Xianglan extraction and nano-assembly collaborative optimization method based on reinforcement learning driving

Abstract

This invention discloses a reinforcement learning-driven synergistic optimization method for lotus root extraction and nanoassembly, relating to the field of artificial intelligence technology. During system operation, the concentration of active ingredients and interaction indices in the whole-organ extract of lotus root are collected to generate an extraction state set Sta; the sensitivity of each active ingredient to nanoparticle size, encapsulation efficiency, and stability is calculated to form a sensitivity matrix Sen; based on Sen and real-time Sta, the optimal control parameter set Ctrl for extraction tank temperature, solvent ratio, enzyme dosage, and microfluidic assembly flow rate is calculated; according to Ctrl, the flow rate ratio of the aqueous phase to the lipid phase, the mixing order, and the component addition order are adjusted to form a nanoparticle product set Par; Par and the corresponding Sta are fed back to update Ctrl to achieve closed-loop synergistic optimization of extraction and assembly. By adopting the above technical solution, this application can achieve synergistic optimization of the extraction and nanoassembly processes, improving nanoparticle size consistency, encapsulation efficiency, and stability.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a reinforcement learning-driven collaborative optimization method for lotus root extraction and nano-assembly. Background Technology

[0002] Lotus seed (Nelumbo nucifera Gaertn.) is an important source of traditional Chinese medicine and natural active substances. Its entire organ is rich in various functional components, such as proanthocyanidins, methyl lotusine, and polysaccharides. These components are widely used in cosmetics and pharmaceuticals for anti-oxidation, anti-inflammation, moisturizing, and anti-aging. With the development of intelligent manufacturing and nanotechnology, researchers have begun to explore efficient preparation methods that combine whole-organ extraction of lotus seed with nano-assembly. The "Reinforcement Learning-Driven Synergistic Optimization Method for Lotus Seed Extraction and Nano-Assembly" is proposed precisely in this direction, aiming to achieve efficient utilization and synergistic efficacy enhancement of active ingredients through computational intelligent optimization of the extraction and nano-assembly processes.

[0003] In actual preparation processes, complex interactions exist between different active ingredients. For example, proanthocyanidins from lotus receptacles and methyl lotusine from lotus hearts may compete with each other in the extract, affecting the encapsulation efficiency and particle size distribution of the final nanoparticles. Currently, most processes still rely on empirical parameters or single optimization objectives, often neglecting trace interactions between components. This leads to significant differences in the extraction and encapsulation rates of different active ingredients within the same batch, failing to guarantee product consistency and maximize efficacy.

[0004] The root cause of this situation lies in the fact that traditional processes lack systematic analysis and dynamic control of the interactions between active ingredients, making it impossible to monitor the concentration of each component and their interactions in the extract in real time. When these interactions are not effectively regulated, a series of abnormal effects occur: for example, hydrophobic components are preferentially adsorbed, leading to the loss of hydrophilic polysaccharides; uneven particle size distribution causes a decrease in transdermal absorption; or there are large fluctuations in encapsulation efficiency between batches, seriously affecting product stability and efficacy. Therefore, there is an urgent need for an intelligent preparation method that can quantify the interactions of active ingredients and achieve synergistic optimization through reinforcement learning. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a reinforcement learning-driven collaborative optimization method for lotus root extraction and nano-assembly, which solves the problems mentioned in the background technology.

[0006] To achieve the above objectives, this invention provides the following technical solution: a reinforcement learning-driven collaborative optimization method for lotus root extraction and nano-assembly, comprising the following steps: Step S1: Collect the concentrations of active ingredients proanthocyanidins, active ingredients methyl lotusine and active ingredients polysaccharides in the whole organ extract of lotus seeds, as well as the interaction indicators between components, and generate the extraction state set Sta. The extraction state set Sta is used to quantify the interaction effects of each active ingredient and as input for subsequent strategies. Step S2: Calculate the sensitivity of each active ingredient (proanthocyanidins, methyl lotusine, and polysaccharides) to nanoparticle size, encapsulation efficiency, and stability using the extracted state set Sta, forming a sensitivity matrix Sen. The sensitivity matrix Sen is used to guide reinforcement learning strategies to optimize the synergistic utilization of active ingredients. Step S3: Based on the sensitivity matrix Sen and the real-time extraction state set Sta, calculate the optimal set of control parameters for extraction tank temperature, solvent ratio, enzyme dosage and microfluidic assembly flow rate, and form the control parameter set Ctrl to regulate the extraction and nano-assembly process to achieve synergistic optimization. Step S4: Adjust the flow rate ratio of the aqueous phase to the lipid phase, the mixing order, and the component addition order according to the control parameter set Ctrl to form the nanoparticle product set Par, and monitor the particle size and encapsulation rate to ensure that the nanoparticle product set Par achieves the optimization goals in terms of particle size consistency, encapsulation efficiency, and stability. Step S5: Feed back the nanoparticle product set Par and the corresponding extraction state set Sta to the strategy calculation module to update the control parameter set Ctrl for the next batch, so as to realize closed-loop collaborative optimization of extraction and assembly.

[0007] Preferably, step S1 includes: S11: Collect concentration data of active ingredients proanthocyanidins, active ingredients methyl lotusine and active ingredients polysaccharides through a spectral analysis device and online detection equipment, and simultaneously determine the interaction indicators between the components. S12: The collected data is transmitted to the data processing module through the data interface, and the module performs data integrity checks, normalization and error correction. S13: Based on the processed data and interaction indicators, an extraction status set Sta is generated, in which each record contains the concentration of three active ingredients and interaction indicators.

[0008] Preferably, step S1 further includes: S14: Perform range verification and repeatability testing on the concentration of active ingredients in the extraction state set Sta to ensure data consistency; S15: Perform error correction and correlation analysis on the interaction indicators to form a complete dataset that can be used for sensitivity calculation.

[0009] Preferably, step S2 includes: S21: Calculate the sensitivity of each active ingredient to nanoparticle size, encapsulation efficiency and stability using the extracted state set Sta, and quantify the interaction effects between ingredients. S22: Integrate the sensitivity data of each component to form a sensitivity matrix Sen, which records the quantitative impact of each active ingredient and its interactions on particle size, encapsulation efficiency and stability.

[0010] Preferably, step S3 includes: S31: Calculate the extraction tank temperature, solvent ratio, and enzyme dosage based on the sensitivity matrix Sen and the real-time extraction state set Sta; S32: Calculates the microfluidic assembly flow rate and integrates it into a set of control parameters Ctrl, which is used to regulate the extraction and nanoassembly process.

[0011] Preferably, step S4 includes: S41: Adjust the flow rate ratio of the aqueous phase to the lipid phase, the mixing order, and the order of component addition according to the set of control parameters (Ctrl) to perform microfluidic online assembly; S42: Real-time monitoring of particle size, encapsulation efficiency, and stability, and integration of data to form a nanoparticle product set Par, with each particle recording particle size, encapsulation efficiency, and stability parameters.

[0012] Preferably, step S4 further includes: S43: Standardize and statistically analyze the particle size, encapsulation efficiency, and stability parameters in the nanoparticle product set Par to generate a quantifiable product data set; S44: Correlate product data with the sensitivity matrix Sen to assess the impact of control parameters on the interaction of active ingredients.

[0013] Preferably, step S5 includes: S51: Input the nanoparticle product set Par and the corresponding extraction state set Sta into the strategy calculation module; S52: Update the next batch of control parameter set Ctrl based on the analysis results.

[0014] Preferably, step S5 further includes: S53: Fine-tune the set of control parameters (Ctrl) to adjust the extraction rate, particle size consistency, and encapsulation rate for the next batch.

[0015] Preferably, step S5 further includes: S54: Record the updated set of control parameters Ctrl and the corresponding set of nanoparticle products Par to ensure data integrity for continuous batch extraction and assembly closed-loop optimization.

[0016] This invention provides a reinforcement learning-driven collaborative optimization method for lotus root extraction and nanoassembly, which has the following beneficial effects: (1) During system operation, the Xianglian extraction and nano-assembly synergistic optimization method provided by this invention dynamically regulates the extraction and assembly process through reinforcement learning strategies, achieving efficient extraction of active ingredients and precise assembly of nanoparticles. Overall, this method can integrate the concentration and interaction information of multiple active ingredients in the extract, quantify the mutual influence between components, and thus scientifically guide the optimization of extraction conditions and assembly parameters. Through real-time data acquisition and strategy updates, extraction efficiency, nanoparticle size, encapsulation rate, and stability can be improved simultaneously, significantly improving the utilization rate of active ingredients and product consistency, and solving the problems of insufficient synergistic utilization of components and large fluctuations in product performance in traditional extraction and nano-assembly processes.

[0017] (2) In the data acquisition and state quantification stages, this method combines spectral analysis and online detection technology to accurately determine the concentrations of proanthocyanidins, methyl lotusine, and polysaccharides, and obtains the binding constants and competition relationships between components through molecular interaction analysis. After integrity checks, normalization, and error correction, this information forms a reliable set of extraction states that reflects the dynamics of component concentrations and their interactions. This processing method ensures high data consistency and repeatability, provides an accurate basis for sensitivity analysis, effectively reveals the synergistic effects between active ingredients, provides a scientific basis for optimizing extraction conditions and assembly strategies, significantly improves extraction efficiency, and ensures the stable performance of components in subsequent nano-assembly.

[0018] (3) In terms of sensitivity calculation and parameter optimization, this method uses a gradient boosting algorithm to quantify the effects of each active ingredient and its interactions on the particle size, encapsulation efficiency, and stability of nanoparticles, and forms a three-dimensional sensitivity matrix accordingly to guide the optimal setting of extraction temperature, solvent ratio, enzyme dosage, and microfluidic assembly flow rate. The deep reinforcement learning strategy continuously optimizes the operating parameters during the dynamic control process, realizing the synergistic utilization of active ingredients and the uniform production of nanoparticles. Finally, the product data is standardized and statistically analyzed, and correlated with sensitivity information to evaluate the effectiveness of the control strategy, which can significantly improve the particle size consistency, encapsulation efficiency, and stability of nanoparticles, while ensuring efficient synergy between extraction and assembly processes, thereby improving product quality and industrial feasibility. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall technical solution architecture of the reinforcement learning-driven collaborative optimization method for lotus root extraction and nano-assembly proposed in this invention. Figure 2This is a schematic diagram of the core principle framework of the present invention, which is based on the sensitivity matrix Sen to extract the state set Sta and the control parameter set Ctrl, and realizes closed-loop collaborative optimization through feedback of the nanoparticle product set Par. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0021] Example 1 To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0022] In the reinforcement learning-driven synergistic optimization method for the extraction and nano-assembly of *Ligusticum sinense*, the entire technical solution revolves around the extraction and nanoparticle assembly of proanthocyanidins, methyl lotusine, and polysaccharides from the whole organs of *Ligusticum sinense*. Reinforcement learning is used to achieve synergistic optimization of the extraction and nano-assembly processes. This embodiment details the specific implementation methods of each step.

[0023] Step S1 involves collecting the concentrations and interaction indices of active ingredients in the whole-organ extract of *Ligusticum striatum*, generating an extraction state set Sta. In this step, concentration data of the active ingredients proanthocyanidins, methyl lotusine, and polysaccharides are first collected using a spectroscopic analysis device and online detection equipment, while simultaneously determining the interaction indices between each component. Specifically, the spectroscopic analysis device can employ a UV-Vis spectrophotometer for the quantitative analysis of proanthocyanidins and polysaccharides, and a high-performance liquid chromatography (HPLC) instrument for the quantitative analysis of methyl lotusine. The online detection equipment may include an online infrared spectral probe and an online turbidity sensor for real-time monitoring of the dynamic changes in component concentrations in the extract. The determination of the interaction indices between components is performed using a molecular interaction analyzer. This instrument, based on surface plasmon resonance (SPR) technology, can quantitatively analyze the binding constants and competition relationships between proanthocyanidins and methyl lotusine, between proanthocyanidins and polysaccharides, and between methyl lotusine and polysaccharides in real time.

[0024] Furthermore, the collected data is transmitted to the data processing module via a data interface, where data integrity checks, normalization, and error correction are performed. The data interface employs a standardized transmission protocol to ensure high-speed and reliable data transmission between the spectral analysis device, online detection equipment, and the data processing module. The data integrity check mechanism includes data frame verification, timestamp consistency checks, and numerical range verification. Normalization maps concentration data obtained from different detection methods to a uniform zero-to-one interval to eliminate the influence of dimensional differences on subsequent analysis. The error correction algorithm uses an adaptive Kalman filter to effectively eliminate interference from sensor noise and baseline drift.

[0025] The processed data and interaction metrics are integrated to generate an extraction state set Sta, where each record contains the concentrations of three active ingredients and their interaction metrics. The extraction state set Sta is stored in a structured data format. Each record includes the concentration values ​​of proanthocyanidins, methyl lotusine, and polysaccharides, as well as the interaction strengths between proanthocyanidins and methyl lotusine, proanthocyanidins and polysaccharides, and methyl lotusine and polysaccharides, along with the corresponding timestamp and batch number. This extraction state set Sta serves as the foundational input data for subsequent sensitivity calculations and reinforcement learning strategy optimization.

[0026] Step S1 further includes performing range verification and repeatability testing on the concentrations of active ingredients in the extraction state set Sta to ensure data consistency. The range verification mechanism sets reasonable upper and lower limits for the concentrations of each active ingredient. When the detected value exceeds the preset range, an abnormal alarm is triggered and the cause of the deviation is recorded. Repeatability testing calculates the coefficient of variation through multiple parallel measurements of the same sample. When the coefficient of variation exceeds 5%, the system automatically marks the batch of data and suggests retesting.

[0027] Error correction and correlation analysis were performed on the interaction indices to form a complete dataset that can be used for sensitivity calculations. Error correction employed a multiple linear regression model to separate systematic errors from the original interaction indices. Correlation analysis quantified the statistical relationship between the concentrations of each active ingredient and the interaction indices by calculating the Pearson correlation coefficient, providing a basis for the subsequent construction of the sensitivity matrix.

[0028] Step S2 involves calculating the sensitivity of each active ingredient to nanoparticle size, encapsulation efficiency, and stability using the extracted state set Sta, forming a sensitivity matrix Sen. In this step, the sensitivity of each active ingredient to nanoparticle size, encapsulation efficiency, and stability is first calculated using the extracted state set Sta, while simultaneously quantifying the interaction effects between components. The specific calculation process employs a gradient boosting algorithm, using the concentration of each active ingredient and interaction indicators as input features, training a first prediction model with nanoparticle size as the target variable, a second prediction model with encapsulation efficiency as the target variable, and a third prediction model with stability as the target variable. By analyzing the importance of the input features to each target variable, the sensitivity values ​​of each active ingredient to the three output indicators are obtained.

[0029] The Sensitivity Matrix Sens is constructed to structure the above calculation results. The rows of the matrix correspond to the three active ingredients and three interaction effects, while the columns correspond to the three output metrics: particle size, encapsulation efficiency, and stability. Each matrix element records the numerical value of the influence of the corresponding active ingredient or interaction effect on the target metric. The Sensitivity Matrix Sens is used to guide reinforcement learning strategies in optimizing the synergistic utilization of active ingredients, providing crucial input for subsequent control parameter optimization.

[0030] Step S3 involves calculating the optimal set of control parameters for the extraction tank temperature, solvent ratio, enzyme dosage, and microfluidic assembly flow rate based on the sensitivity matrix Sen and the real-time extraction state set Sta, forming the control parameter set Ctrl. In this step, the extraction tank temperature, solvent ratio, and enzyme dosage are calculated based on the sensitivity matrix Sen and the real-time extraction state set Sta. The optimization of the extraction tank temperature employs a deep Q-network algorithm in reinforcement learning, using the sensitivity matrix Sen and the real-time extraction state set Sta as state inputs and the temperature setpoint as the action output. The optimal temperature strategy is learned through continuous interaction with the simulation environment. The optimization of the solvent ratio is also based on a deep Q-network. The state input includes the current concentration of each active ingredient and its interaction index, and the action space is defined as the set of possible values ​​for the volume ratio of water to organic solvent. The optimization of the enzyme dosage uses a policy gradient algorithm, dynamically adjusting the amount of enzyme added based on the influence of each component on particle size in the sensitivity matrix Sen.

[0031] The microfluidic assembly flow rate is calculated and integrated into a control parameter set Ctrl, which is used to regulate the extraction and nanoassembly processes. The optimization of the microfluidic assembly flow rate uses the influence of each component in the sensitivity matrix Sen on the encapsulation efficiency as a reward signal, and a reinforcement learning algorithm is used to determine the optimal setpoints for the aqueous and lipid phase flow rates. The control parameter set Ctrl includes key process parameters such as extraction tank temperature setpoint, solvent ratio setpoint, enzyme dosage setpoint, aqueous phase flow rate setpoint, and lipid phase flow rate setpoint. This control parameter set Ctrl serves as the adjustment command, which is issued to the extraction and microfluidic assembly equipment for execution.

[0032] Step S4 involves adjusting the flow rate ratio of the aqueous phase to the lipid phase, the mixing order, and the component addition order according to the control parameter set Ctrl to form a nanoparticle product aggregate Par. The particle size and encapsulation efficiency are monitored to ensure that the nanoparticle product aggregate Par achieves optimized performance in terms of particle size consistency, encapsulation efficiency, and stability. In this step, the flow rate ratio of the aqueous phase to the lipid phase, the mixing order, and the component addition order are adjusted according to the control parameter set Ctrl to perform microfluidic online assembly. The flow rates of the aqueous phase and the lipid phase are precisely controlled by a high-precision injection pump of the microfluidic chip, with the flow rate ratio adjustable within the range of 0.5 to 2 times. The mixing order is defined as either adding the lipid phase first or adding the aqueous phase first. The component addition order specifies the timing and interval of adding the active ingredient extract, lipid material, and surfactant.

[0033] Real-time monitoring of particle size, encapsulation efficiency, and stability is performed, and the data is integrated to form a nanoparticle product aggregate (Par). Each particle's particle size, encapsulation efficiency, and stability parameters are recorded. Particle size monitoring utilizes a dynamic light scattering particle size analyzer with a sampling frequency of once per second. Encapsulation efficiency is determined by separating free drug from encapsulated drug through ultrafiltration centrifugation, followed by quantitative analysis using a UV spectrophotometer. Stability assessment quantifies the aggregation degree of nanoparticles and drug leakage rate through accelerated centrifugation and isothermal incubation tests. The nanoparticle product aggregate (Par) is stored in a list structure, with each record containing batch number, particle number, particle size value, encapsulation efficiency value, stability value, and detection timestamp.

[0034] Step S4 further includes standardizing and statistically analyzing the particle size, encapsulation efficiency, and stability parameters of the nanoparticle product set Par to generate a quantifiable product data set. Standardization unifies the test data from different batches to the same dimensions. Statistical analysis includes calculating descriptive statistics such as mean, standard deviation, coefficient of variation, and distribution histogram to comprehensively assess product quality. The product data is correlated with the sensitivity matrix Sen to evaluate the impact of control parameters on the interaction of active ingredients. By comparing the differences in product performance under different control parameters, the predictive accuracy of the sensitivity matrix Sen is verified, and learning samples are provided for the next round of parameter optimization.

[0035] Step S5 involves feeding back the nanoparticle product set Par and the corresponding extraction state set Sta to the policy calculation module to update the control parameter set Ctrl for the next batch, achieving closed-loop collaborative optimization of extraction and assembly. In this step, the nanoparticle product set Par and the corresponding extraction state set Sta are input into the policy calculation module. The policy calculation module is built on a deep reinforcement learning architecture and includes a feature extraction network, a value evaluation network, and a policy output network. The feature extraction network encodes the raw data of the extraction state set Sta and the nanoparticle product set Par into high-dimensional feature vectors. The value evaluation network predicts the long-term cumulative reward in the current state. The policy output network outputs the optimal control parameters based on the state features.

[0036] The control parameter set Ctrl for the next batch is updated based on the analysis results. The update algorithm employs an experience replay mechanism, storing the state-action pairs of historical batches in an experience pool and randomly selecting small batches of samples for network parameter updates. The update strategy includes two modes: full update and incremental update. Full update is suitable for situations where product performance fluctuates significantly between batches, while incremental update is suitable for situations where product performance is relatively stable.

[0037] Step S5 further includes fine-tuning the control parameter set Ctrl to adjust the extraction rate, particle size consistency, and encapsulation rate for the next batch. The fine-tuning mechanism is based on model predictive control theory, applying a small-amplitude perturbation to the current optimal control parameters. The perturbation amplitude is dynamically adjusted according to the gap between the current product performance and the target performance; a larger perturbation is used to accelerate convergence when the gap is large, and a smaller perturbation is used for fine-tuning when the gap is small.

[0038] The updated control parameter set Ctrl and the corresponding nanoparticle product set Par are recorded to ensure data integrity for continuous batch extraction and assembly closed-loop optimization. Data recording employs a distributed storage architecture, with the primary data node stored on a local server and backup data nodes stored on a cloud storage service. Recorded content includes the original setpoints, actual execution values, and feedback correction values ​​of the control parameter set Ctrl, as well as complete detection data for the nanoparticle product set Par. The data retention period is set to at least thirty-six months to ensure the integrity of long-term data analysis.

[0039] Example 2 This embodiment provides a detailed description of the spectral analysis device and online detection equipment in step S1, offering specific technical means to further ensure the accuracy and real-time nature of active ingredient concentration acquisition.

[0040] In step S11, concentration data of the active ingredients proanthocyanidins, methyl lotusine, and polysaccharides are collected using a spectroscopic analysis device and an online detection device, while the interaction indicators between the components are measured. The spectroscopic analysis device includes three independent detection channels. The first detection channel is equipped with a UV spectrophotometer for the quantitative analysis of proanthocyanidins, with detection wavelengths set at 280 nm and 320 nm, and the linear range of the standard curve is 0.1 to 50 μg / mL. The second detection channel is equipped with a high-performance liquid chromatograph with a fluorescence detector for the quantitative analysis of methyl lotusine. The chromatographic column uses octadecylsilane-bonded silica gel packing, and the mobile phase is a mixed solution of acetonitrile and potassium dihydrogen phosphate buffer, with the detection wavelength set at 280 nm. The third detection channel is equipped with a phenol-sulfuric acid method detection device for the quantitative analysis of polysaccharides, with the detection wavelength set at 490 nm.

[0041] The online detection equipment includes an online infrared spectroscopy probe, an online turbidity sensor, and an online pH meter. The online infrared spectroscopy probe uses attenuated total reflectance mode, directly immersed in the extract to collect infrared spectra in real time, and calculates the relative concentration of each active ingredient based on the characteristic peak area. The online turbidity sensor uses a 90-degree scattering light detection method to monitor the turbidity change trend during extraction. The online pH meter monitors the pH of the extract to ensure that the enzymatic hydrolysis reaction proceeds under optimal pH conditions.

[0042] The determination of intercomponent interaction indices was performed using a molecular interaction analyzer equipped with four detection channels, capable of simultaneously detecting interactions between proanthocyanidins and methylcaenine, between proanthocyanidins and polysaccharides, and between methylcaenine and polysaccharides. The detection principle is based on surface plasmon resonance (SPR) technology, where one component is immobilized on the sensor chip surface, while another component flows through the chip surface. The interaction strength is calculated using binding rate constants and dissociation rate constants. The detection results are output as numerical values ​​of the binding constants, expressed as the reciprocal of moles per liter.

[0043] The collected data is transmitted to the data processing module via a data interface. The data interface supports both Ethernet and serial port transmission modes. Ethernet transmission uses the TCP / IP protocol, while serial port transmission uses the RS232 protocol. The data transmission frequency is set to ten frames per second, with each frame containing a timestamp, detection channel number, and concentration value. The data processing module uses an industrial-grade server equipped with a multi-core CPU and solid-state drive to ensure high-speed and reliable data processing.

[0044] The module performs data integrity checks, normalization, and error correction. Data integrity checks include frame checksum verification, timestamp continuity checks, and numerical range checks. When data loss or anomalies are detected, the module automatically triggers a data retransmission mechanism. Normalization maps concentration data obtained from different detection methods to the zero-to-one range using a minimum-maximum normalization method. Error correction employs a moving average filtering method with a filtering window set to five sampling points, effectively smoothing fluctuations caused by random noise.

[0045] The processed data and interaction indicators are integrated to generate an extraction status set Sta, where each record contains the concentrations of three active ingredients and their interaction indicators. The extraction status set Sta is stored in a structured query language database, with a table structure including fields such as record number, batch number, collection time, proanthocyanidin concentration, methyl lotusine concentration, polysaccharide concentration, interaction strength between proanthocyanidins and methyl lotusine, interaction strength between proanthocyanidins and polysaccharides, and interaction strength between methyl lotusine and polysaccharides. At least one hundred records are generated for each batch of extract to ensure statistical representativeness of the data.

[0046] The concentrations of proanthocyanidins, methyl lotusine, and polysaccharides in the extract were collected using a spectroscopic analysis device and an online detection equipment.

[0047] Specifically: UV spectrophotometer with dual wavelengths for proanthocyanidin determination, linear range 0.1–50 μg / mL; high performance liquid chromatography with fluorescence detector for determination of methylnephrine; phenol-sulfuric acid method for polysaccharide detection, wavelength 490 nm.

[0048] It should be noted that: an online infrared probe monitors the dynamic concentration of components in the extract in real time, a turbidity sensor monitors changes in turbidity, and a pH meter ensures the enzymatic hydrolysis reaction occurs at the optimal pH. Interaction indices are determined using surface plasmon resonance (SPR) technology to measure the binding constants and competition relationships between components.

[0049] During data processing: The data interface supports Ethernet or RS232 transmission at 10 frames per second, including timestamps, channel numbers, and concentration values. An industrial-grade server performs integrity checks, minimum-maximum normalization, and moving average filter error correction.

[0050] In this embodiment: the extraction status set Sta stores structured data, including the concentrations of the three components, interaction strength, collection time, and batch number for each record. ≥100 records are generated for each batch of extract to ensure statistical representativeness.

[0051] Specifically: upper and lower limit verification and repeatability testing were performed on the concentration, and multiple regression error correction and Pearson correlation analysis were conducted on the interaction indicators to form a complete dataset for sensitivity calculation.

[0052] Example 3 This embodiment provides a detailed explanation of the construction and updating mechanism of the sensitivity matrix Sen in step S2, illustrating how to improve the prediction accuracy of the sensitivity matrix through continuous learning.

[0053] In step S21, the sensitivity of each active ingredient to nanoparticle size, encapsulation efficiency, and stability is calculated using the extracted state set Sta, while simultaneously quantifying the interaction effects between components. Sensitivity calculation employs a gradient boosting decision tree algorithm, which uses a decision tree as the base learner and constructs a strong predictive model through iterative optimization of the loss function. Input features include eight characteristics: proanthocyanidin concentration, methyl lotusine concentration, polysaccharide concentration, interaction strength between two components, and interaction strength between three components. Output targets are particle size, encapsulation efficiency, and stability.

[0054] The hyperparameters of the gradient boosting model are set as follows: maximum decision tree depth is set to six layers, learning rate is set to 0.1, number of iterations is set to one hundred, and subsampling ratio is set to 0.8. The training data for each model comes from the historical batch extraction state set Sta and the corresponding nanoparticle product set Par. Feature importance is obtained by calculating the weighted sum of the number of splits for each feature across all decision trees, and after normalization, the percentage influence of each feature on the target variable is obtained.

[0055] Quantification of inter-component interaction effects is achieved by introducing interaction features into the gradient boosting model. The interaction feature is constructed by multiplying the concentrations of the two active components, and its importance reflects the contribution of synergistic or competitive effects between the components to the target variable. When the importance of the interaction feature is significantly higher than that of individual component features, it indicates the existence of a strong inter-component interaction effect.

[0056] In step S22, the sensitivity data of each component are integrated to form a sensitivity matrix Sen. This matrix records the quantitative impact of each active ingredient and its interactions on particle size, encapsulation efficiency, and stability. The sensitivity matrix Sen adopts a three-dimensional matrix structure, with row indices corresponding to active ingredients and interaction types, column indices corresponding to target indicators, and matrix elements being normalized sensitivity values. Normalization involves dividing each sensitivity value by the maximum value of all sensitivity values ​​in that row to ensure that matrix elements are distributed within the range of zero to one.

[0057] The sensitivity matrix Sensing is updated using an online learning mechanism. After each batch of nanoparticles is produced, the new batch of data is added to the training sample pool. When the training sample pool exceeds one hundred batches, the gradient boosting model is retrained using all samples, and the sensitivity matrix Sensing is updated. This online learning mechanism ensures that the sensitivity matrix Sensing can track process drift such as changes in raw materials and equipment status, maintaining the accuracy of sensitivity predictions.

[0058] The sensitivity of each active ingredient to nanoparticle size, encapsulation efficiency, and stability was calculated using the extracted state set Sta.

[0059] Specifically: A gradient boosting decision tree algorithm is used. The input features include the concentrations of three components and their interaction strengths. The output targets are particle size, encapsulation efficiency, and stability. Hyperparameter settings: tree depth 6 layers, learning rate 0.1, 100 iterations, subsampling 0.8.

[0060] In this embodiment: interactive features are introduced to quantify the synergistic or competitive effects between components, and the importance of features reflects their contribution to the target index.

[0061] The Sensitivity Matrix Sens uses a three-dimensional matrix structure, with row indices corresponding to components and interaction types, column indices corresponding to target indicators, and elements normalized to the 0–1 range.

[0062] It should be noted that the sensitivity matrix update adopts online learning. After each batch of nanoparticles is produced, the data is added to the training pool. When the number of samples exceeds 100 batches, the model is retrained to ensure adaptability to raw material and process drift.

[0063] Example 4 This embodiment provides a detailed explanation of the optimization algorithm and execution mechanism of the control parameter set Ctrl in step S3, illustrating how reinforcement learning can achieve synergistic optimization of extraction and nano-assembly.

[0064] In step S31, the extraction tank temperature, solvent ratio, and enzyme dosage are calculated based on the sensitivity matrix Sen and the real-time extraction state set Sta. The optimization of the control parameters employs a deep reinforcement learning algorithm. The specific architecture includes a state encoder composed of a convolutional neural network, a policy network composed of a multilayer perceptron, and a value network composed of a multilayer perceptron. The state encoder encodes the extraction state set Sta and the sensitivity matrix Sen into a state vector. The policy network takes the state vector as input and outputs the mean and variance of each control parameter. The value network evaluates the expected cumulative reward in the current state.

[0065] The optimization objective for the extraction tank temperature is to maximize the extraction rates of proanthocyanidins and methyl lotusine while minimizing energy consumption. The state space includes the current extract temperature, the concentrations of each active ingredient, interaction indices, and historical temperature setpoints. The action space is defined as a temperature setpoint range of 40 to 80 degrees Celsius, discretized into 81 discrete actions. The reward function is designed as a weighted sum of extraction rates minus an energy consumption penalty term, balancing extraction efficiency and energy consumption through weighting coefficients.

[0066] The optimization objective of the solvent ratio is to maximize the solubility of the active ingredient and the subsequent encapsulation rate of nanoparticles. The state space includes the current solvent ratio, the concentration of each active ingredient, and interaction indices. The action space is defined as the volume ratio of water to ethanol, with a selectable range of 3:7 to 7:3, discretized into forty-one discrete actions. The reward function is designed as a weighted combination of the encapsulation rate value and the predicted solubility value.

[0067] The optimization objective for enzyme dosage is to minimize enzyme cost while ensuring extraction efficiency. The state space includes the current enzyme dosage, concentrations of each active ingredient, interaction indices, and extraction time. The action space is defined as the enzyme dosage, with a selectable range of 0.1 to 2 grams per liter, discretized into twenty discrete actions. The reward function is designed as a weighted sum of extraction rates minus an enzyme cost penalty term.

[0068] In step S32, the microfluidic assembly flow rate is calculated and integrated into a control parameter set Ctrl, which is used to regulate the extraction and nanoassembly process. The optimization of the microfluidic assembly flow rate also employs a deep reinforcement learning algorithm, but the reward function focuses on the particle size consistency and encapsulation efficiency of the nanoparticles. The state space includes the sensitivity matrix Sen, the real-time extraction state set Sta, the raw material tank level, and the microfluidic chip inlet pressure. The action space is defined as the aqueous phase flow rate and the lipid phase flow rate; the aqueous phase flow rate can be selected from 100 to 1000 μL / min, and the lipid phase flow rate can be selected from 50 to 500 μL / min.

[0069] The integration of the control parameter set Ctrl adopts a parameter server architecture, storing the extracted parameters and nano-assembly parameters in their respective subsets. The parameter server receives the optimization results from the reinforcement learning module, performs parameter validity checks, and then sends them to the execution device. Validity checks include parameter range verification, parameter consistency verification, and parameter conflict detection. When an abnormal parameter is detected, the system automatically triggers a safety shutdown procedure and issues an alarm to notify the operator.

[0070] The extraction vessel temperature, solvent ratio, enzyme dosage, and microfluidic assembly flow rate were calculated based on Sen and Sta.

[0071] Specifically: extraction temperature optimization uses a deep Q-network with state inputs of Sta and Sen, and action outputs of 40–80°C, resulting in 81 discrete actions; solvent ratio optimization includes the current component concentration in the state, with actions of water:ethanol volume ratios of 3:7–7:3, resulting in 41 discrete actions; enzyme dosage optimization uses a strategy gradient algorithm with an action range of 0.1–2 g / L.

[0072] Microfluidic assembly flow rate optimization uses the influence of each component in the sensitivity matrix on encapsulation efficiency as the reward signal, with aqueous phase flow rate of 100–1000 μL / min and lipid phase flow rate of 50–500 μL / min.

[0073] In the scenario set in Example 1: the control parameter set Ctrl is integrated and its validity is verified by the parameter server, and then sent to the extraction tank and microfluidic chip for execution.

[0074] Example 5 This embodiment provides a detailed explanation of the real-time monitoring, feedback adjustment, and closed-loop optimization mechanisms in steps S4 and S5, illustrating how to achieve continuous synergistic optimization of extraction and nano-assembly.

[0075] In step S41, the flow rate ratio of the aqueous phase to the lipid phase, the mixing order, and the component addition order are adjusted according to the control parameter set Ctrl to perform microfluidic online assembly. Flow rate adjustment is achieved using a high-precision syringe pump driven by a stepper motor, with a flow control accuracy better than one percent. The setpoints for the aqueous phase and lipid phase flow rates are derived from the corresponding fields in the control parameter set Ctrl. The flow rate adjustment response time is less than one hundred milliseconds, ensuring rapid tracking of control commands.

[0076] The mixing sequence is controlled via an array of solenoid valves. These solenoid valves are fast-response pneumatic valves with a switching time of less than fifty milliseconds. The system has two preset mixing sequence modes: Mode 1 defines the addition of the lipid phase before the aqueous phase, and Mode 2 defines the addition of the aqueous phase before the lipid phase. The mixing sequence is selected based on the setpoints in the control parameter set Ctrl, and the control program instructs the solenoid valve array to execute the predefined switching sequence.

[0077] The order of component addition is controlled via a multi-channel dispensing pump. The pump has four independent channels for adding the active ingredient extract, lipid material, surfactant, and auxiliary solvent, respectively. The amount, timing, and interval of each addition are precisely controlled according to the settings in the control parameter set Ctrl. The standard procedure for the addition order is as follows: first, add the lipid material to the main phase container; second, add the surfactant; third, add the active ingredient extract; and finally, add the auxiliary solvent to the target volume.

[0078] In step S42, particle size, encapsulation efficiency, and stability are monitored in real time, and the data are integrated to form a nanoparticle product aggregate Par. For each particle, particle size, encapsulation efficiency, and stability parameters are recorded. Particle size monitoring uses a dynamic light scattering particle size analyzer, specifically a nanoparticle size analyzer, with a measurement range of one nanometer to ten micrometers and a measurement accuracy better than two percent. The sampling frequency is set to once per second, and each measurement collects three thousand scattered light intensity data points. The average particle size and polydispersity index are calculated using a cumulative distribution function.

[0079] The encapsulation efficiency was determined using ultrafiltration centrifugation. First, the nanoparticle suspension was added to an ultrafiltration centrifuge tube, and separation was performed under a set centrifugal force. After separation, the concentrate and supernatant at the bottom of the centrifuge tube were collected, and the total drug content and free drug content were measured, respectively. The encapsulation efficiency was calculated as: encapsulation efficiency equals the drug content in the concentrate divided by the total drug content multiplied by 100%. The measurement results were transmitted to the data processing module in real time via a data interface.

[0080] Stability assessment employed a comprehensive scoring method, including accelerated centrifugation, isothermal incubation, and potential measurement. The accelerated centrifugation test involved centrifuging the nanoparticle suspension at a set centrifugal force for 30 minutes, observing for precipitation or flocculation, and assigning a score from 0 to 100 based on the amount of precipitation. The isothermal incubation test involved placing the nanoparticle suspension at 40 degrees Celsius for 72 hours, measuring the particle size and encapsulation efficiency before and after incubation, and assigning a score from 0 to 100 based on the magnitude of change. The potential measurement test measured the Zeta potential of the nanoparticles using a laser particle size analyzer, and assigned a stability score based on the absolute value of the potential. The comprehensive score was a weighted average of the three scores, with weights of 40% for accelerated centrifugation, 40% for isothermal incubation, and 20% for potential measurement.

[0081] In step S43, the particle size, encapsulation efficiency, and stability parameters in the nanoparticle product set Par are standardized and statistically analyzed to generate a quantifiable product dataset. The standardization process uses Z-score standardization to transform each parameter into a normal distribution with a mean of zero and a standard deviation of one. Statistical analysis includes calculating the mean, standard deviation, coefficient of variation, skewness, kurtosis, and correlation coefficient matrix between parameters. The statistical results are stored in the metadata area of ​​the product dataset for subsequent analysis and visualization.

[0082] Product data were correlated with the sensitivity matrix Sen to assess the impact of control parameters on the interactions of active ingredients. The correlation analysis employed a multiple linear regression model, using each parameter in the control parameter set Ctrl as the independent variable and the product performance index as the dependent variable. The magnitude of the regression coefficients reflects the degree of influence of each control parameter on product performance. The consistency between the sensitivity of each component in the sensitivity matrix Sen and the regression coefficients was verified by comparing their correlation.

[0083] In step S51, the nanoparticle product set Par and the corresponding extraction state set Sta are input into the strategy calculation module. The strategy calculation module adopts a distributed computing architecture, including a data preprocessing node, a feature engineering node, a model inference node, and a decision output node. The data preprocessing node is responsible for receiving and verifying the integrity of the input data. The feature engineering node is responsible for encoding the raw data into feature vectors required for model input. The model inference node is responsible for performing the forward propagation calculation of the deep neural network. The decision output node is responsible for converting the model output into specific control parameter suggestion values.

[0084] In step S52, the control parameter set Ctrl for the next batch is updated based on the analysis results. The update mechanism adopts a policy iteration framework, which includes two alternating phases: policy evaluation and policy improvement. The policy evaluation phase calculates the state value function and action advantage function under the current policy. The policy improvement phase adjusts the policy network parameters based on the evaluation results, updates the control parameters, and generates the policy. The convergence condition for policy iteration is set as the coefficient of variation of product performance being less than 3% for three consecutive batches.

[0085] In step S53, the control parameter set Ctrl is fine-tuned to adjust the extraction rate, particle size consistency, and encapsulation rate for the next batch. Fine-tuning employs a model predictive control method, applying Gaussian noise perturbation to the current optimal control parameters. The perturbation amplitude is dynamically adjusted based on the gap between the current product performance and the target performance, with the initial perturbation amplitude set at 10% of the optimal parameters, gradually narrowing to 1% as the performance gap decreases. After a feasibility check, the fine-tuned parameters replace the corresponding fields in the original control parameter set Ctrl.

[0086] In step S54, the updated control parameter set Ctrl and the corresponding nanoparticle product set Par are recorded to ensure data integrity for continuous batch extraction and assembly closed-loop optimization. Data recording uses a time-series database, stored in key-value pair format. The primary key consists of the batch number and record timestamp, and the key value is a JSON string containing all control and product parameters. Data storage employs a distributed redundancy strategy, synchronously storing data on three data nodes to ensure that a single point of failure does not affect data integrity. Nanoparticle assembly is achieved by adjusting the water / lipid phase flow rate ratio, mixing order, and component addition order according to Ctrl.

[0087] Specifically: the water / lipid flow rate ratio is adjustable from 0.5 to 2 times, the injection pump accuracy is better than 1%, and the response time is <100 ms. The mixing order is controlled by a fast-response solenoid valve; in Mode 1, the lipid phase is added first, and in Mode 2, the aqueous phase is added first. The component delivery order is controlled by a multi-channel dispensing pump, with each channel corresponding to the active ingredient, lipid material, surfactant, and auxiliary solvent.

[0088] Real-time particle size was monitored using a dynamic light scattering instrument, encapsulation efficiency was quantified using ultrafiltration centrifugation, and stability was assessed by a comprehensive evaluation of accelerated centrifugation, constant temperature placement, and zeta potential.

[0089] In this embodiment: data is standardized (Z-score), statistically analyzed (mean, standard deviation, coefficient of variation, skewness, kurtosis, and correlation matrix), and stored in the product dataset. The product data is correlated with the sensitivity matrix, and multiple linear regression is used to evaluate the impact of control parameters on component interactions, providing a basis for the next round of optimization.

[0090] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A reinforcement learning-driven collaborative optimization method for lotus root extraction and nano-assembly, characterized by: Includes the following steps: S1: Collect the concentrations of active ingredients proanthocyanidins, active ingredient methyl lotusine, and active ingredient polysaccharides in the whole organ extract of lotus seeds, as well as the interaction indicators between components, and generate the extraction state set Sta. The extraction state set Sta is used to quantify the interaction effects of each active ingredient and as input for subsequent strategies. S2: Calculate the sensitivity of each active ingredient, proanthocyanidin, methyl lotusine, and polysaccharide, to nanoparticle size, encapsulation efficiency, and stability using the extracted state set Sta, forming a sensitivity matrix Sen. The sensitivity matrix Sen is used to guide reinforcement learning strategies to optimize the synergistic utilization of active ingredients. S3: Based on the sensitivity matrix Sen and the real-time extraction state set Sta, calculate the optimal set of control parameters for extraction tank temperature, solvent ratio, enzyme dosage and microfluidic assembly flow rate, and form the control parameter set Ctrl to regulate the extraction and nano-assembly process to achieve synergistic optimization. S4: Adjust the flow rate ratio of the aqueous phase to the lipid phase, the mixing order, and the component addition order according to the control parameter set Ctrl to form the nanoparticle product set Par, and monitor the particle size and encapsulation rate to ensure that the nanoparticle product set Par achieves the optimization goals in terms of particle size consistency, encapsulation efficiency, and stability. S5: Feed back the nanoparticle product set Par and the corresponding extraction state set Sta to the strategy calculation module to update the control parameter set Ctrl for the next batch, thereby achieving closed-loop collaborative optimization of extraction and assembly.

2. The method for synergistic optimization of lotus root extraction and nano-assembly based on reinforcement learning as described in claim 1, characterized in that: Step S1 includes: Concentration data of the active ingredients proanthocyanidins, methyl lotusine, and polysaccharides were collected using a spectral analysis device and an online detection device, and the interaction indicators between the components were measured simultaneously. The collected data is transmitted to the data processing module through the data interface, and the module performs data integrity checks, normalization and error correction. The processed data and interaction indicators are integrated to generate an extraction status set Sta, in which each record contains three active ingredient concentrations and interaction indicators.

3. The method for synergistic optimization of lotus root extraction and nano-assembly based on reinforcement learning as described in claim 2, characterized in that: The spectral analysis device includes an ultraviolet spectrophotometer, a high-performance liquid chromatograph, and a phenol-sulfuric acid method detection device, which are used to quantitatively analyze proanthocyanidins, methyl lotusine, and polysaccharides, respectively. The online detection equipment includes an online infrared spectroscopy probe, an online turbidity sensor, and an online pH meter, which are used to monitor the dynamic changes in the component concentration, turbidity, and pH of the extract in real time. The inter-component interaction indices were determined using a molecular interaction analyzer, and the binding constants and competition relationships between proanthocyanidins and methyl lotusine, proanthocyanidins and polysaccharides, and methyl lotusine and polysaccharides were analyzed in real time based on surface plasmon resonance technology.

4. The method for synergistic optimization of lotus root extraction and nano-assembly based on reinforcement learning as described in claim 2, characterized in that: Step S1 further includes: Perform range validation and repeatability testing on the concentration of active ingredients in the extraction state set Sta to ensure data consistency; Error correction and correlation analysis are performed on the interaction indicators to form a complete dataset that can be used for sensitivity calculations.

5. The method for synergistic optimization of lotus root extraction and nano-assembly based on reinforcement learning as described in claim 1, characterized in that: Step S2 includes: The sensitivity of each active ingredient to nanoparticle size, encapsulation efficiency, and stability was calculated using the extracted state set Sta, while the interaction effects between ingredients were quantified. Sensitivity data of each component are integrated to form a sensitivity matrix Sen, which records the quantitative effects of each active ingredient and its interactions on particle size, encapsulation efficiency and stability.

6. The reinforcement learning-driven collaborative optimization method for lotus root extraction and nano-assembly according to claim 5, characterized in that: The sensitivity calculation adopts the gradient boosting algorithm, using the concentration of each active ingredient and interaction index as input features, the nanoparticle size as the target variable to train the first prediction model, the encapsulation rate as the target variable to train the second prediction model, and the stability as the target variable to train the third prediction model. The sensitivity value is obtained by analyzing the importance of the input features to each target variable. The sensitivity matrix Sen adopts a three-dimensional matrix structure, with row indices corresponding to active ingredients and interaction types, column indices corresponding to target indicators, and matrix elements being normalized sensitivity values.

7. The method for synergistic optimization of lotus root extraction and nano-assembly based on reinforcement learning as described in claim 1, characterized in that: Step S3 includes: The extraction tank temperature, solvent ratio, and enzyme dosage are calculated based on the sensitivity matrix Sen and the real-time extraction state set Sta. The microfluidic assembly flow rate is calculated and integrated into a set of control parameters Ctrl, which is used to regulate the extraction and nanoassembly process.

8. The reinforcement learning-driven collaborative optimization method for lotus root extraction and nano-assembly according to claim 7, characterized in that: The optimization of the extraction tank temperature, solvent ratio, and enzyme dosage is achieved using a deep reinforcement learning algorithm. The sensitivity matrix Sen and the real-time extraction state set Sta are used as state inputs, and the temperature set value, solvent ratio, and enzyme dosage are used as action outputs. The optimal strategy is learned through continuous interaction with the simulation environment. The optimization of the microfluidic assembly flow rate uses the influence of each component in the sensitivity matrix Sen on the encapsulation rate as a reward signal, and determines the optimal set values ​​of the aqueous phase flow rate and the lipid phase flow rate through a reinforcement learning algorithm.

9. The method for synergistic optimization of lotus root extraction and nano-assembly based on reinforcement learning as described in claim 1, characterized in that: Step S4 includes: Based on the set of control parameters Ctrl, the flow rate ratio of the aqueous phase to the lipid phase, the mixing order, and the order of component addition are adjusted to perform microfluidic online assembly. Real-time monitoring of particle size, encapsulation efficiency, and stability is performed, and the data is integrated to form a nanoparticle product set Par, with each particle recording its particle size, encapsulation efficiency, and stability parameters.

10. The reinforcement learning-driven collaborative optimization method for lotus root extraction and nano-assembly according to claim 1, characterized in that: Step S4 further includes: The particle size, encapsulation efficiency, and stability parameters of the nanoparticle product set Par were standardized and statistically analyzed to generate a quantifiable product data set. By correlating product data with the Sensitivity Matrix (Sen), the impact of control parameters on the interaction of active ingredients was assessed.

Images