A digital twin and virtual-real synchronous control method for fluidized bed granulation process of traditional Chinese medicine

Abstract

This invention relates to a digital twin and virtual-real synchronous control method for the fluidized bed granulation process of traditional Chinese medicine (TCM) in the field of TCM pharmaceutical technology. The method includes: constructing a digital twin model; deploying multi-source sensors to collect data in real time; updating the model state through data fusion; predicting particle quality through rolling time-domain simulation; solving for optimal commands using nonlinear model predictive control; and issuing commands to form a virtual-real synchronous closed loop. The physical model employs CFD-DEM coupling, simulating particle aggregation, breakage, and solidification through particle identification and a multi-sphere model; correcting the initial NMPC value based on online detection of extract properties through feedforward compensation; and constructing a multi-model library and achieving adaptive model switching based on operating condition identification. This invention achieves a leap from offline pre-simulation to online control of digital twins, solving technical problems such as difficulty in modeling particle growth, large fluctuations in physical properties, and adaptation to diverse varieties in the TCM granulation process, significantly improving the real-time performance, accuracy, and adaptability of the control.

Description

Technical Field

[0001] This invention relates to the field of traditional Chinese medicine manufacturing technology, and more specifically, to a digital twin and virtual-real synchronous control method for the fluidized bed granulation process of traditional Chinese medicine. Background Technology

[0002] Fluidized bed granulation is a key unit operation in the pharmaceutical process of traditional Chinese medicine. It involves mixing extracts of traditional Chinese medicine with excipients in a fluidized bed and spraying in atomized liquid. The particles undergo aggregation, crushing, drying, and solidification under the action of airflow, ultimately forming particles with a specific particle size distribution and mechanical strength. This process directly affects the product quality of subsequent processes such as tableting and capsule filling, as well as the dissolution rate and bioavailability of the final traditional Chinese medicine formulation.

[0003] Fluidized bed granulation is a typical complex gas-solid-liquid multiphase physical process characterized by strong nonlinearity, time-varying nature, and multivariate coupling. Particularly for traditional Chinese medicine systems, the physical properties (such as viscosity, surface tension, and glass transition temperature) of extracts from different varieties of medicinal herbs vary significantly, and even extracts from different batches of the same herb exhibit fluctuations, making it difficult to guarantee the repeatability and consistency of the granulation process. Traditional PID control or experience-based operating modes often struggle to cope with this complexity, resulting in large batch-to-batch quality variations and low yield rates.

[0004] In recent years, digital twin technology has been introduced into the field of process control. Digital twins achieve state monitoring, simulation, and decision optimization by constructing virtual mappings of physical entities. In the field of fluidized bed granulation, some related research has been conducted. For example, Chinese patent application CN117669368A discloses a method and system for fluidized bed simulation of a traditional Chinese medicine drying and forming unit based on digital twins. This method uses a CFD-DEM coupled model to simulate the gas-solid two-phase flow within the fluidized bed and fuses multi-source sensor data through a particle filtering algorithm to achieve accurate simulation of the drying process.

[0005] However, the aforementioned existing technologies have the following main shortcomings: they only remain at the "pre-simulation" stage, that is, predicting future states through simulation, without using the prediction results for real-time control, and failing to form a control closed loop of "perception-prediction-decision-execution". This limits the value of digital twins to offline simulation or auxiliary analysis, making it impossible to optimize the production process in real time, dynamically adjust process parameters, and cope with the material fluctuations and operating condition changes commonly encountered in fluidized bed granulation.

[0006] Furthermore, while existing research has made progress in areas such as accelerated CFD-DEM computation and online learning for fault diagnosis, none of it has addressed how to integrate these technologies with the real-time control of the granulation process to form a complete digital twin control closed loop.

[0007] In summary, how to upgrade digital twins from "pre-rehearsal" to "real-time control" and achieve closed-loop optimization of virtual and real synchronization is a technical problem that urgently needs to be solved in this field.

[0008] In addition, the following technical challenges are faced in realizing digital twin control of fluidized bed granulation:

[0009] The problem of accurate modeling of particle growth process: Existing CFD-DEM models mostly focus on macroscopic simulation of gas-solid two-phase flow, and lack microscopic modeling of particle aggregation, breakage and solidification effects during granulation, resulting in insufficient prediction accuracy.

[0010] Adaptability issues of physical property fluctuations in traditional Chinese medicine extracts: The viscosity, surface tension and other physical property parameters of traditional Chinese medicine extracts fluctuate significantly between batches, which is difficult to effectively control using traditional methods, affecting process stability;

[0011] Model adaptation issues due to the diversity of Chinese herbal medicine varieties: The granulation characteristics of different Chinese herbal medicine varieties vary greatly, making it difficult for a single model to cover all working conditions, and the model update efficiency is low when new working conditions appear.

[0012] The problem of adaptive optimization of threshold parameters: The fixed threshold in particle recognition algorithm cannot adapt to changes in working conditions, which affects the recognition accuracy.

[0013] Coordination issues between feedforward compensation and model switching: If property feedforward and model adaptive switching operate independently, it may lead to control detuning;

[0014] The incremental update problem in online learning: Traditional online learning requires a lot of storage and retraining, which is inefficient. Summary of the Invention

[0015] To address the shortcomings of the existing technologies, the first technical problem this invention aims to solve is: how to deeply integrate digital twin technology into the real-time control of the fluidized bed granulation process of traditional Chinese medicine, and construct a virtual-real synchronous control closed loop from "offline pre-simulation" to "online control," thus overcoming the limitation of existing technologies that only remain at the simulation stage and fail to form a control closed loop.

[0016] Based on this, the present invention further addresses the following technical challenges:

[0017] How to accurately simulate the particle aggregation, breakup and solidification process to improve the prediction accuracy of CFD-DEM models;

[0018] How to effectively address batch-to-batch fluctuations in the physical properties of traditional Chinese medicine extracts and improve the robustness of the control system;

[0019] How to adapt to the granulation characteristics of different Chinese herbal medicine varieties and solve the problem that a single model cannot cover diverse working conditions;

[0020] How to achieve adaptive optimization of threshold parameters in particle recognition algorithms to improve the accuracy of particle recognition;

[0021] How to coordinate feedforward compensation and model adaptive switching to avoid control misalignment that may occur when the two operate independently;

[0022] How to efficiently implement an online learning mechanism to achieve dynamic updates of the working condition recognition model and expansion of multiple model libraries.

[0023] To address the above problems, this invention provides a digital twin and virtual-real synchronous control method for the fluidized bed granulation process of traditional Chinese medicine, comprising the following steps:

[0024] S1: Construct a digital twin model of the fluidized bed granulation process of traditional Chinese medicine. The digital twin model includes a geometric model, a physical model, a behavioral model, and a rule model.

[0025] S2: Deploy a multi-source sensor cluster on the physical fluidized bed granulation equipment to collect process parameters and quality status data in real time;

[0026] S3: Based on multi-source sensor data, the state parameters of the digital twin model are updated in real time through data fusion algorithms, so that the virtual model and the physical device remain dynamically synchronized.

[0027] S4: Use the updated digital twin model to perform rolling time-domain simulation and predict the particle quality indicators at N future times under the current process parameters. The particle quality indicators include particle size distribution and moisture content.

[0028] S5: Based on the prediction results, the optimal control command is solved by nonlinear model predictive control algorithm. The objective function of the nonlinear model predictive control algorithm is to minimize the deviation between the particle size distribution and the target value and the deviation between the moisture content and the target value, and to consider the penalty for changes in the control quantity. The constraints include equipment safety constraints and upper and lower limits of process parameters.

[0029] S6: Sends the optimal control command to the actuator of the physical fluidized bed granulation equipment to realize real-time adjustment of air inlet temperature, liquid spraying rate, atomization pressure and fan frequency, forming a virtual-real synchronous control closed loop.

[0030] The above technical solution, by coupling rolling time-domain simulation with nonlinear model predictive control algorithm, realizes the leap from "offline pre-simulation" to "online control" of digital twins in the field of fluidized bed granulation, and solves the core technical problem that "digital twins remain in the pre-simulation stage and fail to form a control closed loop".

[0031] As a further improvement to the present invention, the construction of the digital twin model in S1 also includes the following sub-steps:

[0032] S11: The physical model uses a CFD-DEM coupling method to simulate the gas-solid two-phase flow and particle growth process in a fluidized bed. The particle growth process is simulated by a particle recognition algorithm and a multi-sphere model to simulate particle aggregation, breakup and solidification effects. The particle recognition algorithm identifies particle aggregates that meet the wettability threshold and contact time threshold as new particles based on particle wettability and contact time threshold. The wettability threshold is when the liquid-solid ratio on the particle surface reaches 0.35 to 0.55, and the contact time threshold is when the continuous contact time between particles exceeds 0.15 to 0.4 seconds.

[0033] S12: Before granulation begins, the physical properties of the herbal extract are detected online, including dynamic viscosity, surface tension, and glass transition temperature. Based on the detected physical properties, the correction amount of the initial process parameters is calculated through a pre-trained feedforward compensation model. The initial process parameters include the initial values ​​of the inlet air temperature, the initial value of the spray rate, and the initial value of the atomization pressure. The corrected initial process parameters are used as the initial values ​​of the nonlinear model predictive control algorithm in S5, forming a composite control based on physical property feedforward and state feedback.

[0034] S13: Construct a multi-model library, which contains multiple CFD-DEM coupled sub-models for different types of Chinese herbal medicines and different granulation stages.

[0035] S11-S13 provide solutions to three technical challenges: particle growth modeling, adaptation to property fluctuations, and variety diversity adaptation.

[0036] As a further improvement of the present invention, after S3 and before S4, a model adaptive switching step is also included. The model adaptive switching step includes: extracting the feature parameters of the current working condition in real time, including the physical property parameters of the extract, particle size distribution, and bed pressure difference fluctuation characteristics; inputting the working condition feature parameters into a pre-trained working condition identification model and outputting the category to which the current working condition belongs; and dynamically selecting the sub-model with the highest matching degree with the working condition from the multi-model library according to the working condition identification result, and using it as the main model in the current control cycle for the simulation and deduction of S4.

[0037] Among them, the sub-models in the multi-model library are built based on transfer learning: using the basic CFD-DEM model as the source model, the sub-models of the target varieties are obtained by fine-tuning the model parameters based on a small amount of experimental data for different Chinese medicine varieties; the working condition recognition model adopts an online learning mechanism, and automatically updates the decision boundary of the recognition model when a new working condition is detected.

[0038] As a further improvement of the present invention, the multi-source sensor cluster in S2 includes: a near-infrared spectral sensor deployed above the fluidized bed material layer for real-time monitoring of particle moisture content changes; an online particle size analyzer connected to the fluidized bed via a sampling pipeline for real-time monitoring of particle size distribution; multi-point temperature sensors deployed at the air inlet, material bed, and exhaust outlet for collecting temperature field distribution data; humidity sensors deployed at the air inlet and exhaust outlet for monitoring inlet and exhaust humidity; a differential pressure sensor for monitoring bed pressure difference and determining fluidization state; and a high-speed camera system for acquiring particle fluidization images through a viewing window.

[0039] As a further improvement of the present invention, the data fusion algorithm in S3 includes: preprocessing and feature extraction of near-infrared spectral data; establishing a correlation model between spectral data and particle moisture content and particle size distribution using partial least squares regression or neural network methods; fusing the particle size data retrieved from the spectrum with the measured data from the online particle size analyzer using Kalman filtering; and using an adaptive covariance adjustment based on model prediction error as the fusion strategy to obtain the fused particle size estimate. Temperature, humidity, and pressure difference data are input as boundary conditions into the digital twin model to drive the update of the model's state parameters.

[0040] As a further improvement of the present invention, the nonlinear model predictive control algorithm in S5 uses the CFD-DEM coupled particle growth model constructed in S1 as the prediction model. In each control cycle, it re-solves the optimization problem based on the current state of the digital twin model, outputs the optimal control command sequence in the next control cycle, and only issues and executes the first control command.

[0041] As a further improvement of the present invention, the nonlinear model predictive control algorithm in S5 also introduces the feedforward correction in S12 as the initial value, and performs prediction based on the master model after adaptive switching in S13, so as to realize the collaborative optimization of feedforward-feedback composite control and multi-model adaptive switching.

[0042] As a further improvement of the present invention, for the adaptive optimization problem of the threshold parameter, the method for determining the wettability threshold in the particle identification algorithm includes: based on the distribution characteristics of the liquid-solid ratio on the particle surface in historical batch data, the Gaussian mixture model clustering algorithm is used to divide the particle state into three categories: "unwetted", "moderately wetted" and "overly wetted"; with the goal of maximizing the inter-class variance of the clustering results, the wettability threshold of the current batch is dynamically optimized so that the threshold is adaptive to the fluctuation of extract properties and changes in the granulation stage.

[0043] As a further improvement of the present invention, in response to the coordination problem of feedforward compensation and model switching, the feedforward compensation model and the model adaptive switching step in S13 are coordinated and linked: when the working condition identification model detects the switching of Chinese medicine varieties or changes in the granulation stage, the parameter update of the feedforward compensation model is triggered synchronously; the feedforward compensation model adopts the Bayesian online learning algorithm, based on the physical property detection data under the new working condition and the historical optimal process parameters, and updates the model weight in real time, so that the feedforward correction amount is dynamically matched with the main model after adaptive switching.

[0044] As a further improvement of the present invention, for the incremental update problem of online learning, the online learning mechanism of the working condition recognition model includes: constructing an incremental support vector machine classifier; when a new working condition is detected and the accumulated sample size reaches a preset threshold, using an incremental learning algorithm based on kernel function approximation to update the decision boundary; and simultaneously, incorporating the optimal sub-model corresponding to the new working condition into a multi-model library to achieve dynamic expansion of the model library.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] 1. This invention couples rolling time-domain simulation with a nonlinear model predictive control algorithm and issues optimized control commands for real-time execution, achieving a leap from "offline pre-simulation" to "online control" in the field of fluidized bed granulation. Compared to technical solutions that only remain at the simulation stage, this invention forms a complete "sensing-prediction-decision-execution" control closed loop, which can dynamically optimize process parameters based on prediction results, significantly improving the real-time performance and accuracy of process control.

[0047] 2. This invention achieves accurate simulation of particle aggregation, crushing and solidification processes through particle recognition algorithms and multi-sphere models, overcoming the limitations of existing technologies that can only simulate drying processes or general two-phase flows.

[0048] 3. This invention effectively solves the problem of batch-to-batch fluctuations in the physical properties of traditional Chinese medicine extracts affecting process stability through a feedforward compensation mechanism.

[0049] 4. This invention solves the modeling challenge of diverse Chinese medicinal herbs by constructing a multi-model library and combining it with a working condition recognition model to achieve adaptive switching. Through transfer learning, sub-models for new varieties can be quickly constructed using a small amount of experimental data.

[0050] 5. This invention uses a dynamic threshold optimization method to enable the threshold of the particle identification algorithm to adapt to fluctuations in extract properties and changes during the granulation stage.

[0051] 6. This invention, through a collaborative linkage mechanism, enables the material property feedforward compensation and model adaptive switching to work together organically, avoiding the control detuning problem that may occur when the two operate independently, and further improving the smoothness and stability of the control system during variety switching and stage changes.

[0052] 7. This invention achieves efficient updating of the working condition identification model and dynamic expansion of the multi-model library through an incremental online learning mechanism, enabling the entire control system to continuously learn and evolve during operation and cope with the emergence of unknown new working conditions. Attached Figure Description

[0053] Figure 1 This is a flowchart of the overall method of the present invention. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of protection of this invention.

[0055] Example 1:

[0056] Figure 1 This paper presents a digital twin and virtual-real synchronous control method for a fluidized bed granulation process of traditional Chinese medicine. The core of this method is to construct a digital twin model that maps in real time to the physical granulation system. This involves acquiring the physical system state through multi-source sensors, fusing data using Kalman filtering, predicting particle growth status using a CFD-DEM coupled model, and combining this with an NMPC controller to achieve real-time optimization and adjustment of process parameters. Furthermore, it adapts to the characteristics of multi-variety, small-batch production of traditional Chinese medicine through operating condition identification, multi-model switching, and incremental learning. The specific steps include:

[0057] S1: Construct a digital twin model of the fluidized bed granulation process of traditional Chinese medicine. The digital twin model includes a geometric model, a physical model, a behavioral model, and a rule model.

[0058] S2: Deploy a multi-source sensor cluster on the physical fluidized bed granulation equipment to collect process parameters and quality status data in real time;

[0059] S3: Based on multi-source sensor data, the state parameters of the digital twin model are updated in real time through data fusion algorithms, so that the virtual model and the physical device remain dynamically synchronized.

[0060] S4: Use the updated digital twin model to perform rolling time-domain simulation and predict the particle quality indicators at N future times under the current process parameters. The particle quality indicators include particle size distribution and moisture content.

[0061] S5: Based on the prediction results, the optimal control command is solved by nonlinear model predictive control algorithm. The objective function of the nonlinear model predictive control algorithm is to minimize the deviation between the particle size distribution and the target value and the deviation between the moisture content and the target value, and to consider the penalty for changes in the control quantity. The constraints include equipment safety constraints and upper and lower limits of process parameters.

[0062] S6: Sends the optimal control command to the actuator of the physical fluidized bed granulation equipment to realize real-time adjustment of air inlet temperature, liquid spraying rate, atomization pressure and fan frequency, forming a virtual-real synchronous control closed loop.

[0063] 1. Fluidized bed granulation system configuration

[0064] This embodiment was implemented on a fluidized bed granulation production line of a traditional Chinese medicine pharmaceutical company. The main equipment and parameters of the production line are as follows:

[0065] Fluidized bed main unit: 200L volume, processing capacity 30-50kg / batch;

[0066] Liquid supply system: peristaltic pump + dual-fluid atomizing nozzle, liquid spray rate adjustable from 0-500 mL / min;

[0067] Air supply system: variable frequency fan, air volume 0-2000 m³ / h, heating power 30kW;

[0068] Actuators: Inlet air temperature regulating valve, liquid injection pump frequency converter, atomizing pressure regulating valve, fan frequency converter, all of which support 4-20mA analog quantity control.

[0069] 2. Deployment of multi-source sensor clusters

[0070] Deploy the following sensors, with specific configurations shown in Table 1.

[0071] Table 1 Multi-source sensor cluster configuration

[0072] Sensor type Deployment location sampling frequency Monitoring parameters Near-infrared spectral sensor The viewing window is 50mm in diameter and located 30cm above the material bed. 1 Hz Particle moisture content and particle size distribution characteristics Online particle size analyzer The system is connected to the fluidized bed via an automatic sampling pipeline, which automatically samples every 30 seconds. 0.033 Hz Particle size distribution (D10, D50, D90) Temperature sensor (air intake) Inside the air inlet duct 1 Hz Inlet air temperature (°C) Temperature sensor (bed) material bed middle 1 Hz Material temperature (°C) Temperature sensor (exhaust air) Inside the exhaust duct 1 Hz Exhaust air temperature (°C) Humidity sensor (air intake) Inside the air inlet duct 1 Hz Relative humidity of incoming air (%) Humidity sensor (exhaust fan) Inside the exhaust duct 1 Hz Relative humidity of exhaust air (%) differential pressure sensor Pressure taps at the top and bottom of the bed 10 Hz Bed pressure difference (Pa) High-speed camera system Captured via viewport, resolution 1024×1024, frame rate 500fps. Continuous acquisition, saving one image every 1 second. Particle fluidization image

[0073] All sensor data is acquired by the PLC and transmitted in real time to the industrial control computer (IPC) for processing.

[0074] 3. Digital Twin Model Construction

[0075] 3.1 Geometric Model

[0076] Based on the actual dimensions of the fluidized bed equipment, a 3D geometric model was created using SolidWorks. The model includes: an air inlet distribution plate (600mm diameter, 8% opening ratio), a material bed region (800mm height, 30° cone angle), an expansion section (500mm height), and an exhaust vent (150mm diameter). ANSYS ICEM was used for mesh generation, producing approximately 520,000 hexahedral meshes, with 5 boundary layer layers and finer meshing near the walls. .

[0077] 3.2 Physical Model

[0078] Gas phase model:

[0079] Governing equations: The Euler method is used to solve the continuity equation and the Navier-Stokes equation;

[0080] Continuity equation: ;

[0081] Momentum equation: ;

[0082] Turbulence model: adopted Model, standard wall functions;

[0083] Solver: ANSYS Fluent, pressure-velocity coupling uses SIMPLE algorithm, time step 0.001s.

[0084] Solid-phase model:

[0085] The discrete element method (DEM) was used to track the motion trajectory of each particle. The initial number of particles was 80,000, which changed dynamically during the granulation process.

[0086] Translational motion: ;

[0087] Rotational motion: ;

[0088] Contact force model: Hertz-Mindlin no-slip model was adopted;

[0089] Solver: EDEM, bidirectionally coupled with Fluent, exchanging data every 0.001s.

[0090] Gas-solid coupling:

[0091] Momentum exchange source term: ;

[0092] Coupling frequency: Data exchange occurs once every 10 fluid time steps (0.01s).

[0093] 3.3 Particle growth model (S11)

[0094] Moisture level calculation:

[0095] The liquid-to-solid ratio (LSR) of a particle is defined as the ratio of the volume of the attached liquid to the volume of the particle solid.

[0096] ;

[0097] in, The drying rate is calculated by tracking the amount of liquid adhering to the particle surface, initially zero, increasing with each encounter with droplets, and decreasing as the particle dries. The drying rate is calculated based on particle temperature and surrounding gas humidity.

[0098] Contact time calculation:

[0099] The contact history between particles is traced using a DEM (Digital Emission Model), recording the continuous contact time for each pair of particles. When two particles separate, the contact time is reset to zero; the timing is restarted when they re-engage.

[0100] Particle recognition algorithm:

[0101] At each time step, all particle pairs are scanned, and a particle pair is marked as a coalescence candidate when the following condition is met:

[0102] and (All particles reached the lower limit of the wetting threshold).

[0103] Contact time ;

[0104] In this embodiment, the lower limit of the wetting threshold is taken. upper limit Contact time threshold As initial values.

[0105] When a group of particles (possibly more than 2) are interconnected and all meet the conditions, they are identified as a new particle aggregate.

[0106] Multi-sphere model:

[0107] The identified particle aggregates are represented by multiple spheres:

[0108] Calculate the equivalent diameter of the aggregate: ,in The total volume of the aggregate;

[0109] Determine the number and location of the main sphere and the auxiliary spheres based on the shape of the aggregate;

[0110] The diameter of the main sphere is taken as 0.6-0.8 times the size of the attached sphere, with the diameter taken as... 0.3-0.5 times;

[0111] The spheres are connected by virtual bonds, which have elastic modulus and fracture strength.

[0112] Curing effect:

[0113] When the liquid on the particle surface dries ( When the virtual bond strength increases by 5 times, it simulates a solidification effect. The solidified aggregate may still break when subjected to a sufficiently large impact force.

[0114] Fracture effect:

[0115] When the particle collision energy exceeds the virtual bond strength, the aggregate breaks into multiple sub-particles. The broken sub-particles inherit some properties of the original particles and participate in the DEM calculation again.

[0116] To verify the rationality of the wetting threshold of 0.35-0.55 and the contact time threshold of 0.15-0.4 seconds, a comparative experiment was conducted. For the wetting threshold, with a fixed contact time threshold of 0.15 seconds, tests were conducted at lower limits of 0.30, 0.35, and 0.40 seconds and upper limits of 0.50, 0.55, and 0.60 seconds. The results showed that when the lower limit was below 0.35 (e.g., 0.30), the particle recognition accuracy decreased to 78.3% (insultingly wetted particles being mistakenly identified as aggregates, leading to false positives); when the upper limit was above 0.55 (e.g., 0.60), the accuracy decreased to 79.6% (over-wetted particles were missed, leading to false negatives). For the contact time threshold, with a fixed wetting threshold of 0.35-0.55 seconds, tests were conducted at 0.10, 0.15, and 0.20 seconds (lower limit) and 0.35, 0.40, and 0.45 seconds (upper limit). The results showed that when the lower limit was below 0.15 seconds (e.g., 0.10 seconds), the accuracy dropped to 81.2% (false positives were made after brief contact); when the upper limit was above 0.4 seconds (e.g., 0.45 seconds), the accuracy dropped to 80.5% (some particles that should have coalesced were missed due to insufficient contact time). Therefore, 0.35-0.55 seconds and 0.15-0.4 seconds are the optimal ranges; exceeding these ranges will lead to a significant decrease in recognition accuracy.

[0117] 3.4 Behavioral Model

[0118] Establish a model relating particle wettability to agglomeration probability:

[0119] ;

[0120] in , The result was obtained by fitting experimental data.

[0121] 3.5 Rule Model

[0122] This includes equipment safety rules and process constraints:

[0123] Upper limit of bed temperature: 80℃ (if exceeded, an alarm will be triggered and the inlet air temperature will be reduced);

[0124] Upper limit of bed pressure difference: 2000Pa (exceeding this limit will trigger a bed collapse warning);

[0125] Maximum spray rate: 400 mL / min (exceeding this will reduce the atomization effect);

[0126] 4. Data fusion and status update

[0127] 4.1 Near-infrared spectral data processing

[0128] Preprocessing of the acquired near-infrared spectra:

[0129] The effect of particle size was eliminated by standard normal variable transformation (SNV);

[0130] Baseline drift was eliminated using the Savitzky-Golay first derivative (11-point window, second-order polynomial);

[0131] Select characteristic bands: moisture content related bands 1450nm and 1940nm, particle size related bands 1100-1400nm;

[0132] Establish a partial least squares regression (PLS) model:

[0133] Input: Preprocessed spectral data (100 wavelength points);

[0134] Output: Particle moisture content (%), median particle size D50 (μm);

[0135] Training set: 80 batches of historical data, with 50 samples in each batch;

[0136] Validation set: 20 batches of data, RMSECV=0.32% (moisture content), RMSECV=8.5μm (D50).

[0137] 4.2 Kalman Filter Fusion

[0138] State vector: ;

[0139] Observation vector: (Near-infrared inversion D50, online particle size analyzer measured D50);

[0140] System Model:

[0141] State transition matrix: (Identity matrix, assuming slowly changing states);

[0142] Control input matrix: ;

[0143] Observation matrix: (Only D50 was observed; moisture content was obtained directly via near-infrared spectroscopy);

[0144] Adaptive covariance adjustment:

[0145] Observation noise covariance matrix Dynamically adjust based on model prediction error:

[0146] ;

[0147] in:

[0148] (Basic covariance, in μm²);

[0149] (Adaptive coefficients);

[0150] (Prediction error);

[0151] Kalman filter update steps:

[0152] predict: ;

[0153] Predicting covariance: ( For the process noise covariance, take );

[0154] Kalman gain: ;

[0155] renew: ;

[0156] Update covariance: .

[0157] 4.3 Boundary Condition Input

[0158] Input temperature, humidity, and pressure difference data into the digital twin model in real time:

[0159] Inlet air temperature and humidity are used as gas phase inlet boundary conditions;

[0160] Bed pressure difference is used to verify the accuracy of the model. When the deviation between the model-predicted pressure difference and the measured pressure difference exceeds 10%, model parameter correction is triggered.

[0161] 5. NMPC Controller Design

[0162] 5.1 Prediction Model

[0163] The aforementioned CFD-DEM coupled particle growth model was used as the prediction model for NMPC. To improve computational efficiency, the intrinsic orthogonal decomposition (POD) method was employed to reduce the order of the CFD-DEM model.

[0164] 50 sets of CFD-DEM simulation results under typical working conditions were collected as snapshots;

[0165] POD decomposition was performed, and the first 20 modes (99.2% of the energy) were retained.

[0166] Establish a radial basis function (RBF) neural network to predict mode coefficients;

[0167] After the model was downgraded, the computation time was reduced from 30 minutes to 3 seconds, meeting the requirements for real-time control.

[0168] 5.2 Control Variables and State Variables

[0169] Control variables: ;

[0170] Inlet air temperature (°C), range 50-90;

[0171] Spray rate (mL / min), range 100-400;

[0172] Atomization pressure (MPa), range 0.1-0.4;

[0173] Fan frequency (Hz), range 20-50;

[0174] State variables: ;

[0175] D10, D50, D90: Particle size distribution characteristic values ​​(μm);

[0176] moisture: particle moisture content (%)

[0177] 5.3 Objective Function

[0178] ;

[0179] in:

[0180] Prediction Time Domain (Corresponding to 100s, each control cycle is 10s);

[0181] Control Time Domain (Optimize the control amount in the first 5 steps, and maintain it in the last 5 steps).

[0182] (Target median particle size);

[0183] (Target moisture content);

[0184] , (Weighting coefficients);

[0185] (Penalty coefficient for changes in control quantity);

[0186] .

[0187] 5.4 Constraints

[0188] Equipment safety constraints:

[0189] bed temperature ;

[0190] Bed pressure difference ;

[0191] Upper and lower limits of process parameters:

[0192] ;

[0193] ;

[0194] ;

[0195] ;

[0196] Control quantity change rate constraint:

[0197] ;

[0198] ;

[0199] ;

[0200] .

[0201] 5.5 Optimization Solution

[0202] The above optimization problem is solved using the Sequential Quadratic Programming (SQP) algorithm:

[0203] Initial solution: Using the initial values ​​corrected by S12;

[0204] Iteration termination condition: change in objective function < 0.1% or number of iterations > 100;

[0205] The optimization problem is solved once per control cycle (10s);

[0206] Output the optimal control command sequence for the next control cycle. ;

[0207] Only the first instruction Issued and implemented.

[0208] 6. Virtual-real synchronous control closed loop

[0209] The optimal control command obtained from NMPC solution Distributed to all implementing agencies:

[0210] Inlet air temperature regulating valve: Receives 4-20mA signal to adjust the opening degree;

[0211] Injection pump frequency converter: Receives 0-10V signal to adjust pump speed;

[0212] Atomizing pressure regulating valve: Receives a 4-20mA signal to adjust the opening degree;

[0213] Fan frequency converter: Receives 0-10V signals and adjusts the fan frequency.

[0214] The entire process cycles every 10 seconds.

[0215] Collect sensor data (S2);

[0216] Data fusion updates the model state (S3);

[0217] Model adaptive switching;

[0218] Rolling time-domain simulation derivation (S4);

[0219] NMPC solves for the optimal control instruction (S5);

[0220] Instruction issuance and execution (S6);

[0221] This forms a closed-loop control system that synchronizes virtual and real systems, encompassing "perception-prediction-decision-execution".

[0222] 7. Composite control of material property feedforward (S12)

[0223] 7.1 Online detection of physical property parameters

[0224] Before each batch of granulation begins, a sample of the herbal extract is taken from the material tank and subjected to the following tests:

[0225] Dynamic viscosity: Measured using a rotational viscometer at 60°C, rotor S61, rotation speed 20 rpm.

[0226] Surface tension: Measured using a pendant drop surface tension meter at 60℃.

[0227] Glass transition temperature: determined using a differential scanning calorimeter at a heating rate of 10℃ / min under a nitrogen atmosphere.

[0228] 7.2 Feedforward Compensation Model

[0229] Establish a 3-layer BP neural network as the feedforward compensation model:

[0230] Input layer: 3 nodes (viscosity) Surface tension Glass transition temperature );

[0231] Hidden layer: 6 nodes, activation function tanh;

[0232] Output layer: 3 nodes (inlet air temperature correction) Spray rate correction amount Atomization pressure correction amount );

[0233] Training data: Collect 100 batches of historical data, each batch containing:

[0234] Physical property parameters: ;

[0235] Optimal process parameters: obtained through offline optimization ;

[0236] Correction amount: ,in This is an empirical benchmark value;

[0237] Training parameters:

[0238] Loss function: MSE;

[0239] Optimized algorithm: Adam, learning rate 0.001;

[0240] Training rounds: 500;

[0241] Validation set: 20 batches of data, R²=0.85;

[0242] 7.3 Initial Value Correction

[0243] Add the correction value output by the feedforward compensation model to the basic process parameters:

[0244] ;

[0245] ;

[0246] ;

[0247] in , , .

[0248] The corrected initial value is used as the initial solution for NMPC optimization, so that the optimization starts from a position closer to the optimum, which speeds up the convergence speed and avoids getting trapped in local optima.

[0249] 8. Construction of a multi-model library (S13)

[0250] 8.1 Initial Model Library

[0251] Construct a multi-model library containing 5 sub-models, each corresponding to one of the 5 commonly used Chinese medicinal herbs:

[0252] Model Number Chinese medicine varieties Particle density (g / cm³) Collision recovery coefficient rolling friction coefficient Adhesion energy (J / m²) M1 Coptis chinensis extract 1.20 0.30 0.15 0.05 M2 Honeysuckle extract 1.10 0.35 0.10 0.03 M3 Astragalus extract 1.25 0.25 0.20 0.06 M4 Salvia miltiorrhiza extract 1.15 0.30 0.12 0.04 M5 Licorice extract 1.18 0.28 0.18 0.05

[0253] Each sub-model is based on the basic CFD-DEM model, obtained by adjusting the above parameters, and verified by 10 batches of experimental data, with a particle size distribution prediction error of <10%.

[0254] 8.2 Model Adaptive Switching

[0255] Extraction of operating condition characteristic parameters:

[0256] At the beginning of each control cycle, the following feature parameters are extracted:

[0257] Extract physical properties: viscosity Surface tension Glass transition temperature (Testing begins at the start of the batch; results remain unchanged within the batch)

[0258] Particle size distribution: D10, D50, D90 (measured in real time)

[0259] Bed pressure differential fluctuation characteristics: mean ,variance The main peak frequency of the power spectral density in the range of 0-5Hz ;

[0260] Operating condition identification model:

[0261] A random forest classifier is used, with the following parameters:

[0262] Number of decision trees: 100;

[0263] Maximum depth: 10;

[0264] Minimum number of leaf samples: 5;

[0265] Input features: the above 8-dimensional feature vector;

[0266] Output: 5 categories (corresponding to 5 types of traditional Chinese medicine);

[0267] Training data: 200 batches of historical data, each batch was manually labeled by process experts according to variety and granulation stage.

[0268] Adaptive handover process:

[0269] Each control cycle :

[0270] Extracting operating condition feature vectors ;

[0271] Input a random forest model to obtain the class probability distribution. ;

[0272] Choose the category with the highest probability. ;

[0273] Select the corresponding sub-model from the multi-model library. ;

[0274] Will It serves as the master model for the current cycle and is used for simulation and deduction of S4.

[0275] 9. Collaborative Optimization

[0276] The feedforward correction in S12 is coordinated with the adaptively switched master model in S13:

[0277] Feedforward correction As initial values ​​for NMPC optimization solution;

[0278] The adaptively switched master model serves as the prediction model for NMPC.

[0279] Both participate in the calculation and optimization of the objective function;

[0280] Experiments show that this collaborative optimization improves the control accuracy of particle size distribution by 15% and reduces the settling time by 20% compared to using feedforward or adaptive switching alone.

[0281] Example 2:

[0282] Based on Example 1, this embodiment further optimizes the wettability threshold in the particle recognition algorithm, enabling it to adapt to fluctuations in extract properties and changes during the granulation stage.

[0283] 1. Problem Background

[0284] In Example 1, a fixed threshold was used (wetting degree 0.35-0.55, contact time 0.15-0.4 seconds). However, in actual operation, it was found that the optimal thresholds differed for different varieties and different granulation stages.

[0285] High-viscosity extracts (such as Astragalus membranaceus) require a high wetting threshold to identify aggregation;

[0286] In the initial stage of granulation, the particles are relatively small, requiring a shorter contact time threshold.

[0287] In the later stages of granulation, the particles are larger, so it is necessary to appropriately increase the wetting threshold to avoid over-identification.

[0288] The fixed threshold resulted in an average particle recognition accuracy of only 82.3%, which affected the model's prediction accuracy.

[0289] 2. Gaussian Mixture Model Clustering

[0290] 2.1 Data Acquisition

[0291] At the end of each control cycle, the liquid-to-solid ratio data of all particles within that cycle are collected to form a dataset. , The number of particles recorded during this period (approximately) (magnitude).

[0292] 2.2 GMM Model

[0293] Assume the liquid-to-solid ratio data follows a mixture of three Gaussian distributions:

[0294] ;

[0295] in:

[0296] : Mixing coefficient, satisfying ;

[0297] : No. The mean of the Gaussian components;

[0298] : No. The variance of each Gaussian component.

[0299] The three components correspond to:

[0300] Component 1: Unwetted particles (low liquid-to-solid ratio, minimum $\mu_1$);

[0301] Component 2: Moderately moistened particles (with a suitable liquid-to-solid ratio, ideal for agglomeration);

[0302] Component 3: Overly wetted particles (high liquid-to-solid ratio, which can easily lead to collapse).

[0303] 2.3 EM Algorithm Solution

[0304] The GMM parameters are estimated using the Expectation-Maximization (EM) algorithm:

[0305] Step E: Calculate the... The sample belongs to the first Posterior probability (responsibility) of each component

[0306] ;

[0307] M-step: Update parameters

[0308] ;

[0309] ;

[0310] ;

[0311] Iterate through E-steps and M-steps until the parameter change is less than a threshold (e.g., ...). ).

[0312] 2.4 Classification

[0313] After convergence, the means of the three components are sorted to obtain... ,correspond:

[0314] Component 1: Unwetted granules;

[0315] Component 2: Moderately moisten the granules;

[0316] Component 3: Overly moistened particles.

[0317] 3. Threshold optimization

[0318] 3.1 Calculation of inter-class variance

[0319] Calculate the inter-class variance to measure the degree of separation among the three classes:

[0320] ;

[0321] .

[0322] 3.2 Threshold Determination

[0323] Determine the humidity threshold based on clustering results:

[0324] Lower limit of wetting threshold ;

[0325] Upper limit of the wetting threshold ;

[0326] To ensure the threshold is within a reasonable range, the calculation results are limited:

[0327] The lower limit is set between 0.30 and 0.40 (below 0.30 may lead to over-identification, and above 0.40 may lead to under-identification).

[0328] The upper limit is limited to between 0.50 and 0.60;

[0329] A similar method was used for the contact time threshold, with GMM clustering and optimization based on the distribution of particle contact time.

[0330] 3.3 Dynamic Update Process

[0331] After each batch ends:

[0332] Collect liquid-to-solid ratio data for all particles in this batch;

[0333] Run the EM algorithm to solve for the GMM parameters;

[0334] calculate , , ;

[0335] Update the lower and upper limits of the humidity threshold;

[0336] Use the new threshold for particle identification in the next batch.

[0337] 4. Comparison of experimental results

[0338] In experiments involving 10 batches of each of 5 different types of traditional Chinese medicine, the recognition performance of fixed thresholds and dynamic thresholds was compared:

[0339] variety Fixed threshold accuracy Dynamic threshold accuracy promote Coptis chinensis 81.5% 95.8% +14.3% honeysuckle 84.2% 97.3% +13.1% Astragalus 79.8% 94.5% +14.7% Salvia miltiorrhiza 83.6% 96.2% +12.6% licorice 82.4% 96.1% +13.7% average 82.3% 96.0% +13.7%

[0340] Dynamic thresholds improved particle recognition accuracy by an average of 13.7 percentage points, significantly enhancing the model's prediction accuracy.

[0341] Example 3:

[0342] This embodiment further optimizes the collaborative linkage mechanism between the feedforward compensation model and the adaptive switching of the model, based on Embodiment 1.

[0343] 1. Problem Background

[0344] In Example 1, the feedforward compensation in S12 and the model adaptive switching in S13 operate independently. When the operating condition identification model detects a variety change, the main model switches to the sub-model of the new variety, but the feedforward compensation model still uses the old parameters, resulting in:

[0345] The feedforward correction is mismatched with the master model after the switch.

[0346] The first batch pass rate after the product variety switch was only 68.5%.

[0347] Eight batches of adjustments are needed to reach steady state.

[0348] 2. Collaborative Mechanism

[0349] 2.1 Triggering Conditions

[0350] When the operating condition category output by the operating condition recognition model changes (i.e., a switch in Chinese medicine varieties or a significant change in the granulation stage is detected), a collaborative linkage is triggered.

[0351] 2.2 Feedforward Compensation Model Parameter Update

[0352] The weights of the feedforward compensation model are updated using a Bayesian online learning algorithm.

[0353] 2.3 Bayesian Linear Regression Model

[0354] The feedforward compensation model is simplified to a linear model (for easier Bayesian updating):

[0355] ;

[0356] in:

[0357] (6-dimensional features, including interaction terms);

[0358] Weight vector;

[0359] Gaussian noise.

[0360] Prior distribution:

[0361] Assume the weights follow a Gaussian prior:

[0362] ;

[0363] in Take the weights obtained from training in Example 1. (Indicates the confidence level regarding the prior).

[0364] Likelihood function: for the new operating condition Observation samples The likelihood function is:

[0365] ;

[0366] Posterior distribution:

[0367] According to Bayes' theorem, the posterior distribution is:

[0368] ;

[0369] in:

[0370] ;

[0371] ;

[0372] Design matrix, number Behavior ;

[0373] : Observation vector.

[0374] 2.4 Update Timing

[0375] Immediate Update: When a variety change is detected, immediately perform a Bayesian update using the small number of samples (at least 3 batches) already accumulated under the new operating conditions;

[0376] Regular updates: Under new operating conditions, perform a Bayesian update every 5 batches of accumulated data;

[0377] Convergence criterion: When the posterior covariance When the trace is less than a threshold (e.g., 0.01), the model is considered to have converged and updates are stopped.

[0378] 3. Comparison of experimental results

[0379] In 10 experiments switching from variety A to variety B, the control effects of independent operation and coordinated operation were compared:

[0380] index Independent operation Collaboration and Linkage promote First batch pass rate after switch 68.5% 92.3% +23.8% Number of batches required to reach steady state 8 batches 2 batches -75% Steady-state particle size RSD 5.8% 3.2% -44.8%

[0381] The collaborative linkage mechanism enables the feedforward correction amount to be dynamically matched with the master model after the switch, which significantly improves the smoothness and stability of the control system when switching product types.

[0382] Example 4:

[0383] Based on Example 3, this embodiment further optimizes the online learning mechanism of the working condition identification model to achieve efficient incremental updates and dynamic expansion of the model library.

[0384] 1. Problem Background

[0385] In Example 1, the working condition identification model uses a random forest classifier. When a new working condition occurs:

[0386] The entire model needs to be retrained after collecting enough data.

[0387] Retraining takes 2 hours (30 batches of data);

[0388] It consumes a lot of computing resources and cannot respond to new operating conditions in real time.

[0389] 2. Incremental Support Vector Machine

[0390] 2.1 Basic SVM Model

[0391] The C-SVM classifier is used, with the RBF kernel function:

[0392] ;

[0393] Optimization issues:

[0394] ;

[0395] ;

[0396] The initial model was trained using 200 batches of historical data. , .

[0397] 2.2 Incremental Learning Algorithm

[0398] When new samples Upon arrival, the following incremental learning algorithm is used:

[0399] Step 1: Calculate the decision function value of the new sample under the current model.

[0400] ;

[0401] Step 2: Check KKT conditions

[0402] if and If the new sample satisfies the KKT conditions, it is discarded without changing the model.

[0403] Otherwise, the new sample violates the KKT conditions and needs to be added to the model.

[0404] Step 3: Add new samples

[0405] initialization ;

[0406] Add new samples to the training set;

[0407] The Sequence Minimum Optimization (SMO) algorithm is used to update all $\alpha_i$ until the KKT conditions are met;

[0408] Step 4: Model Compression

[0409] When the number of support vectors exceeds a threshold (e.g., 500), model compression is performed.

[0410] All support vectors are retained;

[0411] Prune non-support vectors;

[0412] The representative point selection technique is used to approximate the original distribution with a small number of points.

[0413] 2.3 Kernel Function Approximation

[0414] To reduce the computational complexity of incremental learning, the Stochastic Fourier Feature (RFF) method is used to approximate the RBF kernel:

[0415] ;

[0416] in It is a random Fourier eigenvector:

[0417] ;

[0418] (Fourier transform of the corresponding RBF kernel); ;

[0419] We selected 500 features (tests showed that 500-dimensional features can achieve 95% of the accuracy of the original kernel function).

[0420] After adopting RFF, the SVM optimization problem is transformed into a linear problem, and the incremental learning complexity is reduced from... Down to .

[0421] 3. Dynamic expansion of the model library

[0422] 3.1 New Operating Condition Testing

[0423] When the incremental SVM identifies a new sample whose class confidence is below a threshold (e.g., 0.6), it is marked as a potential new working condition.

[0424] When the cumulative sample size of potential new operating conditions reaches a threshold (e.g., 10 batches), the emergence of a new operating condition is confirmed.

[0425] 3.2 Construction of New Sub-model

[0426] Based on samples from the new working conditions, a new sub-model is constructed through transfer learning:

[0427] Source model: Existing sub-models of the most similar varieties;

[0428] Fine-tuning method: Freeze the bottom-level parameters of the CFD-DEM model (such as geometry and mesh), and only fine-tune the top-level parameters (such as particle density and collision recovery coefficient).

[0429] Fine-tuning data: 10 batches of experimental data under the new operating conditions;

[0430] Fine-tuning algorithm: Bayesian optimization, with the objective function being to minimize the prediction error.

[0431] 3.3 Model Library Update

[0432] Add the newly constructed sub-model to the multi-model library;

[0433] Update the number of categories in the working condition recognition model, from Classes added to kind;

[0434] Generate representative samples for the new category (based on new working condition data) and add them to the training set of the incremental SVM;

[0435] Continue incremental learning to adapt the model to new categories.

[0436] 4. Comparison of experimental results

[0437] Comparing the traditional retraining strategy with the incremental learning strategy of this embodiment:

[0438] index Retraining strategy Incremental learning strategy promote New operating condition response time 2 hours (30 batches of data need to be collected) 5 minutes (updated in real time) -96% Computational resource consumption CPU 8 cores x 2 hours CPU 1 core x 5 minutes -99% New working condition recognition accuracy 94.5% 93.8% -0.7% (basically unchanged) Model library expansion capabilities Human intervention is required. Automatic expansion Significant improvement

[0439] The incremental learning strategy reduces the response time to new operating conditions from 2 hours to 5 minutes with almost no loss of accuracy. It enables real-time updates of the operating condition identification model and dynamic expansion of the multi-model library, giving the entire control system the ability to continuously learn and evolve during operation.

[0440] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Various changes made within the scope of knowledge possessed by those skilled in the art without departing from the concept of the present invention still fall within the scope of protection of the present invention.

Claims

1. A digital twin and virtual-real synchronous control method for a fluidized bed granulation process of traditional Chinese medicine, characterized in that, Includes the following steps: S1: Construct a digital twin model of the fluidized bed granulation process of traditional Chinese medicine, wherein the digital twin model includes a geometric model, a physical model, a behavioral model and a rule model; S2: Deploy a multi-source sensor cluster on the physical fluidized bed granulation equipment to collect process parameters and quality status data in real time; S3: Based on the multi-source sensor data, the state parameters of the digital twin model are updated in real time through a data fusion algorithm, so that the virtual model and the physical device remain dynamically synchronized. S4: Use the updated digital twin model to perform rolling time-domain simulation and predict the particle quality indicators at N future times under the current process parameters. The particle quality indicators include particle size distribution and moisture content. S5: Based on the prediction results, the optimal control command is solved by a nonlinear model predictive control algorithm. The objective function of the nonlinear model predictive control algorithm is to minimize the deviation between the particle size distribution and the target value and the deviation between the moisture content and the target value, and to consider the penalty for changes in the control quantity. The constraints include equipment safety constraints and upper and lower limits of process parameters. S6: The optimal control command is sent to the actuator of the physical fluidized bed granulation equipment to realize real-time adjustment of the air inlet temperature, liquid spraying rate, atomization pressure and fan frequency, forming a virtual-real synchronous control closed loop.

2. The digital twin and virtual-real synchronous control method for the fluidized bed granulation process of traditional Chinese medicine according to claim 1, characterized in that, The construction of the digital twin model in S1 also includes the following sub-steps: S11: The physical model uses a CFD-DEM coupling method to simulate the gas-solid two-phase flow and particle growth process in a fluidized bed. The particle growth process is simulated by a particle recognition algorithm and a multi-sphere model to simulate particle aggregation, breakup, and solidification effects. The particle recognition algorithm identifies particle aggregates that meet the wettability threshold and contact time threshold as new particles based on particle wettability and contact time threshold. The wettability threshold is a liquid-to-solid ratio of 0.35 to 0.55 on the particle surface, and the contact time threshold is a continuous contact time between particles exceeding 0.15 to 0.4 seconds. S12: Before granulation begins, the physical properties of the herbal extract are detected online, including dynamic viscosity, surface tension, and glass transition temperature. Based on the detected physical properties, the correction amount of the initial process parameters is calculated through a pre-trained feedforward compensation model. The initial process parameters include the initial values ​​of the inlet air temperature, the initial value of the spray rate, and the initial value of the atomization pressure. The corrected initial process parameters are used as the initial values ​​of the nonlinear model predictive control algorithm in S5, forming a composite control based on physical property feedforward and state feedback. S13: Construct a multi-model library, which contains multiple CFD-DEM coupled sub-models for different types of Chinese herbal medicines and different granulation stages.

3. The digital twin and virtual-real synchronous control method for the fluidized bed granulation process of traditional Chinese medicine according to claim 2, characterized in that, After S3 and before S4, there is also a model adaptive switching step, which includes: extracting the feature parameters of the current working condition in real time, including the physical properties of the extract, particle size distribution, and bed pressure difference fluctuation characteristics; inputting the working condition feature parameters into a pre-trained working condition identification model and outputting the category to which the current working condition belongs; and dynamically selecting the sub-model with the highest matching degree with the working condition from a multi-model library based on the working condition identification result, and using it as the main model in the current control cycle for the simulation and deduction of S4. The sub-models in the multi-model library are constructed based on transfer learning: using the basic CFD-DEM model as the source model, and fine-tuning the model parameters to obtain the sub-models for the target varieties based on a small amount of experimental data for different Chinese medicine varieties; the working condition recognition model adopts an online learning mechanism, and automatically updates the decision boundary of the recognition model when a new working condition is detected.

4. The digital twin and virtual-real synchronous control method for the fluidized bed granulation process of traditional Chinese medicine according to claim 1, characterized in that, The multi-source sensor cluster in S2 includes: a near-infrared spectral sensor deployed above the fluidized bed material layer for real-time monitoring of particle moisture content changes; an online particle size analyzer connected to the fluidized bed via sampling pipelines for real-time monitoring of particle size distribution; multi-point temperature sensors deployed at the air inlet, material bed, and exhaust outlet for collecting temperature field distribution data; humidity sensors deployed at the air inlet and exhaust outlet for monitoring inlet and exhaust humidity; a differential pressure sensor for monitoring bed pressure differential and determining fluidization state; and a high-speed camera system for acquiring particle fluidization images through a viewing window.

5. The digital twin and virtual-real synchronous control method for the fluidized bed granulation process of traditional Chinese medicine according to claim 1, characterized in that, The data fusion algorithm in S3 includes: preprocessing and feature extraction of near-infrared spectral data; establishing a correlation model between spectral data and particle moisture content and particle size distribution using partial least squares regression or neural network methods; fusing the particle size data retrieved from the spectrum with the measured data from the online particle size analyzer using Kalman filtering; and using an adaptive covariance adjustment based on model prediction error as the fusion strategy to obtain the fused particle size estimate. Temperature, humidity, and pressure difference data are used as boundary conditions input into the digital twin model to drive the update of the model's state parameters.

6. The digital twin and virtual-real synchronous control method for the fluidized bed granulation process of traditional Chinese medicine according to claim 2, characterized in that, The nonlinear model predictive control algorithm in S5 uses the CFD-DEM coupled particle growth model constructed in S1 as the prediction model. In each control cycle, it re-solves the optimization problem based on the current state of the digital twin model, outputs the optimal control command sequence for the next control cycle, and issues and executes only the first control command.

7. The digital twin and virtual-real synchronous control method for the fluidized bed granulation process of traditional Chinese medicine according to claim 6, characterized in that, The nonlinear model predictive control algorithm in S5 also introduces the feedforward correction in S12 as an initial value, and performs prediction based on the master model after adaptive switching in S13, thereby realizing the coordinated optimization of feedforward-feedback composite control and multi-model adaptive switching.

8. The digital twin and virtual-real synchronous control method for the fluidized bed granulation process of traditional Chinese medicine according to claim 2, characterized in that, The method for determining the wettability threshold in the particle identification algorithm includes: based on the distribution characteristics of the liquid-solid ratio on the particle surface in historical batch data, a Gaussian mixture model clustering algorithm is used to divide the particle state into three categories: "unwetted", "moderately wetted" and "overly wetted"; with the goal of maximizing the inter-class variance of the clustering results, the wettability threshold of the current batch is dynamically optimized so that the threshold adapts to the fluctuation of extract properties and changes in the granulation stage.

9. The digital twin and virtual-real synchronous control method for the fluidized bed granulation process of traditional Chinese medicine according to claim 3, characterized in that, The feedforward compensation model works in tandem with the model adaptive switching step in S13: when the working condition identification model detects a change in the variety of traditional Chinese medicine or a change in the granulation stage, it synchronously triggers the parameter update of the feedforward compensation model. The feedforward compensation model adopts a Bayesian online learning algorithm, which updates the model weights in real time based on the physical property test data under the new operating conditions and the historical optimal process parameters, so that the feedforward correction amount is dynamically matched with the main model after adaptive switching.

10. The digital twin and virtual-real synchronous control method for the fluidized bed granulation process of traditional Chinese medicine according to claim 3, characterized in that, The online learning mechanism of the working condition recognition model includes: constructing an incremental support vector machine classifier; when a new working condition is detected and the accumulated sample size reaches a preset threshold, using an incremental learning algorithm based on kernel function approximation to update the decision boundary; and simultaneously, incorporating the optimal sub-model corresponding to the new working condition into a multi-model library to achieve dynamic expansion of the model library.

Images