Cola production process-oriented deep learning optimization extraction control method for active ingredients of Chinese herbal medicines

Abstract

The invention provides a cola production process-oriented Chinese herbal medicine active component deep learning optimization extraction control method, and belongs to the technical field of adaptive control systems. The method comprises the following steps: establishing a staged sequential extraction model, and respectively designing distillation and water decoction processes for volatile and non-volatile components; a hybrid intelligent controller based on LSTM and a feedforward neural network is constructed, and real-time optimization control of process parameters is realized; designing a multi-objective comprehensive reward function, fusing extraction efficiency, product quality, energy consumption and taste indexes, and performing model training through reinforcement learning; a layered online detection system is configured, and dynamic monitoring and data fusion of key indexes are achieved; executing staged sequential optimization control, and adjusting process parameters based on real-time data to realize efficient extraction of active components; and establishing an edge and cloud collaboration mechanism, and continuously optimizing a control strategy through federal learning. According to the method, the extraction efficiency, the control precision and the product quality are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of adaptive control system technology, specifically relating to a deep learning-based optimization control method for the extraction of active ingredients from traditional Chinese medicines in the cola production process. Background Technology

[0002] With the popularization of health concepts in the beverage industry, health-oriented cola containing herbal ingredients has become an important development direction for the industry. These health-oriented cola drinks retain the traditional cola flavor while adding specific active ingredients from traditional Chinese herbs, endowing them with health-promoting functions such as aiding digestion, reducing inflammation, and boosting energy, thus meeting consumer demand for healthy functional beverages. The extraction of active ingredients from these herbal raw materials is crucial to ensuring the quality of these healthy functional beverages, and optimizing and controlling the extraction process is essential.

[0003] Traditional methods employ single-stage extraction or mixed feeding, failing to separate volatile and non-volatile components, leading to loss or mutual inhibition of active ingredients. Existing technologies rely on manual experience or fixed PID control, lacking dynamic response capabilities to multi-variable coupled processes such as temperature, time, and component concentration, and cannot adjust parameters according to real-time extraction progress, resulting in misjudgment of the extraction endpoint. Traditional methods use offline detection with manual sampling, leading to feedback delays, making it difficult to timely control the concentration of bitter components and quantify the synergistic extraction effect of multiple components. Single control models struggle to cope with fluctuations in raw material properties and changes in equipment status, requiring frequent manual intervention to adjust strategies, increasing operational complexity. Traditional methods use fixed-duration extraction, ignoring differences in component dissolution kinetics and lacking a synergistic optimization mechanism for comprehensive objectives such as extraction rate, energy consumption, and taste.

[0004] Therefore, there is an urgent need for a deep learning-based optimization and control method for the extraction of active ingredients from traditional Chinese medicines in cola production processes that integrates temporal feature learning and multi-objective optimization, in order to solve problems such as poor process adaptability, insufficient control precision, and difficulty in synergistic optimization of multiple indicators. Summary of the Invention

[0005] To address the problems existing in the background technology, this invention provides a deep learning-based optimization and control method for the extraction of active ingredients from traditional Chinese medicines in cola production processes, comprising the following steps: S1: clean, dry, and crush the raw materials of Chinese herbal medicine, establish a phased sequential extraction model, set the volatile component distillation extraction process parameters of Chinese herbal medicine, complete the extraction of volatile components, and set the non-volatile component water decoction extraction process parameters of distillation residues after the extraction of volatile components, and perform phased sequential extraction of volatile component Chinese herbal medicine and non-volatile component Chinese herbal medicine; wherein the volatile component Chinese herbal medicine at least includes clove and amomum villosum; the non-volatile component Chinese herbal medicine at least includes radix angelicae dahuricae and high ginger; S2: construct an intelligent controller based on a single-layer LSTM- feedforward neural network hybrid, including a distillation process controller and a water decoction process controller, extract the time sequence features of each extraction process through a single-layer LSTM network, and use an independent feedforward neural network to perform process parameter optimization control; S3: design a multi-objective comprehensive reward function, take the total extraction rate and product quality indicators as positive incentive factors, and take the energy consumption level and bitter component concentration as negative punishment factors, construct a comprehensive reward function by step-by-step calculation of each single reward value, setting of a weight coefficient, and execution of weighted summation, and perform gradient back propagation training optimization on the deep learning model; S4: configure a hierarchical online detection system, use basic sensors to perform real-time monitoring in combination with timed component detection, and establish an online calculation and dynamic monitoring mechanism for volatile oil yield, total flavonoid extraction rate, total polysaccharide extraction rate, and radix angelicae dahuricae extraction rate; S5: perform phased sequential optimization control, adjust the heating temperature and distillation time of the distillation process to complete the extraction of volatile components based on the fused sensor data state, adjust the extraction temperature and extraction time of the water decoction process, real-time monitor the extraction progress and bitter intensity of each component, and realize the optimized extraction of active components; S6: establish an edge computing and cloud collaborative control strategy optimization mechanism, the cloud platform collects historical extraction data to perform offline training of the deep learning model, regularly pushes updated control strategies to the edge controller, and realizes continuous optimization of the control algorithm. In the preferred scheme, the establishment of the phased sequential extraction model in step S1 includes the following specific operation steps: S11: establish a water vapor distillation four-stage precise control model, including: a preheating stage: set the temperature to 80-90 DEG C, the duration is 15 minutes, realize the initial softening of the cell wall of the medicinal material; a main distillation stage: set the temperature range to 95-125 DEG C, use a dynamic temperature adjustment strategy, the duration is 30-60 minutes, realize the sufficient distillation of volatile components; a rectification stage: set the temperature to 105-115 DEG C, the duration is 15 minutes, realize the recovery of high-boiling-point aromatic compounds; a finishing stage: set the temperature to 90-100 DEG C, the duration is 10 minutes, ensure that the light volatile components are completely collected; S12: establish a water decoction extraction multi-section variable temperature precise control model, including: a soaking stage: set the temperature to 45-55 DEG C, the duration is 30 minutes, realize the full wetting and cell wall activation of the medicinal material; a low-temperature extraction stage: set the temperature range to 60-75 DEG C, the duration is 60 minutes, preferentially extract heat-sensitive flavonoid components;The middle temperature extraction stage: set the temperature range of 80-95℃, adopt dynamic temperature adjustment, realize the dissolution of polysaccharide components; the high temperature strengthening stage: set the temperature of 100-110℃, the duration of 20 minutes, strengthen the extraction of difficult soluble components; S13: establish a phased time coordination mechanism, first start the distillation process and real-time monitor the extraction progress, when the volatile oil yield reaches 90% of the preset target, start the decoction process, use the distillation residue for decoction extraction, according to the real-time detection data of volatile oil yield and flavonoid extraction rate, adopt feedback control algorithm to dynamically adjust the temperature and time parameters of the decoction process, ensure that the total extraction time is controlled within 4-5 hours, realize the sequential connection and component complementary optimization between the two processes; S14: execute the targeted extraction control of volatile components, including: clove volatile oil component extraction: control the distillation temperature in the range of 95-110℃, which is consistent with the main distillation stage 95-125℃, preferentially extract eugenol and acetyl eugenol, when the gas chromatography detects that the eugenol concentration reaches 80% of the target value, adjust the temperature to 105-115℃ to enter the rectification stage, extract β-caryophyllene and other sesquiterpenes; amomum volatile oil component extraction: control the distillation temperature in the range of 100-120℃, which is coordinated with the main distillation stage temperature, extract borneol acetate, borneol and camphor, when the online detection reaches the set threshold value of the characteristic aroma intensity of amomum, extend the distillation time to ensure that the terpenes are fully distilled out; S15: execute the segmented extraction control of non-volatile components, including: angelica sinensis effective component extraction: after wetting for 30 minutes at 45-55℃ in the soaking stage, preferentially extract angelica sinensis in the low temperature stage of 60-75℃, when the extraction rate reaches 60%, enter the middle temperature stage of 80-95℃ to extract angelica sinensis polysaccharide and water-soluble flavonoid components, finally, the high temperature stage of 100-110℃ strengthens the extraction of difficult soluble components; galangal effective component extraction: extract galangol in the middle temperature stage of 80-90℃, real-time monitor the bitterness intensity by electronic tongue, when the bitterness component concentration exceeds 15mg / L, reduce the temperature to 75℃; S16: establish the component balance control equation of double medicinal material decoction liquid: ; wherein, is the comprehensive evaluation function value; is the angelica sinensis weight coefficient; is the angelica sinensis extraction rate, which is obtained by near-infrared spectrum detection and standard sample calibration curve calculation; is the galangol weight coefficient; is the galangol extraction rate, which is obtained by high performance liquid chromatography detection calculation; is the polysaccharide weight coefficient; is the polysaccharide extraction rate, which is obtained by phenol-sulfuric acid method detection calculation; is the bitterness punishment weight coefficient; is the bitterness component concentration, which is obtained by electronic tongue detection and quinine standard solution calibration.

[0006] In a preferred scheme, the single-layer LSTM-feedforward neural network hybrid controller in step S2 is constructed including the following specific steps: S21: constructing a distillation process-specific LSTM-feedforward neural network controller, including: input layer configuration: setting 8 distillation-specific parameter input ports, including distillation column temperature, distillation column pressure, heating power, cooling water flow, steam flow, condenser temperature, oil-water separator liquid level, volatile oil collection amount; time series feature extraction layer: configuring a single-layer LSTM hidden layer containing 64 memory units to learn time series patterns, outputting a 64-dimensional distillation time series feature vector; decision network layer: constructing a three-layer fully connected feedforward neural network, the first layer has 128 neurons, the second layer has 64 neurons, and the third layer has 10 neurons, corresponding to 10 discrete distillation control actions and their number mapping: action 1 is temperature increase 1.0℃, action 2 is temperature decrease 1.0℃, action 3 is pressure increase 0.1kPa, action 4 is pressure decrease 0.1kPa, action 5 is flow increase 10L / min, action 6 is flow decrease 10L / min, action 7 is heating power increase 1kW, action 8 is heating power decrease 1kW, action 9 is switching the condenser on and off state, action 10 is switching the separator on and off drainage state; S22: constructing a water decoction process-specific LSTM-feedforward neural network controller, including: input layer configuration: setting 8 water decoction-specific parameter input ports, including water decoction tank temperature, water decoction tank pressure, stirring speed, extract pH value, liquid level, heating jacket temperature, circulating pump flow, filtration pressure; time series feature extraction layer: configuring a single-layer LSTM hidden layer containing 64 memory units to learn time series patterns, outputting a 64-dimensional water decoction time series feature vector; decision network layer: constructing a three-layer fully connected feedforward neural network, the first layer has 128 neurons, the second layer has 64 neurons, and the third layer has 10 neurons, corresponding to 10 discrete water decoction control actions and their number mapping: action 1 is temperature increase 1.0℃, action 2 is temperature decrease 1.0℃, action 3 is stirring speed increase 10rpm, action 4 is stirring speed decrease 10rpm, action 5 is pH value increase 0.1, action 6 is pH value decrease 0.1, action 7 is heating power increase 1kW, action 8 is heating power decrease 1kW, action 9 is switching the circulating pump on and off state, action 10 is switching on and off vacuum concentration; S23: establishing a dual-controller collaborative learning mechanism, setting independent experience replay buffers each containing 5000 training samples, using a priority experience replay strategy combined with random sampling to select 32 samples from each buffer to form a training batch, using an ε-greedy exploration strategy, setting the learning rate to 0.001, and achieving temperature control accuracy ±0.5℃, time control accuracy ±30 seconds, and pressure control accuracy ±0.1kPa adjustment capability through phased training.

[0007] In the preferred scheme, the multi-objective comprehensive reward function design and step-by-step calculation and weighted summation training optimization process in step S3 includes the following specific operation steps: S31: Establishing a quantitative calculation method for each index, including: total extraction rate calculation: ; wherein, is the total extraction rate; is the volatile oil yield; is the total flavonoid extraction rate; is the total polysaccharide extraction rate; is the total polysaccharide extraction rate; ; wherein, is the product quality index; is the aroma weight; is the aroma score; is the taste weight; is the taste score; is the stability weight; is the stability score; Energy consumption level calculation: ; wherein, is the total energy consumption; is the heating power; is the cooling power; is the total extraction time; is the auxiliary equipment power consumption; Bitterness component concentration: is the bitterness component concentration, obtained by electronic tongue detection; S32: Setting the index weight distribution strategy, according to the efficacy requirements of cola products to determine: total extraction rate weight reflects extraction efficiency, product quality index weight reflects product quality, energy consumption level weight reflects economy, bitterness component concentration weight reflects taste control; S33: Performing the specific process of step-by-step calculation and weighted summation, including: S331: Calculating the positive incentive reward value: ; wherein, is the positive incentive reward value; is the total extraction rate weight coefficient; is the total extraction rate; is the product quality weight coefficient; is the product quality index; S332: Calculating the negative penalty reward value: ; wherein, is the negative penalty reward value; is the energy consumption weight coefficient; is the actual energy consumption; is the dynamic reference energy consumption, dynamically calculated according to the raw material input quantity and equipment power; is the bitterness weight coefficient; This refers to the concentration of bitter components. S333: Calculate the overall reward function; (Bitterness threshold) ;in, The value of the comprehensive reward function; This is the positive incentive reward value, calculated by S331; The negative penalty reward value is calculated in step two; S334: Perform reward normalization processing: ;in, This is the normalized reward value; This is the historical average reward. S335: Model training based on the Actor-Critic algorithm framework, calculating the policy loss function: (The original text appears to be incomplete and contains several typos. A more accurate translation would require the full context.) ;in, The value of the loss function; The reward value is negative normalized. Let be the logarithmic probability of an action in a given state; For policy network functions; Number the control action at time t; S336: Perform gradient backpropagation to update network parameters. ;in, For the updated network parameters; These are the current network parameters; The learning rate; For the loss function with respect to parameters The gradient is obtained through backpropagation; S34: Establish a decision-making mechanism based on the comprehensive reward function: when When it is determined to be an excellent control strategy, it continues to be executed; when When it is determined to be a good control strategy, parameter fine-tuning is performed; when It triggers significant adjustments to the control strategy to achieve multi-objective balance optimization.

[0008] In the preferred embodiment, the configuration and monitoring mechanism of the stratified online detection system in step S4 includes the following specific operational steps: S41: Configure a basic sensor real-time monitoring system, including a Pt100 platinum resistance temperature sensor, a piezoresistive pressure sensor, an electromagnetic flowmeter, an ultrasonic level gauge, a glass electrode pH meter, etc., to monitor the basic process parameters in real time at a data acquisition cycle of 1 second, providing real-time feedback signals to the control system; S42: Configure an online monitoring system for volatile oil yield, by installing a Coriolis mass flowmeter at the oil-water separator outlet to monitor the volatile oil collection amount in real time. Combined with the preset raw material weight The yield of volatile oil is calculated in real time using a sliding window averaging process with a data acquisition cycle of 5 minutes. ; wherein, is the volatile oil yield at time t; represents the ratio of the volatile oil weight at time t to the dry weight of raw materials; is the cumulative volatile oil weight at time t, obtained by integrating the mass flow meter; is the dry weight of raw materials, and is the preset feeding weight; S43: configure a gas chromatography-mass spectrometry detection system, set the DB-5ms capillary column specifications to length 30 m, inner diameter 0.25 mm, film thickness 0.25 μm, sample inlet temperature 250℃, ion source temperature 230℃, monitor the concentration of volatile oil components such as eugenol, borneol acetate, camphor, etc. every 20 minutes, establish a standard curve by external standard method for quantitative analysis, detection limit ≤1 mg / L, relative standard deviation ≤3%; S44: configure a near-infrared spectrum detection system, set the scanning wavelength range to 10000-4000 cm -1 , resolution 4 cm -1 , integral time 32 times, use the partial least squares method to establish a quantitative calibration model of flavonoids, polysaccharides and angelica lactone content, collect the extract sample every 5 minutes through the online sampling system for spectrum scanning, detect the concentration of each component , and calculate the extraction rate of each component combined with the real-time extract volume ; S45: configure an electronic nose and an electronic tongue detection system, the electronic nose is set to 8 metal oxide sensor arrays with different selectivity, working temperature 200-400℃, response time 60 seconds, used for monitoring the change of aromatic profile; the electronic tongue is set to 7 different types of taste sensors, including sweet, sour, bitter, salty, umami, astringent and pungent sensors, response time 90 seconds, using quinine solution for bitter taste calibration, monitoring the bitter taste intensity every 10 minutes ; S46: design a layered data fusion algorithm, use time sequence alignment preprocessing, unify the data time stamp of different detection cycles by interpolation method, and then use weighted fusion mechanism, the weight of basic sensor data is set to 0.6, and the weight of component detection data is set to 0.4, and the weighted average is calculated as follows: ; wherein, is the fused data; is the standardized value of the basic sensor data; is the standardized value of the component detection data; output comprehensive monitoring information, realize effective integration and complementary verification of multi-sensor information.

[0009] In a preferred solution, the step-by-step sequence optimization control in step S5 includes the following specific operation steps: S51: Perform distillation stage optimization control, by monitoring the distillation temperature, pressure and volatile oil collection in real time, when the volatile oil yield reaches 90% of the target value or the distillation time exceeds 90 minutes, end the distillation process, collect the distillation residue for water decoction stage; S52: Perform volatile oil-flavonoids synergistic optimization control in water decoction stage, use the distillation residue for water decoction extraction, by monitoring the water decoction temperature in real time, when the total flavonoid extraction rate reaches , appropriately reduce the water decoction temperature to 60-70℃, wherein, is the current total flavonoid extraction rate; 0.7 is the trigger threshold coefficient; is the target flavonoid extraction rate, which is preset according to product demand; preferentially extract heat-sensitive flavonoid components to achieve full utilization of residue resources; S53: Perform polysaccharide-angelica glycoside-bitterness balance control, establish a multi-index joint monitoring temperature adjustment equation: ; wherein, is the temperature adjustment amount; is the bitterness adjustment coefficient; is the current bitterness concentration, obtained by electronic tongue detection; is the bitterness threshold; is the positive regulation coefficient; is the polysaccharide extraction rate deviation value, which is the difference between the target value and the actual value; is the angelica glycoside extraction rate deviation value, which is the difference between the target value and the actual value; when the bitterness concentration exceeds the standard, reduce the extraction temperature, when the polysaccharide or angelica glycoside extraction rate is insufficient, increase the temperature; S54: Perform LSTM-based sequence prediction end point control, use the trained LSTM prediction model to calculate the future extraction rate: ; wherein, is the predicted extraction rate; is the long short-term memory network function; is the historical extraction rate sequence, containing data of the previous 30 time points; is the LSTM model parameter; is the prediction time interval; when the predicted extraction rate increment of each main component is less than 2% and the comprehensive reward function is less than 0.95, consider the 95% confidence interval and start the termination program.

[0010] In a preferred solution, the edge computing in step S6 cooperates with the cloud to establish a strategy optimization mechanism including the following specific operation steps: S61: Establish an edge intelligent control architecture, the edge controller is responsible for the real-time control task of 50 ms period, and a built-in lightweight single-layer LSTM-feedforward neural network model is used for fast response control; S62: Establish a cloud distributed federated learning service, use the FedAvg algorithm framework to train the distillation process and water decoction process data in batches respectively, set the local training rounds of each participating node to 10 rounds, divide the local data set into training set and validation set according to the ratio of 8:2, use the differential privacy mechanism to protect data privacy, and the global model parameter update formula is: ; wherein, is the distillation process global model parameter; is a summation operator, which represents the summation of all k device nodes; k is the device node number, from 1 to K; K is the total number of device nodes; represents the proportion of the data volume of the kth node to the total data volume; is the distillation data volume of the kth node; is the total distillation data volume; is the distillation local model parameter of the kth node; and ; wherein, is the water decoction process global model parameter; is the water decoction data volume of the kth node; is the total water decoction data volume; is the water decoction local model parameter of the kth node; S63: Design an incremental learning update mechanism, set the model update trigger condition as the cumulative new data volume reaching 50 batches or the control effect evaluation index decreasing by more than 3%, establish a model version rollback mechanism and an abnormality detection strategy, rollback to the previous version when the performance of the new model is lower than 95% of the historical best performance, and use a phased dynamic learning rate adjustment strategy: when the update number , , wherein, is the learning rate of the tth update, and 0.001 is the initial learning rate value; when the update number , , wherein, is an exponential operation, and 0.9 is the decay factor; is a floor operation, and the learning rate decay is performed every 20 rounds of update; when the update number , , wherein, 0.0001 is the minimum learning rate value to avoid overfitting; the updated control model is pushed to the edge controller through a wireless network or an Ethernet wired network, and a model deployment confirmation mechanism is established to ensure the success of the update.

[0011] In a preferred scheme, the training and verification method S7 of the deep learning model is also included: S71: Establish a training data set construction method, collect at least 1000 batches of historical extraction data as training samples, and expand the training set to 5000 samples by using data enhancement techniques including noise injection and time series disturbance; S72: Set up a model verification index system, including control accuracy index, temperature control error MAE≤0.5℃, extraction efficiency index, total extraction rate prediction error MAE≤3%, energy consumption prediction index, energy consumption prediction relative error≤5%; S73: Establish a cross-validation mechanism, use 5-fold cross-validation to evaluate the generalization ability of the model, ensure that the performance index R² on the validation set is ≥0.90, and avoid overfitting.

[0012] Compared with the prior art, the beneficial effects of the present application are: the present application fuses a staged sequential extraction model, an LSTM-feedforward neural network hybrid controller, a hierarchical online detection system and an edge cloud collaborative mechanism to build a complete intelligent closed-loop control system, realizing efficient and optimized extraction of active ingredients of Chinese herbal medicines; the staged processing of volatile and non-volatile components avoids extraction conflicts and fully utilizes residual resources; the hybrid controller learns time series characteristics and makes real-time decision outputs to accurately regulate temperature, pressure and other parameters; multi-sensor fusion monitoring provides dynamic feedback of volatile oil yield, flavonoid polysaccharide extraction rate and bitterness concentration; cloud federated learning continuously optimizes edge model parameters, enabling the system to adapt to changes in raw materials and equipment. Experimental results show that this scheme is significantly better than traditional methods in terms of extraction integrity, control accuracy, energy economy and taste coordination; specific performance is as follows: Firstly, the present application establishes a staged sequential extraction model, realizes targeted optimization extraction of volatile and non-volatile components, and effectively improves the extraction efficiency and integrity of active ingredients. The model divides the extraction process into two stages of distillation and decoction, adopts precise control of steam distillation process for volatile components such as clove and amomum, including preheating, main distillation, rectification and tailing stage, and dynamically adjusts temperature and time parameters to ensure that aromatic compounds are fully released; at the same time, for non-volatile components such as white peony root and high-quality ginger, the distillation residue is used for decoction extraction, and through multi-section variable temperature control strategy, heat-sensitive components are preferentially extracted and the dissolution of polysaccharide and white peony root glycoside is balanced, which avoids the conflict of different medicinal material characteristics, reduces the loss of components, and realizes efficient utilization of resources.

[0013] Secondly, the application constructs a hybrid intelligent controller based on single-layer LSTM and feedforward neural network, which significantly improves the control accuracy and response speed by learning the time sequence features of the extraction process and optimizing the process parameters in real time. The controller is designed with special networks for distillation and decoction processes, the input layer integrates key parameters such as temperature, pressure, and flow, the LSTM layer extracts dynamic change patterns, and the feedforward layer outputs discrete control actions such as temperature fine-tuning or equipment state switching. The total extraction rate and product quality are used as positive incentives, while energy consumption and bitterness concentration are used as negative penalties. Through step-by-step calculation and weighted summation, gradient backpropagation training is achieved. This intelligent control mechanism enables the system to adapt to the fluctuations of the extraction process, optimizes temperature and time adjustment, balances extraction efficiency, product taste, and energy economy, and ensures stable improvement of overall performance.

[0014] Thirdly, the application designs a hierarchical online detection method, which combines real-time monitoring with periodic component detection, establishes a dynamic monitoring mechanism for volatile oil yield, flavonoid extraction rate, polysaccharide extraction rate, and bitterness concentration, and improves the reliability and timeliness of feedback data. High-precision temperature sensors, pressure sensors, electromagnetic flowmeters, and other devices are used to continuously collect process parameters. Gas chromatography-mass spectrometry, near-infrared spectroscopy, electronic nose, and electronic tongue are used for periodic analysis of component concentration. Through time alignment and weighted fusion algorithm, multi-source data is integrated to capture subtle changes in extraction progress and provide accurate state input for the controller, supporting dynamic adjustment and end-point prediction.

[0015] Fourthly, the application establishes an edge computing and cloud collaborative control strategy optimization mechanism. The cloud platform trains deep learning models offline and regularly pushes updates to the edge controller, enabling continuous improvement of control algorithms and long-term efficient operation of the system. Lightweight models are deployed on the edge layer to perform real-time control tasks, while the cloud uses a federated learning framework to aggregate historical data and train global models, protects data security with differential privacy, and designs an incremental learning update strategy to dynamically adjust the learning rate to prevent overfitting. Meanwhile, the deep learning training and verification method improves model generalization ability through data augmentation and cross-validation, adapts to different raw materials and equipment changes, maintains advanced control accuracy and extraction effectiveness, and realizes efficient and sustainable optimization of Chinese herbal medicine active ingredient extraction process. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 is a comparison chart of the extraction rate of each component of Example 1 and Comparative Examples 1-5.

[0017] Figure 2 is a comparison chart of the comprehensive extraction rate of Example 1 and Comparative Examples 1-5.

[0018] Figure 3is a control precision and product quality comparison chart of Example 1 and Comparative Examples 1-5.

[0019] Figure 4 is a bitter component concentration and extraction time comparison chart of Example 1 and Comparative Examples 1-5.

[0020] Figure 5 is a comprehensive reward function value comparison chart of Example 1 and Comparative Examples 1-5.

[0021] Figure 6 is a time series performance comparison chart of Example 1 and Comparative Examples 1-2.

[0022] Figure 7 is an improvement degree comparison chart of the average level of component extraction rate of Example 1 relative to Comparative Examples 1-5.

[0023] Figure 8 is a flowchart of the Chinese herbal medicine active ingredient deep learning optimization extraction control method for the cola production process of the present application.

[0024] Figure 9 is a component structure diagram of the Chinese herbal medicine active ingredient deep learning optimization extraction control system for the cola production process of the present application. DETAILED DESCRIPTION

[0025] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0026] With reference to Figure 8 , the present application designs a Chinese herbal medicine active ingredient deep learning optimization extraction control method for the cola production process. The method realizes intelligent control through a hybrid architecture of a long short-term memory network (LSTM network) and a feedforward neural network. The LSTM, which stands for Long Short-Term Memory Networks, is a recurrent neural network that solves the gradient vanishing problem of traditional recurrent neural networks in processing long sequences through the gating mechanism of the forget gate, input gate, and output gate. The feedforward neural network is the simplest type of artificial neural network, in which information is only propagated forward and there is no loop or feedback connection. The specific implementation process is as follows: S1, cleaning, drying, and crushing the raw materials of Chinese herbal medicine, establishing a phased sequential extraction model, setting the volatile component distillation extraction process parameters of Chinese herbal medicine, after the extraction of volatile components, setting the non-volatile component water decoction extraction process parameters of Chinese herbal medicine, and sequentially extracting the volatile component Chinese herbal medicine and the non-volatile component Chinese herbal medicine, wherein the volatile component Chinese herbal medicine at least includes clove and amomum, and the non-volatile component Chinese herbal medicine at least includes Baizhi and Gaoliangjiang.

[0027] The phased sequential extraction model includes the following specific operation steps: S11, establishing a water vapor distillation four-stage precise control model, which includes four continuous temperature control stages of preheating stage, main distillation stage, rectification stage and end stage. The purpose of the preheating stage is to realize the preliminary softening of the cell wall of medicinal materials, and to prepare for the subsequent extraction. The temperature control range is set to 80-90℃, and the duration is 15 minutes. The relatively low temperature in this stage can avoid the loss of heat-sensitive components while relaxing the cell wall structure to create conditions for the release of volatile components. The main distillation stage is the key period for the full distillation of volatile components. The temperature range is set to 95-125℃, and the dynamic temperature adjustment strategy is adopted, with a duration of 30-60 minutes. The temperature range of this stage covers the boiling point range of most volatile organic compounds. Through dynamic adjustment, the extraction effect can be optimized according to the volatilization characteristics of different components. The rectification stage is specially used to realize the recovery of high-boiling aromatic compounds, with a temperature of 105-115℃ and a duration of 15 minutes. The temperature in this stage is at a medium level, which can ensure the volatilization of high-boiling compounds and prevent excessive heating from causing decomposition of components. The end stage is used to ensure the complete collection of light volatile components, with a temperature of 90-100℃ and a duration of 10 minutes. Through moderate cooling, all volatile components can be completely collected.

[0028] S12, a multi-stage variable temperature precise control model for water decoction extraction is established. The model is designed for the extraction characteristics of non-volatile components with four different temperature stages. The purpose of the soaking stage is to achieve full wetting of the medicinal materials and activation of the cell wall. The temperature is set to 45-55°C, and the duration is 30 minutes. The lower temperature makes the medicinal material tissue fully absorb water and swell, and the cell wall permeability increases, laying a foundation for subsequent efficient extraction. The low-temperature extraction stage is specially designed for the preferential extraction of heat-sensitive flavonoids. The temperature range is set to 60-75°C, and the duration is 60 minutes. Flavonoids are usually sensitive to temperature, and excessive temperature may cause structural damage and loss of activity, so a relatively low temperature is used for long-time extraction. The medium-temperature extraction stage adopts a dynamic temperature adjustment strategy to achieve the dissolution of polysaccharide components. The temperature range is set to 80-95°C, and the polysaccharide molecular structure is relatively stable, which requires moderate temperature to promote its release from the plant matrix and dissolve. The high-temperature strengthening stage is used to strengthen the extraction of difficult-to-dissolve components, with a temperature of 100-110°C and a duration of 20 minutes. The high temperature in this stage destroys the stubborn cell wall structure and releases the deep active components.

[0029] S13, a phased time coordination mechanism is established, which realizes the coordination of the two extraction processes through precise time control. First, the distillation process is started and the extraction progress is monitored in real time. The online detection system continuously tracks the change of volatile oil yield. The formula for calculating the volatile oil yield is the percentage of the mass of volatile oil that can be extracted from unit mass of dry medicinal materials to the mass of raw materials. When the volatile oil yield reaches 90% of the preset target value, the water decoction process is automatically started. At this time, the distillation process is basically completed, but about 10% of the improvement space is still reserved to ensure sufficient extraction. The water decoction extraction of the distillation residue realizes the full utilization of the raw materials. The residue still contains rich non-volatile active components such as flavonoids and polysaccharides. According to the real-time detection data of volatile oil yield and flavonoid extraction rate, the system uses a feedback control algorithm to dynamically adjust the temperature and time parameters of the water decoction process. The feedback control algorithm is based on the proportional-integral-derivative control (PID) principle, and its mathematical expression is: ; where, is the proportional gain, is the integral gain, is the differential gain, is the error signal. Ensure that the total extraction time is controlled within 4-5 hours, which ensures sufficient extraction and avoids excessive processing leading to component loss, realizing the sequential connection and component complementary optimization between the two processes.

[0030] S14, the target extraction control of volatile components is performed, which formulates a special extraction strategy for the characteristic components of different medicinal materials. For the extraction of clove volatile oil components, the system controls the distillation temperature in the range of 95-110°C, which is consistent with the temperature range of 95-125°C in the main distillation stage, to preferentially extract clove phenol and acetyl clove phenol, the two main active ingredients. Clove phenol is the characteristic component of clove, with strong aromatic odor and antibacterial activity. When the GC-MS system detects that the concentration of clove phenol reaches 80% of the target value, the system automatically adjusts the temperature to 105-115°C to enter the rectification stage, which is specially designed to extract β-caryophyllene and other sesquiterpenes. The GC-MS system combines the high separation ability of gas chromatography with the accurate identification ability of mass spectrometry to accurately identify and quantify various volatile compounds. For the extraction of volatile oil components of Amomum villosum, the system controls the distillation temperature in the range of 100-120°C, which is coordinated with the temperature of the main distillation stage, to specially extract borneol acetate, borneol and camphor, which are the characteristic components of Amomum villosum. These components have unique aromatic properties and pharmacological activities. When the online detection system detects that the intensity of the characteristic aroma of Amomum villosum reaches the set threshold, the system automatically extends the distillation time to ensure that the terpenes are fully distilled. Terpenes usually have high boiling points and require sufficient distillation time for complete extraction.

[0031] S15, the segmented extraction control of non-volatile components is performed, which is specially designed for the extraction characteristics of water-soluble active ingredients. For the extraction of Angelica dahurica active ingredients, the system first fully wets the Angelica dahurica in the soaking stage at 45-55°C for 30 minutes to create good conditions for subsequent extraction by swelling the Angelica dahurica tissue. Then, it enters the low-temperature stage of 60-75°C to preferentially extract Angelica dahurica glycosides. Angelica dahurica glycosides, also known as Angelicin, belong to furan coumarins and have anti-inflammatory, analgesic and antibacterial effects. When the extraction rate of Angelica dahurica glycosides reaches 60%, the system enters the medium-temperature stage of 80-95°C to extract Angelica dahurica polysaccharides and water-soluble flavonoids. Polysaccharides require relatively high temperatures to be effectively released from plant matrices. Finally, it enters the high-temperature stage of 100-110°C to intensify the extraction of difficult-to-dissolve components. The high temperature in this stage destroys the strong intermolecular bonding force to release deep active ingredients. For the extraction of high-quality ginger active ingredients, the system extracts high-quality gingerol at 80-90°C in the medium-temperature stage. High-quality gingerol is the main active ingredient of high-quality ginger and has the effect of warming the stomach and dispelling cold. The system monitors the bitterness intensity in real time through an electronic tongue, which is a detection device that simulates the human taste system using an array of taste sensors. When the concentration of bitter components exceeds 15 mg / L, the system automatically reduces the temperature to 75°C to control the further extraction of bitter substances.

[0032] S16, establish the component balance control equation of the double medicinal material water decoction liquid, which realizes the coordinated optimization of multi-component extraction through a mathematical model. The control equation is: ; wherein, is the comprehensive evaluation function value for evaluating the overall extraction effect. is the weight coefficient of angelica lactone, preferably 0.3, which is determined based on the importance of angelica lactone in the cola efficacy formula. Angelica lactone, as the main efficacy component, has an important influence on product quality. is the extraction rate of angelica lactone expressed in percentage, which is obtained by near-infrared spectroscopy detection and calculation using a standard sample calibration curve. Near-infrared spectroscopy technology uses near-infrared light with a wavelength range of 780-2526 nm for qualitative and quantitative analysis, with advantages such as non-destructive detection, rapid analysis, and online monitoring. is the weight coefficient of high curcumen, preferably 0.3, which is determined based on the efficacy contribution of high curcumen. is the extraction rate of high curcumen expressed in percentage, which is obtained by high performance liquid chromatography (HPLC) detection and calculation. This method has the characteristics of high separation efficiency and high detection accuracy. is the weight coefficient of polysaccharide, preferably 0.2, which is determined based on the auxiliary efficacy of polysaccharide. is the extraction rate of polysaccharide expressed in percentage, which is obtained by phenol-sulfuric acid detection and calculation. This method is a classic chemical analysis method for detecting polysaccharide content. is the weight coefficient of bitter punishment, preferably 0.2, which is determined based on the need for taste control. Excessive bitter ingredients can affect the taste acceptance of the product. is the concentration of bitter ingredients in mg / L, which is obtained by electronic tongue detection and calibration using quinine standard solution. Quinine is a standard bitter substance commonly used as a reference standard for bitter intensity.

[0033] S2, construct an intelligent controller based on a single-layer LSTM-feedforward neural network hybrid, including a distillation process controller and a water decoction process controller. The single-layer LSTM network extracts the time sequence features of each extraction process, and the independent feedforward neural network is used for process parameter optimization control. The single-layer LSTM-feedforward neural network hybrid controller in step S2 is constructed including the following specific steps.

[0034] S21, a distillation process-specific LSTM-feedforward neural network controller is constructed, which is designed specifically for the characteristics of the distillation process. The input layer is configured with 8 distillation-specific parameter input ports, including distillation column temperature, distillation column pressure, heating power, cooling water flow, steam flow, condenser temperature, oil-water separator level, and volatile oil collection amount, which comprehensively reflect the state information of the distillation process. The time series feature extraction layer is configured with a single-layer LSTM hidden layer containing 64 memory cells to learn the time series pattern, and the number of hidden cells is determined based on the experience rule of input dimension 8x8=64. The LSTM network determines what information to discard from the cell state through the forget gate, what new information to store in the cell state through the input gate, and what part of the cell state to output through the output gate. A 64-dimensional distillation time series feature vector is output for subsequent decision-making. The decision network layer is constructed with a three-layer fully connected feedforward neural network. The first layer contains 128 neurons, which is twice the number of hidden cells, the second layer contains 64 neurons, which is equal to the number of hidden cells, and the third layer contains 10 neurons, which is determined based on the number of control actions. Corresponding to 10 discrete distillation control actions and their number mapping. Action 1 is to increase the temperature by 1.0°C, action 2 is to reduce the temperature by 1.0°C, action 3 is to increase the pressure by 0.1 kPa, action 4 is to reduce the pressure by 0.1 kPa, action 5 is to increase the flow by 10 L / min, action 6 is to reduce the flow by 10 L / min, action 7 is to increase the heating power by 1 kW, action 8 is to reduce the heating power by 1 kW, action 9 is to switch the condenser on and off, and action 10 is to switch the separator on and off to drain water. These control actions cover the adjustment requirements of all key parameters in the distillation process.

[0035] S22, a water decoction process specific LSTM-feedforward neural network controller is constructed, which is specifically designed for the characteristics of the water decoction process. The input layer is configured with 8 water decoction specific parameter input ports, including water decoction tank temperature, water decoction tank pressure, stirring speed, extract pH, liquid level, heating jacket temperature, circulating pump flow, and filtration pressure, which comprehensively monitor the key states of the water decoction extraction process. The time series feature extraction layer is configured with a single-layer LSTM hidden layer containing 64 memory units to learn the time series pattern, outputting a 64-dimensional water decoction time series feature vector, and the LSTM network captures the complex rules of the change of each parameter over time in the water decoction process. The decision network layer is constructed with a three-layer fully connected feedforward neural network, the first layer contains 128 neurons, the second layer contains 64 neurons, and the third layer contains 10 neurons, corresponding to 10 discrete water decoction control actions and their number mapping. Action 1 is to increase the temperature by 1.0℃, action 2 is to reduce the temperature by 1.0℃, action 3 is to increase the stirring speed by 10rpm, action 4 is to reduce the stirring speed by 10rpm, action 5 is to increase the pH value by 0.1, action 6 is to reduce the pH value by 0.1, action 7 is to increase the heating power by 1kW, action 8 is to reduce the heating power by 1kW, action 9 is to switch the circulating pump on and off, and action 10 is to switch on and off vacuum concentration, which accurately adjusts the process parameters of the water decoction process.

[0036] S23, a double-controller collaborative learning mechanism is established, which improves the overall control performance through distributed learning. The system sets up independent experience replay buffer to accommodate 5000 training samples each, experience replay is an important technique in reinforcement learning, which improves sample utilization efficiency and breaks the time correlation between data. A priority experience replay strategy combined with random sampling is used to select 32 samples from each buffer to form a training batch, priority experience replay prioritizes samples with large time difference errors for training to improve learning efficiency. An ε-greedy exploration strategy is used, which is an exploration strategy in reinforcement learning, with a probability of ε selecting a random action for exploration and a probability of 1-ε selecting the current optimal action for utilization, the initial value of ε is set to 0.9, indicating that exploration is the main at the beginning, and the decay rate is set to 0.995, indicating that ε gradually decreases from exploration to utilization as the training progresses. The learning rate is set to 0.001, which controls the step size of network parameter update, and a proper learning rate ensures the stability and convergence of the training. Through phased training, the temperature control accuracy is ±0.5℃, the time control accuracy is ±30 seconds, and the pressure control accuracy is ±0.1kPa, which ensures that the system meets the requirements of precise control.

[0037] S3, design a multi-objective comprehensive reward function, take the total extraction rate and product quality index as positive incentive factors, take the energy consumption level and bitter component concentration as negative punishment factors, construct the comprehensive reward function through step-by-step calculation of each single reward value, setting of weight coefficients, and execution of weighted summation, and perform gradient back propagation training optimization on the deep learning model. The multi-objective comprehensive reward function design and step-by-step calculation and weighted summation training optimization process in step S3 include the following specific operation steps.

[0038] S31, establish a quantitative calculation method for each index, which provides a quantitative basis for multi-objective optimization. The total extraction rate calculation uses the formula: wherein, is the total extraction rate expressed in percentage, is the volatile oil yield expressed in percentage, is the total flavonoid extraction rate expressed in percentage, is the total polysaccharide extraction rate expressed in percentage, is the total polysaccharide extraction rate expressed in percentage, and the formula reflects the overall extraction effect by the arithmetic mean of the extraction rates of the four main components. The product quality index calculation uses the formula: wherein, is the product quality index, is the aroma weight, preferably 0.4, is the aroma score using a 1-10 point evaluation standard, is the taste weight, preferably 0.4, is the taste score using a 1-10 point evaluation standard, is the stability weight, preferably 0.2, is the stability score using a 1-10 point evaluation standard, and the formula takes into account the sensory quality and physical stability of the product. The energy consumption level calculation uses the formula: wherein, is the total energy consumption in kWh, is the heating power in kW, is the cooling power in kW, is the total extraction time in h, is the auxiliary equipment power consumption in kWh, and the formula comprehensively calculates the energy consumption of the entire extraction process. The bitter component concentration is the bitter component concentration in mg / L, obtained by electronic tongue detection, and the electronic tongue system is configured with 7 different types of taste sensors to accurately quantify the concentration of various taste components.

[0039] ​​​S32, set the index weight allocation strategy, which determines the importance of each index according to the efficacy requirements of cola products. Total extraction rate weight Reflect the importance of extraction efficiency, which directly affects the content of efficacy ingredients and economic benefits of products. Product quality index weight Reflect the importance of product quality, which determines the acceptance of consumers and market competitiveness. Energy consumption level weight Reflect the importance of economy, and energy consumption control directly affects production cost and environmental friendliness. Bitterness component concentration weight Reflect the importance of taste control, and moderate bitterness control improves the taste acceptance of products.

[0040] S33, execute the specific process of step-by-step calculation and weighted summation, which realizes multi-objective optimization through systematic mathematical operations. S331, calculate the positive incentive reward value, using the formula: ; wherein, is the positive incentive reward value, is the total extraction rate weight coefficient, preferably 0.4, is the total extraction rate expressed in percentage, is the product quality weight coefficient, preferably 0.3, is the product quality index, which takes extraction efficiency and product quality as positive incentive factors to promote the system to develop in the direction of optimization. S332, calculate the negative punishment reward value, using the formula: ; wherein, is the negative punishment reward value, is the energy consumption weight coefficient, preferably 0.2, is the actual energy consumption in kWh, is the dynamic benchmark energy consumption in kWh, which is dynamically calculated according to the raw material input and equipment power, is the bitterness weight coefficient, preferably 0.1, is the bitterness component concentration in mg / L, is the bitterness threshold in mg / L, preferably 15, which takes high energy consumption and excessive bitterness as negative punishment factors to guide the system to optimize in the direction of energy saving and taste improvement. S333, calculate the comprehensive reward function, using the formula: ; wherein, is the comprehensive reward function value, is the positive incentive reward value calculated by S331, is the negative punishment reward value calculated by S332, which realizes multi-objective coordinated optimization through the balance of positive incentive and negative punishment. S334, execute reward normalization processing, using the formula: ; wherein, is the normalized reward value, is the historical reward mean, is the historical reward standard deviation, the normalization eliminates the magnitude difference between different batches of data to improve the stability of training. S335, model training is performed based on the Actor-Critic algorithm framework, and the policy loss function is calculated. The Actor-Critic algorithm is a reinforcement learning algorithm that combines value-based methods and policy-based methods. The Actor is responsible for selecting actions, and the Critic is responsible for evaluating the value of actions. The loss function calculation uses the formula: ; wherein, is the loss function value, is the negative normalized reward value, is the log probability of the action under the given state, is the policy network function, is the control action number at time , and is the system state vector at time . The loss function optimizes the policy network by maximizing the expected reward. S336, gradient backpropagation is performed to update the network parameters, using the formula: ; wherein, is the updated network parameter, is the current network parameter, is the learning rate, preferably 0.001, is the gradient of the loss function with respect to the parameter , which is calculated by the backpropagation algorithm. Backpropagation is a core algorithm in deep learning that efficiently calculates the parameter gradients of complex networks.

[0041] S34, a decision-making mechanism based on the comprehensive reward function is established. This mechanism realizes adaptive control strategy adjustment through hierarchical judgment. When , it is determined that the control strategy is excellent and the current control parameters are continued to be executed, indicating that the system is in good running state and all indicators have reached the expected target. When 0.85 ≥ , it is determined that the control strategy is good and parameter fine-tuning is performed, further optimizing system performance through small adjustments to control parameters. When , major control strategy adjustment is triggered, and the system will re-evaluate the current control strategy and make significant parameter adjustments or strategy switching, achieving multi-objective balance optimization to ensure the system always maintains the best running state.

[0042] S4, configure a hierarchical online detection system, which uses real-time monitoring of basic sensors combined with timed component detection to establish online calculation and dynamic monitoring mechanisms for volatile oil yield, total flavonoid extraction rate, total polysaccharide extraction rate, and angelica lactone extraction rate. The hierarchical online detection system configuration and monitoring mechanism in step S4 includes the following specific operation steps.

[0043] S41 is equipped with a basic sensor real-time monitoring system. This system achieves real-time monitoring of process parameters through various high-precision sensors. The system includes a Pt100 platinum resistance temperature sensor, which uses platinum as the thermistor element and has a resistance of 100Ω at 0℃. It features a measurement accuracy of ±0.1℃, a temperature coefficient of 0.003851Ω / Ω / ℃, and a measurement range of -200℃ to +850℃. A piezoresistive pressure sensor has a measurement accuracy of ±0.25%FS, where FS represents full scale. This sensor converts pressure changes into resistance changes and then into an electrical signal through the piezoresistive effect. An electromagnetic flowmeter operates based on Faraday's law of electromagnetic induction; when a conductive fluid passes through a magnetic field, an induced electromotive force is generated, the magnitude of which is proportional to the fluid velocity. An ultrasonic level gauge calculates the liquid level by emitting ultrasonic waves and receiving the reflected echoes, featuring non-contact measurement, high accuracy, and good stability. A glass electrode pH meter measures the pH value of a solution by the potential difference between a glass electrode and a reference electrode, featuring fast response and high measurement accuracy. These sensors monitor basic process parameters in real time with a data acquisition cycle of 1 second. High-frequency sampling promptly captures minute changes in process parameters, providing real-time feedback signals to the control system to ensure the accuracy and timeliness of control decisions.

[0044] The S42 is equipped with an online monitoring system for volatile oil yield. This system achieves real-time monitoring of the volatile oil collection rate through precision flow measurement. The system monitors the volatile oil collection rate in real time by installing a Coriolis mass flow meter at the oil-water separator outlet. The Coriolis mass flow meter measures fluid mass flow rate based on the Coriolis force principle. When fluid passes through a vibrating measuring tube, it experiences a phase difference due to the Coriolis force. This phase difference is proportional to the mass flow rate. This flow meter offers advantages such as direct measurement of mass flow rate, insensitivity to fluid density and viscosity, accuracy up to ±0.1%, and simultaneous density measurement. (Combined with preset raw material weight...) The volatile oil yield is calculated in real time using a sliding window averaging method with a 5-minute data acquisition cycle. Sliding window averaging is a data processing method that uses a fixed-length window to average time-series data, reducing noise and smoothing data fluctuations. The calculation formula is as follows: ;in, for The yield of volatile oil at any given time is expressed as a percentage. express The ratio of the weight of the volatile oil to the dry weight of the raw material at any given time. for The cumulative weight of the volatile oil at any given time, expressed in grams, is obtained by integration using a mass flow meter. The raw material dry weight is set in grams as the preset feeding weight, and this calculation method reflects the extraction progress of volatile oil in real time.

[0045] Then S43, configure gas chromatography-mass spectrometry detection system, the system through high precision analysis technology to realize the qualitative and quantitative analysis of volatile components. The system is provided with a DB-5ms capillary column, the chromatographic column has good thermal stability and chemical inertness with 5% phenyl-95% methyl polysiloxane as a stationary phase, and the specification is 30m in length, 0.25mm in inner diameter, 0.25μm in film thickness, and 325℃ in maximum use temperature. The inlet temperature is set to 250℃ to ensure complete vaporization of the sample, and the ion source temperature is set to 230℃ to ensure ionization efficiency, and the concentration of volatile oil components such as eugenol, borneol acetate and camphor is monitored in a 20-minute detection cycle. The system establishes a standard curve by external standard method for quantitative analysis, the external standard method is a classical method for chromatographic quantitative analysis, a standard solution with a known concentration is prepared to establish a linear relationship between the concentration and the peak area, and then the sample concentration is calculated according to the sample peak area. The detection limit ≤1mg / L represents the minimum concentration of the measured substance detected by the method, and the relative standard deviation ≤3% represents the precision of the detection method, and the relative standard deviation calculation formula is: ; wherein, is the standard deviation, is the average value.

[0046] S44, configure near infrared spectrum detection system, the system realizes rapid detection of multiple components through spectral analysis technology. The system is provided with a scanning wavelength range of 10000-4000cm -1 corresponding to a wavelength range of 780-2500nm, a resolution of 4cm -1 represents the minimum wave number difference resolved by the spectrometer, and the integral time is 32 times, which means that the average value is taken 32 times for each measurement to improve the measurement accuracy. The system uses partial least squares (PLS) to establish a quantitative calibration model of flavonoids, polysaccharides and angelica lactone content, and the partial least squares is a multivariate statistical analysis method, which is used to establish a quantitative relationship model between spectral data and component content, and a regression model is established by extracting the principal components of spectral variables and concentration variables. The number of calibration samples ≥100 ensures the representativeness of the model, and the number of verification sets ≥0.95 indicates that the proportion of explained variance of the model reaches more than 95%, ≤5% indicates that the root mean square error, i.e. the standard deviation of prediction, is controlled within 5%, The calculation formula is: ; wherein, is the actual value, is the predicted value, is the sample number. The system collects the sample of the extraction liquid every 5 minutes through an online sampling system to perform spectral scanning, detects the concentration of each component , and calculates the extraction rate of each component combined with the real-time extraction liquid volume , and the detection frequency timely tracks the change trend of the component concentration.

[0047] S45, configure the electronic nose and electronic tongue detection system, which realizes the objective quantification of aroma and taste through biomimetic sensing technology. The electronic nose system sets up 8 different selective metal oxide sensor arrays, and the electronic nose is a device that uses gas sensor arrays to simulate the human olfactory system to identify and detect volatile compounds. The sensor array includes TGS2600 for detecting hydrogen and ethanol vapor, TGS2602 for detecting ammonia and volatile amines, TGS2610 for detecting butane and propane, TGS2611 for detecting methane, TGS2620 for detecting alcohol and organic solvent vapor, MQ-3 for detecting alcohol concentration, MQ-135 for detecting air quality, MQ-138 for detecting formaldehyde and other aldehyde compounds, which cover various volatile components that may be produced during the extraction process of Chinese herbal medicine. The working temperature is set to 200-400℃ to ensure the sensitivity and stability of the sensor, and the response time is controlled within 60 seconds to ensure the real-time detection of the system, which is used to monitor the change of aroma profile and identify and quantify the change of aroma components through the differential response mode of the sensor array. The electronic tongue system sets up 7 different types of taste sensors, and the electronic tongue is a detection device that simulates the human taste system using an array of taste sensors. The sensors include sweet sensors for detecting sugars and sweeteners, sour sensors for detecting organic acids, bitter sensors for detecting alkaloids and bitter compounds, salty sensors for detecting inorganic salts, umami sensors for detecting amino acids and nucleotides, astringent sensors for detecting tannins, and spicy sensors for detecting capsaicin compounds, with a response time controlled within 90 seconds. The system uses quinine solution to calibrate the bitterness, and quinine is a standard bitter substance commonly used as a reference standard for bitterness intensity, with a calibration concentration range of 0.1-100mg / L, and a 10-minute detection period for monitoring bitterness intensity , which can detect abnormal changes in the concentration of bitter components in time.

[0048] S46, design a layered data fusion algorithm to improve the reliability and accuracy of the detection system through multi-sensor information integration. The system uses time alignment preprocessing, as different sensors have different detection periods and need to be time synchronized, the data timestamps of different detection periods are unified through interpolation method, and the interpolation method estimates the intermediate value between known data points to realize the time alignment of data. The system then uses a weighted fusion mechanism, with the basic sensor data weight set to 0.6 to reflect the importance of its real-time and stability, and the component detection data weight set to 0.4 to reflect the importance of its accuracy and professionalism, and the weighted average calculation is as follows: ; wherein, is the fused data, is the standardized value of the basic sensor data, To standardize the values of the ingredient detection data, the normalization process eliminates the magnitude differences between different types of data, and the output comprehensive monitoring information realizes the effective integration and complementary verification of multi-sensor information, and improves the robustness and reliability of the entire detection system.

[0049] S5, performing phased sequential optimization control, based on the fused sensor data state, adjusting the heating temperature and distillation time of the distillation process to complete the extraction of volatile ingredients, adjusting the extraction temperature and extraction time of the decoction process, monitoring the extraction progress and bitterness intensity of each ingredient in real time, and realizing the optimized extraction of active ingredients. The phased sequential optimization control in step S5 includes the following specific operation steps.

[0050] S51, performing distillation phase optimization control, which realizes automatic management of the distillation process through real-time monitoring and intelligent judgment. The system continuously tracks the progress of the distillation process by monitoring key parameters such as distillation temperature, pressure, and volatile oil collection amount, and automatically ends the distillation process when the volatile oil yield reaches 90% of the target value or the distillation time exceeds 90 minutes. The 90% extraction rate threshold ensures the sufficient extraction of main volatile ingredients, while the 10% safety margin prevents over-distillation, and the 90-minute time limit prevents excessive heating that may cause ingredient decomposition or excessive energy consumption. After the distillation process is completed, the system automatically collects the distillation residue for the next decoction phase. The distillation residue still contains rich non-volatile active ingredients, which are extracted through subsequent decoction to maximize the utilization of raw materials.

[0051] S52, performing volatile oil-flavonoid synergistic optimization control in the decoction phase, which realizes the coordinated optimization of the extraction of two types of ingredients. The system uses the distillation residue for decoction extraction to realize resource recycling, and dynamically adjusts the decoction temperature based on real-time monitoring and online detection data of flavonoid extraction rate. When the total flavonoid extraction rate reaches , the system automatically reduces the decoction temperature to 60-70°C, where is the current total flavonoid extraction rate expressed in percentage, and 0.7 is the trigger threshold coefficient indicating that the extraction rate reaches 70% of the target value, is the target flavonoid extraction rate expressed in percentage and preset according to product requirements. The purpose of the temperature reduction measure is to preferentially extract heat-sensitive flavonoid components to avoid damage to the structure of these compounds at high temperatures. Flavonoids are generally sensitive to temperature, and high temperatures may cause changes in molecular structure and loss of biological activity, thus realizing the full utilization of residue resources and maximizing the recovery of valuable ingredients.

[0052] S53, performing polysaccharide-bai-ting-ting bitterness balance control, which realizes complex balance optimization through multi-index joint monitoring. The system establishes a multi-index joint monitoring temperature adjustment equation: ; wherein, is the temperature adjustment amount in ℃, is the bitterness adjustment coefficient in ℃·L / mg, preferably taking the value 2.5, which determines the sensitivity of the temperature adjustment to the change of the bitterness concentration. is the current bitterness concentration in mg / L, obtained by electronic tongue detection, is the bitterness threshold in mg / L, preferably taking the value 15, which is determined based on the product taste acceptance. is the positive adjustment coefficient in ℃, preferably taking the value 1.2, which controls the temperature compensation amplitude when the active ingredients are insufficient, is the polysaccharide extraction rate deviation value in percentage, which is the difference between the target value and the actual value, is the angelica lactiflorin extraction rate deviation value in percentage, which is the difference between the target value and the actual value. The design logic of this equation is to reduce the extraction temperature to reduce the further release of bitter substances when the bitterness concentration exceeds the standard, and to increase the temperature to promote the dissolution of these beneficial ingredients when the polysaccharide or angelica lactiflorin extraction rate is insufficient, achieving a dynamic balance between multiple goals.

[0053] S54, execute the LSTM-based sequence prediction end-point control, which is achieved by artificial intelligence technology to intelligently judge the extraction end-point. The system uses the trained LSTM prediction model to calculate the future extraction rate, and the LSTM model predicts the future trend by learning the time sequence rules in the historical data, and the prediction formula is: ; wherein, is the predicted extraction rate in percentage, is the long short-term memory network function, is the historical extraction rate sequence containing data of the previous 30 time points, and the 30-time-point historical data window can capture enough time sequence information without causing high computational complexity, is the LSTM model parameter obtained by offline training, is the prediction time interval in minutes, preferably taking the value 10, and a 10-minute prediction interval can provide sufficient warning time and maintain the accuracy of the prediction. When the prediction of the extraction rate increment of each main component is less than 2% and the comprehensive reward function , the system considers the 95% confidence interval and starts the termination program, the 2% increment threshold indicates that the extraction process is close to the equilibrium state and the benefit of continuing extraction is limited, the 95% confidence interval is a statistical concept indicating that there is 95% confidence in the credibility of the prediction result, and the comprehensive reward function greater than 0.85 indicates that the current state has reached an excellent level.

[0054] S6, establish an edge computing and cloud collaborative control strategy optimization mechanism, the cloud platform collects historical extraction data for offline training of a deep learning model, and regularly pushes updated control strategies to the edge controller to realize continuous optimization of the control algorithm. The edge computing and cloud collaborative control strategy optimization mechanism established in step S6 includes the following specific operation steps.

[0055] S61, establish an edge intelligent control architecture that realizes low-latency real-time control through edge computing technology. The edge controller is responsible for real-time control tasks with a 50 ms cycle, and the 50 ms control cycle meets the strict requirements of industrial control systems for real-time performance, ensuring that control instructions respond promptly to changes in process parameters. The edge controller has a built-in lightweight single-layer LSTM-Feedforward Neural Network model for fast response control. Lightweight design reduces the number of network layers and the number of parameters to reduce computational complexity while ensuring control accuracy, enabling the model to run efficiently on resource-constrained edge devices. The edge controller uses an industrial-grade embedded computer with an ARM Cortex-A72 quad-core processor running at 1.8 GHz, 8GB of LPDDR4 memory, and 128GB of BeMMC storage, running a Linux real-time operating system. These hardware configurations ensure the reliability and real-time performance of edge computing. S62, establish a cloud distributed federated learning service that realizes collaborative optimization of the model through distributed machine learning technology. The system uses the FedAvg algorithm framework, which is a federated learning algorithm proposed by Google in 2017. It allows multiple clients to train models locally and then upload model parameters to the server for aggregation, enabling distributed machine learning without sharing raw data. The system trains in batches for distillation and decoction processes, with 10 local training rounds per participating node. Local training takes full advantage of the data characteristics of each node, with the local dataset divided into training and validation sets in an 8:2 ratio. This allocation ensures both the adequacy of training data and the effectiveness of validation. The system uses differential privacy mechanisms to protect data privacy. Differential privacy is a strict privacy protection definition that protects individual privacy by adding noise to data. The noise parameter ε = 1.0 represents the privacy budget, and a smaller ε value means higher privacy protection but may affect data usability. The global model parameter update formula is: ; wherein, is the global model parameter for the distillation process, is the summation operator representing the sum of all device nodes, is the device node number from 1 to , is the total number of device nodes in units of , and represents the proportion of the data volume of the th node to the total data volume.For the first The amount of distillation data per node is measured in units of 1. The total amount of distillation data is expressed in units of [number]. For the first The local model parameters for distillation at each node. And: ;in, These are the global model parameters for the decoction process. For the first The amount of data for each node is measured in units of 1. The total data volume for water decoction is expressed in units of [number]. For the first The local model parameters of each node are updated using these update formulas, which achieve model parameter aggregation based on data volume weighting.

[0056] S63 introduces an incremental learning and update mechanism that continuously improves the model through an intelligent update strategy. The system sets model update trigger conditions to include 50 batches of new data or a performance degradation threshold exceeding 3%. The 50-batch threshold ensures the necessity of updates, while the 3% performance degradation threshold promptly detects model degradation. The system establishes a model version rollback mechanism and anomaly detection strategy. When the performance of a new model falls below 95% of its historical best performance, it rolls back to the previous version. This rollback mechanism prevents severe performance degradation due to new data quality issues or training anomalies. The system employs a phased dynamic learning rate adjustment strategy. The learning rate controls the step size of network parameter updates. An appropriate learning rate adjustment strategy achieves rapid convergence in the early stages of training and fine-tunes in the later stages. When the update count... hour, ,in For the first The learning rate for each update is set to 0.001, with the initial learning rate value determined based on experience and pre-experiments. A higher initial learning rate is beneficial for rapid learning. As the number of updates... hour, ,in For exponential operations, 0.9 is the decay factor, determined based on optimization using the learning curve. To perform floor function rounding, the learning rate decays every 20 updates. This gradual decay of the learning rate promotes stable convergence of the model. When the update count... hour, The minimum learning rate of 0.0001 is used to avoid overfitting, and a smaller learning rate ensures the stability of the model during long-term training. The system pushes the updated control model to the edge controller via either a wireless network or a wired Ethernet network, and establishes a model deployment confirmation mechanism to ensure successful updates. This mechanism verifies the integrity and correctness of the model transmission through a handshake protocol.

[0057] S7, a training and verification method for the deep learning model, which ensures the reliability and generalization ability of the model through a systematic training and verification process.

[0058] S71, a method for establishing a training data set construction method, which provides a solid foundation for model training through sufficient data preparation. The system collects at least 1000 batches of historical extraction data as training samples, and the data volume of 1000 batches covers various working conditions and operation changes, providing a rich sample space for model learning. The system uses data augmentation techniques including noise injection, time series disturbance, etc. to expand the training set to 5000 samples. Data augmentation is an important technique in machine learning, which generates new training samples by appropriately transforming the original data. Noise injection improves the robustness of the model by adding random noise to the data. Time series disturbance increases the diversity of data by adjusting the time scale of the time series. The 5000-sample training set size meets the data volume requirements of the deep learning model.

[0059] S72, set up a model verification index system, which ensures the quality of the model through multi-dimensional performance evaluation. The verification index includes control accuracy index, temperature control error MAE≤0.5℃, MAE is Mean Absolute Error, the calculation formula is: ; this index evaluates the accuracy level of temperature control. Extraction efficiency index, total extraction rate prediction error MAE≤3%, this index evaluates the accuracy of the model in predicting the extraction process. Energy consumption prediction index, energy consumption prediction relative error≤5%, relative error eliminates the influence of different batch energy benchmark value differences, more accurately reflects the prediction accuracy. These indicators comprehensively evaluate the model performance from control accuracy, process effect, economy and other dimensions.

[0060] S73, establish a cross-validation mechanism, which evaluates the generalization ability of the model through statistical methods. The system uses 5-fold cross-validation to evaluate the generalization ability of the model. Cross-validation is a classic technique for model verification. The data set is divided into 5 subsets, and 4 subsets are used to train the model and 1 subset is used to verify the model, repeated 5 times. 5-fold cross-validation makes full use of limited data while providing reliable performance evaluation. Ensure that the performance indicators on the validation set , R-squared indicates the proportion of variance explained by the model, which means that the model explains more than 90% of the data variation, has a good fitting effect, avoids overfitting problem and ensures the reliability of the model in practical application.

[0061] Reference Figure 9The Chinese herbal medicine active ingredient deep learning optimization extraction control method for a cola production process of the present application is realized by a Chinese herbal medicine active ingredient deep learning optimization extraction control system. The system is an automated system integrating multi-stage extraction equipment, an intelligent controller, a multi-sensor detection network, and a cloud-edge collaborative computing architecture.

[0062] The core extraction equipment of the system includes two main units of a water vapor distillation device and a water decoction extraction device. The water vapor distillation device is composed of a distillation column, a heating jacket, a condenser, an oil-water separator, and a volatile oil collector, wherein the distillation column is made of stainless steel 316L. The inner diameter of the distillation column is 300 mm, the height is 1200 mm, and multiple sieve plate structures are arranged in the column to enhance the mass transfer effect. The sieve plate structure can increase the gas-liquid contact area and improve the mass and heat transfer efficiency. The heating jacket surrounds the bottom and side wall of the distillation column, uses electric heating, is equipped with a power-adjustable PID temperature controller, and has a heating power range of 0-15 kW, which can realize accurate temperature control in the range of 80-125℃. The PID controller realizes accurate temperature control through proportional-integral-derivative adjustment. The condenser uses a tube-type heat exchanger structure, with a tube length of 2000 mm and a heat exchange area of 8 square meters. The cooling medium is circulating cooling water, and the inlet temperature is controlled within the range of 15-20℃. The tube-type heat exchanger has the characteristics of high heat transfer efficiency and compact structure. The oil-water separator uses gravity separation principle, with a volume of 50L, and is internally provided with guide plates and overflow weirs to ensure effective separation of volatile oil and water. Gravity separation is a separation method based on density difference, and volatile oil usually has a density less than water, so it can float on the water surface to realize separation.

[0063] The water decoction extraction device is composed of a water decoction tank, a stirring system, a heating system, a circulating pump, and a filtration system. The water decoction tank is a double-jacketed stainless steel reactor with an effective volume of 500L. The double-jacketed structure can achieve uniform heating and improve thermal efficiency. It is equipped with a mechanical stirrer, and the stirring paddle uses a paddle blade with a adjustable speed range of 30-200rpm. Mechanical stirring can promote mass transfer and improve extraction efficiency. The heating system uses a jacketed steam heating method with a steam pressure of 0.3-0.8MPa, and cooperates with a PID controller to realize accurate temperature control in the range of 45-110℃. Steam heating has the advantages of uniform heating and stable temperature. The circulating pump is a stainless steel centrifugal pump with a flow range of 5-50L / min, which is used for circulating stirring and forced convection heat transfer of the extraction liquid. The circulating pump can enhance liquid mixing and improve heat and mass transfer effect. The filtration system includes a coarse filter and a fine filter. The coarse filter uses a 100-mesh stainless steel screen for preliminary filtration, and the fine filter uses a 0.45μm membrane filter for precise filtration. The graded filtration can effectively remove impurities and obtain clear extraction liquid.

[0064] The intelligent control core of the system is composed of two special LSTM-feedforward neural network hybrid controllers, which are responsible for the automatic control of the distillation process and the water decoction process respectively. The distillation process controller is configured with 8 input ports to receive real-time data signals from the distillation column temperature sensor, pressure sensor, heating power meter, cooling water flow meter, steam flow meter, condenser temperature sensor, oil-water separator liquid level meter, and volatile oil mass flow meter. The LSTM network layer inside the controller contains 64 memory units for learning the time series variation law of the distillation process. The LSTM network can effectively process long time series data through the gating mechanism, solving the gradient vanishing problem of traditional neural networks, and output a 64-dimensional feature vector to the three-layer fully connected feedforward neural network. The first hidden layer of the feedforward network contains 128 neurons, which is twice the number of hidden units, the second hidden layer contains 64 neurons, which is equal to the number of hidden units, and the output layer contains 10 neurons, corresponding to temperature regulation, pressure regulation, flow regulation, heating power regulation, and device switch control, etc. 10 discrete control actions.

[0065] The water decoction process controller is also configured with 8 input ports connected to the water decoction tank temperature sensor, pressure sensor, stirring speed meter, pH meter, liquid level meter, jacket temperature sensor, circulating pump flow meter, and filter pressure meter. The controller uses the same LSTM-feedforward network architecture as the distillation controller, with the LSTM layer containing 64 memory units outputting a 64-dimensional time series feature vector, and the feedforward network using a three-layer structure of 128-64-10, outputting 10 control action signals, including temperature regulation, stirring speed regulation, pH value regulation, heating power regulation, and device state switching operation instructions. These control actions can accurately adjust the process parameters of the water decoction process to ensure the optimization of the extraction effect.

[0066] The system is configured with a multi-level sensor detection network to achieve comprehensive monitoring of the extraction process. The basic sensor layer includes Pt100 platinum resistance temperature sensors with a measurement accuracy of ±0.1°C, installed at key positions on the top and bottom of the distillation column and inside the water decoction tank. The Pt100 sensors have good linearity and high stability, allowing them to work stably in harsh industrial environments for a long time. The pressure sensor uses a piezoresistive sensor with an accuracy of ±0.25% FS and a measurement range of 0-1 MPa, installed at the top of the distillation column and the water decoction tank. The piezoresistive sensor converts pressure changes into resistance changes through the piezoresistive effect of semiconductor materials, with fast response speed and strong anti-interference ability. The flow meter uses electromagnetic flow meters and Coriolis mass flow meters. The electromagnetic flow meter measures the flow of conductive fluids based on Faraday's law of electromagnetic induction and is used to measure cooling water and steam flow. The Coriolis mass flow meter is specifically designed for accurate measurement of volatile oil collection, with advantages such as direct measurement of mass flow, no influence of fluid properties, and high precision. The liquid level sensor uses an ultrasonic liquid level meter with a measurement accuracy of ±2mm, installed on the side wall of the oil-water separator and the water decoction tank. The ultrasonic liquid level meter calculates the liquid level height by measuring the propagation time of ultrasonic waves, with non-contact measurement and no influence of the medium.

[0067] The component detection system consists of gas chromatography-mass spectrometry, near-infrared spectrometer, electronic nose, and electronic tongue, etc. The gas chromatography-mass spectrometer is equipped with a DB-5ms capillary column with a length of 30m, an inner diameter of 0.25mm, and a film thickness of 0.25μm. The DB-5ms column is a widely used general chromatographic column in gas chromatography, with 5% phenyl-95% methyl polysiloxane stationary phase having good separation selectivity and thermal stability. The inlet temperature is set to 250°C to ensure complete vaporization of the sample, and the ion source temperature is 230°C to ensure ionization efficiency. It is used for quantitative detection of the concentration of volatile components such as eugenol, borneol acetate, and camphor. The near-infrared spectrometer is set to a scanning wavelength range of 10000-4000cm -1 , a resolution of 4cm -1 , and an integration time of 32 times. Near-infrared spectroscopy is based on the frequency doubling and frequency mixing absorption of molecular vibration, and the quantitative relationship between the spectrum and the component content is established through chemometrics. The online sampling system automatically collects the extraction liquid sample every 5 minutes for spectral scanning, realizing real-time quantitative analysis of components such as flavonoids, polysaccharides, and white peony glycosides.

[0068] The electronic nose system is configured with 8 different selective metal oxide sensor arrays, including TGS2600, TGS2602, TGS2610, TGS2611, TGS2620, MQ-3, MQ-135, MQ-138, etc. The TGS series is a metal oxide semiconductor sensor produced by Japan's Fagron Company, and the MQ series is a gas sensor produced by China's Hanwei Company. These sensors have different selectivity and sensitivity to different types of gases. The sensor operating temperature is set to 200-400°C, which can improve the sensitivity and selectivity of the sensor. The response time is controlled within 60 seconds to ensure real-time detection. The electronic nose system can simulate the human olfactory system's ability to recognize complex odors through the differential response mode of the sensor array to identify and quantify changes in aromatic components.

[0069] The electronic tongue system is configured with 7 different types of taste sensors, including sweet, sour, bitter, salty, umami, astringent, and spicy sensors. Each sensor has high selectivity for specific taste components. The sweet sensor mainly responds to sugars and sweeteners, the sour sensor mainly responds to organic acids, the bitter sensor mainly responds to alkaloids and bitter compounds, the salty sensor mainly responds to inorganic salts, the umami sensor mainly responds to amino acids and nucleotides, the astringent sensor mainly responds to tannins, and the spicy sensor mainly responds to capsaicin compounds. The system uses a quinine standard solution for bitter calibration, with a concentration range of 0.1-100mg / L covering a range of weak to strong bitter intensity. The response time is controlled within 90 seconds, and the bitter intensity is detected automatically every 10 minutes. The detection frequency can timely detect abnormal changes in the concentration of bitter components and trigger corresponding control adjustments.

[0070] The data acquisition and processing system consists of an edge computing controller and a cloud server distributed architecture. The edge computing controller uses an industrial-grade embedded computer with an ARM Cortex-A72 quad-core processor. The ARM architecture has low power consumption and high performance, suitable for long-term stable operation in industrial environments. The main frequency of 1.8GHz can meet the computing needs of real-time control. The memory is 8GB LPDDR4, which is a low-power double-data-rate memory with high bandwidth and low latency. The storage is 128G eMMC, which is an embedded multimedia card with high reliability and shock resistance. The system runs the Linux real-time operating system, which has the advantages of open source, stability, and customization. The real-time kernel can ensure the deterministic response of control tasks. The controller has a lightweight LSTM-Feedforward Neural Network model that can execute real-time control tasks with a control period of 50ms, including sensor data acquisition, control algorithm calculation, and actuator command output. The 50ms control period can meet the strict real-time requirements of industrial control.

[0071] The cloud server adopts a distributed cluster architecture and is configured with multiple high-performance computing nodes, each of which is equipped with an Intel Xeon processor, an Intel Xeon processor specially designed for servers, with the characteristics of multi-core high frequency, 64 GB DDR4 memory, DDR4 memory with high bandwidth and low delay, 2 TB NVMe SSD with ultra-high read and write speed, and NVIDIA Tesla V100 GPU for deep learning model training, Tesla V100 is a professional computing card with powerful parallel computing capability, especially suitable for deep learning training tasks. The server runs the FedAvg federated learning algorithm framework, supports collaborative learning of multiple edge nodes, adopts differential privacy mechanism to protect data privacy, and sets the noise parameter epsilon to 1.0 to balance the demand for privacy protection and data availability. The cloud regularly collects the training data of each edge node, performs batch model training and optimization, and then pushes the updated control model to the edge controller through wireless network or gigabit Ethernet, which has the characteristics of high bandwidth and low delay, and can quickly transmit model parameters. The actuators of the system include electric regulating valves, frequency converters, solenoid valves and servo motors, etc. The electric regulating valve selects an intelligent electric actuator with an adjustment accuracy of ±0.5% and a response time of less than 30 seconds, which is used to control the steam flow and cooling water flow. The electric regulating valve drives the valve opening degree through the electric motor to realize accurate control of the flow. The frequency converter selects a vector control type frequency converter, which is an advanced motor control technology that can realize decoupling control of motor torque and magnetic flux, with an output frequency range of 0-50 Hz, which is used to control the stirring motor speed and circulating pump speed. Frequency control can realize stepless speed regulation and energy-saving operation of the motor. The solenoid valve is made of stainless steel with a working pressure of 1.6 MPa, which is used to control the on-off of each pipeline. The solenoid valve has the characteristics of fast response speed and high control accuracy. The servo motor is configured with an encoder feedback, which can feedback the position information of the motor in real time, with a positioning accuracy of ±0.1°, which is used to accurately control the opening of the sampling valve and the shunt valve. The servo control system has the characteristics of high precision and high response speed.

[0072] The whole system realizes data communication and coordinated control among subsystems through field bus network. The sensor and actuator devices are connected by Profibus-DP field bus protocol. Profibus-DP is a distributed peripheral protocol of process field bus, which is specially used for communication at sensor and actuator level in automation system. The communication rate is 12 Mbps, which can meet the real-time communication needs of a large number of devices. The bus length is up to 1200 m, which can cover the entire production site. It supports up to 126 device nodes. The upper computer monitoring system uses SCADA software platform. SCADA is a supervisory control and data acquisition system, which provides human-machine interface and historical data storage function. The operator can monitor the entire production process through the graphical interface, supports remote monitoring and fault diagnosis, and the remote monitoring function enables experts to provide technical support for the system in different places. The system is also equipped with UPS uninterruptible power supply and redundant communication link. UPS can provide temporary power protection when the power fails. The redundant communication link can automatically switch to the backup link when the main communication link fails, ensuring reliable operation in abnormal conditions.

[0073] The system realizes efficient extraction and quality control of active ingredients of Chinese herbal medicine in cola production process through the synergistic optimization of multi-stage extraction process, real-time adjustment of intelligent control algorithm, fusion processing of multi-sensor information and continuous learning of cloud-edge collaboration. The whole system integrates modern control theory, artificial intelligence technology, precision detection technology and industrial automation technology, forming a complete intelligent extraction control solution, which can significantly improve extraction efficiency, improve product quality, reduce production cost and reduce manual dependence, providing important technical support for modern production of Chinese herbal medicine.

[0074] In Example 1, the production process of a health cola product containing Chinese herbal ingredients of an enterprise is taken as an example, and the intelligent extraction control method for Chinese herbal medicine active ingredients in cola production process is used to control the sequential extraction of four kinds of Chinese herbal medicine, i.e., clove, amomum villosum, radix angelicae dahuricae and alpinia officinarum.

[0075] High-quality Chinese herbal medicine raw materials, clove, amomum villosum, radix angelicae dahuricae and alpinia officinarum, were selected. In the raw material pretreatment stage, the Chinese herbal medicine was cleaned to remove impurities, then dried in a constant temperature drying oven at 60°C for 24 hours to a water content of less than 8%, and finally crushed to 40-60 mesh particle size with a universal crusher. 50 kg of clove, 30 kg of amomum villosum were taken as volatile components, 40 kg of radix angelicae dahuricae and 35 kg of alpinia officinarum were taken as non-volatile components, and the total amount of raw materials was 155 kg.

[0076] The system first starts the phased sequence extraction model, and establishes the four-stage precise control model of water vapor distillation according to the technical solution of claim 2. The preheating stage is set to 85℃ for 15 minutes, and the intelligent controller monitors the distillation tower temperature in real time through the Pt100 platinum resistance temperature sensor, and the temperature control accuracy reaches ±0.3℃. The main distillation stage adopts a dynamic temperature adjustment strategy, with an initial temperature setting of 95℃, and the system detects the volatile oil component concentration every 5 minutes through a gas chromatography-mass spectrometry system, and when the eugenol concentration reaches 85% of the target value, the temperature is automatically adjusted to 108℃ to enter the rectification stage, and the whole main distillation stage lasts for 45 minutes. The rectification stage controls the temperature to 110℃ for 15 minutes to extract β-caryophyllene and other sesquiterpenes. The end stage is cooled to 95℃ for 10 minutes to ensure complete collection of light volatile components.

[0077] During the distillation process, the LSTM-feedforward neural network hybrid controller continuously learns the time sequence features, and the input layer receives 8 parameters such as distillation tower temperature 119.8℃, pressure 0.15kPa, heating power 8.5kW, cooling water flow 25L / min, steam flow 2.3kg / h, condenser temperature 18℃, oil-water separator liquid level 65%, and volatile oil collection amount cumulative 1.85kg. After 64 memory units of single-layer LSTM network extract time sequence features, three-layer feedforward network outputs control action, and the system automatically executes temperature increase 0.5℃ control instruction, and the whole decision process takes 42 milliseconds.

[0078] When the volatile oil yield reaches 91% of the preset target, i.e. 1.89kg, the system automatically starts the water decoction process. A multi-section variable temperature precise control model is established for water decoction extraction, and the soaking stage is set to 50℃ for 30 minutes to fully wet the distillation residue. The low-temperature extraction stage controls the temperature to 68℃ for 60 minutes to preferentially extract heat-sensitive flavonoids, and the near-infrared spectroscopy detection system automatically collects extraction liquid samples every 5 minutes, and when the total flavonoid extraction rate reaches 42%, it enters the medium-temperature stage. The medium-temperature extraction stage adopts a dynamic temperature adjustment strategy to control the temperature to 88℃, and the electronic tongue system monitors the bitterness intensity in real time, and when the bitterness component concentration reaches 12mg / L, the system automatically reduces the temperature to 82℃ according to the multi-index joint monitoring temperature adjustment equation. The high-temperature strengthening stage controls the temperature to 105℃ for 20 minutes to strengthen the extraction of difficult-to-dissolve components.

[0079] The whole extraction process, the on-line detection system continues to work, the basic sensor to collect temperature and pressure parameters at 1 second cycle, volatile oil yield monitoring system through the coriolis mass flowmeter real-time tracking collection, the final volatile oil yield reaches 1.92%. Gas chromatography-mass spectrometry system detects eugenol concentration 284 mg / L, camphor glycol acetate concentration 156 mg / L, camphor concentration 98 mg / L, all reach the expected goal. Near infrared spectroscopy detection shows that the total flavone extraction rate is 78.5%, the total polysaccharide extraction rate is 82.3%, and the angelica lactone extraction rate is 75.8%. The electronic tongue detects the final bitter component concentration 13.2 mg / L within the threshold range.

[0080] The edge computing and cloud coordination mechanism collects 1200 training samples during the experiment, and the cloud server runs the FedAvg algorithm for model training. After 150 iterations, the model parameters converge, and the control accuracy is further improved to temperature ±0.2℃, time ±25 seconds, and pressure ±0.08kPa. The total extraction time is 4 hours and 18 minutes, which is completed within the preset time window, realizing efficient and high-quality extraction of active ingredients of Chinese herbal medicine.

[0081] Comparative Example 1, this comparative example uses the traditional PID control method to extract the same Chinese herbal medicine. The system is configured with the same hardware devices and raw materials, but the controller uses the conventional PID control algorithm instead of the LSTM-Feedforward Neural Network hybrid controller. The PID parameters are determined by the Ziegler-Nichols tuning method, the proportional gain , the integral gain , and the differential gain .

[0082] The distillation stage adopts a fixed temperature control strategy, the preheating stage is 85℃ for 15 minutes, the main distillation stage is fixed at 105℃ for 60 minutes, the rectification stage is 110℃ for 15 minutes, and the end stage is 95℃ for 10 minutes. Due to the lack of intelligent prediction and dynamic adjustment capability, the temperature control accuracy is only ±1.2℃, the pressure control accuracy is ±0.3kPa, and the time control accuracy is ±2 minutes. The water decoction stage also adopts a fixed temperature program, the soaking stage is 50℃ for 30 minutes, the low temperature extraction stage is 70℃ for 60 minutes, the medium temperature extraction stage is 90℃ for 90 minutes, and the high temperature strengthening stage is 105℃ for 20 minutes. Lack of real-time component detection feedback, unable to dynamically adjust parameters according to extraction progress.

[0083] Comparative Example 2, this comparative example uses the traditional manual experience control method for extraction experiment. The operator manually adjusts each parameter according to the traditional Chinese medicine extraction process experience, the temperature adjustment is based on the observation of the steam state and the color change of the liquid, and the time control is mainly judged by experience. The operator manually sets the preheating temperature to 80°C in the distillation stage, and adjusts it to 100°C for main distillation after observing sufficient steam. Whether to enter the rectification stage is judged according to the change of the volatile oil collection speed. The whole distillation process lacks precise temperature and time control. The water decoction stage also relies on experience to judge, and the extraction endpoint is determined by observing the color depth and tasting the bitterness of the extraction liquid. The temperature fluctuation range during the experiment is ±3.5°C, the time control error is ±8 minutes, and the pressure control error is ±0.8kPa. Due to the lack of precise control and real-time monitoring, the volatile oil yield is only 1.58%, and the concentration of each volatile component is generally low. The total flavone extraction rate is 65.3%, the total polysaccharide extraction rate is 69.2%, the angelica lactone extraction rate is 61.7%, and the comprehensive extraction rate is 49.2%. The bitterness component concentration reaches 22.4mg / L, which seriously exceeds the standard and affects the taste of the product. The energy consumption is as high as 89.3kWh, and the total extraction time is 6 hours and 48 minutes.

[0084] Comparative Example 3, this comparative example uses a single neural network control method for extraction experiment. The control system only uses a three-layer feedforward neural network with a network structure of 8-64-32-10, and does not contain an LSTM time series feature extraction layer. The training data uses the same historical extraction data, but the network cannot learn the time series change rule. Lack of time series memory ability, the controller has poor response to the dynamic changes of the extraction process, the temperature control precision is ±0.8°C, the pressure control precision is ±0.2kPa, and the time control precision is ±45 seconds. In the case of frequent fluctuations in process parameters, the control effect is obviously worse than that of the LSTM-feedforward hybrid controller.

[0085] Comparative Example 4, this comparative example uses a fixed stage extraction method for experiment, and does not use a staged sequential extraction model. The system simultaneously extracts four kinds of Chinese herbal medicines, first performs unified steam distillation and then performs unified water decoction extraction, without considering the differences in the most suitable extraction conditions of different medicinal materials. The distillation stage uses a compromise temperature program, preheating at 85°C, main distillation at 100°C for 50 minutes, and rectification at 108°C for 15 minutes. The water decoction stage is soaked at 50°C, low temperature at 65°C for 45 minutes, medium temperature at 85°C for 75 minutes, and high temperature at 105°C for 25 minutes.

[0086] Comparative Example 5: This comparative example employed a control method without an online detection system. The system was only equipped with basic temperature and pressure sensors and did not include component detection equipment such as gas chromatography-mass spectrometry, near-infrared spectroscopy, electronic nose, or electronic tongue, thus failing to monitor the extraction progress of each component in real time. The controller executed the extraction process using a preset fixed program and could not dynamically adjust based on real-time detection data. The distillation and decoction processes ran according to preset time programs, lacking intelligent endpoint determination based on component concentration.

[0087] To verify the effectiveness of the method of the present invention, statistical analysis was performed on the key indicators of Example 1 and each comparative example. For example... Figure 1 As shown, the experimental process for comparing the extraction rates of each component was as follows: First, a Coriolis mass flow meter was used to monitor the amount of volatile oil collected in real time. A precision flow meter was installed at the outlet of the oil-water separator, and the volatile oil yield was calculated based on the preset raw material weight. In Example 1, an intelligent adjustment using an LSTM-feedforward neural network hybrid controller achieved a volatile oil yield of 1.92%. Comparative Example 1 used a traditional PID control method, which, due to a lack of dynamic adjustment capability, only achieved a volatile oil yield of 1.76%. In Comparative Example 2, the manual experience control method resulted in a volatile oil yield dropping to 1.58% due to excessive temperature fluctuations. The total flavonoid extraction rate was determined using a near-infrared spectroscopy detection system, with a scanning wavelength range of 10000-4000 cm⁻¹. -1 Real-time detection was performed using a quantitative calibration model established by partial least squares method. The intelligent temperature control strategy in Example 1 achieved a total flavonoid extraction rate of 78.5%, while the comparative examples, lacking precise temperature control and real-time monitoring feedback, all had lower total flavonoid extraction rates than Example 1. The total polysaccharide extraction rate was calculated using the phenol-sulfuric acid method. Example 1 achieved an 82.3% total polysaccharide extraction rate through precise multi-segment temperature control using a staged sequential extraction model. Comparative example 3, although using neural network control, lacked temporal memory capabilities, resulting in a total polysaccharide extraction rate of only 77.5%. The angelica glycoside extraction rate was quantitatively analyzed using high-performance liquid chromatography. Example 1, through a segmented extraction control strategy targeting the active components of Angelica dahurica, preferentially extracted angelica glycosides at a low temperature of 60-75℃, achieving a 75.8% angelica glycoside extraction rate. Comparative example 4, using a fixed-stage extraction method, could not optimize control based on the characteristics of different medicinal materials, resulting in an angelica glycoside extraction rate of only 64.3%.

[0088] like Figure 2 As shown, the experimental procedure for comparing the comprehensive extraction rates involved comprehensively calculating and statistically analyzing the extraction rates of the four main components. The experiment used the following formula: ; the extraction rate of each component is obtained by the corresponding detection method. In Example 1, the extraction control method is optimized by deep learning, and a four-stage precise control model is used in the distillation stage. The temperature and time parameters of each stage, including preheating, main distillation, rectification, and ending, are optimized by a feedforward neural network after learning the time sequence characteristics by an LSTM network. The final comprehensive extraction rate reaches 59.63%. In Comparative Example 1, the traditional PID control method is used. Since the PID parameters are fixed and lack adaptive adjustment ability, the extraction rate of each component is lower than that of Example 1, and the comprehensive extraction rate is 54.4%. In Comparative Example 2, the artificial experience control method is used. Since the operator manually adjusts the parameters based on traditional process experience, and lacks precise control and scientific basis, the comprehensive extraction rate is only 49.2%. In Comparative Example 3, a single feedforward neural network control is used. Although it has certain intelligent characteristics, it lacks the time sequence memory ability of LSTM and cannot effectively handle the dynamic changes in the extraction process. The comprehensive extraction rate is 55.8%. In Comparative Example 4, a fixed stage extraction method is used. Four kinds of Chinese herbal medicines are uniformly processed without considering the characteristic differences of each medicinal material. The comprehensive extraction rate is 52.7%. In Comparative Example 5, the online detection system is lacking, and it cannot real-time monitor the extraction progress of each component and dynamically adjust the parameters. The comprehensive extraction rate is 53.1%.

[0089] As shown in Figure 3 , the experimental process of control accuracy and product quality comparison is divided into two parts: temperature control accuracy test and product quality evaluation. The temperature control accuracy is measured by Pt100 platinum resistance temperature sensor with an accuracy of ±0.1℃. The temperature fluctuation range in the extraction process is continuously monitored to evaluate the accuracy of the control system. The LSTM-feedforward neural network hybrid controller of Example 1 achieves a temperature control accuracy of ±0.3℃ by learning the historical temperature variation law and adjusting the control parameters in real time, which is much better than other comparative examples. The traditional PID controller of Comparative Example 1 has a temperature control accuracy of ±1.2℃ due to fixed parameters and response lag. The temperature control accuracy of the artificial experience control method of Comparative Example 2 is the worst, reaching ±3.5℃ due to subjective judgment of the operator and manual adjustment delay. The product quality evaluation uses a comprehensive scoring method, including aroma score, taste score, and stability score. A professional sensory evaluation team conducts blind evaluation, and each dimension uses a 1-10 point evaluation standard. Example 1 maintains the stability of active ingredients and the consistency of products through precise temperature control and optimized extraction process parameters, with a comprehensive quality index of 8.5 points. Comparative Example 1 has a product quality index of 7.2 points due to larger temperature fluctuations that may lead to degradation of some heat-sensitive components. Comparative Example 2 has the lowest product quality of 6.1 points due to the worst control accuracy. Comparative Example 3 has a product quality index of 7.8 points despite using neural network control due to lack of time sequence optimization ability. Comparative Examples 4 and 5 have product quality indexes of 6.9 points and 6.8 points respectively due to insufficient process optimization.

[0090] As Figure 4 shown, the experimental process of bitterness concentration and extraction time comparison is to monitor the bitterness intensity by electronic tongue detection system and record the total extraction time of each experiment. The determination of bitterness component concentration adopts the electronic tongue system with 7 different types of taste sensors, including a special bitterness sensor for detecting alkaloids and bitter compounds, with a response time of 90 seconds, and a quinine standard solution is used for bitterness calibration to establish a calibration curve. Example 1 uses a multi-index joint monitoring temperature adjustment equation to automatically reduce the extraction temperature when the bitterness component concentration approaches the threshold value, and finally controls the bitterness component concentration at 13.2 mg / L, successfully below the bitterness threshold of 15 mg / L. Due to the lack of real-time bitterness monitoring and dynamic adjustment capability, the bitterness component concentration of Comparative Example 1 reaches 18.7 mg / L, exceeding the threshold range. The artificial control method of Comparative Example 2 lacks objective bitterness quantification means, and the bitterness component concentration is as high as 22.4 mg / L, seriously affecting the taste of the product. The recording of the total extraction time starts from the completion of the raw material pretreatment and ends until the extraction process is completely finished and the final product is collected. Example 1 uses LSTM-based sequence prediction end-point control, which starts the termination program when the predicted increment of each main component extraction rate is less than 2% and the comprehensive reward function is greater than 0.85, and the total extraction time is controlled at 4.3 hours. Comparative Example 1 uses a fixed time program, and the lack of intelligent end-point judgment leads to an extension of the total extraction time to 5.6 hours. The artificial experience control of Comparative Example 2 relies on the subjective judgment of the operator, and the total extraction time is the longest, reaching 6.8 hours. Although Comparative Example 3 has certain intelligent features, it lacks timing prediction capability, and the total extraction time is 4.9 hours. Comparative Examples 4 and 5 are 5.1 hours and 5.4 hours, respectively, both of which exceed the target time upper limit of 5 hours.

[0091] As Figure 5 shown, the experimental process of comprehensive reward function value comparison is to establish a multi-objective comprehensive reward function and calculate the reward value of each experiment step by step. The experiment first establishes a quantitative calculation method for each index. The total extraction rate is calculated by the arithmetic mean of the extraction rates of the four main components, the product quality index is calculated by the weighted score of aroma, taste, and stability, and the bitterness component concentration is obtained by electronic tongue detection.

[0092] Example 1 has a comprehensive reward function value of 24.7, which exceeds the excellent control strategy threshold of 21.25. The comprehensive reward function value of Comparative Example 1 is 18.3, which is in the good control strategy interval but does not reach the excellent level. The comprehensive reward function value of Comparative Example 2 is only 12.8, which is close to the lower limit of the good control strategy. The comprehensive reward function value of Comparative Example 3 is 20.5, which is relatively good in the comparison but still lower than the excellent control strategy threshold. The comprehensive reward function values of Comparative Example 4 and Comparative Example 5 are 16.2 and 15.6, respectively, which are in the middle and lower levels of the good control strategy interval. Reward normalization processing was also performed during the experiment, using the formula to eliminate the magnitude difference between different batches of data and ensure the objectivity and comparability of the evaluation results.

[0093] As shown in Figure 6 , the experimental process of time series performance comparison is to simulate and record the dynamic performance changes of each experimental method during the entire extraction process. The experiment sets a 7-hour observation time window, covering the complete extraction process of all experimental methods, with a time sampling interval of 0.1 hour to ensure sufficient data density. The time series performance curve of Example 1 shows rapid convergence and high performance maintenance characteristics through the time series learning ability of the LSTM network and the real-time optimization decision of the feedforward network. The traditional PID control method of Comparative Example 1 has a slower performance convergence speed due to response lag and fixed parameters, reflecting a lower final performance level and longer adjustment time. The artificial experience control method of Comparative Example 2 shows obvious performance fluctuations and convergence difficulties, reflecting the instability and subjectivity of manual operation. The experiment marks key markers at the completion time points of each method. Example 1 maintains the highest relative performance value at the completion point of 4.3 hours, while Comparative Examples 1 and 2 complete the extraction process at 5.6 hours and 6.8 hours, respectively, with lower performance levels.

[0094] As shown in Figure 7 , the experimental process of component extraction rate improvement degree comparison is to calculate the improvement degree of Example 1 relative to the average level of all comparative examples and perform statistical analysis. The experiment first calculates the arithmetic mean of each comparative example in terms of the extraction rates of the four main components, namely the comparative average value of volatile oil yield, the comparative average value of total flavonoid extraction rate, the comparative average value of total polysaccharide extraction rate, and the comparative average value of angelica lactone extraction rate. The relative improvement rate formula is used to calculate the improvement degree.

[0095] From Figure 1The comparison of extraction rates of various components shows that Example 1 exhibits significant advantages in the extraction of all four main components. The volatile oil yield of Example 1 reached 1.92%, which is 0.16, 0.34, 0.24, 0.20, and 0.25 percentage points higher than that of Comparative Example 1 (1.76%), Comparative Example 2 (1.58%), Comparative Example 3 (1.68%), Comparative Example 4 (1.72%), and Comparative Example 5 (1.67%), respectively. The total flavonoid extraction rate of Example 1 was 78.5%, significantly higher than that of Comparative Example 1 (71.2%), Comparative Example 2 (65.3%), Comparative Example 3 (73.8%), Comparative Example 4 (69.7%), and Comparative Example 5 (71.0%). The total polysaccharide extraction rate of Example 1 reached 82.3%, exceeding the extraction levels of all comparative examples. The angelicin extraction rate of Example 1 was 75.8%, also the best among all experimental groups. This demonstrates that the LSTM-feedforward neural network hybrid controller of the present invention achieves efficient extraction of various active ingredients through intelligent parameter adjustment and temporal feature learning.

[0096] from Figure 2 The comparison of the overall extraction rates shows that the overall extraction rate of Example 1, at 59.63%, is significantly higher than that of all comparative examples. Specifically, it is 5.23 percentage points higher than that of Comparative Example 1, 10.43 percentage points higher than that of Comparative Example 2, 3.83 percentage points higher than that of Comparative Example 3, 6.93 percentage points higher than that of Comparative Example 4, and 6.53 percentage points higher than that of Comparative Example 5. This demonstrates the significant advantage of the method of the present invention in terms of overall extraction efficiency, indicating that the phased sequential extraction model and the multi-objective comprehensive optimization strategy effectively coordinate the extraction needs of different components and maximize overall performance.

[0097] from Figure 3 The temperature control accuracy comparison shows that Example 1 achieves a temperature control accuracy of ±0.3℃, which is significantly better than Comparative Example 1's ±1.2℃, Comparative Example 2's ±3.5℃, Comparative Example 3's ±0.8℃, Comparative Example 4's ±1.0℃, and Comparative Example 5's ±1.5℃. This demonstrates that the temporal memory capability of the LSTM network combined with the decision optimization capability of the feedforward neural network achieves more precise temperature control. Figure 3 The comparison of product quality indicators shows that the product quality indicator of Example 1 reached 8.5 points, which is significantly higher than that of Comparative Example 1 (7.2 points), Comparative Example 2 (6.1 points), Comparative Example 3 (7.8 points), Comparative Example 4 (6.9 points), and Comparative Example 5 (6.8 points), indicating the advantages of the method of the present invention in maintaining the stability of active ingredients and product consistency.

[0098] from Figure 4The comparison of bitter component concentrations shows that the bitter component concentration in Example 1 was 13.2 mg / L, successfully controlled within the threshold range of 15 mg / L. In contrast, the concentrations in Comparative Examples 1 (18.7 mg / L), 2 (22.4 mg / L), 4 (19.8 mg / L), and 5 (17.9 mg / L) all exceeded the bitterness threshold. Only Comparative Example 3 (16.3 mg / L) slightly exceeded the limit but was still higher than that in Example 1. Figure 4 The comparison of total extraction times shows that Example 1 achieved a total extraction time of 4.3 hours, which is within the target time limit of 5 hours. Compared with Comparative Example 1 (5.6 hours), Comparative Example 2 (6.8 hours), Comparative Example 3 (4.9 hours), Comparative Example 4 (5.1 hours), and Comparative Example 5 (5.4 hours), this example achieves a shorter extraction time. This indicates that the intelligent control method of the present invention effectively controls bitter components that affect the taste of the product while ensuring extraction efficiency, thus optimizing time efficiency.

[0099] from Figure 5 The comparison of the comprehensive reward function values ​​reveals that the comprehensive reward function value of Example 1 is 24.7, exceeding the excellent control strategy threshold of 21.25, indicating that the system has reached an excellent control level. Comparative Example 1's 18.3 and Comparative Example 3's 20.5 fall within the good control strategy range of 12.5-21.25, while Comparative Example 2's 12.8, Comparative Example 4's 16.2, and Comparative Example 5's 15.6 are all below the excellent control strategy threshold. This demonstrates the superior performance of the method of the present invention in balancing multiple objectives such as extraction rate, product quality, and bitterness control, reflecting the effectiveness of the comprehensive reward function design.

[0100] from Figure 6 The time-series performance comparison shows that Example 1 maintained a higher performance level and faster convergence speed throughout the extraction process. Example 1's performance curve showed a higher relative performance value when it reached the completion point at 4.3 hours, compared to Comparative Example 1 at 5.6 hours and Comparative Example 2 at 6.8 hours. Example 1 not only completed the extraction process earlier but also maintained superior performance metrics. This advantage in time-series performance demonstrates the superiority of LSTM networks in processing time-series data and dynamic optimization.

[0101] from Figure 7The comparison of the improvement in component extraction rates shows that Example 1 achieved significant improvements in the extraction of all four main components compared to the average levels of the comparative examples. The relative improvement in volatile oil yield reached 14.2%, total flavonoid extraction rate improved by 8.6%, total polysaccharide extraction rate improved by 10.8%, and angelica glycoside extraction rate improved by 12.3%. This indicates that the comprehensive application of technologies such as deep learning optimization, staged sequential extraction, multi-objective comprehensive optimization, online detection feedback, and edge-cloud collaboration has achieved significant performance improvements based on traditional extraction processes, providing an advanced technical solution for the industrial-scale and efficient extraction of active ingredients from traditional Chinese medicine.

[0102] For those skilled in the art, the present invention is not limited to the details of the exemplary embodiments described above, and may be implemented in other specific forms without departing from the spirit or scope of the invention. Therefore, the embodiments of the present invention are exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than by the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims should be covered within the claims. Any reference numerals in the claims should not be construed as limiting the scope of the claims. It is evident that those skilled in the art can make various modifications and variations to the present invention without departing from its spirit and scope. Therefore, if such modifications and variations fall within the scope of the claims and their equivalents, the present invention also intends to include such modifications and variations.

Claims

1. A deep learning-based optimization and control method for the extraction of active ingredients from traditional Chinese medicines in cola production processes, characterized in that... Includes the following steps: S1: The raw materials of Chinese herbal medicines are cleaned, dried, and pulverized. A staged sequential extraction model is established, and the process parameters for the distillation extraction of volatile components are set. After the extraction of volatile components, the process parameters for the decoction extraction of non-volatile components from the distillation residue are set, and the volatile and non-volatile components are extracted in a staged sequential manner. Among them, the volatile components of the Chinese herbal medicines include at least cloves and cardamom; the non-volatile components of the Chinese herbal medicines include at least angelica and galangal. S2: Construct an intelligent controller based on a hybrid single-layer LSTM-feedforward neural network, including a distillation process controller and a decoction process controller. Extract the temporal features of each extraction process through a single-layer LSTM network, and use independent feedforward neural networks to optimize and control each process parameter. S3: Design a multi-objective comprehensive reward function, taking the total extraction rate and product quality indicators as positive incentive factors, and energy consumption level and bitter component concentration as negative penalty factors. Construct the comprehensive reward function by calculating the individual reward values ​​step by step, setting weight coefficients, and performing weighted summation, and then use gradient backpropagation to train and optimize the deep learning model. S4: Configure a layered online detection system, which adopts a combination of real-time monitoring by basic sensors and timed component detection to establish an online calculation and dynamic monitoring mechanism for volatile oil yield, total flavonoid extraction rate, total polysaccharide extraction rate and angelica glycoside extraction rate; S5: Perform phased sequential optimization control. Based on the fused sensor data status, adjust the heating temperature and distillation time of the distillation process to complete the extraction of volatile components, adjust the extraction temperature and extraction time of the decoction process, and monitor the extraction progress and bitterness intensity of each component in real time to achieve optimized extraction of active ingredients. S6: Establish a control strategy optimization mechanism that integrates edge computing and cloud collaboration. The cloud platform collects historical extracted data for offline training of deep learning models and regularly pushes updated control strategies to the edge controller to achieve continuous optimization of the control algorithm.

2. The method according to claim 1, characterized in that, The phased sequential extraction model establishment in step S1 includes the following specific steps: S11: Establish a precise four-stage control model for steam distillation, including: Preheating stage: Set the temperature to 80-90℃ for 15 minutes to achieve initial softening of the cell walls of the medicinal materials; Main distillation stage: Set the temperature range to 95-125℃, adopt a dynamic temperature control strategy, and last for 30-60 minutes to achieve full distillation of volatile components; Distillation stage: Set temperature 105-115℃, duration 15 minutes, to achieve recovery of high-boiling-point aromatic compounds; Final stage: Set the temperature to 90-100℃ for 10 minutes to ensure complete collection of light volatile components; S12: Establish a multi-stage temperature-controlled precision model for water decoction extraction, including: Infiltration stage: Set the temperature to 45-55℃ and the duration to 30 minutes to fully wet the medicinal materials and activate the cell walls; Low-temperature extraction stage: Set the temperature range to 60-75℃ for 60 minutes, prioritizing the extraction of heat-sensitive flavonoids; Medium-temperature extraction stage: The temperature range is set to 80-95℃, and dynamic temperature control is used to achieve the dissolution of polysaccharide components; High-temperature enhancement stage: Set the temperature to 100-110℃ and the duration to 20 minutes to enhance the extraction of poorly soluble components; S13: Establish a phased time-series coordination mechanism. First, start the distillation process and monitor the extraction progress in real time. When the volatile oil yield reaches 90% of the preset target, start the decoction process and use the distillation residue for decoction extraction. Based on the real-time detection data of volatile oil yield and flavonoid extraction rate, use a feedback control algorithm to dynamically adjust the temperature and time parameters of the decoction process to ensure that the total extraction time is controlled within 4-5 hours, thereby achieving sequential connection and component complementarity optimization between the two processes. S14: Perform targeted extraction control of volatile components, including: Extraction of volatile components from clove: The distillation temperature is controlled within the range of 95-110℃, which is consistent with the range of 95-125℃ in the main distillation stage. Eugenol and acetyleugenol are extracted first. When the gas chromatography detects that the concentration of eugenol reaches 80% of the target value, the temperature is adjusted to 105-115℃ to enter the rectification stage to extract β-caryophyllene sesquiterpenoid components. Extraction of volatile oil components from Amomum villosum: The distillation temperature is controlled within the range of 100-120℃ and coordinated with the temperature of the main distillation stage to extract borneol acetate, borneol and camphor. When the intensity of the characteristic aroma of Amomum villosum is detected to reach the set threshold online, the distillation time is extended to ensure that the terpenoid components are fully distilled out. S15: Implement segmented extraction control of non-volatile components, including: Extraction of Angelica dahurica active ingredients: After fully wetting at 45-55℃ for 30 minutes in the soaking stage, Angelica dahurica glycosides are extracted preferentially at a low temperature stage of 60-75℃. When the extraction rate reaches 60%, the medium temperature stage of 80-95℃ is entered to extract Angelica dahurica polysaccharides and water-soluble flavonoids. Finally, the high temperature stage of 100-110℃ is used to enhance the extraction of insoluble components. Extraction of active ingredients from galangal: Galangin is fully extracted at a medium temperature of 80-90℃. The bitterness intensity is monitored in real time by an electronic tongue. When the concentration of bitter components exceeds 15mg / L, the temperature is reduced to 75℃. S16: Establish the component balance control equation for the aqueous decoction of the two medicinal materials: ； in, To comprehensively evaluate the function value; The weighting coefficient for angelica glycosides; The angelicin extraction rate was obtained by near-infrared spectroscopy detection and calculation using a standard sample calibration curve. The weighting coefficient for galangin; The extraction rate of galangin was obtained by high-performance liquid chromatography (HPLC) detection and calculation. The polysaccharide weighting coefficient; The polysaccharide extraction rate was calculated using the phenol-sulfuric acid method. The bitterness penalty weighting coefficient; The concentration of bitter components was obtained by electronic tongue detection and calibration using quinine standard solution.

3. The method according to claim 1, characterized in that, The construction of the single-layer LSTM-feedforward neural network hybrid controller in step S2 includes the following specific steps: S21: Constructing a dedicated LSTM-feedforward neural network controller for the distillation process, including: Input layer configuration: Set up 8 dedicated distillation parameter input ports, including distillation column temperature, distillation column pressure, heating power, cooling water flow rate, steam flow rate, condenser temperature, oil-water separator level, and volatile oil collection rate; Temporal feature extraction layer: Configure a single-layer LSTM hidden layer containing 64 memory units to learn temporal patterns and output a 64-dimensional distillation temporal feature vector; Decision network layer: A three-layer fully connected feedforward neural network is constructed, with 128 neurons in the first layer, 64 neurons in the second layer, and 10 neurons in the third layer, corresponding to 10 discrete distillation control actions and their number mappings: Action 1 is to increase the temperature by 1.0℃, Action 2 is to decrease the temperature by 1.0℃, Action 3 is to increase the pressure by 0.1kPa, Action 4 is to decrease the pressure by 0.1kPa, Action 5 is to increase the flow rate by 10L / min, Action 6 is to decrease the flow rate by 10L / min, Action 7 is to increase the heating power by 1kW, Action 8 is to decrease the heating power by 1kW, Action 9 is to switch the condenser on and off state, and Action 10 is to switch the separator on and off state and drain state. S22: Constructing a dedicated LSTM-feedforward neural network controller for the decoction process, including: Input layer configuration: Set up 8 dedicated parameter input ports for decoction, including decoction tank temperature, decoction tank pressure, stirring speed, extract pH value, liquid level, heating jacket temperature, circulating pump flow rate, and filtration pressure; Temporal feature extraction layer: Configure a single-layer LSTM hidden layer containing 64 memory units to learn temporal patterns and output a 64-dimensional temporal feature vector; Decision network layer: A three-layer fully connected feedforward neural network is constructed, with 128 neurons in the first layer, 64 neurons in the second layer, and 10 neurons in the third layer, corresponding to 10 discrete water decoction control actions and their number mappings: Action 1 is to increase the temperature by 1.0℃, Action 2 is to decrease the temperature by 1.0℃, Action 3 is to increase the stirring speed by 10 rpm, Action 4 is to decrease the stirring speed by 10 rpm, Action 5 is to increase the pH value by 0.1, Action 6 is to decrease the pH value by 0.1, Action 7 is to increase the heating power by 1kW, Action 8 is to decrease the heating power by 1kW, Action 9 is to switch the circulation pump on and off, and Action 10 is to switch the vacuum concentration on and off. S23: Establish a dual-controller collaborative learning mechanism, set up independent experience replay buffers to accommodate 5000 training samples each, adopt a priority experience replay strategy combined with random sampling to select 32 samples from each buffer to form a training batch, adopt an ε-greedy exploration strategy, set the learning rate to 0.001, and achieve the adjustment capability of temperature control accuracy ±0.5℃, time control accuracy ±30 seconds, and pressure control accuracy ±0.1kPa through phased training.

4. The method according to claim 1, characterized in that, The multi-objective comprehensive reward function design, step-by-step calculation, and weighted summation training optimization process in step S3 includes the following specific operational steps: S31: Establish quantitative calculation methods for each indicator, including: Total extraction rate calculation: ; in, Total extraction rate; For volatile oil yield; Total flavonoid extraction rate; Total polysaccharide extraction rate; The extraction rate of angelica glycosides; Product quality index calculation: ; in, For product quality indicators; Aroma weighting; Rate the aroma; Weighted by taste; Rate the taste; For stability weights; Assess stability score; Energy consumption level calculation: ; in, Total energy consumption; This refers to the heating power. Cooling power; Total extraction time; For auxiliary equipment power consumption; Concentration of bitter components: The concentration of bitter components was obtained through electronic tongue detection. S32: Set the weighting strategy for each indicator, determined based on the efficacy requirements of the cola product: Total Extraction Rate Weighting Reflecting extraction efficiency, product quality indicators have weight. Energy consumption level is a key indicator of product quality. To reflect economic efficiency, the concentration weight of bitter components is used. It reflects the control of taste; S33: The specific process of performing step-by-step calculations and weighted summation includes: S331: Calculate the positive incentive reward value: ； in, This is a positive incentive reward value; This is the weighting coefficient for the total extraction rate; Total extraction rate; This is the product quality weighting coefficient; For product quality indicators; S332: Calculate the negative penalty reward value: ； in, This is a negative penalty reward value; Energy consumption weighting coefficient; This represents actual energy consumption. The dynamic benchmark energy consumption is calculated dynamically based on the amount of raw materials input and equipment power. This is the bitterness weighting coefficient; This refers to the concentration of bitter components. The bitterness threshold; S333: Calculate the comprehensive reward function: ； in, The value of the comprehensive reward function; This is the positive incentive reward value, calculated by S331; This is the negative penalty reward value, calculated in the second step. S334: Perform reward normalization processing: ； in, This is the normalized reward value; This is the historical average reward. The standard deviation of historical rewards; S335: Model training is performed based on the Actor-Critic algorithm framework, and the policy loss function is calculated. ； in, The value of the loss function; The reward value is negative normalized. Let be the logarithmic probability of an action in a given state; For policy network functions; Number the control action at time t; Let be the system state vector at time t; S336: Perform gradient backpropagation to update network parameters: ； in, For the updated network parameters; These are the current network parameters; The learning rate; For the loss function with respect to parameters The gradient is obtained by calculating the backpropagation algorithm; S34: Establish a decision-making mechanism based on a comprehensive reward function: when When it is determined to be an excellent control strategy, it continues to be executed; when When it is determined to be a good control strategy, parameter fine-tuning is performed; when It triggers significant adjustments to the control strategy to achieve multi-objective balance optimization.

5. The method according to claim 1, characterized in that, The hierarchical online detection system configuration and monitoring mechanism in step S4 includes the following specific operational steps: S41: Configured with a basic sensor real-time monitoring system, including a Pt100 platinum resistance temperature sensor, a piezoresistive pressure sensor, an electromagnetic flow meter, an ultrasonic level meter, and a glass electrode pH meter, to monitor basic process parameters in real time with a data acquisition cycle of 1 second, and provide real-time feedback signals for the control system. S42: Configure an online monitoring system for volatile oil yield, and monitor the amount of volatile oil collected in real time by installing a Coriolis mass flow meter at the oil-water separator outlet. Combined with the preset raw material weight The yield of volatile oil is calculated in real time using a sliding window averaging process with a data acquisition cycle of 5 minutes. ； in, Let t be the yield of volatile oil. This represents the ratio of the weight of the volatile oil to the dry weight of the raw material at time t. The cumulative weight of the volatile oil at time t is obtained by integration using a mass flow meter. The dry weight of the raw material is the preset feed weight. S43: Configure a gas chromatography-mass spectrometry (GC-MS) detection system, setting the DB-5ms capillary column to a length of 30m, an inner diameter of 0.25mm, and a film thickness of 0.25μm. Set the injection port temperature to 250℃ and the ion source temperature to 230℃. Monitor the concentrations of eugenol, bornyl acetate, and camphor volatile oil components at 20-minute detection intervals. Establish a standard curve using the external standard method for quantitative analysis. The detection limit is ≤1mg / L, and the relative standard deviation is ≤3%. S44: Equipped with a near-infrared spectroscopy detection system, with a scanning wavelength range of 10000-4000 cm⁻¹. -1 4cm resolution -1 The integration time was 32 times. A quantitative calibration model for the content of flavonoids, polysaccharides, and angelica glycosides was established using partial least squares method. Extract samples were collected every 5 minutes using an online sampling system for spectral scanning to detect the concentration of each component. Combined with real-time extraction liquid volume Calculate the extraction rate of each component; S45: Equipped with an electronic nose and electronic tongue detection system. The electronic nose features an array of eight metal oxide sensors with different selectivity, operating at 200-400℃ with a response time of 60 seconds, used to monitor changes in aroma profile. The electronic tongue features seven different types of taste sensors, including sweet, sour, bitter, salty, umami, astringent, and spicy sensors, with a response time of 90 seconds. Bitterness is calibrated using a quinine solution, and bitterness intensity is monitored in 10-minute detection cycles. ; S46: Design a hierarchical data fusion algorithm, employing time-series alignment preprocessing, unifying data timestamps from different detection periods through interpolation, and then using a weighted fusion mechanism. The weight of the basic sensor data is set to 0.6, and the weight of the component detection data is set to 0.

4. The weighted average is then calculated. ； in, For the merged data; Standardized values ​​for basic sensor data; It standardizes the component detection data; outputs comprehensive monitoring information, and realizes effective integration and complementary verification of information from multiple sensors.

6. The method according to claim 1, characterized in that, The phased sequential optimization control execution in step S5 includes the following specific operational steps: S51: Perform optimized control of the distillation stage. By monitoring the distillation temperature, pressure and volatile oil collection in real time, the distillation process is ended when the volatile oil yield reaches 90% of the target value or the distillation time exceeds 90 minutes. The distillation residue is collected to prepare for the decoction stage. S52: Implement synergistic optimization control of volatile oils and flavonoids during the decoction stage, using distillation residue for decoction extraction. By monitoring the decoction temperature in real time, when the total flavonoid extraction rate reaches... When necessary, the water temperature should be appropriately lowered to 60-70℃, among which, The current total flavonoid extraction rate is 0.7; the trigger threshold coefficient is 0.

7. To achieve the target flavonoid extraction rate, preset parameters are taken based on product requirements; priority is given to extracting heat-sensitive flavonoids to fully utilize residual resources. S53: Implement polysaccharide-angelica glycoside-bitterness balance control and establish a multi-index joint monitoring temperature regulation equation: ； in, This refers to the temperature regulation amount; This is the bitterness adjustment coefficient; The current bitterness concentration was obtained through electronic tongue detection. The bitterness threshold; It is a positive adjustment coefficient; The deviation value for polysaccharide extraction rate is the difference between the target value and the actual value. The deviation value for angelica glycoside extraction rate is the difference between the target value and the actual value. When the bitterness concentration exceeds the standard, the extraction temperature should be reduced; when the extraction rate of polysaccharides or angelica glycosides is insufficient, the temperature should be increased. S54: Perform LSTM-based sequence prediction endpoint control, and calculate the future extraction rate using the trained LSTM prediction model: ； in, To predict the extraction rate; For Long Short-Term Memory (LSTM) network functions; This is a historical extraction rate sequence, containing data from the first 30 time points; For LSTM model parameters; For the prediction time interval; When the predicted increase in the extraction rate of each major component is less than 2% and the comprehensive reward function When considering a 95% confidence interval, the termination procedure is initiated.

7. The method according to claim 1, characterized in that, The establishment of the edge computing and cloud collaborative control strategy optimization mechanism in step S6 includes the following specific operational steps: S61: Establish an edge intelligent control architecture. The edge controller is responsible for real-time control tasks with a 50ms cycle. It has a built-in lightweight single-layer LSTM-feedforward neural network model for fast response control. S62: Establish a cloud-based distributed federated learning service, using the FedAvg algorithm framework to batch train the distillation and decoction process data respectively. Each participating node has 10 local training rounds, and the local dataset is divided into training and validation sets in an 8:2 ratio. Differential privacy mechanism is used to protect data privacy. The global model parameter update formula is: ； in, These are global model parameters for the distillation process; This is the summation operator, representing the summation over all k device nodes; k is the device node number, from 1 to K; K is the total number of device nodes; This represents the proportion of data in the k-th node to the total data volume; The amount of distillation data at the k-th node; This represents the total amount of data from distillation. Here are the distillation local model parameters for the k-th node; and: ； in, These are the global model parameters for the decoction process; Let be the amount of water-boiled data in the k-th node; This represents the total data volume from the water-boiling process; The parameters of the local model for boiling water at the k-th node; S63: Design an incremental learning and update mechanism, setting the model update trigger condition as 50 batches of cumulative new data or a decrease in the control effect evaluation index exceeding 3%. Establish a model version rollback mechanism and anomaly detection strategy. When the performance of the new model is lower than 95% of the historical best performance, it will roll back to the previous version. A phased dynamic learning rate adjustment strategy will be adopted. When the number of updates hour, ,in, Let t be the learning rate for the t-th update, and 0.001 be the initial learning rate value; When the number of updates hour, ,in, For exponential operations, 0.9 is the decay factor; To perform floor operations, the learning rate decays every 20 rounds of updates; When the number of updates hour, Here, 0.0001 is the minimum learning rate value to avoid overfitting; The updated control model is pushed to the edge controller via either a wireless network or a wired Ethernet network, and a model deployment confirmation mechanism is established to ensure successful updates.

8. The method according to claim 1, characterized in that, It also includes S7, a method for training and validating deep learning models: S71: Establish a training dataset construction method, collect at least 1000 batches of historical extracted data as training samples, and use data augmentation techniques, including noise injection and temporal perturbation methods, to expand the training set to 5000 samples; S72: Set up a model validation index system, including control accuracy index (temperature control error MAE ≤ 0.5℃), extraction efficiency index (total extraction rate prediction error MAE ≤ 3%), and energy consumption prediction index (energy consumption prediction relative error ≤ 5%). S73: Establish a cross-validation mechanism, use 5-fold cross-validation to evaluate the model's generalization ability, ensure that the performance index R² on the validation set is ≥0.90, and avoid overfitting.

Images