Intelligent dynamic monitoring method and system for manufacturing process of food-medicine homologous plant concentrate

Abstract

The application discloses a near-infrared spectrum intelligent dynamic monitoring method and system for a food and medicine homologous plant concentrate manufacturing process, establishes a multi-batch sample database covering a complete concentration track, takes SG smoothing as an optimal spectrum pretreatment scheme, creates an index self-adaptive variable selection strategy, uses competitive self-adaptive reweighted sampling for total polysaccharide, uses a continuous projection algorithm for soluble solid content, and uses full-spectrum PLS for total polyphenol; the overall superiority of PLS to SVR is used to introduce a slope-intercept confidence ellipse as a statistical unbiasedness verification tool to establish an optimal prediction model; the optimal prediction model is deployed to an online system to sequentially execute SG pretreatment, variable selection on real-time spectrum flow of a diffuse reflection optical fiber probe, and then input the optimal prediction model to obtain prediction values of three quality indexes, and the preset threshold is combined to realize automatic judgment of a concentration end point. The application realizes real-time online synchronous quantitative prediction of multiple key quality indexes and automatic judgment of a concentration end point.

Description

Technical Field

[0001] This invention belongs to the field of food quality testing and process analysis technology, specifically relating to a near-infrared spectroscopy intelligent dynamic monitoring method and system applicable to the manufacturing process of concentrated extracts from food and medicinal plants. It can perform online, real-time, non-destructive synchronous quantitative detection of multiple key quality indicators such as total polysaccharide, total polyphenol, and soluble solids (SSC) content, and automatically determine the concentration endpoint during the concentration manufacturing process. Background Technology

[0002] Food and medicinal plants are a group of plants that, according to traditional Chinese medicine theory, can be used as food and also have certain medicinal value. These include wolfberry (Lycium barbarum L.), polygonatum (Polygonatum sibiricum F. Delaroche), raspberry (Rubus chingii Hu), licorice (Glycyrrhiza uralensis Fisch.), astragalus (Astragalus membranaceus Bge.), jujube (Ziziphus jujuba Mill.), longan pulp (Dimocarpus longanLour.), codonopsis pilosula (Franch.) Nannf., and angelica sinensis (Oliv.) Diels. With the rapid expansion of the health food and functional beverage market, the industrial production scale of concentrated liquid products made from a combination of various food and medicinal plants has continued to grow. Quality control in the manufacturing process has become a core element in ensuring product safety and efficacy stability.

[0003] During the extraction and concentration process of edible and medicinal plant concentrates, key quality indicators such as total polysaccharides, total polyphenols, and solute concentration (SSC) continuously and dynamically change with the increasing concentration level. Total polysaccharide content reflects the concentration state of the formulation and the enrichment degree of polysaccharide active ingredients; total polyphenol content characterizes the antioxidant activity level of the product and the batch-to-batch consistency of phenolic components; SSC comprehensively reflects the total solute amount and is one of the main reference indicators for determining the concentration endpoint in industry. The synergistic changes of these three indicators constitute a complete description of the quality status of the concentration process.

[0004] However, current industrial production still relies on traditional offline wet chemical analysis methods for detecting these indicators: total polysaccharides are analyzed using the phenol-sulfuric acid colorimetric method, which takes about 2 hours per analysis; total polyphenols are analyzed using the Folin-Ciocalteu method, which requires 90 minutes of incubation in the dark before measurement; and SSC is measured using a digital refractometer with manual sampling. All three methods require periodic manual sampling, making continuous monitoring impossible. Each quality decision lags behind the production process by at least 2 hours, severely restricting the ability to finely control the concentration process and increasing the risk of batch-to-batch product quality fluctuations.

[0005] Near-infrared (NIR) spectroscopy, based on molecular vibrations, detects the overtone and combination frequency absorptions of OH, CH, and NH bonds in the 780-2500 nm range. It can acquire multi-component fingerprint information of a sample within seconds, requiring no sample pretreatment, and has been established as a core sensing method within the Process Analytical Technology (PAT) framework by regulatory agencies such as the US FDA and EMA. Combining NIR spectroscopy with chemometric models allows for simultaneous real-time prediction of multiple quality indicators without interrupting production, fundamentally overcoming the lag inherent in traditional detection methods.

[0006] In existing technologies, near-infrared spectroscopy has been used to detect single indicators in single herbal preparations, but research on simultaneous online monitoring of multiple indicators covering the complete concentration trajectory in concentrated solutions of various mixed edible and medicinal plants is still lacking. Furthermore, existing research generally suffers from the following shortcomings: First, variable selection algorithms lack specific adaptation to the spectral characteristics of different analytes. Using a uniform variable selection strategy for indicators with wide and narrow dynamic ranges leads to overfitting of wide-range indicator models and loss of characteristic signals for narrow-range indicators, resulting in suboptimal model accuracy. Second, model evaluation relies solely on scalar indicators such as root mean square error and coefficient of determination, lacking rigorous verification of the statistical unbiasedness of the slope-intercept joint distribution, thus failing to meet the requirements of the PAT framework. Third, there is a lack of specialized processing methods for online dynamic diffuse reflectance signals (affected by flow rate fluctuations, temperature gradients, and bubble interference), as well as a complete technical implementation plan from offline calibration to online deployment.

[0007] Therefore, there is an urgent need to develop a near-infrared spectroscopy intelligent dynamic monitoring method and system for the manufacturing process of concentrated extracts from food and medicinal plants, so as to realize real-time online synchronous quantitative prediction of multiple key quality indicators and automatic determination of concentration endpoint. Summary of the Invention

[0008] The purpose of this invention is to address the shortcomings of existing technologies by providing a near-infrared spectroscopy-based intelligent dynamic monitoring method and system for the manufacturing process of concentrated extracts from food and medicinal plants. This invention employs an adaptive variable selection strategy, combined with confidence ellipse statistical unbiasedness verification, to achieve real-time online synchronous prediction of three key quality indicators: total polysaccharides, total polyphenols, and SSCs. It also provides a complete online monitoring system solution from spectral acquisition to quality decision-making.

[0009] The present invention achieves the above-mentioned technical objectives through the following technical means.

[0010] Near-infrared spectroscopy intelligent dynamic monitoring method for the manufacturing process of concentrated extracts from food and medicinal plants:

[0011] Samples were systematically collected from multiple key production stages of the compound concentrate production line of food and medicine homologous plants, and supplemented with random representative samples from different batches. Reference values ​​of three quality indicators, namely total polysaccharide, soluble solids content and total polyphenols, were determined to construct a multi-batch sample database.

[0012] Near-infrared diffuse reflectance spectra of each sample were collected, and various preprocessing strategies were compared using 10-fold cross-validation to confirm that SG smoothing was the optimal spectral preprocessing method.

[0013] An adaptive variable selection strategy was adopted, and competitive adaptive reweighted sampling, continuous projection algorithm and full-spectrum PLS were used to select variables for quality indicators with different dynamic ranges and spectral response characteristics.

[0014] For each quality indicator, a slope-intercept joint 95% confidence ellipse is constructed using the partial least squares regression model. The optimal prediction model for each quality indicator is determined by using the inclusion of the ideal point in the confidence ellipse as the statistical unbiasedness criterion.

[0015] The optimal prediction model is deployed to the online system. SG preprocessing and variable selection are performed sequentially on the real-time spectral stream of the diffuse reflection fiber optic probe. Then, the optimal prediction model is input to obtain the predicted values ​​of three quality indicators. Combined with preset thresholds, the concentration endpoint is automatically determined.

[0016] Furthermore, the edible and medicinal plants include, but are not limited to, four or more of the following compound combinations: wolfberry, polygonatum, raspberry, licorice, astragalus, jujube, longan pulp, peach kernel, codonopsis, and angelica; the key production stages are nine stages from the initial water extract to the final concentrated product, with no fewer than 10 samples collected at each stage; the reference values ​​of the quality indicators are confirmed in the following ways: the total polysaccharide content is determined by the phenol-sulfuric acid colorimetric method at 490 nm, the total polyphenol content is determined by the Folin-Ciocalteu method at 765 nm, and the SSC is determined by a digital refractometer.

[0017] Furthermore, the near-infrared diffuse reflectance spectrum was collected in the wavelength range of 780 to 1080 nm, and 921 effective spectral variables were extracted; the window length of the SG smoothing was 9 points and the polynomial order was 2; the lowest root mean square error was used as the evaluation criterion, and 10-fold cross-validation confirmed that the SG smoothing performance was superior to SNV, MSC, mean centering and SG+SNV combined processing.

[0018] Furthermore, the adaptive variable selection strategy for the indicator is specifically as follows:

[0019] For the total polysaccharide index, with a relative standard deviation (RSD) of 37.7%, a competitive adaptive reweighted sampling algorithm was used to screen feature variables: spectral variables were screened by combining Monte Carlo sampling with adaptive weighted partial least squares regression iteration, and the optimal variable combination was determined based on the cross-validation error, achieving a variable reduction of 88.6%.

[0020] For the SSC index, RSD=36.8%, the continuous projection algorithm was used to screen feature variables: information duplication between variables was eliminated by orthogonal projection, and the predictive performance of different variable combinations was evaluated by cross-validation, achieving a 92.9% variable reduction.

[0021] For the total polyphenol index, with an RSD of 7.5%, a full-spectrum partial least squares regression model was used for modeling, with all effective spectral variables as feature variables.

[0022] Furthermore, the multi-batch sample database was randomly divided into a calibration set and a prediction set at 70% and 30% respectively. Nine models were established for the three quality indicators: Full-PLS, CARS-PLS, and SPA-PLS, for a total of nine models. Simultaneously, a corresponding PLS-SVR model was also established. RMSEC, RMSEP, and R... 2 Based on a comprehensive evaluation of predictive performance using both PLS and RPD, the PLS model outperforms the SVR model in all indicator and variable configurations. Therefore, the partial least squares regression model is chosen to construct the confidence ellipse.

[0023] Furthermore, principal component analysis (PCA) was performed on the SNV-normalized spectral matrix of the near-infrared diffuse reflectance spectra of all samples, using Hotelling's T... 2 Statistical screening and elimination of spectral anomalies were performed. Then, kernel density estimation was used to visualize the continuous and ordered distribution of samples in the PCA score space for each production stage, verifying the monotonicity of the spectral trajectories of all key production stages and confirming that the global single partial least squares model can cover the complete condensed trajectory.

[0024] Furthermore, by combining preset thresholds, the concentration endpoint is automatically determined, specifically as follows:

[0025] The soluble solids content prediction value is the core judgment indicator. When the soluble solids content prediction value reaches the preset endpoint threshold, the endpoint prediction signal is triggered. The confirmation condition is that three consecutive sampling values ​​are not lower than the threshold. After confirmation, the endpoint judgment signal is output to the distributed control system, which supports automatic shutdown or transfer to the next process. At the same time, the predicted values ​​of total polysaccharide and total polyphenols are monitored in real time to see if they exceed the preset threshold. If either indicator exceeds the limit, an audible and visual alarm is triggered to prompt the operator to check the process parameters.

[0026] Furthermore, for different food and medicinal plant formulation systems, slope-bias correction was performed using a small number of new formulation samples to achieve lightweight cross-formulation optimal prediction model transfer.

[0027] A near-infrared spectroscopy intelligent dynamic monitoring system for the manufacturing process of concentrated extracts from edible and medicinal plants includes:

[0028] The online spectral acquisition module consists of a near-infrared spectrometer, a halogen lamp light source, and an industrial control acquisition card. The near-infrared spectrometer is connected to a diffuse reflection fiber optic probe, which is installed at the online detection position of the circulation pipeline of the concentration equipment to acquire the near-infrared diffuse reflection spectrum of the flowing concentrate in real time. The spectral acquisition is carried out in conjunction with the halogen lamp light source.

[0029] The spectral preprocessing module automatically performs dark current correction, nonlinear correction, and SG smoothing on the single-frame spectrum acquired online, and outputs a preprocessed spectral vector.

[0030] The indicator prediction module includes prediction models based on different quality indicators, used to calculate the predicted values ​​of total polysaccharide, total polyphenol and soluble solids content based on preprocessed spectral data; the prediction model includes a feature variable selection module and a partial least squares regression prediction model, which outputs the corresponding quality indicator prediction results in real time based on the input spectral information.

[0031] The quality decision and alarm module compares the predicted value of the indicator with the preset threshold in real time. When the soluble solids content reaches the endpoint threshold or any indicator exceeds the threshold, it triggers an audible and visual alarm and outputs a concentration endpoint judgment signal to the distributed control system (DCS).

[0032] The data management and visualization module records spectral data streams, predicted value time series, and alarm events in real time, provides dynamic curve visualization of quality indicators in the concentration process, supports data comparison between batches and model drift diagnosis, and saves complete electronic batch records.

[0033] In the above technical solution, the data management and visualization module includes a model maintenance submodule, which supports lightweight model updates through slope-bias correction. When the slope deviation of the dynamic curve exceeds ±5% or the intercept deviation exceeds ±3%, an update prompt is automatically triggered.

[0034] Compared with the prior art, the present invention has the following beneficial effects:

[0035] 1. This invention enables simultaneous online detection of multiple indicators, reducing the detection cycle from several hours to sub-seconds. This invention collects near-infrared spectra of multiple batches of samples covering the complete concentration trajectory (S1 to S9, a total of 9 key stages), achieving simultaneous real-time prediction of three indicators: total polysaccharides, total polyphenols, and SSCs in concentrated extracts of four or eight mixed edible and medicinal plants. The optimal model RPD values ​​for the three indicators reach 7.02, 3.63, and 7.60, respectively. Total polysaccharides and SSCs meet the requirements for high-precision online process control (RPD not less than 7.0), and total polyphenols meet the requirements for inter-batch trend monitoring (RPD not less than 3.5), overcoming the inefficient mode of traditional wet chemical methods that take more than 2 hours per batch.

[0036] 2. This invention establishes an adaptive feature variable selection strategy for indicators, systematically revealing the indicator specificity of the optimal model and providing methodological guidance for the design of multi-indicator near-infrared calibration schemes. This invention demonstrates that: CARS is suitable for indicators with wide dynamic range (RSD > 30%) and multi-band contributions, improving RPD to 7.02 with 88.6% variable reduction; SPA is suitable for non-collinear feature indicators, achieving the highest RPD (7.60) with 92.9% variable reduction; Full-PLS is suitable for weak signal indicators with narrow dynamic range (RSD < 10%). The specific applicability conditions of the three strategies are systematically verified within the same multi-indicator system, avoiding the accuracy loss caused by a unified variable selection strategy.

[0037] 3. This invention introduces slope-intercept confidence ellipse as a key tool for screening narrow dynamic range index models, solving the methodological challenge of selecting low variance index models. For low variance indices such as total polyphenols with RMSEP differences of only 0.0003 mg / mL, traditional scalar error indices cannot distinguish between superior and inferior models. This invention reveals through confidence ellipse analysis that Full-PLS has a configuration that meets the requirement of statistical unbiasedness (compact ellipse containing ideal points), while CARS-PLS suffers from systematic underestimation (slope less than 1) and SPA-PLS exhibits intercept bias. This enables a decisive distinction between the three models, meeting the statistical validation requirements of the PAT framework.

[0038] 4. This invention verifies that the linear PLS model is the optimal framework for computationally efficient online deployment. PLS outperforms SVR in RPD across all metrics and variable configurations (improving by 0.4 to 5.5), confirming that the spectrum-concentration relationship is dominated by linear Beer-Lambert behavior. It supports sub-millisecond real-time feedback for standard embedded processors, saves approximately 80% of computing resources compared to SVR, and significantly reduces the threshold for online deployment and hardware costs.

[0039] 5. This invention provides a complete PAT deployment roadmap, covering multiple formulation systems and exhibiting good cross-formulation applicability. The method of this invention has been validated through dual-formulation experiments: static transmission experiments with four herbs (goji berries, polygonatum, raspberries, and licorice) and dynamic diffuse reflection online experiments with eight herbs (astragalus, jujube, longan pulp, goji berries, peach kernel, codonopsis, angelica, and licorice). Lightweight cross-formulation model transfer is achieved through slope-bias correction of a small number of samples, providing a low-cost technical path for the rapid deployment of products with different formulations. Simultaneously, it provides a clear technical interface for subsequent integration with DCS to achieve automatic endpoint determination. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 A schematic diagram of the industrial field installation of the online near-infrared monitoring system of the present invention is shown;

[0042] Figure 2 A flowchart of the intelligent dynamic monitoring method for near-infrared spectroscopy of the present invention is shown;

[0043] Figure 3(a) is a scatter plot (including the ideal line) of the predicted values ​​against the reference values ​​of the calibration set and prediction set of the total polysaccharide CARS-PLS model of this invention.

[0044] Figure 3(b) shows the residual distribution of the total polysaccharide CARS-PLS model of this invention;

[0045] Figure 3(c) is a slope-intercept 95% confidence ellipse plot of the total polysaccharide CARS-PLS model of the present invention (ideal points are marked with crosses).

[0046] Figure 3(d) is a scatter plot (including the ideal line) of the predicted values ​​of the calibration set and prediction set of the SSC SPA-PLS model of the present invention against the reference values.

[0047] Figure 3(e) shows the residual distribution of the SSC SPA-PLS model of the present invention;

[0048] Figure 3(f) is a slope-intercept 95% confidence elliptic plot of the SSC SPA-PLS model of the present invention (ideal points are marked with crosses).

[0049] Figure 3(g) is a scatter plot (including the ideal line) of the predicted values ​​against the reference values ​​of the calibration set and prediction set of the total polyphenol Full-PLS model of this invention.

[0050] Figure 3(h) is a residual distribution diagram of the total polyphenol Full-PLS model of the present invention;

[0051] Figure 3(i) is a slope-intercept 95% confidence ellipse plot of the total polyphenol Full-PLS model of the present invention (ideal points are marked with crosses). Detailed Implementation

[0052] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0053] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments of the present invention. When the terms include or comprise are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0054] The near-infrared spectroscopy intelligent dynamic monitoring system for the manufacturing process of concentrated extracts of food and medicinal plants of the present invention includes the following functional modules:

[0055] (1) Online spectral acquisition module (e.g.) Figure 1 As shown): It consists of a halogen lamp light source (LS-HA), a near-infrared spectrometer (collecting wavelengths of 780 to 1080 nm, with a resolution of no more than 1 nm), and an industrial control acquisition card. The near-infrared spectrometer is connected to a diffuse reflection fiber optic probe. The diffuse reflection fiber optic probe is installed in a non-contact manner through the optical window of a steam-resistant and clean-in-liquid (CIP) optics at the online detection position of the circulation pipeline of the concentration equipment (such as the circulation outlet pipe section). The near-infrared diffuse reflection spectrum of the flowing concentrate is collected in real time. The spectral collection is carried out in conjunction with the halogen lamp light source. The collection frequency is set according to production requirements (recommended once every 30 seconds to 5 minutes), and the collection time for each frame of spectrum does not exceed 2 seconds. Figure 1 The diagram showcases the production line equipment layout (including process equipment for extraction, filtration, concentration, secondary concentration, vacuum degassing, and tertiary concentration), the installation location of the fiber optic diffuse reflection probe on the circulation pipeline, the process flow, and probe installation details. The nine sampling stages, S1 to S9, are distributed across various process steps in the concentration process, covering the complete concentration trajectory from the initial water extract to the final concentrated product.

[0056] (2) Spectral preprocessing module: Automatically performs dark current correction, nonlinear correction and SG (Savitzky-Golay) smoothing on the single-frame spectrum acquired online, and outputs preprocessed spectral vectors; no reference spectrum is required, and real-time streaming spectral data can be processed directly.

[0057] (3) Indicator prediction module: This module includes prediction models based on different quality indicators, used to calculate the predicted values ​​of total polysaccharide, total polyphenol, and soluble solids (SSC) content based on the preprocessed spectral data. The prediction model includes a feature variable selection module and a partial least squares regression (PLS) prediction model, which can output the corresponding quality indicator prediction results in real time based on the input spectral information.

[0058] (4) Quality decision and alarm module: The predicted value of the indicator is compared with the preset threshold in real time. When the SSC reaches the end threshold or any indicator exceeds the threshold, an audible and visual alarm is triggered. The condensed end-point judgment signal is output to the distributed control system DCS through the standard OPC-UA interface, which supports automatic shutdown or transfer to the next process.

[0059] (5) Data Management and Visualization Module: Real-time recording of spectral data stream, predicted value time series and alarm events, providing visualization of dynamic curves of quality indicators in the concentration process, supporting data comparison between batches and model drift diagnosis, and saving complete electronic batch records; includes a model maintenance sub-module, supporting lightweight model updates through slope-bias correction, and automatically triggering update prompts when the slope deviation of the dynamic curve exceeds ±5% or the intercept deviation exceeds ±3%.

[0060] like Figure 2 As shown, the near-infrared spectroscopy intelligent dynamic monitoring method for the manufacturing process of concentrated plant extracts that are both food and medicine of the present invention includes the following steps:

[0061] Step 1: Construct a multi-batch sample database. Systematically collect samples from nine key production stages of the concentration process of the food and medicinal plant compound concentrate production line, from S1 (initial aqueous extract) to S9 (final concentrated product), with no fewer than 10 replicates for each stage. S1 to S9 are nine key sampling stages in the concentration process of the food and medicinal plant compound concentrate, divided according to the concentration progress. S1 corresponds to the initial aqueous extract stage, S9 corresponds to the final concentrated product stage, and S2 to S8 are seven intermediate gradient sampling stages between S1 and S9. These stages are evenly distributed according to the concentration process at equal time intervals or equal density gradients, maintaining a clear and identifiable concentration difference between adjacent stages (e.g., sampling one stage for every approximately 5°Brix increase; the specific interval can be adjusted according to the actual material characteristics). Together, they form a complete and smooth transition sequence from low to high concentration; each sampling stage is determined based on the time progression and changes in physicochemical indicators during the actual production process, used to characterize the complete concentration trajectory rather than corresponding to independent process equipment nodes; at the same time, random representative samples from different batches and different production dates are added to introduce inter-batch variability; the total polysaccharide content (mg / mL) of each sample (including collected samples and supplementary samples) is determined by the phenol-sulfuric acid colorimetric method, the total polyphenol content (mg / mL) by the Folin-Ciocalteu method, and the SSC (°Brix) by the digital refractometer; a multi-batch sample database covering the complete concentration trajectory and inter-batch variability is constructed from each sample and its corresponding measurement indicators.

[0062] Step 2: Near-infrared diffuse reflectance spectral acquisition, determination of optimal preprocessing method, and sample quality check. Near-infrared diffuse reflectance spectra (i.e., 921 effective spectral variables) of each sample were acquired using a near-infrared spectrometer within the wavelength range of 780 to 1080 nm. A 10-fold cross-validation system was selected to compare five near-infrared spectral preprocessing strategies: SG smoothing (setting the sliding window length to 9 data points and performing second-order polynomial least squares fitting on the data points within the sliding window), standard normal variable transformation (SNV), multivariate scattering correction (MSC), mean centering, and SG+SNV combined processing. The optimal preprocessing scheme was determined based on the lowest root mean square error (RMSECV) of the cross-validation of each sample's measurement indicators. Validation showed that SG smoothing was the optimal preprocessing scheme for all three indicators. This method requires only single-frame spectral processing, does not require reference spectra or overall sample statistics, and can be directly applied to real-time spectral stream processing.

[0063] Simultaneously, principal component analysis (PCA) was performed on the SNV-normalized spectral matrix of the near-infrared diffuse reflectance spectra of all samples, using Hotelling's T... 2The statistical measure (95% confidence margin) was used to screen and remove spectral anomalous samples. Then, the continuous and ordered distribution of samples in the PC score space of each production stage was visualized by kernel density estimation (KDE) to verify the monotonicity of the spectral trajectories of S1 to S9 and to confirm that the global single partial least squares (PLS) model can cover the complete condensed trajectory.

[0064] Step 3: Adaptive Variable Selection for Indicators. A differentiated variable selection strategy is constructed to address the differences in the dynamic range of various quality indicators and their near-infrared diffuse reflectance spectral response characteristics, as detailed below:

[0065] (1) For the total polysaccharide index with a wide dynamic range (relative standard deviation RSD = 37.7%), the competitive adaptive reweighted sampling algorithm (CARS) was used to screen feature variables: spectral variables were screened by combining Monte Carlo sampling with adaptive weighted partial least squares regression iteration, and the optimal variable combination was determined based on the cross-validation error. In this embodiment, 105 feature variables were finally obtained from 921 valid spectral variables (reduced by 88.6%), mainly distributed in the characteristic absorption regions of 790–830 nm, 850–890 nm, and approximately 970 nm.

[0066] (2) For the SSC index (RSD=36.8%) with a wide dynamic range and strong collinearity among spectral variables, the continuous projection algorithm (SPA) was used to screen the feature variables: the information duplication among variables was eliminated by orthogonal projection, and the prediction performance of different variable combinations was evaluated by cross-validation. Finally, 65 feature variables were screened from 921 effective spectral variables (reduced by 92.9%), which are mainly distributed in the characteristic absorption regions of 770-890 nm and 960-1050 nm.

[0067] (3) For the total polyphenol index (RSD=7.5%) with a narrow dynamic range and weak spectral response, a full-spectrum partial least squares regression model (Full-PLS) was used for modeling, with 921 effective spectral variables as feature variables. The prediction stability of the model under weak signal conditions was improved by comprehensively utilizing the full-band spectral information.

[0068] Step 4: Establishment and Comparison of Indicator-Specific PLS Models with PLS-SVR. The database is randomly divided into a calibration set (used for data calibration during model building) and a prediction set (the prediction set is completely archived throughout the modeling process) at a ratio of 70% to 30%. Full-PLS, CARS-PLS, and SPA-PLS models are established for the three indicators, totaling nine models. Corresponding SVR (Support Vector Regression) models are also established. RMSEC (Root Mean Square Error of Calibration), RMSEP (Root Mean Square Error of Prediction), and R... 2The predictive performance was comprehensively evaluated using both the coefficient of determination (COD) and the residual prediction deviation (RPD). The dedicated PLS model outperformed the PLS-SVR model in all indicators and variable configurations (RPD increased from 0.4 to 5.5), confirming that the spectrum-concentration relationship is dominated by linear Beer-Lambert behavior. This established linear PLS as the optimal framework for computationally efficient online deployment. An RPD of not less than 5.0 was used as the qualified standard for online deployment, and an RPD of not less than 7.0 was used as the standard for high-precision online process control.

[0069] Step 5: Validation of the statistical unbiasedness of the confidence ellipse and determination of the optimal prediction model for the index. For the prediction set of the PLS model specific to each index, construct a slope-intercept joint 95% confidence ellipse of the predicted regression line. The inclusion of the ideal point (slope = 1, intercept = 0) in the confidence ellipse is used as the criterion for satisfying statistical unbiasedness. For low-variance quality indicators with RMSEP differences less than 0.001 mg / mL, this step is an effective tool for decisively distinguishing models with similar RMSEP, meeting the requirements of the PAT framework for statistical validation of online analysis methods.

[0070] For total polyphenols, the RMSEP of the Full-PLS, CARS-PLS, and SPA-PLS models were 0.0024, 0.0027, and 0.0027 mg / mL, respectively, with a difference of only 0.0003 mg / mL. Traditional scalar error indices could not effectively distinguish the superiority or inferiority of the models. Confidence ellipse analysis provided a decisive distinguishing method: the characteristic variables screened by CARS did not include the 890–940 nm phenolic hydroxyl (OH) stretching region, resulting in the loss of key characteristic signals, and its confidence ellipse slope was less than 1, indicating a systematic underestimation; although the characteristic variables screened by SPA were concentrated in the spectral characteristic region, they produced an intercept shift in PLS regression; the confidence ellipse of Full-PLS was the most compact and contained ideal points, satisfying the requirement of statistical unbiasedness, and was therefore determined to be the optimal model configuration for online monitoring of total polyphenols.

[0071] Confidence ellipse validation determined the specific model configurations for the three indices as follows: Total polysaccharides: CARS-PLS (105 feature variables, LV=7); SSC: SPA-PLS (65 feature variables, LV=7); Total polyphenols: Full-PLS (921 full-spectrum variables, LV=9). The confidence ellipses of all three optimal models satisfy the requirement of statistical unbiasedness (including ideal points) and can be used as the final models for online deployment.

[0072] Step Six: Online Dynamic Monitoring Implementation. The optimal prediction model is deployed to the online monitoring system. For the real-time spectral stream from the diffuse reflection fiber optic probe installed on the circulation pipeline of the concentration process, SG smoothing preprocessing and selection of index-specific feature variables are performed sequentially. The selected spectral feature variables are then input into the optimal prediction model, which synchronously outputs the predicted values ​​of three indicators within seconds after each spectral acquisition. Combined with a preset endpoint threshold, automatic determination of the concentration endpoint is achieved: the system uses the SSC predicted value as the core judgment indicator. When the SSC predicted value reaches the preset endpoint threshold (e.g., the SSC threshold is set to 62.52°Brix), an endpoint prediction signal is triggered. Three consecutive sampled values ​​not lower than the threshold are used as confirmation conditions. After confirmation, an endpoint judgment signal is output to the distributed control system via the OPC-UA interface, supporting automatic shutdown or transition to the next process. Simultaneously, the system monitors in real time whether the predicted values ​​of total polysaccharides and total polyphenols exceed preset thresholds (e.g., the upper limit threshold for total polysaccharides is set at 195.97 mg / mL, and the upper limit threshold for total polyphenols is set at 0.128 mg / mL; the specific thresholds can be adjusted according to actual process specifications). If any indicator exceeds the limit, an audible and visual alarm is triggered, prompting the operator to check the process parameters. For different food and medicinal plant formulation systems, slope-bias correction is performed using a small number of new formulation samples to achieve lightweight cross-formulation optimal prediction model transfer.

[0073] Example 1:

[0074] This embodiment uses standardized water extracts of four edible and medicinal plants—Lycium barbarum L., Polygonatum sibiricum F. Delaroche, Rubus chingii Hu, and Glycyrrhiza uralensis Fisch.—as research objects to establish a multi-index near-infrared prediction model covering the complete concentration trajectory.

[0075] I. Sample Collection and Reference Value Determination

[0076] 227 samples were collected from the production line of the cooperating manufacturer (Jiangdayuan Ecological Biotechnology Co., Ltd.): 135 systematic gradient samples (9 sampling stages from S1 to S9, 15 replicates per stage, covering the complete concentration trajectory from initial aqueous extract to final concentrated product); and 92 supplementary representative samples from random batches (different production dates, different raw material batches). The 9 sampling stages were distributed across... Figure 1 In the concentration, secondary concentration, vacuum degassing, and tertiary concentration processes shown, data were collected systematically according to the time-concentration gradient. Reference values ​​for three indicators were determined: Total polysaccharides were measured using the phenol-sulfuric acid colorimetric method (centrifuged at 12000 r / min for 15 min, diluted 25 times, and measured at 490 nm, compared with the glucose standard curve, R...). 2=0.9976); total polyphenols were measured using the Folin-Ciocalteu method (50-fold dilution, 765 nm, 90 min incubation in the dark, with gallic acid as the reference curve, R²=0.9977); SSC was measured directly using a digital refractometer. Reference ranges for the three indicators are shown in Table 1.

[0077] Table 1. Statistical analysis of reference values ​​for three quality indicators in the sample.

[0078]

[0079] II. Near-infrared diffuse reflectance spectral acquisition

[0080] An Ocean Optics USB2+ near-infrared spectrometer with a halogen lamp light source (LS-HA) was used to acquire the spectra of each sample in transmission mode: the acquisition wavelength was 780 to 1080 nm, the resolution was 0.39 nm, corresponding to the extraction of 921 effective spectral variables from pixel numbers 500 to 1420, the integration time was 0.5 s, and an average of 3 scans were performed per measurement. Dark current correction and nonlinear correction were applied simultaneously during acquisition. It should be noted that this embodiment uses transmission mode for spectral acquisition to obtain a stable and reliable spectrum-quality index correspondence; in subsequent online deployment (Example 2), it will be switched to diffuse reflection mode to adapt to the non-contact installation requirements of industrial stainless steel pipelines. The model transfer between the two is achieved through slope-bias correction.

[0081] The collected spectral data will serve as the basis for subsequent preprocessing and optimization.

[0082] III. Preprocessing Optimization

[0083] Five preprocessing strategies (SG smoothing, SNV, MSC, mean centering, and SG+SNV combination) were systematically evaluated. Using the lowest RMSECV in the 10-fold cross-validation of all-variable PLS as the criterion, all three quality indicators confirmed that SG smoothing (9-point window, second-order polynomial) was the optimal solution. This result is consistent with the conclusions of the preprocessing optimization in step two. Analysis suggests that the concentrated liquid system studied in this example has high viscosity and a large range of solid content (SSC range from 15.87 to 62.52°Brix). While the introduction of SNV eliminates scattering effects, it amplifies the noise introduced by local optical path fluctuations between high-viscosity samples, which is detrimental to model stability. SG smoothing only performs spectral denoising, avoiding additional variation introduced by overcorrection. After SG smoothing, the ordered spectral structure of S1 to S9 is preserved. The OH stretching vibration absorption near 970 nm systematically increases with solute concentration, and the CH overtones and combination peaks from 820 to 900 nm provide complementary concentration-sensitive information.

[0084] IV. PCA Structural Analysis

[0085] The KDE density surface of PCA reveals a continuously stretched distribution in the PC1-PC2 score space, and the nine production stages form a clear and ordered trajectory from S1 to S9 along PC1, confirming that a single global PLS model can cover the entire production range; Hotelling's T 2 The tests confirmed that all 227 samples were within the 95% control limit and there were no abnormalities.

[0086] V. Selection of Adaptive Variables for Indicators

[0087] Total polysaccharides (CARS): After 100 iterations, 105 characteristic variables were selected at the global minimum point of RMSECV in stage 3 (reduction of 88.6%), mainly concentrated in three regions: 790-830 nm (CH second harmonic), 850-890 nm (CH combination), and approximately 970 nm (OH stretching and polysaccharide hydration related). The number of optimal latent variables decreased from 12 to 7. SSC (SPA): The RMSECV minimum point was reached at 65 characteristic variables (reduction of 92.9%), with selected variables concentrated in the 770-890 nm and 960-1050 nm sugar characteristic absorption regions. Total polyphenols (Full-PLS, 921 variables): A PLS model was established using 921 effective spectral variables as input and 9 latent variables. CARS and SPA showed inferior screening performance compared to the full-spectrum approach; a detailed comparative analysis is provided in step six, the confidence ellipse validation section.

[0088] VI. Comparison of PLS ​​and SVR and Model Performance Verification

[0089] The samples were divided into a calibration set (159 samples) and an independent prediction set (68 samples) at 70% and 30% respectively. For all model configurations of the three quality indicators, PLS outperformed SVR (RPD improved from 0.4 to 5.5), confirming that the spectrum-concentration relationship is dominated by linear Beer-Lambert behavior. The complete performance comparison of each model configuration for the three quality indicators is shown in Table 2. The PLS model results are shown in Figures 3(a), 3(d), and 3(g), and the corresponding residual distributions are shown in Figures 3(b), 3(e), and 3(h).

[0090] Table 2 Summary of Predictive Performance of Each Model Configuration for the Three Quality Indicators

[0091]

[0092] VII. Statistical Unbiasedness Verification of Confidence Ellipse

[0093] The confidence ellipses of the CARS-PLS model for total polysaccharides and the SPA-PLS model for SSCs were compact and contained the ideal point (Figure 3(c), Figure 3(f)), confirming statistical unbiasedness. The RMSEP difference among the three models for total polyphenols was only 0.0003 mg / mL, and the scalar error index could not distinguish between them; the confidence ellipses provided a decisive distinction (Figure 3(i)): the Full-PLS confidence ellipse was the most compact and closest to the ideal point (acceptable); the CARS-PLS confidence ellipse was biased, with a slope less than 1 (systematic underestimation, unacceptable); the SPA-PLS model had an intercept bias (biased, unacceptable); thus, the Full-PLS model was confirmed as the online monitoring model for total polyphenols that met the requirement of statistical unbiasedness.

[0094] Example 2:

[0095] This embodiment, based on Embodiment 1, implements an online dynamic near-infrared spectroscopy monitoring system. The overall system configuration is as follows: Figure 1 As shown.

[0096] I. Hardware System Configuration

[0097] The diffuse reflection fiber optic probe is fixed to the concentration process circulation pipeline in a non-contact installation manner (e.g., Figure 1 As shown, the probe is directly optically coupled to the flowing concentrate through a steam-resistant and liquid-resistant (CIP) optical window located near the secondary concentrator. A near-infrared spectrometer (780 to 1080 nm, resolution no greater than 1 nm) and a halogen lamp light source (LS-HA) are placed in a protective control cabinet. The near-infrared spectrometer is connected to the diffuse reflection fiber optic probe via optical fiber. The industrial control acquisition card is connected to the industrial control computer via a bus interface, responsible for the analog-to-digital conversion and data acquisition of the near-infrared spectrometer output signal, and transmitting the digital spectral data to the industrial control computer in real time. The system uses diffuse reflection mode instead of transmission mode for online spectral acquisition, eliminating the need for customized transparent pipe sections and adapting to the on-site installation requirements of industrial stainless steel pipelines. Example 1 uses transmission mode to establish the prediction model to obtain a stable and reliable spectral-quality index correspondence; this example uses diffuse reflection mode for online deployment, and combines model migration and online verification to realize the application of the prediction model to industrial online monitoring scenarios.

[0098] II. Online Monitoring Software Process

[0099] The online monitoring software is deployed on an industrial control computer (pre-installed with Windows or Linux operating system, equipped with no less than 4GB of memory and solid-state drive, and with Ethernet communication interface). It is responsible for real-time processing of spectral data, model prediction, and communication with DCS. It executes the following process in a loop: First, it receives the raw near-infrared diffuse reflectance spectrum output by the near-infrared spectrometer in real time at a set frequency (adjustable from 30s to 5min); Second, it automatically performs dark current correction, nonlinear correction, and SG smoothing preprocessing; Third, after performing the selection of total polysaccharide CARS feature variables (105 variables), SSC SPA feature variables (65 variables), and total polyphenols all variables (921 variables), it inputs the selected spectral variable vectors into the optimal model. The prediction of the three indicators is completed in the sub-millisecond level (referring to the PLS matrix operation stage); Fourth, it displays the predicted values ​​as time series curves on the operation interface in real time; Fifth, when the SSC predicted value reaches the preset endpoint threshold or any indicator exceeds the threshold, it triggers an audible and visual alarm and sends an endpoint judgment signal to DCS through the OPC-UA interface, supporting automatic shutdown or transfer to the next process.

[0100] III. Cross-Formulation Model Transfer

[0101] When changing the formula (e.g., switching from four herbs to eight herbs: Astragalus membranaceus, jujube, longan pulp, wolfberry, peach kernel, codonopsis pilosula, angelica sinensis, and licorice), using Example 1 as a template, only a small number of samples (recommended at least 5 per stage) of the new formula are needed for slope-bias correction, achieving lightweight model updates without requiring full recalibration, significantly reducing the cost of deploying and debugging the new formula. It is recommended to take at least 5 samples from each of the three intermediate stages S2, S5, and S8, covering the low, medium, and high concentration ranges to ensure the robustness of the correction.

[0102] IV. Long-term operation and maintenance

[0103] The system's data management and visualization module continuously records the deviation between online predicted values ​​and synchronously sampled offline measured values. When the slope deviation exceeds ±5% or the intercept deviation exceeds ±3%, a model update prompt is automatically triggered. It supports monitoring the long-term statistical unbiasedness of the model through confidence ellipses and regularly generates batch consistency reports to meet the electronic batch record requirements of GMP regulations and the PAT framework.

[0104] Example 3:

[0105] This embodiment uses a compound aqueous extract concentrate of eight edible and medicinal plants—Astragalus membranaceus Bge., Ziziphus jujuba Mill., Dimocarpus longan Lour., Lycium barbarum L., Prunus persica (L.) Batsch, Codonopsis pilosula (Franch.) Nannf., Angelica sinensis (Oliv.) Diels, and Glycyrrhiza uralensis Fisch.—as the research object. The method and system described in this invention are used to achieve real-time online monitoring of multiple indicators in industrial applications, verifying the cross-formulation applicability of the method.

[0106] After the concentration process begins, the system automatically acquires diffuse reflectance spectra every 2 minutes, with an integration time of 0.5 seconds per frame and an average of 3 scans. Following acquisition, dark current correction, nonlinear correction, and SG smoothing preprocessing are automatically performed. After preprocessing, the system performs selection of total polysaccharide CARS characteristic variables (105 variables), SSC SPA characteristic variables (65 variables), and total polyphenols all variables (921 variables). The selected spectral variable vectors are input into the optimal model, and the predicted values ​​of total polysaccharide, total polyphenols, and SSC are output synchronously in sub-milliseconds and displayed as time-series curves on the user interface in real time. When the predicted SSC value reaches the preset endpoint threshold, the system automatically triggers an audible and visual alarm and sends an endpoint signal to the DCS. Operators can also predict the concentration progress and adjust process parameters in advance based on the dynamic trend curves of the three quality indicators. The spectral data streams, predicted value time series, and alarm events for each batch are automatically saved to electronic batch records for batch-to-batch quality consistency analysis and regulatory audit traceability. When switching between different formulations, the slope-bias correction of the new formulation samples is used to complete the lightweight model migration. The entire switching and debugging process does not require the re-collection of a large number of samples or the reconstruction of the model.

[0107] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0108] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A near-infrared spectroscopy intelligent dynamic monitoring method for the manufacturing process of concentrated extracts from food and medicinal plants, characterized in that: Samples were systematically collected from multiple key production stages of the compound concentrate production line of food and medicine homologous plants, and supplemented with random representative samples from different batches. Reference values ​​of three quality indicators, namely total polysaccharide, soluble solids content and total polyphenols, were determined to construct a multi-batch sample database. Near-infrared diffuse reflectance spectra of each sample were collected, and various preprocessing strategies were compared using 10-fold cross-validation to confirm that SG smoothing was the optimal spectral preprocessing method. An adaptive variable selection strategy was adopted, and competitive adaptive reweighted sampling, continuous projection algorithm and full-spectrum PLS were used to select variables for quality indicators with different dynamic ranges and spectral response characteristics. For each quality indicator, a slope-intercept joint 95% confidence ellipse is constructed using the partial least squares regression model. The optimal prediction model for each quality indicator is determined by using the inclusion of the ideal point in the confidence ellipse as the statistical unbiasedness criterion. The optimal prediction model is deployed to the online system. SG preprocessing and variable selection are performed sequentially on the real-time spectral stream of the diffuse reflection fiber optic probe. Then, the optimal prediction model is input to obtain the predicted values ​​of three quality indicators. Combined with preset thresholds, the concentration endpoint is automatically determined.

2. The near-infrared spectroscopy intelligent dynamic monitoring method for the manufacturing process of concentrated edible and medicinal plant extracts according to claim 1, characterized in that, The medicinal and edible plants include, but are not limited to, four or more of the following: wolfberry, polygonatum, raspberry, licorice, astragalus, jujube, longan pulp, peach kernel, codonopsis, and angelica. The key production stages are nine stages from the initial aqueous extract to the final concentrated product, with no fewer than 10 samples collected at each stage. The reference values ​​for the quality indicators are confirmed in the following ways: the total polysaccharide content is determined by the phenol-sulfuric acid colorimetric method at 490 nm, the total polyphenol content is determined by the Folin-Ciocalteu method at 765 nm, and the SSC is determined by a digital refractometer.

3. The near-infrared spectroscopy intelligent dynamic monitoring method for the manufacturing process of concentrated edible and medicinal plant extracts according to claim 1, characterized in that, The near-infrared diffuse reflectance spectrum was collected in the wavelength range of 780 to 1080 nm, and 921 effective spectral variables were extracted. The SG smoothing window length was 9 points and the polynomial order was 2. Using the lowest calibration root mean square error as the evaluation criterion, 10-fold cross-validation confirmed that SG smoothing had the best performance among the five preprocessing strategies.

4. The near-infrared spectroscopy intelligent dynamic monitoring method for the manufacturing process of concentrated extracts of food and medicinal plants according to claim 1, characterized in that, The specific strategy for selecting adaptive variables for the indicators is as follows: For the total polysaccharide index, with a relative standard deviation (RSD) of 37.7%, a competitive adaptive reweighted sampling algorithm was used to screen feature variables: spectral variables were screened by combining Monte Carlo sampling with adaptive weighted partial least squares regression iteration, and the optimal variable combination was determined based on the cross-validation error, achieving a variable reduction of 88.6%. For the SSC index, RSD=36.8%, the continuous projection algorithm was used to screen feature variables: information duplication between variables was eliminated by orthogonal projection, and the predictive performance of different variable combinations was evaluated by cross-validation, achieving a 92.9% variable reduction. For the total polyphenol index, with an RSD of 7.5%, a full-spectrum partial least squares regression model was used for modeling, with all effective spectral variables as feature variables.

5. The near-infrared spectroscopy intelligent dynamic monitoring method for the manufacturing process of concentrated extracts of food and medicinal plants according to claim 1, characterized in that, Multiple batches of sample databases were randomly divided into calibration and prediction sets at 70% and 30% respectively. Nine models were established for the three quality indicators: Full-PLS, CARS-PLS, and SPA-PLS, totaling nine models. Corresponding PLS-SVR models were also established. RMSEC, RMSEP, and R... 2 Based on a comprehensive evaluation of predictive performance using both PLS and RPD, the PLS model outperforms the SVR model in all indicator and variable configurations. Therefore, the partial least squares regression model is chosen to construct the confidence ellipse.

6. The near-infrared spectroscopy intelligent dynamic monitoring method for the manufacturing process of concentrated extracts of food and medicinal plants according to claim 5, characterized in that, Principal component analysis was performed on the SNV-normalized spectral matrix of the near-infrared diffuse reflectance spectra of all samples, using Hotelling's T... 2 Statistical screening and elimination of spectral anomalies were performed. Then, kernel density estimation was used to visualize the continuous and ordered distribution of samples in the PCA score space for each production stage, verifying the monotonicity of the spectral trajectories of all key production stages and confirming that the global single partial least squares model can cover the complete condensed trajectory.

7. The near-infrared spectroscopy intelligent dynamic monitoring method for the manufacturing process of concentrated extracts of food and medicinal plants according to claim 1, characterized in that, The concentration endpoint is automatically determined by combining preset thresholds, specifically as follows: The soluble solids content prediction value is the core judgment indicator. When the soluble solids content prediction value reaches the preset endpoint threshold, the endpoint prediction signal is triggered. The confirmation condition is that three consecutive sampling values ​​are not lower than the threshold. After confirmation, the endpoint judgment signal is output to the distributed control system, which supports automatic shutdown or transfer to the next process. At the same time, the predicted values ​​of total polysaccharide and total polyphenols are monitored in real time to see if they exceed the preset threshold. If either indicator exceeds the limit, an audible and visual alarm is triggered to prompt the operator to check the process parameters.

8. The near-infrared spectroscopy intelligent dynamic monitoring method for the manufacturing process of concentrated edible and medicinal plant extracts according to claim 1, characterized in that, For different food and medicinal plant formulation systems, slope-bias correction is performed using a small number of new formulation samples to achieve lightweight cross-formulation optimal prediction model transfer.

9. A system for implementing the near-infrared spectroscopy intelligent dynamic monitoring method for the manufacturing process of concentrated extracts of edible and medicinal plants as described in any one of claims 1-8, characterized in that, include: The online spectral acquisition module consists of a near-infrared spectrometer, a halogen lamp light source, and an industrial control acquisition card. The near-infrared spectrometer is connected to a diffuse reflection fiber optic probe, which is installed at the online detection position of the circulation pipeline of the concentration equipment to acquire the near-infrared diffuse reflection spectrum of the flowing concentrate in real time. The spectral acquisition is carried out in conjunction with the halogen lamp light source. The spectral preprocessing module automatically performs dark current correction, nonlinear correction, and SG smoothing on the single-frame spectrum acquired online, and outputs a preprocessed spectral vector. The indicator prediction module includes prediction models based on different quality indicators, used to calculate predicted values ​​of total polysaccharide, total polyphenol, and soluble solids content based on preprocessed spectral data; the prediction model... It includes a feature variable selection module and a partial least squares regression prediction model, which outputs the corresponding quality index prediction results in real time based on the input spectral information. The quality decision and alarm module compares the predicted value of the indicator with the preset threshold in real time. When the soluble solids content reaches the endpoint threshold or any indicator exceeds the threshold, it triggers an audible and visual alarm and outputs a concentration endpoint judgment signal to the distributed control system. The data management and visualization module records spectral data streams, predicted value time series, and alarm events in real time, provides dynamic curve visualization of quality indicators in the concentration process, supports data comparison between batches and model drift diagnosis, and saves complete electronic batch records.

10. The system according to claim 9, characterized in that, The data management and visualization module includes a model maintenance submodule, which supports lightweight model updates through slope-bias correction. When the slope deviation of the dynamic curve exceeds ±5% or the intercept deviation exceeds ±3%, an update prompt is automatically triggered.

Images