speed adjustment amount.

[0187] The initial temperature adjustment amount and the initial speed adjustment amount are combined according to a time sequence to generate an initial temperature-speed adjustment instruction.

[0188] In the embodiment, the enzyme activity prediction value sequence is time sequence data reflecting the change of enzyme activity in the fixation process, which is collected by a sensor in real time, the deviation is obtained by the difference between the current prediction value and the target value, and the deviation change rate represents the change trend of the deviation at adjacent time points. The fuzzification interface is used to convert the continuous numerical variable into a fuzzy language variable, and the conversion process is realized based on a preset membership function, so as to ensure the reasonable mapping of the accurate quantity to the fuzzy concept.

[0189] The temperature adjustment fuzzy rule and the speed adjustment fuzzy rule are pre-set control knowledge bases, and each rule describes the fuzzy relationship between the input variable and the output variable in the form of “if-then”. Preferably, the reasoning operation adopts the Mamdani fuzzy reasoning method, and the results of multiple applicable rules are integrated through fuzzy logic operation. The fuzzy control output information includes a fuzzy set of the temperature adjustment amount and the speed adjustment amount, and the output range is matched with the physical limit of the equipment actuator.

[0190] The defuzzification processing converts the fuzzy output into an accurate control instruction, in which the temperature adjustment amount is calculated by using the barycenter method, and the barycenter position of the area surrounded by the membership function curve and the abscissa is obtained; the speed adjustment amount is calculated by using the maximum membership degree method, and the accurate value corresponding to the maximum point of the membership degree is directly selected. The initial temperature-speed adjustment instruction is combined according to the time sequence to ensure the time sequence coordination of the control action, and the time interval is determined according to the process response characteristics.

[0191] The implementation principle of the embodiment can be understood as follows: the fuzzy PID controller is used to convert the enzyme activity prediction deviation and the change rate thereof into a continuous adjustable control instruction. The fuzzification processing realizes the conversion of the accurate quantity to the fuzzy concept, the fuzzy reasoning simulates the expert control experience, and the defuzzification ensures the executability of the control instruction. The coordinated control of the temperature and the speed ensures the enzyme activity inhibition effect and maintains the stability of the fixation process. The whole control process forms a closed-loop regulation from prediction to execution, and realizes the intelligent optimization of the fixation process parameters.

[0192] The embodiment realizes intelligent perception of fixation process parameters through real-time deviation analysis of enzyme activity prediction value sequence and accurate conversion of fuzzy interface; ensures that control decisions meet process requirements through collaborative reasoning of temperature regulation fuzzy rules and rotation speed regulation fuzzy rules; generates accurate temperature-rotation speed regulation instructions through defuzzification processing, forms a closed-loop control system, effectively suppresses enzyme activity while maintaining process stability, and realizes dynamic optimization control of the fixation process.

[0193] In some embodiments, edge execution nodes of each temperature zone in the fixation chamber are controlled according to the temperature-rotation speed regulation instructions, and the edge execution nodes adjust hot air temperature, fixation cylinder rotation speed and axial segmented wind speed, including:

[0194] The temperature-rotation speed regulation instructions are decomposed into independent control instructions of each temperature zone, and the independent control instructions of each temperature zone include preheating zone temperature instructions, main fixation zone temperature instructions and stable zone temperature instructions;

[0195] The preheating zone temperature instructions are signal converted to generate a gas valve opening degree control signal to adjust the gas combustion intensity of the preheating zone;

[0196] The main fixation zone temperature instructions are power modulation converted to obtain the output power of the heating element to adjust the hot air temperature of the main fixation zone;

[0197] The stable zone temperature instructions are PID operated to obtain a fan rotation speed regulation amount to adjust the temperature interval of the stable zone;

[0198] And the temperature-rotation speed regulation instructions are input into a variable frequency drive to adjust the operating frequency of the drive unit of the fixation cylinder, including:

[0199] According to real-time feedback information of the load current of the drive unit, the output frequency of the variable frequency drive is fine-tuned;

[0200] Actual temperature data of each temperature zone is collected, and a deviation value from the target temperature is calculated;

[0201] The control instruction output of each zone is dynamically corrected according to the temperature deviation value;

[0202] The corrected control instructions are issued to the corresponding edge execution nodes for execution.

[0203] In this embodiment, the control strategy of the temperature zone edge execution node is further optimized by introducing a process parameter coupling analysis model to dynamically correlate the preheating zone gas valve opening degree with the main fixation zone heating power. When changes in the moisture content of the material are detected, the energy distribution ratio of the two zones is automatically coordinated. This ratio coefficient is dynamically calculated based on a regression model trained using historical process data. The temperature control of the main fixation zone uses an adaptive threshold mechanism. The allowable fluctuation range of the temperature is automatically adjusted based on the current material passing speed. When the material passes quickly, the threshold is appropriately relaxed to reduce energy waste. When the material passes slowly, the threshold is tightened to ensure fixation uniformity.

[0204] The self-tuning of the PID parameters in the stabilization zone is achieved through an online identification algorithm. The proportional coefficient, integral coefficient, and derivative coefficient are automatically updated based on the real-time collected temperature response curve characteristics. The adjustment range of the proportional coefficient is positively correlated with the rate of change of the temperature deviation. The integral time constant is dynamically corrected based on the system lag characteristics. Furthermore, the frequency fine-tuning of the frequency converter driver introduces a load torque observer. The load torque variation of the fixation drum is indirectly estimated through harmonic analysis of the motor current. When a sudden increase in torque caused by material accumulation is detected, the frequency output is compensated in advance to prevent a drop in speed.

[0205] Preferably, the dynamic correction of the temperature deviation value uses a fuzzy prediction compensation strategy. Not only the current deviation value is considered, but also the deviation trend of the previous three control periods is predicted and compensated. The compensation weight is determined by the membership function. The response characteristic calibration of the edge execution node introduces digital twin technology. The execution delay characteristics under different working conditions are simulated through a virtual debugging platform to generate a device response time compensation parameter library. In actual control, the corresponding parameters are retrieved and used for advanced control according to the working conditions. Through the multi-level adaptive optimization mechanism, the fixation process control has stronger working condition adaptability and energy optimization potential. While ensuring the retention rate of chlorophyll, more precise energy consumption control is achieved.

[0206] In some embodiments, fixation endpoint determination information is collected, including chlorophyll fluorescence threshold, volatile organic matter ratio, and leaf surface temperature. When the fixation endpoint determination information meets the preset determination condition, the fixation is terminated, including:

[0207] Real-time sampling of the chlorophyll fluorescence threshold is performed to obtain the current fluorescence intensity value.

[0208] The current fluorescence intensity value is compared with the preset fluorescence threshold to calculate the fluorescence intensity decay rate.

[0209] When the fluorescence intensity decay rate reaches the preset enzyme inactivation threshold, a first termination signal is obtained.

[0210] and, monitoring the concentration of each component in the volatile organic compounds in real time, calculating a volatile organic compound ratio, the volatile organic compound ratio including a real-time concentration ratio of linalool and jasmone;

[0211] when the volatile organic compound ratio reaches a preset aroma balance threshold, a second termination signal is obtained;

[0212] and, collecting the leaf surface temperature in real time, calculating the average value of the leaf surface temperature to obtain a leaf surface temperature average value;

[0213] when the leaf surface temperature average value is lower than a preset surface temperature threshold, a third termination signal is obtained;

[0214] The method further comprises:

[0215] when the first termination signal and the second termination signal are received at the same time, a kill-enzyming termination instruction is triggered immediately, and the kill-enzyming termination operation comprises:

[0216] turning off the hot air supply system;

[0217] maintaining the rotation of the drum for residual heat dissipation;

[0218] turning on the cooling fan to accelerate cooling;

[0219] when only the first termination signal or the second termination signal is received, a delay determination program is started, and the third termination signal is detected within a preset time.

[0220] In the embodiment, the dynamic calibration of the chlorophyll fluorescence threshold introduces a machine learning optimization algorithm, and a fluorescence intensity-enzyme activity mapping model of different varieties of tea is established by training historical process data, so that the preset fluorescence threshold can be automatically fine-tuned according to the batch characteristics of the raw materials. The calculation of the fluorescence intensity decay rate uses an adaptive sliding window technology, and the window size is dynamically adjusted according to the current kill-enzyming stage. A larger window is used to smooth noise in the early kill-enzyming stage, and a smaller window is used to improve detection sensitivity in the critical stage.

[0221] The monitoring system of the volatile organic compound ratio integrates electronic nose technology, and distinguishes the characteristic peaks of linalool and jasmone through a pattern recognition algorithm. The preset aroma balance threshold is continuously optimized in combination with expert sensory score data, forming a flavor control curve with self-learning ability. The average value calculation of the leaf surface temperature introduces a spatial weight coefficient, which gives different weights to the temperature measurement points at different positions of the drum, and focuses on monitoring the material accumulation area where temperature deviation is prone to occur.

[0222] Further, when only the first termination signal or the second termination signal is received, a delay determination program is started, and the preset time of the delay determination program uses fuzzy logic control, which comprehensively considers multiple factors such as the current kill-enzyming progress, temperature change rate, and material state, to realize intelligent adjustment of the determination time.

[0223] Preferably, the execution of the termination operation of the fixation introduces a gradient control strategy, the shutdown of the hot air supply system is carried out in stages, the gas pressure is first reduced and then completely cut off, so as to avoid the influence of temperature sudden change on the quality; the rotation speed of the drum is intelligently adjusted according to the residual temperature distribution curve, so as to ensure uniform heat dissipation; the start and stop of the cooling fan is managed by a PID controller, and the air volume is dynamically adjusted according to the real-time cooling rate.

[0224] In this embodiment, by introducing intelligent learning algorithm and adaptive control strategy, the end point determination system has self-optimization capability of process parameters, which improves the adaptability to different raw materials and process conditions while ensuring the accuracy of determination, and realizes the best balance between quality control and energy optimization.

[0225] In the second aspect, the embodiment also provides a tea fixation process adjustment system, which is suitable for the adjustment method of the first aspect. The adjustment system comprises a fixation device, a driving unit, a hot air assembly, a temperature sensor, an olfactory sensor, an image acquisition unit and a control unit. The fixation device comprises a fixation drum, and the fixation drum has a fixation chamber in which the tea to be fixed is contained. The driving unit is in transmission connection with the fixation drum. The hot air assembly comprises at least two fans, and the at least two fans are distributed in the fixation drum in a preset manner. The temperature sensor is arranged in the fixation drum. The olfactory sensor is arranged in the fixation drum. The image acquisition unit is arranged in the fixation drum. The control unit is electrically connected with the driving unit, the hot air assembly, the temperature sensor, the olfactory sensor and the image acquisition unit respectively, and is used for executing the method of the first aspect.

[0226] In this embodiment, the fixation drum of the fixation device is preferably made of stainless steel and has a cylindrical structure, and a flow guide plate is arranged in the fixation chamber for optimizing the turning track of the tea. The driving unit can be a frequency conversion motor cooperating with a speed reduction mechanism, which is used for accurately controlling the rotation speed of the fixation drum, and the transmission connection is realized through a gear set or a belt pulley. The fan in the hot air assembly is a high-temperature-resistant centrifugal fan, which is distributed in the fixation drum in an axial symmetry manner to form a uniform hot air circulation field.

[0227] Preferably, the temperature sensor is an armored thermocouple array, which is distributed along the axial direction and the circumferential direction of the fixation drum, and is used for monitoring the temperature distribution of different regions in real time. The olfactory sensor is a gas detection module based on MEMS technology, which can identify the component change of volatile organic compounds in the fixation process in real time. The image acquisition unit adopts a high-resolution industrial camera cooperating with a near-infrared light source, which is used for capturing the surface morphology and color change of the tea. The control unit is an embedded industrial control system, which integrates a multi-channel signal acquisition card and a motion control card to realize the coordinated control of each execution unit.

[0228] When the system is running, the fixation cylinder rotates at a constant speed under the drive of the driving unit, the hot air assembly generates airflow with controllable temperature, the temperature sensor, the olfactory sensor and the image acquisition unit collect process parameters in real time, and the control unit dynamically adjusts the hot air temperature and the rotation speed of the roller after comprehensively analyzing these parameters, so as to realize precise control of the fixation process. Through multi-sensor information fusion and a closed-loop feedback mechanism, the system can automatically adapt to the fixation requirements of different varieties of tea.

[0229] In this embodiment, a comprehensive fixation process monitoring system is constructed by integrating temperature, odor and visual multi-modal sensing technologies, so that the system can perceive the state changes of tea in all directions. The distributed layout of the hot air assembly cooperates with the intelligent control strategy, effectively solving the problem of uneven temperature existing in traditional fixation equipment. The multi-parameter collaborative adjustment mechanism not only ensures the stability of fixation quality, but also improves energy utilization efficiency, providing an intelligent process solution for tea processing.

[0230] In a third aspect, the embodiment also provides a computer-readable storage medium having computer program instructions stored thereon, wherein the computer program instructions, when executed by a processor, implement the method in the first aspect.

[0231] The computer program involved in the embodiment can be stored in a computer device readable storage medium, including but not limited to magnetic disk, magnetic tape, magnetic card, floppy disk, flash memory, optical disk, optical card, read-only memory (ROM), random access memory (RAM), erasable programmable ROM (EPROM) and electrically erasable programmable ROM (EEPROM) and the like, and also includes other biological, physical or chemical structures that can realize similar or equivalent functions as the above-mentioned storage media, such as DNA, RNA, protein and the like units with information storage ability, etc. In specific embodiments, the storage medium involved can be one of the above-mentioned medium types, or a combination of the above-mentioned medium types. In different embodiments, the computer program involved in the embodiment can be centrally stored in a single medium, or can be distributedly stored in multiple media. The storage medium containing the computer device readable storage medium can be a non-volatile memory or a random access memory. These computer device readable storage media can be built-in in the device, or connected with the device as an external device or part of the external device. In some embodiments, the storage medium with the computer device readable storage medium is deployed locally; in other embodiments, the storage medium can also be deployed remotely from the processor, such as network-attached storage accessed via RF circuit or external port and communication network, wherein the communication network can be Internet, one or more intranets, local area network (LAN), wide area network (WAN), storage area network (SAN) and the like, or appropriate combination thereof, as long as the access of the computer device to the storage medium can be realized. In addition, the computer program involved in the embodiment can be stored in plaintext / ciphertext form, or can be designed as training data, and integrated and reorganized by means of model training to be implicitly saved in the parameter state of the deep neural network or other machine learning model.

[0232] Compared with the prior art, the above technical scheme has the following beneficial effects: through real-time collection of multi-modal sensing information of infrared temperature time sequence information, chlorophyll fluorescence intensity and volatile organic matter concentration, a multi-dimensional feature space of enzyme activity attenuation curve, aroma volatile characteristic spectrum and temperature fluctuation characteristics is constructed, and real-time coupling analysis of enzyme activity dynamics and aroma substance generation in the fixation process is realized. Based on the multi-task learning architecture of the LSTM prediction model, the enzyme activity prediction value sequence, the chlorophyll degradation trend and the aroma generation potential are synchronously predicted, and the problem of lag in monitoring the dynamic change of enzyme activity in the traditional fixation process is solved. The fuzzy PID controller optimized by reinforcement learning dynamically updates the adjustment parameters through the reward function of chlorophyll retention rate and energy consumption ratio, and realizes intelligent generation of temperature-rotation speed regulation instructions. The edge execution node accurately adjusts the hot air temperature and the rotation speed of the fixation cylinder of each temperature zone according to the axial segmented wind speed control model, and effectively improves the uneven temperature distribution phenomenon of the traditional fixation. The end point determination system judges through the multi-condition fusion of chlorophyll fluorescence threshold, volatile organic matter ratio and leaf surface temperature, establishes the determination standard of enzyme activity inhibition and aroma quality optimization, and realizes the coordinated optimization of enzyme activity and aroma quality. The process chain stored by the block chain completes the record of the enzyme activity curve, the aroma quality score and the energy consumption distribution, and provides traceable data support for process optimization. The above technical scheme realizes the coordinated and accurate regulation and control of enzyme activity and aroma substances in the fixation process, improves the fixation quality stability, and significantly reduces the energy consumption.

[0233] The above only describes some embodiments of the present application, and does not limit the protection scope of the present application, and any equivalent device or equivalent process transformation using the content of the specification and drawings, or direct or indirect application in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A method for adjusting the withering process of tea leaves, characterized in that, include: Multimodal sensing information is collected during the blanching process, including infrared temperature time-series information, chlorophyll fluorescence intensity, and volatile organic compound concentration. Feature extraction is performed on the multimodal sensing information to obtain multidimensional features, including enzyme activity decay curve, aroma volatility characteristic spectrum and temperature fluctuation characteristics; Based on the multidimensional features, an LSTM prediction model is dynamically constructed to output a sequence of predicted enzyme activity values ​​for the predicted time period. The enzyme activity prediction sequence is input into a reinforcement learning-optimized fuzzy PID controller to generate a temperature-speed regulation command. The regulation parameters of the fuzzy PID controller are configured to be dynamically updated through a reward function of chlorophyll retention rate and energy consumption ratio. According to the temperature-speed adjustment command, the edge execution nodes of each temperature zone in the blanching chamber are controlled, and the edge execution nodes adjust the hot air temperature, blanching cylinder speed and axial segmented wind speed. Collect blanching endpoint determination information, which includes chlorophyll fluorescence threshold, volatile organic compound ratio and leaf surface temperature. When the blanching endpoint determination information meets the preset determination conditions, the blanching process is terminated.

2. The method for adjusting the tea fixing process according to claim 1, characterized in that, Feature extraction is performed on the multimodal sensing information to obtain multidimensional features, including: The infrared temperature time series information is divided into multiple temperature sequence windows by a sliding window. Spectral analysis is performed on the temperature data within each temperature sequence window to extract the dominant frequency components and the energy distribution characteristics corresponding to each dominant frequency component, thereby obtaining the temperature fluctuation characteristics. In addition, digital filtering is performed on chlorophyll fluorescence intensity to remove high-frequency noise and obtain smooth fluorescence intensity information; The instantaneous change rate of enzyme activity is obtained by performing differential calculation on the smoothed fluorescence intensity information; The instantaneous rate of change of enzyme activity was fitted over time using a nonlinear least squares method to obtain the enzyme activity decay curve of fluorescence intensity over time. The concentration of each component in the volatile organic compounds is monitored in real time, and the dynamic concentration ratio of the preset aroma substances is calculated. Based on the dynamic concentration ratio, a transformation relationship model among the components of volatile organic compounds was established using time series regression analysis. Based on the transformation relationship model, the dynamic change characteristics of aroma substances are extracted, and an aroma volatility characteristic spectrum is constructed. The temperature fluctuation characteristics, enzyme activity decay curve, and aroma volatility characteristic spectrum are normalized to generate the multidimensional features.

3. The method for adjusting the tea fixing process according to claim 1, characterized in that, Based on the aforementioned multidimensional features, an LSTM prediction model is dynamically constructed, outputting a sequence of predicted enzyme activity values ​​for the predicted time period, including: The temperature fluctuation characteristics, enzyme activity decay curves, and aroma volatility characteristic spectra are time-aligned to construct a multidimensional feature input sequence. The multidimensional feature data collected in real time is input into the trained LSTM prediction model; The LSTM network is used to perform forward propagation calculations to output the original enzyme activity prediction sequence for the predicted time period. The original enzyme activity prediction value sequence is smoothed to obtain the enzyme activity prediction value sequence. The LSTM prediction model is constructed through the following steps: Initialize the basic structure of the LSTM prediction model, and configure the number of neurons in the input layer, hidden layer, and output layer; The training features of samples of the same tea category are divided into training sample sets using a sliding time window method; The LSTM prediction model is iteratively trained using the training sample set to optimize the network weight parameters until the LSTM prediction model is fully trained.

4. The method for adjusting the tea fixing process according to claim 1, characterized in that, The predicted enzyme activity sequence is input into a reinforcement learning-optimized fuzzy PID controller to generate a temperature-speed regulation command, including: Establish the membership functions of the input variables and the output variables of the fuzzy PID controller. The input variables include the deviation and the rate of change of the enzyme activity prediction value sequence, and the output variables include the temperature regulation and the speed regulation. Construct a reinforcement learning agent, configure the state space, action space and reward function of the reinforcement learning agent. The state space includes the current PID parameters and control effect evaluation index. The action space is the adjustment range of the PID parameters. The reward function is calculated based on chlorophyll retention rate and energy consumption ratio. The interaction information generated by the reinforcement learning agent interacting with the blanching equipment is used to dynamically optimize the parameter correction strategy of the fuzzy PID controller. The enzyme activity prediction value sequence is input into the optimized fuzzy PID controller to calculate the fuzzy control output information of temperature regulation and speed regulation. The fuzzy control output information is defuzzified to generate a precise initial temperature-speed adjustment command; The feasibility of the initial temperature-speed adjustment command is verified based on the current working status in the blanching chamber. Record the verified initial temperature-speed adjustment command as a temperature-speed adjustment command and output it.

5. The method for adjusting the tea fixing process according to claim 4, characterized in that, A reinforcement learning agent is constructed, and its state space, action space, and reward function are configured. The state space includes the current PID parameters and control performance evaluation metrics. The action space represents the adjustment range of the PID parameters. The reward function is calculated based on chlorophyll retention rate and energy consumption ratio, and includes: The state space is represented by formula (1), which is as follows: ； In formula (1), For state space, For the parameter combination of the fuzzy PID controller, The proportional gain of the fuzzy PID controller. The integral coefficients of the fuzzy PID controller are... The derivative coefficients of the fuzzy PID controller are... As indicators for evaluating the effectiveness of control, This is due to chlorophyll retention rate deviation. This refers to the deviation in energy consumption ratio; The action space is represented by formula (2), which is as follows: ； In formula (2), For the action space, This is the adjustment amount for the proportional coefficient. , for Limit adjustment amplitude, This is the adjustment amount for the integral coefficient. , for Limit adjustment amplitude, This is the adjustment amount for the differential coefficient. , for Limit adjustment amplitude; The reward function is expressed by formula (3), which is as follows: ； In formula (3), For the reward function, The weighting coefficient for chlorophyll retention rate. This is a weighting coefficient for energy efficiency. This represents the absolute deviation of chlorophyll retention rate. To determine the energy consumption ratio deviation at the upper limit, ; Initialize the policy network and value network of the reinforcement learning agent, and configure the training hyperparameters, including the learning rate and discount factor; The interaction information generated between the reinforcement learning agent and the blanching equipment is used to dynamically optimize the parameter correction strategy of the fuzzy PID controller, including: The interaction information generated during the real-time interaction between the reinforcement learning agent and the filming equipment includes control state, executed actions, obtained rewards, and new control state. An experience sample library is constructed based on the interactive information, and training data is extracted from the experience sample library using a random sampling method. The training data is used to update the parameter weights of the policy network and the value network to generate a new parameter adjustment strategy; The new parameter adjustment strategy is applied to the parameter update of the fuzzy PID controller, and the parameter update includes proportional coefficient update, integral coefficient update and derivative coefficient update; The updated fuzzy PID controller is value-evaluated, and the fuzzy PID controller that passes the value evaluation is then stability-verified to ensure that the control parameters are within the preset safe range. The strategy will be adjusted and executed based on the new parameters that have passed stability verification.

6. The method for adjusting the tea fixing process according to claim 4, characterized in that, The predicted enzyme activity sequence is input into the optimized fuzzy PID controller to calculate the fuzzy control output information for temperature and speed regulation, including: The deviation of the predicted enzyme activity value sequence is calculated to obtain the predicted enzyme activity value deviation and the rate of change of deviation. The predicted enzyme activity deviation and the rate of change of deviation are input into the fuzzification interface and converted into corresponding fuzzy language variables. The fuzzy PID controller is used to match applicable fuzzy rules for temperature regulation and fuzzy rules for speed regulation. Perform inference operations on the matched temperature regulation fuzzy rules and speed regulation fuzzy rules to generate fuzzy control output information containing temperature regulation amount and speed regulation amount; The fuzzy control output information is defuzzified to generate a precise initial temperature-speed adjustment command, including: The initial temperature regulation value is obtained by calculating the temperature regulation value in the fuzzy control output information using the centroid method. The initial speed adjustment is calculated by using the maximum membership method on the speed adjustment amount in the fuzzy control output information. The initial temperature adjustment amount and the initial speed adjustment amount are combined in a time sequence to generate an initial temperature-speed adjustment command.

7. The method for adjusting the tea fixing process according to claim 1, characterized in that, The edge execution nodes of each temperature zone in the blanching chamber are controlled according to the temperature-speed adjustment command. The edge execution nodes adjust the hot air temperature, the blanching cylinder speed, and the axial segmented wind speed, including: The temperature-speed regulation command is decomposed into independent control commands for each temperature zone. The independent control commands for each temperature zone include the preheating zone temperature command, the main blanching zone temperature command, and the stabilization zone temperature command. The temperature command for the preheating zone is converted into a signal to generate a gas valve opening control signal, so as to adjust the gas combustion intensity in the preheating zone. The temperature command of the main blanching zone is converted by power modulation to obtain the output power of the heating element, so as to adjust the hot air temperature of the main blanching zone; The stable temperature command is subjected to PID calculation to obtain the fan speed adjustment amount, so as to adjust the temperature range of the stable zone; In addition, the temperature-speed adjustment command is input into the frequency converter to adjust the operating frequency of the drive unit of the blanching cylinder.

8. The method for adjusting the tea fixing process according to claim 1, characterized in that, Collect blanching endpoint determination information, including chlorophyll fluorescence threshold, volatile organic compound ratio, and leaf surface temperature. Blanching is terminated when the blanching endpoint determination information meets preset determination conditions, including: The chlorophyll fluorescence threshold is sampled in real time to obtain the current fluorescence intensity value; The current fluorescence intensity value is compared with a preset fluorescence threshold to calculate the fluorescence intensity decay rate. When the fluorescence intensity decay rate reaches the preset enzyme inactivation threshold, a first termination signal is obtained; In addition, the concentration of each component in the volatile organic compounds is monitored in real time, and the volatile organic compound ratio is calculated, including the real-time concentration ratio of linalool to chlorophyll. When the volatile organic compound ratio reaches a preset aroma balance threshold, a second termination signal is obtained; In addition, the blade surface temperature is collected in real time, and the average value of the blade surface temperature is calculated to obtain the average value of the blade surface temperature; A third termination signal is obtained when the average surface temperature of the blade is lower than a preset surface temperature threshold. The method further includes: When the first termination signal and the second termination signal are received simultaneously, the killing termination command is immediately triggered. When only the first or second termination signal is received, a delay determination procedure is started to detect the third termination signal within a preset time.

9. A tea fixing process control system, characterized in that, The adjustment system, applicable to the adjustment method according to any one of claims 1 to 8, comprises: A fixation device, including a fixation cylinder, the fixation cylinder having a fixation chamber, the fixation chamber containing tea leaves to be fixed; The drive unit is connected to the blanching cylinder via a transmission. The hot air assembly includes at least two fans, and the at least two fans are distributed inside the blanching cylinder in a preset manner; A temperature sensor is installed inside the blanching cylinder; An olfactory sensor is installed inside the blanching cylinder; An image acquisition unit is installed inside the blanching cylinder; The control unit is electrically connected to the drive unit, the hot air assembly, the temperature sensor, the olfactory sensor, and the image acquisition unit, respectively, and the control unit is used to execute the method described in any one of claims 1 to 8.

10. A computer-readable storage medium storing computer program instructions thereon, characterized in that, The computer program instructions, when executed by a processor, implement the method as described in any one of claims 1 to 8.

Images