A process online quality monitoring method for near-infrared combined fluidized bed granulation of traditional Chinese medicine

Abstract

The present application provides a kind of online quality monitoring method of traditional Chinese medicine in near infrared combined fluidized bed granulation process, the monitoring method will near infrared spectrum (NIRs) technology be combined with partial least squares (PLSr), and the online monitoring model of key quality attribute (CQAs) in granulation process of Buqi Tongluo Granules (BQTL-G) is constructed.Based on the real-time monitoring data of the model, the key process parameter of peristaltic pump speed is dynamically regulated by negative feedback mechanism through PID controller, the bed collapse problem caused by excessive moisture content in the granulation process is accurately intervened, and the granulation process state is stabilized.The online quality monitoring method of the present application solves the quality uniformity problem commonly existing in the production process of traditional Chinese medicine granules.

Description

Technical Field

[0001] This invention belongs to the field of traditional Chinese medicine detection and analysis, and particularly relates to an online quality monitoring method for traditional Chinese medicine in the near-infrared combined fluidized bed granulation process. Background Technology

[0002] Solid dosage forms of traditional Chinese medicine (TCM) are based on TCM theory and are processed into solid forms using specific techniques. Common types include granules, tablets, and capsules. However, the complex active ingredients, large dosages, high proportion of excipients, and hygroscopic nature of TCM compound preparations lead to low efficiency and high energy consumption in traditional production methods, severely impacting the clinical efficacy, safety, and modernization of TCM preparations. The main differences lie in appearance, active ingredients, and dissolution rate. The content of active ingredients is particularly significant as a cause of inconsistent quality. Taking TCM granules as an example, differences in the properties of the raw materials can lead to problems such as inconsistent hygroscopicity and mixing uniformity.

[0003] The reasons affecting the quality uniformity of solid dosage forms of traditional Chinese medicine can be mainly attributed to the following two dimensions: First, the influence of raw and excipient materials. As the material basis for the preparation of solid dosage forms of traditional Chinese medicine, the fluctuation of the quality attributes of raw and excipient materials directly determines the quality of batch-to-batch consistency of the preparation. At the level of raw materials, Chinese medicinal materials are affected by multiple factors such as germplasm resources, ecological environment (soil, climate, altitude), harvesting season, place of origin, processing and storage conditions, resulting in significant differences in the content of active ingredients, fingerprint characteristics and impurity profiles. At the level of excipients, even for pharmaceutical excipients that meet the pharmacopoeia standards, batch-to-batch differences in physicochemical properties such as particle size distribution, flowability, crystal form and moisture content will be amplified through the "dilution-mixing-forming" transfer effect, ultimately affecting the dissolution behavior and bioavailability of the preparation. Therefore, establishing a raw material and auxiliary material quality traceability system and a supplier audit system is the primary step in achieving quality uniformity control. Systematic influences on the production process and the variability of manufacturing processes are another key source of quality drift, mainly manifested in changes in feeding methods: the traditional "experience-based feeding" model lacks precise metering control, and subtle differences in the proportion of medicinal materials, feeding sequence, and pretreatment methods (such as particle size and extraction solvent ratio) between different batches can lead to significant fluctuations in the concentration of multiple components in the extract; differences in production units: traditional Chinese medicine compound preparations often involve multiple unit operations such as extraction, concentration, drying, granulation, and total mixing. Small deviations in equipment models and process parameters (temperature, pressure, time, speed) at each stage can accumulate into significant quality deviations in the final product through a "cumulative effect"; insufficient control of intermediate uniformity: traditional Chinese medicine extracts have high viscosity and strong heat sensitivity, making it difficult to achieve microscopic uniform dispersion of multiple components (raw powder + extract + excipients) using traditional tank mixing methods, easily forming "enriched phases" and "depleted phases," resulting in unqualified content uniformity.

[0004] A review of existing technologies related to the technical subject of this invention reveals that CN121775738A discloses a control method and apparatus for fluidized bed granulation, granulation equipment, and storage medium. This patent reports that by collecting gas flow rate, gas pressure, inlet temperature, and outlet temperature of a gas fluidized bed, and combining these parameters, the initial residual moisture content of the particles in the fluidized bed can be predicted, thereby achieving accurate control of the process. CN121783780A discloses a detection system, method, and apparatus for the endpoint of fluidized bed granulation of traditional Chinese medicine. The method includes: receiving spectral data of traditional Chinese medicine particles collected at at least three preset sites; converting the spectral data into predicted values ​​of key quality attributes of the traditional Chinese medicine particles at each site based on a quantitative analysis model, wherein the key quality attributes include moisture content and particle size, and the quantitative analysis model is a partial least squares regression model; calculating the relative standard deviation of the key quality attributes at all sites; and generating a granulation endpoint signal if the predicted values ​​of the key quality attributes at all sites reach the preset target range and the relative standard deviation of the predicted values ​​of the key quality attributes is lower than a set threshold, thereby realizing online detection and control of traditional Chinese medicine compound granules.

[0005] Near-infrared spectroscopy (NIR) technology, based on the vibrational absorption characteristics of hydrogen-containing groups in chemical substances within a sample, combined with chemometric algorithms to construct quality control models, enables real-time information acquisition of the production process and is widely used in online monitoring and quality uniformity evaluation in the manufacturing of traditional Chinese medicine (TCM). However, due to the complexity of TCM compound systems and the synergistic effects of multiple targets, relying solely on a single active ingredient, indicative component, or high-content chemical substance as the basis for quality evaluation is insufficient to systematically reflect the overall characteristics of the chemical composition of TCM. In particular, comprehensive monitoring of key quality parameters in the TCM production process is crucial to ensure the stability of effective components at each stage. Therefore, there is an urgent need to introduce an online monitoring system capable of comprehensively characterizing the complex chemical systems of TCM. Fluidized bed granulation, as a semi-continuous pharmaceutical equipment, commonly uses moisture content and particle size as quality indicators during granulation. Appropriate moisture content variation not only ensures the orderly progress of the granulation process but is also a critical prerequisite for the release of TCM solid dosage forms at key stages. Reports indicate that excessively high moisture content in the expansion tank can cause granulation collapse, resulting in unnecessary losses and leading to adhesion and aggregation issues during subsequent tablet compression. Conversely, excessively low moisture content increases granule brittleness, leading to the loss of raw material components. Studies have shown a significant intrinsic correlation between particle growth kinetics and separation mechanisms; different initial particle sizes result in different particle growth trends. Therefore, monitoring particle size and moisture content during the production process is crucial. Summary of the Invention

[0006] This invention provides an online quality monitoring method for traditional Chinese medicine (TCM) granulation using near-infrared spectroscopy combined with partial least squares (PLSr) to construct an online monitoring model for key quality attributes (CQAs) during the granulation process of Buqi Tongluo granules (BQTL-G). Based on the real-time monitoring data from this model, a PID controller is used to dynamically regulate the peristaltic pump speed—a key process parameter—through a negative feedback mechanism. This precisely intervenes in the bed collapse problem caused by excessive moisture content during granulation, stabilizing the granulation process. This online quality monitoring method addresses the common problem of quality uniformity in the production of TCM granules.

[0007] This paper employs near-infrared spectroscopy combined with chemometrics and control algorithms to monitor and control moisture and particle size online during fluidized bed granulation. To realize the production philosophy of "quality originates from design," this approach ensures product quality and automated production through process and workflow assurance, thereby improving the quality analysis and control level of the fluidized bed granulation process. This results in a set of key technologies for intelligent quality control in traditional Chinese medicine production, providing technical guidance for quality monitoring throughout the entire solid dosage form granulation process of traditional Chinese medicine.

[0008] The technical solution of this invention patent application is as follows: A near-infrared combined fluidized bed granulation method for online quality monitoring of traditional Chinese medicine, the monitoring method comprising the following steps: (1) Granulation process and sample collection Weigh out the compound extract of traditional Chinese medicine and the fine powder of mixed traditional Chinese medicine, and granulate them using fluidized bed granulation. First, preheat the fluidized bed drying system for 30 minutes, then add the fine powder of mixed traditional Chinese medicine and continue mixing for 8-12 minutes. Then start the atomization spraying program of traditional Chinese medicine extract. The following key process parameters are controlled during the spraying process: pump flow rate 10-20 mL / min, inlet air temperature 60-70℃, atomization pressure 0.8-1.5 Bar, ambient humidity 2.0-4.0 g / kg, total granulation time 110-130 min, drying time 20-40 min. The criterion for judging the granulation endpoint is that the moisture content of the particles drops to below 6.0%. When comparing control methods, samples are taken after all the binder has been sprayed in to obtain the test samples. (2) Near-infrared spectral acquisition A near-infrared spectrometer was inserted into the fluidized bed cavity, and spectra were acquired in diffuse reflectance mode. The moisture and particle size values ​​were analyzed online using a near-infrared monitoring model. Before acquiring material information, the surrounding air environment was scanned by dark current, and then a polytetrafluoroethylene plate was collected as background spectrum to deduct the interference of the PAT-U device on the spectral information. The absorption wavelength range of the device was set to 908.1-1676.2nm, the integration time was set to 6.6ms, the number of spectral acquisitions was 250, and the air in the viewing window was purged once every 10s. When the model was established, Micro NIR Pro V3.2 was used to collect spectral information online. When the model was predicted, the moisture and particle size models established in Unscrambler X10.4 were used to predict the moisture and particle size values ​​in real time. This invention also adopts a PID control system algorithm, which controls the peristaltic pump speed based on the negative feedback of the difference between the theoretical value and the preset value of particle moisture content, so as to achieve precise control of moisture to the preset value. (3) Measurement of monitoring indicators Establish "moisture content" and "particle size" as quality testing indicators; (4) Spectral preprocessing and band selection The raw spectral preprocessing methods examined include: vector normalization, derivative spectroscopy, smoothing, multivariate scattering correction, standard normal variable transformation, baseline correction, and combined strategies of multiple methods. After removing impurity information from the original NIR spectra, the band selection is performed using a variable iterative spatial shrinkage algorithm, which involves the following steps: ① Based on the weighted binary sampling method, the original 12-band variables are divided into several subsets, each variable is assigned an initial weight of 0.5, and N subsets are generated after N random samplings.

[0009] ② Input each subset of variables into the PLSr model, record the mean squared error of cross-validation, select the model with the lowest RMSECV as the optimal set, calculate the frequency of occurrence of variables and update the weights, and record the mean RMSECV. The formula for calculating the weight of the nth variable is as follows: In the formula, Wn represents the weight, fn represents the frequency of the nth variable in the optimal model set, and Nbest represents the number of optimal model sets. ③ Iterative execution steps (1) and (2): WBMS sampling weights are updated according to the above formula. When RMSECV no longer decreases, the iteration is terminated. At this time, the band with a weight of 1 is defined as the core information band, the band with a weight of 0 is regarded as the interference band, and the band with a weight between 0 and 1 is classified as the weak information band. The three types of bands are arranged in descending order of weight, and the PLSR model is constructed according to the size of the variable set from large to small. The combination corresponding to the minimum value of RMSECV is used as the final feature band. (5) Establishment of chemometric methods Spectral data was recorded simultaneously during sampling and correlated with the quality inspection index data from step (3). After spectral preprocessing, a quantitative analysis model was established using the partial least squares regression (PLSR) algorithm. The model evaluation index was R0. 2 C R 2 P RMSEC, RMSEP, R 2 The optimized model is then applied to the online predictive analysis of the fluidized bed granulation process.

[0010] Preferably, in step (1), the herbal compound extract is a qi-tonifying and meridian-clearing extract, and the herbal mixed fine powder is Panax notoginseng micro powder and dextrin, wherein the ratio of the herbal compound extract to the herbal mixed fine powder is 1:2-2.5.

[0011] Preferably, the fluidized bed granulation peristaltic pump speed control model in step (1) adopts a gradient control mode, with Kp=0.85 and Ti=120 as the optimized control parameters.

[0012] Preferably, the following key process parameters are controlled during the spraying process in step (1): pump flow rate 15 mL / min, inlet air temperature 65℃, atomization pressure 1.2 Bar, and ambient humidity 3.0 g / kg.

[0013] Preferably, the frequency and interval of sampling in step (1) granulation are as follows: sampling once every 5 minutes during the process and once every 5 minutes during the drying stage.

[0014] Preferably, the sample to be tested in step (1) is placed at the inlet of the linear laser particle size analyzer, and compressed air is connected. Using compressed air as a medium, the sample particles are passed through the inlet sequentially and uniformly. This process is repeated two or three times, and the average value is recorded as the D value of the current sample. 10 D 50 and D 90 value.

[0015] Preferably, in step (4), the spectral preprocessing uses the SPXY algorithm to divide the dataset and the VISSA algorithm to filter key variable information, and selects the SD-1+De-trending preprocessing method to establish the PLSR model.

[0016] Preferably, the control system in step (2) is based on the Siemens S7-300 PLC and TP1200 HMI architecture. The PLC obtains near-infrared prediction values ​​through communication via a unified architecture. The human-machine interface (HMI) displays the moisture trend graph in real time and corrects the parameters. The PID module calculates the pump speed based on the feedback deviation. The Profinet communication transmits instructions to the frequency converter, which links the peristaltic pump to adjust the speed, thereby achieving closed-loop precise control of moisture.

[0017] Preferably, in step (3), the "moisture content" is determined by the drying method in the 2025 edition of the Chinese Pharmacopoeia, and the "particle size determination" is performed using an online laser particle size analyzer.

[0018] Preferably, the online quality monitoring method for traditional Chinese medicine processes can be used in the production process of compound granules of traditional Chinese medicine.

[0019] Preferably, the sample in the test group of this invention is a traditional Chinese medicine compound for replenishing qi and unblocking collaterals granules.

[0020] The beneficial effects of the patented technical solution of this invention are as follows: (1) This invention focuses on the granulation process of "compound qi-tonifying and meridian-clearing granules of traditional Chinese medicine," and utilizes NIRs technology to monitor the fluidized bed granulation process online. While collecting near-infrared spectra during the granulation process online, offline analysis is performed on the moisture and particle size information at the corresponding collection points. Through preprocessing of the original spectra, outlier removal, and key band screening, a predictive model for key quality attributes (moisture and particle size) is finally established using PLSr.

[0021] This invention involves 277 samplings during the granulation process with different parameters. Moisture content and particle size were measured online for each sample, and near-infrared spectra were processed. In establishing the moisture model, PCA combined with Mahalanobis distance was first used to remove outlier values; six outlier samples were removed based on the actual measurements. Next, the SPXY algorithm was used to divide the dataset into a 7:3 ratio, resulting in a test set of 54 samples and a calibration set of 217 samples. SNV+SD-2 was selected as the optimal spectral preprocessing method, with 7 latent variables. After establishing the full-band model, the R²c, R²p, RMSEC, and RMSEP values ​​were 0.8875, 0.8450, 0.4390, and 0.5159, respectively. The VISSA algorithm was used to effectively filter 125 near-infrared variables, resulting in 32 variables specifically for moisture. The R²c, R²p, RMSEC, and RMSEP values ​​for these variables were superior to those of the full-band model, with values ​​of 0.8904, 0.8699, 0.4333, and 0.4718, respectively. Finally, for the prediction of two batches of materials with varying process parameters, a paired t-test was performed comparing offline detection and online prediction methods. The result was P = 0.6538 > 0.05, indicating that the moisture model prediction can replace the traditional offline monitoring method. In establishing the D50 online detection model, offline samples were collected using a laser particle size analyzer. Three outlier samples were removed from the actual D50 measurements using PCA combined with Mahalanobis distance. The SPXY algorithm was used to partition the dataset, and the VISSA algorithm was used to screen key variables. Finally, the SD-1+De-trending preprocessing method was selected to establish the PLSR model. The evaluation indices R²c, R²p, RMSEC, and RMSEP values ​​were 0.9799, 0.9741, 27.2929, and 31.5950, respectively. A paired t-test yielded P=0.6538, indicating no significant difference between the predicted and actual D50 values ​​from the established online monitoring model. Near-infrared online monitoring can replace traditional offline sampling and detection in the granulation process, providing a foundation for further research on control models.

[0022] (2) This invention uses the VISSA algorithm combined with PLSr to establish a quantitative model of key quality attributes of BQTL-G, and also establishes an intelligent control model of key production links of BQTL-G, providing a practical and feasible method for the subsequent construction of intelligent continuous production of oral solid dosage forms of traditional Chinese medicine.

[0023] The technological innovation of this invention lies in the online prediction of final moisture and particle size values ​​during fluidized bed granulation using a near-infrared monitoring model, and the automated intelligent control of key process parameters (peristaltic pump speed) achieved by utilizing a PID algorithm combined with a programmable logic controller (PLC). Based on the PID algorithm, a peristaltic pump speed control model was successfully established using engineering experiments. By comparing the particle size evolution patterns when the moisture setpoints were 7.0%, 7.3%, and 7.5%, 7.3% was ultimately selected as the target moisture content for setpoint control. Simultaneously, the PI parameters were optimized for both setpoint and gradient control modes: for the former, the proportional coefficient Kp=1 and the integral time Ti=100; for the latter, Kp=0.85 and Ti=120. Further comparison of the relative standard deviation (RSD) and relative width (RW) of D50 for three batches of samples under manual control, setpoint control, and gradient control showed that the gradient control mode exhibited the best batch-to-batch consistency, thus establishing it as the core control strategy for subsequent research.

[0024] (3) This invention comprehensively verifies the stability of the established control strategy by systematically adjusting key process parameters such as inlet air temperature, atomization pressure, and ambient humidity. The results show that the peristaltic pump control mechanism maintains stable operation under the aforementioned three parameter disturbances, with the residual difference between the measured and set moisture values ​​controlled within 2%. The particle size variation patterns of each batch exhibit good consistency, with no significant fluctuations observed. These results fully demonstrate that the control model possesses high robustness and process adaptability, effectively coping with parameter fluctuations in actual production.

[0025] (4) This invention determined the physical properties of 15 batches of BQTL-G powder. The results showed that the powder properties remained stable among batches, and the product quality was consistent. Furthermore, an analytical method was established using gentianin as a reference peak. The relative standard deviation (RSD) of its precision, repeatability, and stability tests were all below 3%, indicating that the method validation met the requirements. Further similarity analysis was used to evaluate the 15 batches of samples. The results showed good similarity among batches and high quality consistency, confirming that the production process under this control method is stable and controllable, and the product quality is uniform. Attached Figure Description

[0026] Figure 1 Near-infrared raw spectra of 9 batches of samples; Figure 2 Abnormal sample removal chart; Figure 3 PCA analysis diagram after SPXY partitioning of the sample set; Figure 4 VISSA variable selection chart; Figure 5Comparison chart of full-band and VISSA variable screening PLSR models; Figure 6 PCA combined with Mahalanobis distance removes outlier sample values; Figure 7 VISSA algorithm variable result selection chart; Figure 8 A comparison chart of predicted and actual values ​​from the D50 model; Figure 9 Moisture variation curves and residual plots for the two PI parameters when the moisture content is 7.3%; Figure 10 Comparison of particle size growth trends for three fixed moisture values; Figure 11 Gradient control of moisture change trends and residual plots; Figure 12 Graph showing the changes in moisture content and PC1 score over time during manual granulation process; Figure 13 PC1 load diagram of manual granulation process; Figure 14 Figure 1. Changes in moisture content and PC1 score over time during the fixed-value granulation process. Figure 15 Figure 1. Changes in moisture content and PC1 score over time during gradient granulation process. Figure 16 Compare the physical fingerprints of physical spectrum R and 15 batches of BQTL-G; Figure 17 , BQTL-G cluster analysis chart of 15 batches. Detailed Implementation

[0027] Unless otherwise defined, the technical or scientific terms used in this patent application specification and claims shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. Related pronouns used in this invention are explained herein to aid those skilled in the art in better understanding the invention. API - Active pharmaceutical ingredient, CQAs - Key Quality Attributes, DOE - Design of Experiments, MCCG - Constant Moisture Control Granulation, MGCG - Gradient Moisture Control, NIRs - Near-Infrared Technology, PAT - Process Analysis Technology, PCA - Principal Component Analysis, PCR - Principal Component Regression, PID - Proportional-Differential-Integral, PLC - Programmable Logic Controller, PLS - Partial Least Squares Method, PLSR - Partial Least Squares Regression, QbD - Quality by Design, R² - Coefficient of Determination, R²cal - Coefficient of Determination for Correction Set, RG - Conventional Granulation, RMSEC - Corrected Root Mean Square Error, RMSECV - Cross-Validation Root Mean Square Error, RPD - Performance Deviation Rate, Kp - Proportional Coefficient, Ti - Integral Time.

[0028] Example 1: Establishment of an online monitoring model for the BQTL-G fluidized bed granulation process based on NIRs technology 1. Experimental materials and instruments 1.1 Experimental Apparatus LGL002 6L fluidized bed (Shandong Xinma Pharmaceutical Equipment Co., Ltd.), Micro NIR PAT-U (Near Infrared Spectrometer), Micro NIRTM pro v3.2 spectral acquisition software, VIAVI Matlab2021b (Mathworks, USA), PLS_Toolbox (Eigenvector Research, USA), online laser particle size analyzer (PARSUM, Germany), halogen rapid moisture analyzer (Changzhou Co., Ltd.) 300L extraction and concentration line (Shaanxi Traditional Chinese Medicine Modernization Research Company), medicinal herb pulverizer (Zhongnan Pharmaceutical Machinery Factory).

[0029] Unscrambler X10.4 (CAMO, Norway).

[0030] 1.2 Reagents and Tests Astragalus membranaceus, Cicadae Periostracum, Curcuma longa, Alisma plantago-aquatica, Angelica sinensis, Ligusticum chuanxiong, Panax notoginseng, Cinnamomum cassia (Shaanxi Kang Shengtang Pharmaceutical Co., Ltd.), Dextrin (Qufu Tianli Pharmaceutical Excipients Co., Ltd.).

[0031] 2. Experimental Methods 2.1 Extraction and Concentration Process of Medicinal Materials The Panax notoginseng medicinal materials were pulverized, and the remaining medicinal materials were fed into a 300L extraction and concentration line according to the specified ratio to concentrate them into an extract with a final relative density of 1.22-1.24 m³. 3 / kg (measured at 45℃), solid content is 45-55%.

[0032] 2.2 Granulation and Sampling Process 2.2.1 Experimental Design for Granulation Process Nine batches of experiments were designed according to the previously established experimental parameters, as detailed in Table 1. BQTL-G produced in batches 1-5 were used to establish models for moisture content and particle size. During granulation, the rate of change in the peristaltic pump flow rate (mL / min) had the most significant impact on the granulation effect and the moisture content within the granulation chamber; therefore, this parameter was varied within a certain range to expand the model sample size. Batches 6-7 served as the test set to assess the model's feasibility. Batches 8-9 introduced a new process parameter change, namely atomization pressure, primarily to verify the robustness of the established monitoring model.

[0033] Table 1 Experimental Design for Establishing the Near-Infrared Online Monitoring Model

[0034] 2.2.2 Granulation process and sample collection Panax notoginseng powder and dextrin were mixed according to the specified ratio and granulated using a fluidized bed granulation method. Each batch of granulation tank contained 800 g of powder 1 and 1760 g of extract. The fluidized bed equipment was preheated for 30 minutes until the temperature of the expansion tank reached approximately 35 ℃. Powder 1 was then added, and the mixture was fluidized for 10 minutes, for a total granulation time of 120 minutes. Drying was then carried out for 30 minutes. Drying was completed when the moisture content of the particles in the tank was ≤ 6%. Samples were taken at the end of the mixing process and before spraying; samples were taken every 5 minutes during granulation and every 5 minutes during the drying stage, for a total of 30 samples per batch. A total of 9 batches were sampled, resulting in a total of 270 samples.

[0035] 2.3 Near-infrared spectral acquisition A miniature near-infrared spectrometer (PAT-U) was inserted into the fluidized bed chamber, and spectra were acquired in diffuse reflectance mode. Before acquiring material information, a dark current scan of the surrounding air environment was performed, and then a polytetrafluoroethylene plate was collected as background spectrum to subtract interference from the PAT-U device on the spectral information. The absorption wavelength range of the device was set to 908.1-1676.2 nm, the integration time was set to 6.6 ms, and the number of spectral acquisitions was 250. To prevent the probe from being contaminated during granulation and causing information acquisition failure, the air in the viewing window was purged once every 10 seconds. When building the model, Micro NIR Pro V3.2 was used to collect spectral information online. When predicting the model, the moisture and particle size models established in Unscrambler X10.4 were used, and Micro NIR Pro V3.2 was used to predict the moisture and particle size values ​​in real time.

[0036] 2.4 Data Measurement 2.4.1 Moisture content determination The moisture content (%) of the sample was determined using Method II (drying method) in the 2025 edition of the Chinese Pharmacopoeia.

[0037] 2.4.2 Particle size determination The experiment used an online laser particle size analyzer to determine the particle size of the sample. A 10 g sample was placed at the inlet of the online laser particle size analyzer, and compressed air was connected. Using compressed air as a medium, the sample particles were passed through the inlet sequentially and evenly. This process was repeated two or three times, and the average value was recorded as the particle size (D) of the current sample. 10 D 50 and D 90 value.

[0038] 2.5 Spectral Preprocessing and Band Selection The raw spectral preprocessing methods investigated in this experiment include: vector normalization, derivative spectroscopy (first and second derivatives), smoothing methods (Savitzky-Golay smoothing and Norris Derivative smoothing), multivariate scattering correction (MSC), standard normal variable transformation (SNV), baseline correction, and combined strategies of multiple methods. After removing impurity information from the raw NIR spectra, the Variable Iterative Space Shrinkage Approach (VISSA) was used to filter the bands, which involved the following steps: (1) Based on the weighted binary sampling method (WBM), the original 12-band variables are divided into several subsets, each variable is assigned an initial weight of 0.5, and N subsets are generated after N random samplings.

[0039] (2) Input each subset of variables into the PLSr model, record the mean squared error of cross-validation (RMSECV), and select the model with the lowest RMSECV as the optimal set. Calculate the frequency of occurrence of variables and update their weights, while recording the mean RMSECV. The formula for calculating the weight of the nth variable is as follows: In the formula W n For the weight, f n Let N be the frequency of the nth variable in the optimal model set. best This represents the number of optimal model sets.

[0040] (3) Iterate through steps (1) and (2), updating the WBMS sampling weights according to the above formula. The iteration terminates when RMSECV no longer decreases. At this point, the bands with a weight of 1 are defined as core information bands, those with a weight of 0 are considered interference bands, and those with a weight between 0 and 1 are classified as weak information bands. Arrange the three types of bands in descending order of weight, construct the PLSR model according to the size of the variable set from largest to smallest, and use the combination corresponding to the minimum value of RMSECV as the final feature band.

[0041] 2.6 Establishment of Chemometric Methods Spectral data were recorded simultaneously during sampling and correlated with data from sections "2.4.1" and "2.4.2". After spectral preprocessing, a quantitative analysis model was established using the partial least squares regression (PLSR) algorithm. The model evaluation index was R0. 2 C R 2 P RMSEC, RMSEP, R 2 cal The optimized model was applied to online predictive analysis of the fluidized bed granulation process, providing strong data support for subsequent online control of the peristaltic pump speed.

[0042] 3. Experimental Results 3.1 Results of material preparation before pelleting A total of 3.3 kg of Panax notoginseng powder was obtained. The extract was extracted three times, yielding a total of 536 L. After preliminary concentration, 42.5 kg of extract was obtained, bringing the total weight of the concentrated extract to 31.7 kg. The density of the extract was 1.23 g·cm³. -3 .

[0043] 3.2 Data Measurement Results 3.2.1 Establishment of Near-Infrared Moisture Model 3.2.1.1 Near-infrared sample spectrum This experiment involved nine batches of fluidized bed granulation. Data from batches 1-7 were used to establish a quantitative analysis model for water content, while the remaining batches were used to verify the robustness of the model. Figure 1 The raw near-infrared spectra of nine batches of samples were presented, showing that the spectral profiles of each batch were highly similar and the baseline characteristics were consistent. Due to the presence of numerous OH groups in the samples, strong OH absorption peaks were observed in the spectra. Furthermore, the baseline noise in the spectra was significant; therefore, chemometric methods were applied to eliminate baseline noise, environmental influences, and other impact factors.

[0044] 3.2.1.2 Removing outliers from the near-infrared spectrum In near-infrared analysis, some spectra deviate significantly from the overall confidence interval, failing to accurately reflect the data information and strongly interfering with the accuracy of the model. Therefore, outliers need to be removed before modeling.

[0045] like Figure 2 As shown, this experiment uses principal component analysis combined with Mahalanobis distance to remove abnormal samples from the original near-infrared spectrum. It can be seen that some values ​​of the spectral data overflow the set line, so 6 abnormal sample values ​​are removed.

[0046] 3.2.1.3 Partitioning of the Sample Set The sample set was divided in a 7:3 ratio. Five batches were selected to form the calibration set, two batches to form the test set, and the remaining two batches to form the validation set. A total of 210 samples were obtained. This experiment used the SPXY (Side-by-Side) algorithm, which calculates the distance between sample pairs and uses a pairing-only approach to ensure that the statistical characteristics of the divided dataset remain consistent across all groups. Its advantage lies in improving the model's predictive ability while covering a multi-dimensional vector space. The PCA scatter plot of the sample set partitioning is shown below. Figure 3 As shown, the validation set is evenly distributed within the calibration set, indicating that the sample set division is reasonable.

[0047] 3.2.1.4 Selection of Spectral Preprocessing Method Preprocessing methods were examined using RMSECV and RMSEP as indicators. Different preprocessing methods on the original spectra revealed significant differences in the subsequent model results for moisture or particle size. Furthermore, the order of preprocessing also affected the results. In order to eliminate noise and baseline interference from the instrument and environment, a combination of SNV and reciprocal smoothing was often used to eliminate irrelevant effects caused by background interference.

[0048] Table 2 shows the PLSr models under different preprocessing methods. The results show that the R-values ​​of the models differ depending on the preprocessing method. C 2 and R P 2 With some changes, and considering other evaluation metrics such as RMSEC and RMSEP, the combined method of SNV+SD-2 was ultimately chosen as the preprocessing result for the model.

[0049] Table 2 Comparison of modeling results for different pretreatment methods in the moisture model

[0050] Note: MSC: Multivariate scattering correction; SD-2: Second derivative; SNV: Standard normal variable transformation; Detrend: Dynamic trend transformation.

[0051] 3.2.1.5 Band Selection To eliminate the influence of irrelevant information on the quantitative water content model, this study selected the Interval Variable Iterative Spatial Shrinkage (VISSA) method. This method aims to find the optimal combination of variables, maximizing the predictive performance of the model while maintaining its simplicity. After preliminary exploration of the data to understand the relationships between variables, an iterative strategy was adopted. Each iteration required evaluating the importance of variables to select or eliminate them. This process was repeated until a satisfactory number of variables was reached or the maximum number of iterations was achieved. This algorithm is suitable for feature variable selection in solving regression problems. In this experiment, 50 iterations were selected. The variable selection process and results are as follows. Figure 4 As shown. Among them. Figure 4 Figure (a) shows the near-infrared spectrum obtained after preprocessing, while Figure (b) shows that the PLSR model has the lowest RMSECV value when the lowest point is selected, indicating the best variable selection model effect. Figure (c) shows information on 32 effective variables, and Figure (d) shows the variable selection diagram after 50 iterations. Combining the variable information, the first and second overtone absorptions of the OH stretching vibration of water are around 1440 nm and 960 nm, respectively. The variable range selected by the VISSA algorithm mainly reflects changes in water content. Figure 5 A comparison of the modeling plots using full-band modeling and variable selection shows that the accuracy of the model has been improved.

[0052] 3.2.1.6 Moisture Model Establishment and Evaluation Through the above analysis and discussion, a quantitative analysis model for moisture content in the BQTL-G fluidized bed granulation process was successfully constructed. The final optimized modeling conditions were: SNV+SD-2 spectral preprocessing method, a principal component count of 7, and VISSA algorithm for feature variable selection, choosing 32 key variables to establish the PLSr quantitative model. The model evaluation parameter R... C 2 R P 2 The RMSECV and RMSECP values ​​were 0.8904, 0.8699, 0.4333%, and 0.4718%, respectively.

[0053] To verify the predictive power of the moisture content model, paired t-tests were used to analyze the measured moisture content values ​​and model predictions for batches 8 and 9. The results are shown in Table 3. At a 95% confidence level, P =0.9832>0.05 and H=0, indicating that there is no significant difference between the two methods, that is, online NIRs are suitable for moisture monitoring in fluidized bed granulation drying.

[0054] Table 3. Statistical results of paired t-tests

[0055] 3.2.2 Establishment of Near-Infrared Online Monitoring Particle Size Model Particle size variations can be quantitatively analyzed by monitoring spectral baseline changes, but baseline drift in fluidized bed granulation can easily confuse particle size with drift causes. To improve D... 50 To assess the quality of model construction, this experiment used PCA combined with Mahalanobis distance to remove three outlier samples. The results are as follows: Figure 6 As shown in Table 4, the experiment also explored the impact of different data preprocessing methods on the model's prediction performance. The relevant results are shown in Table 4, where SD-1+De-trending is the optimal preprocessing scheme, and the corresponding model evaluation index R0 is [value missing]. C 2 R P 2 The RMSEC and RMSEP values ​​were 0.9799, 0.9741, 27.2929, and 31.5950, respectively. Based on the preprocessed near-infrared data, key variables were screened using the VISSA algorithm, ultimately obtaining information on 40 key variables, as shown in the results. Figure 7 As shown; the actual measured value and the model predicted value of D50 are as follows: Figure 8 As shown.

[0056] Table 4 Comparison of different preprocessing methods

[0057] To verify the predictive power of the particle size model, pairing t Verification of particle size D 50 The differences between the measured values ​​and the predicted values ​​of NIRs are shown in Table 5, at a 95% confidence level. P =0.6538>0.05 and H =0 indicates that there is no significant difference between the two methods, meaning that near-infrared technology can be applied to the D in fluidized bed granulation process. 50 Online monitoring Table 5. Results of paired t-tests

[0058] 4. Summary and Discussion This experiment aims to systematically construct an online near-infrared spectroscopy monitoring system based on process analysis technology. Taking fluidized bed granulation unit operation as the research object, it achieves real-time quantitative analysis of critical quality attributes (CQAs) of solid dosage forms of traditional Chinese medicine (TCM), and explores the feasibility and applicable boundaries of NIRs technology for monitoring the TCM manufacturing process. Specifically, the research focuses on the online quantitative determination of particle moisture content using NIRs—this parameter, as a core CQAs in the fluidized bed granulation process, directly determines the particle forming quality, drying endpoint determination, and subsequent tableting or filling performance. The research design adopts a process parameter perturbation strategy based on the design space concept. Based on the existing factory production process parameter center point, key process parameters (CPPs) such as inlet air temperature, atomization pressure, spray rate, and fluidizing gas velocity are systematically varied to deliberately create diverse operating conditions covering normal fluctuation ranges and edge conditions. Within this framework, NIR spectral data and process parameter trajectories of each experimental batch throughout the granulation process are simultaneously collected to establish a high-temporal-resolution dataset of dynamic moisture changes. To verify and calibrate the prediction accuracy of the online model, a synchronous sampling-offline reference strategy was implemented during the study: sampling valves were set at specific geometric locations in the fluidized bed expansion tank, and representative samples were extracted every 5 minutes. The actual moisture content was immediately measured using a halogen moisture analyzer, and the particle size distribution (PSD) was determined by laser diffraction, serving as reference values ​​for modeling. Finally, partial least squares regression (PLSR) was adopted as the core chemometric algorithm. Through spectral preprocessing (such as standard normal variable transformation SNV, first derivative, Savitzky-Golay smoothing, etc.), characteristic wavelength screening (such as competitive adaptive reweighted sampling CARS, variable importance projection VIP), and model optimization (cross-validation to determine the optimal number of latent variables), a robust, accurate, and interpretable quantitative analysis model was constructed. This laid the methodological foundation for subsequent real-time release testing (RTRT) and closed-loop control of moisture in the fluidized bed granulation process. After optimizing the model parameters, the optimal quantitative analysis results were obtained. The RC², RP², RMSECV, and RMSECP values ​​for the moisture model were 0.8904, 0.8699, 0.4333%, and 0.4718%, respectively. Meanwhile, the RC² of the D50 model... 2 RP 2The RMSEC and RMSEP values ​​were 0.9799, 0.9741, 27.2929, and 31.5950, respectively. The comparison between the predicted values ​​of the validation batch after changing the process parameters and the actual moisture content values ​​indicates that the model prediction results are highly accurate.

[0059] The results of this study show that the quantitative model for multi-index CQAs in the fluidized bed granulation process exhibits good linearity and high accuracy. It can serve as a foundation for establishing control models in future research.

[0060] Example 2: Establishment of BQTL-G Online Control Model Based on PID Algorithm This experiment established a near-infrared online monitoring model and integrated a PID control algorithm to achieve closed-loop negative feedback control of moisture content in the fluidized bed granulation process. Engineering tuning was employed to optimize the PID parameters under both setpoint control and gradient control modes. Simultaneously, the feasibility of control methods under different conditions (conventional control, setpoint control, and gradient control) was compared, and the particle size differences at the granulation endpoint were compared among the three methods.

[0061] 1. Experimental instruments and materials 2. Experimental Methods This experiment employs a PID control algorithm, using negative feedback to control the peristaltic pump speed based on the difference between the theoretical and preset particle moisture content, thus achieving precise moisture regulation to the preset value. The control process mainly consists of three parts: (1) Fluidized bed granulation: Experimental fluidized bed operation, with main control parameters including peristaltic pump flow rate, inlet air temperature, exhaust air ratio and atomization pressure.

[0062] (2) Near-infrared spectral acquisition software and data analysis process: Micro NIR PAT-U was used to acquire spectra, and the moisture and particle size values ​​were analyzed online using a near-infrared monitoring model.

[0063] (3) Control system: Based on Siemens S7-300 PLC and TP1200 HMI architecture, the PLC obtains near-infrared prediction values ​​through unified architecture (OPC Unified Architecture, OPC UA) communication, the human machine interface (HMI) displays the moisture trend map in real time and corrects the parameters, the PID module calculates the pump speed based on the feedback deviation, and the Profinet communication transmits the instructions to the frequency converter to link the peristaltic pump speed adjustment, thereby realizing closed-loop precise control of moisture.

[0064] The PID algorithm integrates proportional (P), integral (I), and derivative (D) components, possessing both transient and steady-state response control capabilities. It is currently widely used in machinery, food processing, agriculture, and other fields. Essentially, it involves P, I, and D function calculations based on the deviation value, with the result used to control the output value.

[0065] The governing equation for particle moisture content is: PID parameter tuning uses the P, I, and D components of the linear combination error to form the control input, optimizing the dynamic and static characteristics of the system to achieve the desired performance. Parameter tuning involves adjusting key controller parameters, including proportional gain (ζ) and integral time (T). i ) and differential time (T) a The process of optimizing PID controller parameters aims to improve the dynamic response characteristics and static stability of the system, thereby achieving ideal control quality. There are numerous methods for optimizing PID controller parameters, primarily theoretical calculation tuning and engineering experimental tuning. The former calculates parameters based on mathematical models, relying on model accuracy and requiring actual debugging and correction; the latter predicts the controlled variable's change trend based on process data, offering simplicity, versatility, and widespread application. Because the derivative element is sensitive to noise, this study employs engineering experimental methods to implement PI control.

[0066] 2.1 Feasibility Analysis 2.2.1 Feasibility Analysis of Setpoint Control The PI tuning parameters of the fixed-value model were optimized, and 7.0%, 7.3%, and 7.5% were set as the moisture constant change values ​​to monitor particle growth and moisture change trends during the granulation process.

[0067] 2.2.2 Feasibility Analysis of Gradient Control Optimize the PI gradient model tuning parameters and monitor the actual moisture changes and particle growth process during granulation under varying moisture gradient conditions.

[0068] 2.2 Control Model Establishment After completing the feasibility analysis, the equipment parameters were adjusted: atomization pressure 1.2 Bar; inlet air temperature 65 ℃; and a control model was established using Table 6.

[0069] Table 6 Experimental Design Diagram for Control Model Establishment

[0070] 2.3 Granulation process Granulation was carried out using an experimental fluidized bed. According to the formulation theory, the material weights were as follows: 800 g for each batch of powder and 1760 g for each batch of extract, with a solid content of 50%.

[0071] Before starting each batch of granulation, the fluidized bed drying system was preheated for 30 minutes to ensure the fluidization state stabilized before starting the granulation process. After the preheating stage, prescription powder 1 was added and mixed continuously for 10 minutes, followed by starting the atomization spraying program for the traditional Chinese medicine extract. The endpoint of granulation was determined when the moisture content of the particles dropped below 6.0%. During the comparative evaluation of control methods, samples were taken after all the binder had been sprayed to determine the median particle size (D). 50 Nine batches of validation experiments yielded a total of nine samples. In the control robustness evaluation experiments, no sampling was performed midway; instead, the moisture content and particle size prediction model established in the previous chapter was used directly for online estimation.

[0072] 2.4 Spectral Acquisition In the fluidized bed granulation process, a Micro NIR PAT-U type near-infrared process analysis probe was installed inside the fluidized bed cavity to continuously acquire spectral signals using diffuse reflectance. Before probe insertion, background correction was performed using a PTFE standard white plate. The instrument parameters were set as follows: integration time 6.6 ms, spectral acquisition range 908.1–1676.1 nm, and 100 scans per spectrum. The moisture content and median particle size (D) parameters constructed in Chapter 3 were used. 50 A quantitative analysis model is used to make real-time predictions for the entire granulation process. The data is analyzed online through the Micro NIR Pro V3.2 software platform, and a prediction result is output approximately every 3.3 seconds. The obtained moisture data is transmitted to the programmable logic controller (PLC) in real time via data cable to support subsequent process control decisions.

[0073] 2.5 Particle size determination The experiment used an online laser particle size analyzer to determine the particle size of the sample. 10 g of the sample was added to the inlet of the online laser particle size analyzer, and compressed air was connected. Using compressed air as a medium, the sample particles were passed through the inlet sequentially and evenly, and the particle size D was measured. 10 D 50 D 90 Perform three parallel measurements and calculate the average value. RW is then calculated using the following formula: 3. Experimental Results 3.1 Feasibility Analysis of Control Methods This experiment employs an engineering experimental method, experimentally adjusting the relationship between the three parameters of a PID controller. Through extensive experiments, two optimized sets of control parameters (Kp=1, Ti=100; Kp=0.85, Ti=120) were applied to two different control methods: setpoint control and gradient control. By comparing the particle growth curves and moisture residuals under different parameters, the optimal parameters were determined.

[0074] 3.1.1 Feasibility Analysis of Setpoint Control Multiple batches of experiments show that, under the constant value control mode, the particle state exhibits stable characteristics with moisture regulation. Figure 9 The effects of the proportionality coefficient (Kp) and integral time (Ti) on the granulation process under different moisture setpoints are shown. Figure 9 Figures A and B present the evolution trend of peristaltic pump speed, the change curves of setpoint and actual moisture content, and the corresponding residual distribution under the parameter combination of Kp=0.85 and Ti=120. The results show that when granulation enters the later stage, the peristaltic pump speed climbs to a peak of 20 rbm, and the actual moisture content fails to track the set trajectory, ultimately leading to bed collapse. Figure 9 C and D represent the granulation process under the parameters Kp=1 and Ti=100. Under these parameters, the process operates stably with minimal fluctuations in moisture residuals. Based on the above analysis, Kp=1 and Ti=100 were selected as the optimized control parameters for subsequent fixed-value control research.

[0075] Given the determined PI parameters, the moisture setpoints were examined, and the moisture change trends under three conditions—7.5%, 7.3%, and 7.0%—were obtained as shown in the graphs. Figure 10 As shown in the figure. The results show that when the moisture content is set at 7.0%, the granulation process takes 139 minutes. The residual difference between the actual moisture content and the set value is small, indicating relatively stable control. However, the long overall time leads to a higher particle breakage rate during the granulation process. 50 The value is 690 μm, which leads to a decrease in D compared to control levels of 7.3% and 7.5%. 50 Smaller ( Figure 10 (A and B); when the moisture content is set at 7.5%, the granulation time is shorter, totaling 116 min. D 50 The value is 814 μm ( Figure 10 (C) and (D) have large residuals between the actual and set moisture values, resulting in high fluctuations and unstable batch control, which can easily lead to granulation bed collapse. When the moisture setpoint is 7.3%, the fluctuation range of the residual value is smaller, and the granulation process is less volatile. 50 The value is 727 μm ( Figure 10 Compared to conditions of 7.5% and 7.0%, 7.3% was ultimately chosen as the setpoint for the constant control moisture content.

[0076] 3.1.2 Feasibility Analysis of Gradient Control Figure 11A and C respectively present the dynamic evolution of moisture under parameter combinations of Kp=1, Ti=100 and Kp=0.85, Ti=120. Due to the characteristics of the gradient control mode, the deviation between the actual value and the set value is small, and the moisture content approaches the target value in about 50 minutes and fluctuates slightly along the preset trajectory. Figure 11 The comparison of residuals for B and D showed that the residual fluctuation amplitude under the conditions of Kp=1 and Ti=100 was significantly higher than that under the conditions of Kp=0.85 and Ti=120. Based on this, Kp=0.85 and Ti=120 were selected as the optimal control parameters for subsequent gradient control research.

[0077] Based on the particle size variation trends of setpoint control and gradient control under various PI parameter conditions, as follows: Figure 12 As shown, the results indicate that when the peristaltic pump speed is controlled by a fixed value, the overall particle size growth path is relatively aggressive, which can easily lead to the collapse of the entire equipment. In contrast, the particle size growth is more moderate in the gradient control mode, which is more suitable for production modes under complex conditions.

[0078] 3.2 Control Model Establishment and Model Selection 3.2.1 Results of manual granulation experiment Granulation experiments were conducted using manual control mode, with the peristaltic pump speed set at 15 mL / min, atomization pressure at 1.2 Bar, and inlet air temperature at 65℃. The inlet air volume was gradually increased from 0 to 70 m³ / min over time. 3 / h. The experiment followed the formulation and process requirements, but the granulation endpoint time could not be controlled. The results showed that the granulation times for the three batches were 118 min, 139 min, and 123 min, respectively, with significant differences between batches. After the liquid spraying was completed, the inlet air temperature was adjusted to 55 ℃ and the inlet air volume to 45 m³ / h. 3 / h, dry for 30 min until moisture content ≤6%. Figure 12 As shown, the evolution of moisture content in the three batches and the time change curve of PC1 score after SNV+SD-2 pretreatment showed poor reproducibility before 60 minutes. However, the process time fluctuations between batches were large under manual control mode, resulting in poor reproducibility.

[0079] PCA analysis was performed on the near-infrared spectral information obtained in real time from three batches of materials. The relationship between the PC1 loading plot and the wavelength is as follows: Figure 13 As shown in the figure. The results show that the significant difference in PC1 loading among the three batches of experimental results is mainly reflected in the 1300 nm-1500 nm wavelength band, the infrared absorption range of water molecules.

[0080] 3.2.2 Results of the setpoint control experiment A PI control algorithm (Kp=1, Ti=100) was used to conduct a batch test with constant moisture content at 7.3%. A total of three batches were produced, with all other granulation process parameters remaining consistent with manual granulation. The granulation times for the three batches were 139 min, 147 min, and 143 min, respectively, with a drying stage lasting 30 min. Figure 14 It can be seen that the moisture content tends to stabilize around 40 min; however, the second batch showed significant fluctuations at 100 min, an anomaly also confirmed by the PC1 score time variation curve in the spectral information. Overall, the moisture evolution pattern under the constant value control mode has a good correspondence with the variation trend of the principal component score in the near-infrared spectrum.

[0081] 3.2.3 Gradient Control Experiment Results A gradient control batch experiment was conducted using a PI control algorithm (Kp=0.85, Ti=120). The moisture setpoint varied according to a preset gradient curve, while all other equipment parameters remained consistent with the setpoint control. A total of three batches of production were completed. Figure 15 As shown, the granulation times for the three batches were 119 min, 120 min, and 121 min, respectively, with the drying stage lasting 30 min. Compared to the constant value control mode, the gradient control strategy significantly reduced the difference in production time between batches, showed good consistency in the trend of moisture content change, and demonstrated a more stable and controllable process, exhibiting better batch reproducibility and gentler operation.

[0082] 3.3 Particle size distribution evaluation and control model Different particle qualities can significantly impact subsequent production processes. One crucial indicator for verifying batch-to-batch consistency during granulation is particle size distribution. Table 7 shows the particle size distribution and RW values ​​for each batch at the end of granulation drying.

[0083] Table 7 Particle size information and RW values ​​for each batch

[0084] 3.4 Summary and Discussion Through statistical analysis of multiple batches of experiments, the trends in moisture content and particle growth under constant moisture content conditions of 7.0%, 7.3%, and 7.5% were compared to analyze the feasibility of constant moisture content granulation. It was found that when 7.0% was chosen as the moisture setpoint, on the one hand, the excessively long granulation time led to uneven particle quality; on the other hand, increased uncontrollable environmental risks could also lead to control failure. When the moisture content was set at 7.5%, although the final particle size was larger, the particle growth process showed a near-linear upward trend, indicating a more aggressive granulation process and a higher risk of bed collapse. Therefore, a moisture content of 7.3% was chosen for constant moisture content control.

[0085] Furthermore, the differences in the effects of PI parameters on actual control were compared. Due to the sensitivity of PID control to noise, the derivative is prone to introducing large errors. Therefore, the derivative part is often eliminated during PID parameter optimization. Through extensive experiments, two sets of control parameters (Kp=1, Ti=100; Kp=0.85, Ti=120) were finally optimized and applied to setpoint control and gradient control experiments, respectively. The optimized settings were Kp=1, Ti=100 for setpoint control and Kp=0.85, Ti=120 for gradient control. Error analysis was used to verify the reliability of the two algorithms, providing a foundation for the three control methods discussed in the next section.

[0086] During manual granulation, the moisture content remained relatively consistent over time, with all three granulation times lasting 120 minutes. The overall trend was an initial increase followed by a decrease, reflecting the gradual transformation of powder 1 into granules after absorbing moisture. Differences in the maximum moisture content led to variations in particle properties between batches after granulation. Compared to the moisture change graph, interpreting the spectral information over time using multivariate statistical projection (MPP) analysis of the PC1 score graph revealed significant differences in PC1 scores between batches. This indicates that even with consistent parameters during granulation, external uncontrollable factors can still affect moisture content, thus impacting quality uniformity. PCA analysis of the collected spectral information showed that the main difference in PC1 scores was between 1300 nm and 1500 nm, precisely the absorption wavelength of the OH bonds in water molecules. This suggests that moisture variation is a significant factor contributing to bed collapse during fluidized bed production.

[0087] In the 7.3% moisture setpoint control model, it was found that the moisture content tended to stabilize 40 minutes after the start of granulation. However, during the latter half of the granulation process, in the particle growth stage, significant fluctuations in moisture content occurred between 100 and 120 minutes due to particle agglomeration and poor fluidization. The results, as shown in the PC1 score plot, indicated that the main area of ​​fluctuation was likely due to overclocking, leading to large fluctuations in the peristaltic pump speed calculated by the PI algorithm, resulting in poor batch-to-batch granulation performance. Therefore, a more stable control method is needed to ensure batch-to-batch consistency.

[0088] Finally, gradient control was adopted to control the fluidized bed granulation process, ensuring that the actual moisture content consistently fluctuated around the setpoint. Results showed that the batch-to-batch moisture variation was smaller compared to other granulation methods, with good batch-to-batch reproducibility, relatively stable overall granulation time, a moderate particle growth trend, and relatively uniform particle quality across batches. Therefore, gradient granulation was ultimately selected as the intelligent control method for the BQTL-G fluidized bed process.

[0089] This experiment also compared the particle size distribution and RW of the three granulation methods, and the results showed that D 10 The RSDs were 5.73%, 22.61%, and 9.12%, respectively. 50 The RSDs were 10.60%, 31.18%, and 5.78%, respectively. 90 The RSD values ​​were 10.92%, 18.46%, and 4.79%, respectively. Under different methods, the gradient granulation method showed smaller batch-to-batch variations, and its RSD value was better than that of manual and constant-value granulation. Finally, we applied the gradient automatic granulation method to replace the traditional manual granulation method.

[0090] Example 3 Robustness Study of BQTL-G Fluidized Bed Granulation Control Model Based on NIRs In the previous chapter, we explored the optimal control method for three granulation processes and ultimately established a gradient control model with control parameters Kp=0.85 and Ti=120. However, the fluidized bed granulation process requires control of many key process parameters, and changes in these parameters can significantly impact the quality of the final product. Therefore, to accurately control the fluidized bed granulation endpoint and meet the required moisture content, further evaluation of the robustness of the control system is needed to ensure the reproducibility of the entire granulation process.

[0091] This experiment comprehensively examines the robustness of the system from both equipment parameter factors and external environmental factors, conducting experiments by adjusting individual parameters for each batch. The inlet air temperature has a significant impact on moisture content throughout the process. Preliminary results showed that under gradient control conditions at an inlet air temperature of 65℃, the actual moisture content fluctuated well around the set value. To investigate whether this control method can still stably control moisture changes when the temperature is changed, this experiment added granulation batch experiments at inlet air temperatures of 60℃ and 70℃ to verify the control model. Secondly, changes in atomization pressure refer to altering the atomizer's spray angle and the size of the sprayed droplets, primarily affecting particle size. Finally, from the perspective of the external environment, the experiment also examined whether increasing the humidity in the air environment affected the granulation process.

[0092] 1.1 Experimental Methods 1.1.1 Experimental Design This experiment focused on varying the intake air temperature, atomization pressure, and external ambient humidity. Two batches of experiments were conducted, and the specific variation values ​​and experimental design are shown in Table 8.

[0093] Table 8 Experimental Design Table

[0094] 1.1.2 Experimental methods for granulation process Granulation was carried out using an experimental fluidized bed. According to the formulation theory, the material weights were as follows: 800 g for each batch of powder and 1760 g for each batch of extract, with a solid content of 50%.

[0095] Before starting each batch of granulation, the fluidized bed drying system is preheated for 30 minutes, and equipment parameters, monitoring models, and control models are set. Granulation can only begin after the fluidization state has stabilized. After the preheating stage, prescription powder 1 is added and mixed continuously for 10 minutes. The moisture monitoring model is then activated, followed by the atomization spraying program for the traditional Chinese medicine extract, and the peristaltic pump automatic control program is started. Drying is completed when the actual moisture content is ≤6%. After granulation, the granules are sieved using a 10-100 mesh pharmacopoeia sieve.

[0096] 1.1.3 Equipment parameter changes Inlet air temperature robustness test Two batches of BQTL-G were produced using the above-mentioned automatic granulation method, with the inlet air temperature set to 60 ℃ and 70 ℃ respectively, the atomization pressure set to 1.2 Bar, and the ambient humidity set to 3.0 g / kg respectively.

[0097] Atomization pressure robustness test The air inlet temperature was set to 65℃, the atomization pressure was reduced to 1.0 Bar, and the ambient humidity was maintained at 3.0 g / kg. Two batches of BQTL-G were produced using the above-mentioned automatic granulation method.

[0098] 1.1.4 Changes in environmental factors The inlet air temperature was set to 65 ℃, the atomization pressure to 1.0 Bar, and the ambient humidity to 4.0 g / kg. Two batches of BQTL-G were produced using the above-mentioned automatic granulation method.

[0099] 1.2 Experimental Results 1.2.1 Temperature robustness test When the inlet air temperature was 60 ℃, and other parameters remained unchanged compared to the original production process, the overall granulation times for low-temperature batches 1 and 2 were 150 min and 128 min, respectively. The results indicate that when the inlet air temperature was set to 65 ℃, significant fluctuations in the peristaltic pump speed were observed. Under this temperature condition, the drying efficiency was significantly lower, causing the rate of moisture increase in the material to exceed the rate of moisture evaporation. The system required high-frequency adjustment of the peristaltic pump speed to effectively control the moisture increase. The deviation between the actual moisture value and the set value was controlled within 2%, indicating that the granulation process remained generally stable.

[0100] When the inlet air temperature was set to 70 ℃ and other equipment parameters remained unchanged, the granulation endpoint times for the two batches were 112 min and 113 min, respectively. At this temperature, the increased airflow volume led to a rise in the heat load per unit heated surface, and the particle dehydration rate exceeded the moisture accumulation rate, causing the extreme moisture content to fall short of the preset curve peak. However, the measured moisture content deviation from the target value was controlled within 2%, indicating good stability of the dual-batch process control and its applicability to industrial production.

[0101] 1.2.2 Robustness test of atomization pressure Under a fluidized bed atomization pressure of 1.2 Bar, the moisture content during granulation can be well tracked along the target trajectory. To verify the applicability of the control strategy at lower atomization pressures, a granulation experiment at 1 Bar was added, with all other process parameters remaining consistent. Granulation times were 138 min and 131 min, respectively. The moisture change curves in both cases followed the set trajectory, and the maximum deviation between the target value and the measured value did not exceed 2%, confirming the excellent robustness of the control system.

[0102] 1.3 Summary and Discussion This study focuses on the fluidized bed granulation process of BQTL-G (air-injected and ductile granules), systematically investigating the robustness of equipment parameters (inlet air temperature, atomization pressure) and environmental factors (ambient humidity) to verify the adaptability and reliability of the automatic granulation control model under different operating conditions. Two temperature gradients, 60 ℃ and 70 ℃, were investigated. Results showed that at the low temperature of 60 ℃, the granulation time was prolonged, with excessively rapid moisture growth in the early stage and frequent adjustments to the peristaltic pump in the later stage, but the moisture deviation was still controlled within 2%. At the high temperature of 70 ℃, the granulation time was significantly shortened, and the accelerated drying rate caused the peak moisture content to deviate from the set curve; however, the control stability was good, suggesting that this temperature range is more suitable for production. Temperature fluctuations significantly affected granulation efficiency but had a limited impact on control accuracy. Reducing the atomization pressure from 1.2 Bar to 1.0 Bar, with a granulation time of 131–138 min, the moisture growth trend matched the set curve well, with a moisture deviation not exceeding 2%. This indicates that the control model has strong robustness to reduced atomization pressure and a relatively wide adjustment range for process parameters.

[0103] Under all experimental conditions, the deviation between the actual and set moisture values ​​was controlled within 2%, indicating that the automatic control system based on moisture feedback has strong anti-interference capabilities and can effectively cope with moderate fluctuations in key process parameters, providing technical support for the continuous and stable operation of the production process. Increasing the temperature from 60℃ to 70℃ shortened the granulation time and significantly improved production efficiency. However, excessively high temperatures may lead to excessively rapid drying and changes in particle density; therefore, the optimal temperature window needs to be comprehensively evaluated in conjunction with physical quality properties. Reducing the atomization pressure to 1.0 Bar did not significantly affect control accuracy; energy consumption can be appropriately reduced while ensuring atomization effect. The environmental humidity investigation provided data support for establishing a humidity control strategy and formulating production environment standards.

[0104] The automatic granulation control model established in this invention exhibits good robustness under moderate fluctuations in equipment parameters and environmental factors, and the key process parameters have a controllable adjustment range, providing a scientific basis for the large-scale production and process parameter optimization of BQTL-G.

[0105] Example 4: Evaluation of BQTL-G Powder Quality Consistency Based on Physical and Chemical Fingerprint Maps Under the optimized process control conditions described above, this study produced 15 batches of granule samples. Referring to the company's existing quality standards, the study systematically evaluated multiple key physical quality parameters of the granules using powder characterization techniques, including particle size distribution, flowability, bulk density, and angle of repose. Based on this, a comprehensive physical fingerprint spectrum was established. Furthermore, a similarity algorithm was used to quantitatively assess the quality uniformity among batches of samples to ensure batch-to-batch consistency of physical properties. Simultaneously, the study further utilized chemical fingerprinting technology to systematically analyze and evaluate the intrinsic chemical composition of the granules. Based on the clear identification of each characteristic chemical component, a similarity evaluation was performed on the chemical fingerprint spectrum of BQTL-G, thus comprehensively verifying and completing the comprehensive study of batch-to-batch quality consistency of this formulation from both physical and chemical dimensions.

[0106] 1. Experimental Methods 2.1 Preparation process of BQTL-G 2.1.1 Pre-treatment of medicinal materials Weigh out the following ingredients according to the prescription: Astragalus membranaceus, Cicadae Periostracum, Curcuma longa, Angelica sinensis, Cinnamomum cassia, Ligusticum chuanxiong, and Alisma plantago-aquatica. After extraction and concentration, an extract is obtained. The Panax notoginseng slices are then ultra-finely pulverized to obtain Panax notoginseng powder.

[0107] 2.1.2 Granulation process The Panax notoginseng powder and excipients were mixed evenly in a certain proportion and placed into a fluidized bed expansion tank. The extract was sprayed onto the powder surface using a peristaltic pump. The equipment parameters were adjusted as follows: inlet air temperature 65 ℃, atomization pressure 1.2 Bar, and inlet air volume 10~75 m³ / h. 3 The peristaltic pump speed is constantly adjusted via a PLC controller during the granulation process. After granulation and extract spraying, during the granule drying stage, the PLC control model needs to be deactivated, and the air intake volume adjusted to 45 m³ / h. 3 The drying process is carried out at a rate of 1000 m³ / h, with an inlet air temperature of 55 ℃. When the moisture content of the particles is less than 6%, the drying process is completed, and the particles are sieved to obtain the final product. These are then numbered S1-S15.

[0108] 2.2 Experimental Method for Establishing Physical Fingerprint Spectra 2.2.1 Methods for determining physical property parameters 2.2.1.1 Moisture absorption rate Open the stoppered glass weighing bottle and place it in a desiccator containing a supersaturated NaCl solution for 24 hours. Take 2.0 g of BQTL-G granules, spread them evenly in the weighing bottle, place it in the desiccator, wait for 24 hours, weigh it accurately, and calculate the moisture absorption rate of the granules.

[0109] 2.2.1.2 Angle of repose (α) The angle of repose was measured using a powder property analyzer at 0°, 120°, and 240°, using the following formula: α=arctanH / R 2.2.1.3 Moisture content (HR) The moisture content was determined according to the moisture determination method (drying method) in Section 0832 of Part IV of the 2025 edition of the Pharmacopoeia. The proportion of water content in the test sample was calculated based on the weight loss of the granules.

[0110] 2.2.1.4 Particle size (D) 10 D 50 D 90 ) This experiment uses an online laser particle size analyzer to determine the particle size of the sample. 10 g of the sample is added to the inlet of the online laser particle size analyzer, and compressed air is connected. Using compressed air as a medium, the sample particles are passed through the inlet sequentially and evenly, and the particle size D is measured. 10 D 50 D 90 The values ​​are calculated, and the particle size distribution width (Span) and particle size range (Width) are determined using the following formulas: Span = D90 - D10 / D50 Width=D90-D10 2.2.1.5 Loose packing density (D) a ) and tap density (D c ) The particle characteristics were determined using a comprehensive powder analyzer. BQTL-G was sieved using a vibrating sieve and then allowed to fill naturally into a standard density container of known volume (V). Feeding was stopped when the material was completely filled and began to overflow the container opening. Excess powder was scraped off along the container opening using a flat scraper. The mass of the empty container was accurately measured and recorded as G1, and the total mass of the container after filling with the sample was recorded as G2. The particle size distribution (Da) was calculated based on these measurements.

[0111] Fill the Dc assembly with BQTL-G, ensuring the upper surface of the sample reaches the middle of the transparent observation window. Tighten the cap and place the assembly on a vibratory compactor for 5 minutes of continuous up-and-down vibration. Then remove the Dc assembly. Following the aforementioned measurement procedure, accurately weigh the container and record the mass as G3, and record the total mass of the container after filling with the sample as G4. Calculate the particle density (Dc) based on these measurements.

[0112] 1.2.1.6 Haunersby (IH), Carr index (I C ) and interparticle porosity (I e ) All by D a D c The calculation is as follows; 2.2.2 Determination and Standardization Conversion of Physical Indicators To more intuitively demonstrate and compare different physical quality indicators, particle flowability, packability, uniformity, and compressibility are used as primary indicators, along with α, HR, IH, and I... C I e D a D c H is a secondary indicator. Based on relevant literature, each parameter is normalized to the same interval 0-10. The conversion formulas for each parameter are shown in Table 9.

[0113] Table 9 Units, numerical ranges, and standardization methods for physical parameters

[0114] 2.2.3 Similarity Analysis Physical fingerprinting can be used to evaluate the quality consistency between batches of particles. The similarity evaluation method is used to evaluate the similarity between batches from an overall perspective. The Pearon correlation coefficient between each batch is calculated using Origin software. The closer the absolute value of the similarity is to 1, the more similar the physical properties are. Cluster analysis is performed using SPSS software.

[0115] 2.2.4 Cluster Analysis To better evaluate the differences in physical properties among the 15 batches of BQTL-G, 10 physical property indicators obtained from the particles were imported as variables into SPSS software for cluster analysis of the samples.

[0116] 2.3 Experimental methods for establishing chemical fingerprint spectra 2.3.1 Chromatographic conditions Waters ACQUITY UPLC ® BEH Shield RP18 column (100×2.1 mm, 1.7 μm); mobile phase acetonitrile (A)-0.05% formic acid (B), gradient elution as shown in Table 10; flow rate 0.2 mL / min; column temperature 30 ℃; detection wavelengths 254 nm and 203 nm; injection volume 2 μL.

[0117] Table 10 Gradient elution program

[0118] 2.3.2 Preparation of reference solution Accurately weigh appropriate amounts of the reference standards gentianin, gentianin, caffeic acid, ferulic acid, ligustrazine I, and cinnamic acid, and add 70% methanol to prepare a mixed reference solution containing 0.080 mg ferulic acid, 0.052 mg ligustrazine I, 0.016 mg caffeic acid, 0.010 mg gentianin, 0.125 mg gentianin, and 0.020 mg cinnamic acid per 1 mL.

[0119] 2.3.3 Preparation of test solution Take approximately 5 g of BQTL-G, accurately weigh it, and transfer it to a 150 mL stoppered conical flask. Accurately add 20 mL of 70% methanol solution and weigh the total mass. Then place it in an ultrasonic extractor and sonicate it for 30 min at a power of 800 W and a frequency of 40 kHz. After the extract has cooled to room temperature, replenish the lost mass with 70% methanol, shake thoroughly, filter through filter paper, and collect the filtrate to obtain the test solution.

[0120] 2.3.4 Methodological Examination 2.3.4.1 Precision Examination Take the same sample solution and inject it under the chromatographic conditions in section “2.3.1” for determination 6 times. Use the gentianin peak as a reference and calculate the RSD value.

[0121] 2.3.4.2 Repeatability Test Take 6 parallel samples of granules from the same batch of production, prepare test solutions according to the method in section "2.3.3", and determine them under the chromatographic conditions in section "2.3.1". Calculate the RSD value using the gentianin peak as a reference.

[0122] 2.3.4.3 Stability Assessment Take granules from the same batch of production, prepare test solution according to the method in section "2.3.3", and inject and determine at room temperature at 0, 2, 4, 8, 12 and 24 h under the chromatographic conditions in section "2.3.1". Calculate the RSD value with the gentianin peak as a reference.

[0123] 2.3.5 Fingerprint Mapping 2.3.5.1 Atlas Generation Fifteen batches of BQTL-G samples were taken, and test solutions were prepared according to the method described in section "2.3.3". The samples were then injected and analyzed under the chromatographic conditions specified in section "2.3.1". The obtained chromatographic data were imported into the "Traditional Chinese Medicine Chromatographic Fingerprint Similarity Evaluation System (2012 Edition)" software platform. A time window width of 0.2 min was set, and a multi-point correction algorithm was used for full-spectrum matching analysis of the chromatographic peaks.

[0124] 2.3.5.2 Similarity Analysis The obtained chromatographic data were imported into the "Traditional Chinese Medicine Chromatographic Fingerprint Similarity Evaluation System (2012 Edition)" software platform. The generated control fingerprint chromatogram was used as a reference standard to calculate the similarity between each batch of samples and the control chromatogram.

[0125] 2. Experimental Results 3.1. Results of Particle Preparation By setting relevant parameters, a total of 15 batches of BQTL-G were produced and numbered S1 to S15. The production results are shown in Table 11.

[0126] Table 11 Batch Numbers of BQTL-G Production (Batch 15)

[0127] 3.2. Quality Consistency Evaluation Based on Physical Fingerprint Spectra 3.2.1 Determination of physical properties of powder and establishment of physical fingerprint spectrum The results of the physical properties of 15 batches of BQTL-G are shown in Table 12.

[0128] Table 12. Physical property values ​​of 15 batches of BQTL-G and control batch R.

[0129] After normalizing the secondary data obtained from the table above, the data was imported into Origin software to obtain 15 batches of physical fingerprint spectra, as shown in the table above. Figure 16 Data analysis using Origin software was performed to calculate the Pearson correlation coefficients among different batches of BQTL-G. The similarity results among the batches all showed >0.94, indicating that the 15 batches of BQTL-G have high consistency in physical properties.

[0130] 3.2.2 Cluster Analysis System cluster analysis was performed using SPSS software, and the results are as follows: Figure 17 As shown, when the squared Euclidean distance is set to 5, the 15 batches of BQTL-G samples can be divided into three clusters: Cluster I includes S1, S2, S12, and S13; Cluster II includes S3, S7, S10, S11, and S15; and Cluster III includes S4, S5, S6, S8, S9, and S14. This cluster analysis indicates that the physical quality attributes of different batches of samples are relatively uniformly distributed.

[0131] 3.3. Quality Consistency Evaluation Based on Chemical Fingerprint Spectra 3.3.1 Methodological Investigation Results 3.3.1.1. Precision Experiment The chromatographic peak of gentianin was selected as the reference peak, and the relative retention time and relative peak area of ​​each common peak were calculated. The results showed that all RSD values ​​were less than 3.0%, indicating good instrument precision. 3.3.1.2. Stability Test The chromatographic peak of gentianin was selected as the reference peak, and the RSD values ​​of the relative retention time and relative peak area of ​​each common peak were calculated. The results showed that all RSD values ​​were less than 3.0%, indicating that the method has good stability.

[0132] 3.3.1.3. Repeatability Experiment The chromatographic peak of gentianin was selected as the reference peak, and the RSD values ​​of the relative retention time and relative peak area of ​​each common peak were calculated. The results showed that all RSD values ​​were less than 3.0%, indicating that the method has good repeatability.

[0133] 3.3.2 BQTL-G fingerprint pattern establishment and similarity evaluation Prepare 15 batches of BQTL-G test solutions according to section “2.3.3”. Analyze each test solution according to the chromatographic conditions in “2.3.1” and record the chromatograms. Establish fingerprint chromatograms and evaluate similarity according to the conditions in “2.3.5”. The results confirm that the number of common peaks in the 15 batches of samples is the same.

[0134] 3.3.3 Common Peak Identification Take the mixed reference solution from section "2.3.2" and determine it according to the chromatographic conditions in "2.3.1" to obtain the chromatogram of the reference solution. Identify the common peaks in BQTL-G by comparing the retention times of each peak with those in the chromatogram of the test solution. The results show that the peaks are caffeic acid, ferulic acid, ligustilide I, gentiopicrin, cinnamic acid, and gentiopicrin.

[0135] 3. Summary and Discussion This chapter's experiments used fingerprinting technology to verify the batch-to-batch quality consistency of BQTL-G under the intelligent control system. The results show that, using the established monitoring method, the physical properties of the 15 batches of BQTL-G produced were relatively stable, and their chemical compositions were relatively similar. It exhibits good batch-to-batch reproducibility, allowing this model to be practically applied to production.

Claims

1. A method for online quality monitoring of traditional Chinese medicine granulation process using near-infrared spectroscopy combined with fluidized bed granulation, characterized in that, The monitoring method includes the following steps: (1) Granulation process and sample collection Weigh out the compound extract of traditional Chinese medicine and the fine powder of mixed traditional Chinese medicine, and granulate them using fluidized bed granulation. First, preheat the fluidized bed drying system for 30 minutes, then add the fine powder of mixed traditional Chinese medicine and continue mixing for 8-12 minutes. Then start the atomization spraying program of traditional Chinese medicine extract. The following key process parameters are controlled during the spraying process: pump flow rate 10-20 mL / min, inlet air temperature 60-70℃, atomization pressure 0.8-1.5 Bar, ambient humidity 2.0-4.0 g / kg, total granulation time 110-130 min, drying time 20-40 min. The criterion for judging the granulation endpoint is that the moisture content of the particles drops to below 6.0%. When comparing control methods, samples are taken after all the binder has been sprayed in to obtain the test samples. (2) Near-infrared spectral acquisition A near-infrared spectrometer was inserted into the fluidized bed cavity, and spectra were acquired in diffuse reflectance mode. The moisture and particle size values ​​were analyzed online using a near-infrared monitoring model. Before acquiring material information, the surrounding air environment was scanned for dark current, and then a polytetrafluoroethylene plate was collected as background spectrum to deduct the interference of the PAT-U device on the spectral information. The absorption wavelength range of the device was set to 908.1-1676.2nm, the integration time was set to 6.6ms, the number of spectral acquisitions was 250, and the air in the viewing window was purged once every 10s. When the model was established, Micro NIR Pro V3.2 was used to collect spectral information online. When the model was predicted, the moisture and particle size models established in UnscramblerX10.4 were used to predict the moisture and particle size values ​​in real time. This invention also adopts a PID control system algorithm, which controls the peristaltic pump speed based on the negative feedback of the difference between the theoretical value and the preset value of particle moisture content, so as to achieve precise control of moisture to the preset value. (3) Measurement of monitoring indicators Establish "moisture content" and "particle size" as quality testing indicators; (4) Spectral preprocessing and band selection The raw spectral preprocessing methods examined include: vector normalization, derivative spectroscopy, smoothing, multivariate scattering correction, standard normal variable transformation, baseline correction, and combined strategies of multiple methods. After removing impurity information from the original NIR spectra, the band selection is performed using a variable iterative spatial shrinkage algorithm, which involves the following steps: ① Based on the weighted binary sampling method, the original 12-band variables are divided into several subsets, each variable is assigned an initial weight of 0.5, and N subsets are generated after N random samplings. ② Input each subset of variables into the PLSr model, record the mean squared error of cross-validation, select the model with the lowest RMSECV as the optimal set, calculate the frequency of occurrence of variables and update the weights, and record the mean RMSECV. The formula for calculating the weight of the nth variable is as follows: In the formula, Wn represents the weight, fn represents the frequency of the nth variable in the optimal model set, and Nbest represents the number of optimal model sets. ③ Iterative execution steps (1) and (2): WBMS sampling weights are updated according to the above formula. When RMSECV no longer decreases, the iteration is terminated. At this time, the band with a weight of 1 is defined as the core information band, the band with a weight of 0 is regarded as the interference band, and the band with a weight between 0 and 1 is classified as the weak information band. The three types of bands are arranged in descending order of weight, and the PLSR model is constructed according to the size of the variable set from large to small. The combination corresponding to the minimum value of RMSECV is used as the final feature band. (5) Establishment of chemometric methods Spectral data was recorded simultaneously during sampling and correlated with the quality inspection index data from step (3). After spectral preprocessing, a quantitative analysis model was established using the partial least squares regression (PLSR) algorithm. The model evaluation index was R0. 2 C R 2 P RMSEC, RMSEP, R 2 The optimized model is then applied to the online predictive analysis of the fluidized bed granulation process.

2. The online quality monitoring method as described in claim 1, characterized in that, The step (1) is a compound extract of traditional Chinese medicine, which is a qi-tonifying and meridian-clearing extract, and a mixed fine powder of traditional Chinese medicine, which is Panax notoginseng micro powder and dextrin. The ratio of compound extract of traditional Chinese medicine to mixed fine powder of traditional Chinese medicine is 1:2-2.

5.

3. The online quality monitoring method as described in claim 1, characterized in that, The fluidized bed granulation peristaltic pump speed control model in step (1) adopts a gradient control mode, with Kp=0.85 and Ti=120 as the optimized control parameters.

4. The online quality monitoring method as described in claim 1, characterized in that, The following key process parameters are controlled during the spraying process in step (1): pump flow rate 15 mL / min, inlet air temperature 65℃, atomization pressure 1.2 Bar, and ambient humidity 3.0 g / kg.

5. The online quality monitoring method as described in claim 1, characterized in that, The frequency and interval of sampling in step (1) granulation are as follows: sampling once every 5 minutes during the process and once every 5 minutes during the drying stage.

6. The online quality monitoring method as described in claim 1, characterized in that, Take the sample to be tested in step (1), place it in the inlet of the linear laser particle size analyzer, connect compressed air, and use compressed air as a medium to make the sample particles pass through the inlet uniformly in sequence. Repeat two or three times, and record the average value as the D of the current sample. 10 D 50 and D 90 value.

7. The online quality monitoring method as described in claim 1, characterized in that, In step (4), the spectral preprocessing uses the SPXY algorithm to divide the dataset and the VISSA algorithm to filter key variable information. The SD-1+De-trending preprocessing method is selected to establish the PLSR model.

8. The online quality monitoring method as described in claim 1, characterized in that, The control system in step (2) is based on the Siemens S7-300 PLC and TP1200 HMI architecture. The PLC obtains near-infrared prediction values ​​through communication via a unified architecture. The human-machine interface (HMI) displays the moisture trend graph in real time and corrects the parameters. The PID module calculates the pump speed based on the feedback deviation. The Profinet communication transmits the command to the frequency converter, which links the peristaltic pump to adjust the speed, thereby realizing closed-loop precise control of moisture.

9. The online quality monitoring method as described in claim 1, characterized in that, In step (3), the "moisture content" was determined by drying method according to the 2025 edition of the Chinese Pharmacopoeia, and the "particle size determination" was performed using an online laser particle size analyzer.

10. The online quality monitoring method for traditional Chinese medicine processes as described in claims 1-9 can be used in the production process of compound granules of traditional Chinese medicine.

Images