A content synchronous online detection system in a plant polyphenol extraction process

Abstract

The application discloses a kind of plant polyphenol extraction process content synchronous online detection system, it is related to online detection and process control technical field, including the online spectrum in extraction process acquisition, dynamic evolution information of characteristic absorption peak is analyzed to construct the dynamic content change model that can reflect dissolution kinetics process.By multidimensional feature comparison and matching between this dynamic model and multiple calibration models in historical database, multiple historical models are weighted and fused according to matching similarity, and online real-time calibration model suitable for current batch is generated, and the polyphenol content instantaneous value is calculated in real time according to it.The system overcomes the problem of poor adaptability of static model, realizes the adaptive optimization of detection model to specific working condition and the essence tracking of dissolution process, improves the accuracy of online detection and the accuracy of process control.

Description

Technical Field

[0001] This invention belongs to the field of online detection and process control technology, specifically a system for synchronous online detection of the content of plant polyphenols during the extraction process. Background Technology

[0002] In the extraction and production of plant polyphenols, achieving real-time and accurate online detection of polyphenol content during the extraction process is crucial for precise process control, ensuring stable product quality, and improving efficiency. Currently, this field mainly relies on online analysis technologies such as near-infrared spectroscopy and ultraviolet-visible spectroscopy, combined with chemometric methods to establish quantitative calibration models between spectral data and offline laboratory measurements. These technologies treat spectral data as a set of static variables, and the model establishment focuses on the global mathematical mapping relationship between spectral absorbance and target analyte concentration.

[0003] Existing models based on static spectral correlations have limitations. Model training relies on historical data from specific batches of raw materials and under fixed process conditions, and their intrinsic variables do not explicitly correlate with the dynamic mechanism of polyphenol dissolution. When the origin or variety of raw materials fluctuates, or when extraction process parameters deviate from the calibration conditions, the static mathematical relationship between spectral characteristics and content may fail, leading to significant biases in model predictions and insufficient generalization ability. Furthermore, conventional applications often employ a single global model or simply select a predetermined model based on limited conditional parameters, lacking the ability to subtly adapt to the unique dynamic trajectory of each specific extraction process.

[0004] There is a need for an online detection method that can more fundamentally reflect the kinetic characteristics of the polyphenol dissolution process and adapt to changes in specific operating conditions, in order to overcome the problems of inaccurate predictions and poor adaptability of static models in dynamic and variable industrial environments. Summary of the Invention

[0005] This invention aims to solve at least one of the technical problems existing in the prior art; Therefore, this invention proposes a system for simultaneous online detection of plant polyphenol content during extraction, comprising: The spectral acquisition module is used to acquire multi-channel online spectral data of plant raw materials in the extraction equipment, and to perform preprocessing and feature extraction to obtain spectral feature vectors; The dynamic modeling module is used to parse multiple characteristic absorption peaks related to the polyphenol structure and their dynamic changes from the spectral feature vector, and to construct a dynamic content change model describing the polyphenol dissolution process based on the characteristic absorption peaks and their dynamic changes. The model matching module is used to call up a set of historical calibration models that match the current plant raw material type and extraction process conditions from the historical extraction database. The dynamic content change model is compared and matched with the set of historical calibration models in multiple dimensions to obtain the matching similarity. Based on the matching similarity, multiple historical calibration models are weighted and fused to generate an online real-time calibration model suitable for the current extraction process. The content calculation module is used to input the spectral feature vector into the online real-time calibration model in real time to calculate the instantaneous detection value of polyphenol content; The process control module is used to establish a dynamic correlation between the instantaneous detection value of the polyphenol content and the process parameters of the extraction process, and generate process parameter control instructions accordingly.

[0006] Furthermore, the process of collecting multi-channel online spectral data of plant raw materials within the extraction device, and performing preprocessing and feature extraction to obtain spectral feature vectors includes: Real-time acquisition of multi-channel online spectral data of plant raw materials within the extraction equipment; The multi-channel online spectral data is subjected to spectral noise filtering and baseline correction to obtain preprocessed standard spectral data; Multi-band feature extraction is performed on the preprocessed standard spectral data to obtain a spectral feature vector containing the characteristic wavelength and the corresponding absorbance; The real-time acquisition of multi-channel online spectral data of plant raw materials in the extraction equipment includes: deploying online spectral probes at multiple key locations in the extraction equipment, including the area near the feed inlet, the main body area of ​​the extraction tank, the action area of ​​the stirring impeller, and the area near the liquid outlet. The online spectral probe is used to synchronously acquire the raw transmission or reflection spectra of plant raw material slurry at multiple characteristic wavelengths at a fixed frequency. The raw signals from each of the online spectral probes are amplified and converted from analog to digital to obtain a multi-channel discrete spectral data stream. The discrete spectral data streams of the multi-channel array are time-stamped and aligned, and integrated to form the multi-channel online spectral data that strictly corresponds to the extraction time axis.

[0007] Further, the step of performing spectral noise filtering and baseline correction on the multi-channel online spectral data to obtain preprocessed standard spectral data includes: The high-frequency random noise in the multi-channel online spectral data is separated and filtered out using the wavelet transform method to obtain the denoised spectral data; For the denoised spectral data, its spectral baseline is estimated by an iterative fitting algorithm. The spectral baseline reflects non-characteristic background signals caused by factors such as light scattering and background absorption. Baseline correction is completed by subtracting the spectral baseline from the denoised spectral data; The baseline-corrected spectral data is vector-normalized to eliminate overall intensity variations caused by factors such as optical path length and sample concentration differences, and finally the preprocessed standard spectral data is output.

[0008] Further, the step of performing multi-band feature extraction on the preprocessed standard spectral data to obtain a spectral feature vector containing characteristic wavelengths and corresponding absorbance includes: Within a preset polyphenol characteristic spectral range, the preprocessed standard spectral data is subjected to continuous first-order or second-order derivative transformations to enhance the fine structural features of the spectrum. In the derivative spectrum, all local extrema are identified and located, and the wavelength position corresponding to each extrema is marked as a potential characteristic wavelength. Based on preset peak width threshold and peak height intensity threshold, the potential characteristic wavelengths are screened to remove false peaks caused by noise; Calculate the absorbance value of each selected characteristic wavelength on the preprocessed standard spectral data; The selected characteristic wavelengths and their corresponding absorbance values ​​are arranged in wavelength order to form the spectral feature vector.

[0009] Furthermore, the construction of a dynamic content change model describing the polyphenol dissolution process based on the characteristic absorption peaks and their dynamic changes includes: With extraction time as the horizontal axis and the absorbance value or the rate of change of its derivative of each characteristic absorption peak as the vertical axis, multiple dynamic change curves of the characteristic absorption peaks are plotted. Calculate the rate of change of the dynamic change curve of each characteristic absorption peak at different extraction stages, including the initial dissolution acceleration stage, the stable dissolution stage, and the dissolution saturation stage; Based on the rate of change, a kinetic curve model of each characteristic absorption peak changing with extraction time is constructed using a piecewise linear or nonlinear function fitting method. The kinetic curve models of multiple characteristic absorption peaks changing with extraction time are weighted and superimposed, with the weighting coefficients set based on the contribution of each characteristic absorption peak to the types and contents of polyphenols, thereby synthesizing a comprehensive dynamic content change model.

[0010] Furthermore, the step of retrieving a set of historical calibration models from the historical extraction database that match the current plant material type and extraction process conditions includes: Receive input current detection task parameters, which include at least the variety, part, particle size, type of extraction solvent, and temperature parameters of the plant raw material; Using the current detection task parameters as query conditions, perform multi-dimensional matching and retrieval in the historical extraction database; The historical extraction database stores multiple sets of historical extraction records. Each set of records includes historical task parameters, historical spectral feature vector sequences, and sequences of true polyphenol content values ​​obtained through offline chemical analysis. From the successfully matched historical extraction records, the trained historical calibration model is extracted. The historical calibration model describes the mapping relationship between the spectral feature vector and the true value of polyphenol content under specific conditions. All successfully matched and extracted historical calibration models are collected to form the historical calibration model set.

[0011] Further, the step of performing multi-dimensional feature comparison and matching between the dynamic content change model and the historical calibration model set to obtain the matching similarity includes: Extract the feature parameters of the dynamic content change model, including the overall trend of the curve, the timing of the appearance of the characteristic peak, the peak intensity change pattern, and the dynamic model parameters; Extract the feature parameters of each historical calibration model in the historical calibration model set, using the same extraction method as the dynamic content change model; Within a unified feature space, calculate the Euclidean distance between the feature parameters of the dynamic content change model and the feature parameters of each of the historical calibration models. The Euclidean distance is mapped to a similarity value between zero and one using a preset similarity conversion function. This similarity value is the matching similarity with the corresponding historical calibration model.

[0012] Furthermore, the step of weightedly fusing multiple historical calibration models based on the matching similarity to generate an online real-time calibration model suitable for the current extraction process includes: The matching similarity is normalized so that its sum is one, and this is used as the weight coefficient of each of the historical calibration models in the fusion process; Obtain the core mapping function for each of the historical calibration models. The input of the core mapping function is a spectral feature vector, and the output is the predicted value of polyphenol content. Using the weighting coefficients, the core mapping functions of multiple historical calibration models are linearly combined or more complexly fused across function spaces to generate an integrated mapping function; Using the spectral feature vectors of the small amount of initial extraction stage data collected at present and the corresponding offline analysis reference values, the integrated mapping function is fine-tuned and optimized to make the parameters more suitable for the current process. The finely tuned and optimized integrated mapping function is established as the online real-time calibration model applicable to the current extraction process.

[0013] Further, the step of inputting the spectral feature vector into the online real-time calibration model in real time to calculate the instantaneous detection value of polyphenol content includes: At the end of each new spectral acquisition and processing cycle, the newly generated spectral feature vector is obtained; The spectral feature vector is used as input and provided to the mapping function in the online real-time calibration model; The mapping function processes the spectral feature vector based on its internal algorithm and directly outputs a value representing the polyphenol concentration, which is the instantaneous detection value of the polyphenol content. The instantaneous detection values ​​of the polyphenol content in the continuous output are subjected to time-series smoothing filtering to eliminate random fluctuations and obtain a more stable instantaneous content sequence.

[0014] Furthermore, establishing a dynamic correlation between the instantaneous detection value of the polyphenol content and the extraction process parameters, and generating process parameter control instructions accordingly, includes: The current process parameters of the extraction equipment are acquired in real time, including at least the extraction temperature, stirring speed, and solvent flow rate. A dynamic correlation model is constructed, with the instantaneous change rate of the polyphenol content as input and the adjustment amount of the process parameters as output. The instantaneous detection value of the polyphenol content at the current moment and its changing trend are input into the dynamic correlation model; The dynamic correlation model calculates the direction and amount of adjustment to each process parameter to achieve the optimal or most stable polyphenol dissolution efficiency based on built-in rules or learned mapping relationships. The adjustment direction and adjustment amount are formatted into an instruction format that the extraction equipment control system can recognize and execute, thereby generating the process parameter control instruction.

[0015] Compared with the prior art, the beneficial effects of the present invention are: The dynamic modeling module not only identifies characteristic absorption peaks related to polyphenol structures from continuously acquired spectra, but also analyzes the dynamic changes of these peaks over time, and constructs a process model based on this. This technology deeply correlates spectral information with the time dimension of the dissolution process and the behavior of chemical groups, enabling the model's internal variables to reflect the kinetic state of the extraction process. Thus, the content prediction results are combined with the physicochemical nature of the process, and the model's tracking of the extraction process no longer relies on a static concentration-absorbance relationship, but rather on an understanding of the temporal evolution of characteristic spectra. This enhances the system's ability to analyze and judge the true state of polyphenol dissolution during non-steady-state operation or process disturbances, thereby reducing prediction biases introduced by dynamic process factors.

[0016] The model matching module, after retrieving multiple relevant calibration models from the historical database, does not directly select one. Instead, it performs multi-dimensional feature comparison between the real-time constructed dynamic model and these historical models, calculates a quantified matching similarity, and uses this as a weight to weight and fuse multiple historical models, generating an online calibration model unique to the current extraction process. This technology makes the final model a dynamically generated hybrid that integrates features from various historical scenarios. Through feature comparison and weighted calculations, the system automatically adjusts and optimizes the parameters and structure of the historical models to adapt to the current operating conditions. This method enables the detection system to autonomously adjust the model output to fit the dynamic trajectory unique to the current batch when facing actual situations such as natural fluctuations in raw material properties and subtle changes in process conditions, thereby maintaining the accuracy and reliability of prediction results under changing production conditions. Attached Figure Description

[0017] Figure 1 This is a time-series diagram of the plant polyphenol content synchronous online detection system during the extraction process described in this invention; Figure 2 A flowchart for spectral data preprocessing; Figure 3 Flowchart for constructing a dynamic content change model; Figure 4 The graph shows the loss variation during the fine-tuning optimization process of the integrated mapping function in the scenario of polyphenol extraction model fusion. Figure 5 This is a curve showing the dynamic control of process parameters in the extraction of plant polyphenols. Detailed Implementation

[0018] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] See Figure 1 The spectral acquisition module deploys probes at key locations within the extraction equipment to acquire multi-channel online spectral data reflecting the material's state in real time. This data is preprocessed and feature extracted, transforming it into spectral feature vectors that characterize the material's spectral properties. The dynamic modeling module focuses on resolving multiple characteristic absorption peaks directly related to the polyphenol's chemical structure from these spectral feature vectors and tracking the dynamic evolution of these peaks over the extraction timeline. This allows for the construction of a dynamic content change model describing the polyphenol's dissolution process from the raw material. To correlate spectral changes with specific content values, the model matching module retrieves a set of similar historical calibration models from a historical database based on the specific conditions of the current extraction task. It calculates the matching similarity through multi-dimensional feature comparison and then weights and fuses these models to generate a dedicated online real-time calibration model applicable to the current batch extraction process. The content calculation module uses this generated online real-time calibration model to instantly convert the acquired spectral feature vectors into instantaneous polyphenol content values. Finally, the process control module establishes a dynamic correlation between the instantaneous detection value and key process parameters such as extraction temperature and stirring speed, and generates process parameter control instructions in real time based on the changing trend of the instantaneous detection value, thus forming a closed loop from detection to control.

[0020] See Figure 2In one embodiment of the present invention, the spectral acquisition module acquires multi-channel online spectral data of plant raw materials in real time within the extraction device. This process takes the extraction of tea polyphenols from green tea leaves using an ethanol-water solution in a stirred extraction tank as an example scenario. Four online spectral probes are installed inside the extraction device. The probes are positioned to cover the area near the feed inlet, the main body of the extraction tank, the action area of ​​the stirring impeller, and the area near the liquid outlet. The online spectral probes synchronously acquire the original transmission spectral signals of the plant raw material slurry at multiple characteristic wavelengths in the wavelength range of 400 nm to 800 nm at a fixed frequency of once per second. The original signals from each online spectral probe are amplified and converted from analog to digital to obtain four-channel discrete spectral data streams. The four-channel discrete spectral data streams are time-stamped and aligned, and integrated to form multi-channel online spectral data that strictly corresponds to the extraction time axis. In practice, multi-channel online spectral data undergoes spectral noise filtering and baseline correction to obtain preprocessed standard spectral data. Wavelet transform is used to separate and filter high-frequency random noise from the multi-channel online spectral data. The wavelet transform uses the Daubechies wavelet basis function to decompose the spectral signal, removing high-frequency detail coefficients and reconstructing the denoised spectral data. For the denoised spectral data, an iterative fitting algorithm is used to estimate its spectral baseline. This algorithm employs an asymmetric least squares method, adjusting weights through multiple iterations to fit low-value points in the spectral data, thereby obtaining a spectral baseline reflecting non-characteristic background signals caused by light scattering, background absorption, and other factors. The spectral baseline is subtracted from the denoised spectral data to complete baseline correction. Vector normalization is then performed on the baseline-corrected spectral data to eliminate overall intensity variations caused by factors such as optical path length and sample concentration differences. Vector normalization is achieved by dividing each spectral vector by its Euclidean norm, expressed by the formula:

[0021] Where: S represents the baseline-corrected spectral data vector, s i S is the absorbance value at the i-th wavelength point in the vector, where n is the total number of wavelength points. norm It is a normalized standard spectral data vector, and the final output is the preprocessed standard spectral data.

[0022] In some embodiments, multi-band feature extraction is performed on the preprocessed standard spectral data to obtain a spectral feature vector containing characteristic wavelengths and corresponding absorbance. Within a preset polyphenol characteristic spectral range, such as 500 nm to 700 nm, a continuous first-order derivative transformation is performed on the preprocessed standard spectral data to enhance the fine structural features of the spectrum. The derivative transformation is achieved by calculating the difference in absorbance values ​​at each wavelength point. On the derivative spectrum, all local extrema are identified and located, and the wavelength position corresponding to each extrema is marked as a potential characteristic wavelength. Based on preset peak width thresholds and peak height intensity thresholds, potential characteristic wavelengths are screened. The peak width threshold is set to 5 nm, and the peak height intensity threshold is set to twice the average amplitude of the derivative spectrum to remove false peaks caused by noise. The absorbance value corresponding to each selected characteristic wavelength on the preprocessed standard spectral data is calculated. The selected characteristic wavelengths and their corresponding absorbance values ​​are arranged in wavelength order to form a spectral feature vector. In the example scenario, three characteristic wavelengths, 520 nm, 580 nm, and 650 nm, are extracted from the spectral data of the green tea extraction process, and their corresponding absorbance values ​​change over time to form a sequence. It can be understood that the multi-band feature extraction process relies on preset threshold parameters, which can be adjusted according to the spectral characteristics of different plant materials.

[0023] Optionally, different wavelet basis functions or filtering algorithms can be used in the spectral noise filtering step, such as using the Symlet wavelet basis function for noise separation to adapt to the noise characteristics of different types of spectral signals. It is understood that the iterative fitting algorithm in the baseline correction step can also be replaced by a polynomial fitting method, estimating the spectral baseline by fitting a low-order polynomial curve, suitable for spectral data with relatively smooth background signals. In some embodiments, the derivative transformation in the feature extraction step can be a second-order derivative transformation to further highlight the inflection point information of the spectrum. The second-order derivative transformation is achieved by calculating the difference between the first-order derivatives. In the example scenario, the second-order derivative transformation is used to enhance the absorption valley characteristics of green tea leaves near 600 nm. Optionally, an adjacent peak merging rule can be introduced during the feature wavelength screening process. When the distance between two potential feature wavelengths is less than a preset minimum peak interval, they will be merged into one feature wavelength to avoid repeatedly detecting similar spectral features.

[0024] See Figure 3In one embodiment of the present invention, the dynamic modeling module parses multiple characteristic absorption peaks related to the polyphenol structure and their dynamic changes from the spectral feature vector. Using extraction time as the horizontal axis and the absorbance value or its derivative change rate of each characteristic absorption peak as the vertical axis, multiple dynamic change curves of the characteristic absorption peaks are plotted. In the example scenario, for the extraction of black tea leaves using hot water as a solvent, two characteristic absorption peaks located at wavelengths of 530 nm and 650 nm are identified from the spectral feature vector. Their absorbance value sequences within the extraction time of 0 to 60 minutes are recorded, thus forming two dynamic change curves of the characteristic absorption peaks. The rate of change of each characteristic absorption peak's dynamic change curve at different extraction stages is calculated, including the initial dissolution acceleration stage, the stable dissolution stage, and the dissolution saturation stage. For the dynamic change curve of the 530 nm characteristic absorption peak, the average slope of its absorbance value relative to time in the three time intervals of 0-10 minutes, 10-40 minutes, and 40-60 minutes is calculated to obtain the rate of change in the initial dissolution acceleration stage, the stable dissolution stage, and the dissolution saturation stage, respectively. Based on the rate of change, a kinetic curve model of each characteristic absorption peak changing with extraction time is constructed using piecewise linear or nonlinear function fitting methods. For the 530 nm characteristic absorption peak, a three-segment function is used for fitting: a quadratic polynomial function is used to describe the rapid upward trend in the initial dissolution acceleration stage; a linear function is used to describe the steady growth in the stable dissolution stage; and an exponential decay function is used to describe the process approaching saturation in the dissolution saturation stage. The kinetic curve models of multiple characteristic absorption peaks changing with extraction time are then weighted and superimposed. The weighting coefficients are set based on the contribution of each characteristic absorption peak to the type and content of polyphenols, expressed by the following formula:

[0025] Where: M(t) represents the output value of the comprehensive dynamic content change model at time t, f j (t) represents the kinetic curve model function of the j-th characteristic absorption peak, α j α represents the weighting coefficient of the j-th characteristic absorption peak, and m represents the total number of characteristic absorption peaks. In the black tea extraction example, the weighting coefficient α1 of the 530 nm characteristic absorption peak is set to 0.7, and the weighting coefficient α2 of the 650 nm characteristic absorption peak is set to 0.3, thus synthesizing a comprehensive dynamic content change model.

[0026] In some embodiments, the model matching module retrieves a set of historical calibration models from the historical extraction database that match the current plant material type and extraction process conditions. It receives the input parameters for the current detection task, which include at least the plant material variety, part, particle size, type of extraction solvent, and temperature. In the black tea extraction example, the specific parameters are: variety "Yunnan Black Tea," part "one bud and two leaves," particle size "40 mesh," solvent "pure water," and temperature "85 degrees Celsius." Using the current detection task parameters as query conditions, a multi-dimensional matching search is performed in the historical extraction database. The matching search process sets a similarity tolerance for each parameter field; for example, a deviation of ±5 degrees Celsius is allowed for the temperature parameter, and the solvent type must be completely consistent. It can be understood that the historical extraction database stores multiple sets of historical extraction records, each containing historical task parameters, historical spectral feature vector sequences, and sequences of true polyphenol content values ​​obtained through offline chemical analysis. From the successfully matched historical extraction records, pre-trained historical calibration models are extracted. These models describe the mapping relationship between spectral feature vectors and the true polyphenol content under specific conditions. In the example, three historical records were successfully matched, and three corresponding historical calibration models were extracted for each. All successfully matched and extracted historical calibration models are then combined to form a historical calibration model set.

[0027] Optionally, when constructing a kinetic curve model of the characteristic absorption peak changing with extraction time, the nonlinear function fitting method can employ a kinetic equation based on the principle of mass transfer, such as using the one-dimensional solution form of Fick's diffusion law for fitting, to describe the process of polyphenol diffusion and dissolution from the solid raw material into the solvent from a physical mechanism perspective. It can be understood that the weighting coefficient α... j The setting can be determined based on the slope of the standard curve of absorbance at various characteristic wavelengths and concentration of specific polyphenol components obtained from offline chemical analysis. The larger the slope, the higher the contribution of the characteristic absorption peak to the total content, and the larger the corresponding weighting coefficient should be. In some embodiments, when calling the historical calibration model set from the historical extraction database, the matching search conditions can include process parameters such as "extraction equipment model" and "stirring rate" to achieve more accurate matching. In the example scenario, the search conditions "extraction tank volume 50 liters" and "stirring rate 200 rpm" are added to further filter historical extraction records performed under the same equipment and stirring conditions. Optionally, for successfully matched historical calibration models, the average prediction error during their historical application can be additionally recorded as a confidence index, and this confidence index can be referenced in subsequent model fusion steps.

[0028] In one embodiment of the present invention, the model matching module performs multi-dimensional feature comparison and matching between the dynamic content change model and the historical calibration model set to obtain the matching similarity. It then extracts the feature parameters of the dynamic content change model, including the overall trend of the curve, the temporal sequence of characteristic peaks, the peak intensity change pattern, and kinetic model parameters. In the grape seed proanthocyanidin extraction example, the dynamic content change model is represented by an absorbance curve that changes over time. The overall trend of the curve is expressed by a quadratic polynomial function y=p1t. 2 The model is obtained by fitting p2t+p3, and its feature parameters include a coefficient vector P=[p1,p2,p3]. The timing of the characteristic peaks is recorded by the extraction time point list T=[t1,t2] for the two main absorption peaks. The peak intensity variation pattern is described by recording the average slope parameter k before the peak reaches the inflection point. The kinetic model parameters are derived from a specific set of coefficients C=[c1,c2] in the piecewise function used when constructing the dynamic content variation model. The feature parameters of each historical calibration model in the historical calibration model set are extracted in the same way as the dynamic content variation model. For each model in the historical calibration model set, the same feature parameter extraction process is performed to obtain the coefficient vector P corresponding to each historical model. (i) Time series table T (i) Slope parameter k (i) and the coefficient set C (i) The superscript i indicates different historical calibration models.

[0029] Within a unified feature space, the Euclidean distance between the feature parameters of the dynamic content change model and the feature parameters of each historical calibration model is calculated. The unified feature space is achieved by concatenating all types of feature parameters into a comprehensive feature vector. For the dynamic content change model, its comprehensive feature vector is constructed as V=[P,T,k,C], and for the i-th historical calibration model, its comprehensive feature vector is constructed as V (i) =[P (i) ,T (i) ,k (i) C (i) ] Calculate the Euclidean distance:

[0030] Where: D (i) This represents the Euclidean distance between the dynamic model and the i-th historical model, characterizing the overall difference between the two models in the feature space. The Euclidean distance is mapped to a similarity value between zero and one using a preset similarity transformation function. This similarity value represents the matching similarity with the corresponding historical calibration model. The preset similarity transformation function uses an exponential decay form, and the formula is defined as follows: , where: S (i)D represents the matching similarity with the i-th historical calibration model. (i) This is the calculated Euclidean distance, where β is a scale adjustment parameter greater than zero used to control the rate at which similarity decays with distance. This function ensures that when the Euclidean distance D... (i) Matching similarity S is zero (i) The value is 1, and the matching similarity decreases towards 0 as the Euclidean distance increases.

[0031] In some embodiments, the overall trend of the curves in the feature parameters can be fitted using higher-order polynomial functions to capture more complex trend changes, such as cubic or quartic polynomials, which also increases the dimension of the coefficient vector. It is understood that when comparing the time-series parameter T of the feature peak occurrence, the time series needs to be aligned first, for example, shifting the time axis based on the occurrence time of the first feature peak, and then calculating the differences between corresponding time points. Optionally, the peak intensity change pattern parameter k can be obtained by calculating the average of the first derivatives within the peak region to more stably describe the rate of intensity change. In some embodiments, when calculating the Euclidean distance, weights can be applied to different sub-parameters in the comprehensive feature vector V to reflect the differences in importance of different feature dimensions in matching, for example, assigning higher weights to the dynamic model parameter C. It is understood that the similarity transformation function is not limited to the exponential decay form and can also use functions such as S... (i) =1 / (1+γ·D (i) The scaling factor β is an inverse proportional function of γ, where γ is the scaling parameter. Optionally, the value of the scaling parameter β can be learned from historical data so that the matching similarity can better distinguish historical models with different levels of similarity.

[0032] In one embodiment of the present invention, the model matching module performs weighted fusion of multiple historical calibration models based on matching similarity to generate an online real-time calibration model suitable for the current extraction process. The matching similarity is normalized so that its sum is one, which is used as the weight coefficient for each historical calibration model in the fusion process. In an example scenario involving the extraction of polyphenols from apple peels, the model matching module obtains matching similarities with three historical calibration models, with values ​​of 0.85, 0.60, and 0.35, respectively. These matching similarities are normalized to obtain the corresponding normalized weight coefficients. The core mapping function of each historical calibration model is obtained. The input of the core mapping function is a spectral feature vector, and the output is the predicted polyphenol content value. In the example, the core mapping functions of the three historical calibration models are model function h1(x) based on partial least squares regression, model function h2(x) based on support vector regression, and model function h3(x) based on ridge regression, respectively, where x represents the input spectral feature vector. By using normalized weight coefficients, the core mapping functions of multiple historical calibration models are linearly combined to generate an integrated mapping function. The linear combination formula is as follows: Where: H(x) represents the fusion-generated ensemble mapping function, N represents the total number of historical calibration models participating in the fusion, and λ i h represents the normalized weight coefficient corresponding to the i-th historical calibration model. i (x) represents the core mapping function of the i-th historical calibration model. Using a small number of spectral feature vectors from the initial extraction stage and their corresponding offline analysis reference values, the ensemble mapping function is fine-tuned to better fit the current process. In the apple peel extraction example, three sets of spectral feature vectors from the first 5 minutes and their actual polyphenol content values ​​measured by high-performance liquid chromatography were collected, forming a fine-tuning dataset {(x1,y1),(x2,y2),(x3,y3)}. The parameter θ in the ensemble mapping function is adjusted by minimizing the mean square error between the predicted values ​​and the offline analysis reference values. The fine-tuned and optimized ensemble mapping function is then established as the online real-time calibration model suitable for the current extraction process. See Table 1.

[0033] Table 1: Example Table of Fusion Parameters for Historical Calibration Models

[0034] In some embodiments, the normalization of weight coefficients can be achieved by directly dividing each matching similarity by the sum of all matching similarities. It can be understood that the core mapping function is obtained by directly loading its pre-trained algorithm structure and parameters from model files stored in the historical database. Optionally, the fusion of multiple core mapping functions can employ a non-linear combination method, for example, using the predicted values ​​of each model as input to train a shallow neural network as a meta-model to construct the ensemble mapping function H(x). In some embodiments, the fine-tuning optimization process can employ the gradient descent algorithm to fine-tune the loss function on the dataset. To guide the iterative updating of the adjustable parameter θ in the ensemble mapping function. It's understood that obtaining offline analysis reference values ​​requires simultaneous sampling with spectral acquisition and standard chemical analysis of the samples to ensure the reliability of the fine-tuning data. Optionally, an early stopping mechanism can be set in the fine-tuning optimization step, stopping optimization when the validation error no longer decreases to prevent overfitting on a small amount of data.

[0035] See Figure 4 This is a graph showing the loss variation during the fine-tuning optimization process of the ensemble mapping function in a polyphenol extraction model fusion scenario. The continuous decline of the red curve indicates that the gradient descent algorithm effectively minimizes the mean square error between the predicted and actual values, and the model converges continuously during iterations. The curve trend shows that the error reduction rate slows significantly after the 35th-40th iteration, consistent with the design logic of the early stopping mechanism, which avoids overfitting on a small amount of fine-tuning data. This graph directly corresponds to the model fine-tuning stage of "apple peel polyphenol extraction," verifying the effectiveness of optimizing the ensemble mapping function based on initial real data, and ensuring the accuracy of subsequent online detection. More accurate model predictions can provide reliable input to the process control module, helping to optimize parameters such as extraction temperature and stirring speed, thereby improving polyphenol dissolution efficiency and product yield.

[0036] In one embodiment of the present invention, the content calculation module inputs the spectral feature vector into the online real-time calibration model in real time to calculate the instantaneous detection value of polyphenol content. At the end of each new spectral acquisition and processing cycle, the newly generated spectral feature vector is obtained. Taking the extraction process of rosmarinic acid from rosemary leaves as an example, the spectral acquisition module outputs a new spectral feature vector containing absorbance of five characteristic wavelengths every 30 seconds. The latest spectral feature vector is used as input to the mapping function in the online real-time calibration model. The mapping function in the online real-time calibration model is an ensemble function H(x) after weighted fusion and fine-tuning optimization. Based on its internal algorithm, the mapping function processes the latest spectral feature vector and directly outputs a numerical value representing the polyphenol concentration, which is the instantaneous detection value of polyphenol content. For example, at the 15th minute of extraction, a spectral feature vector x is input. 15 The mapping function calculates y. 15 =H(x15 Output instantaneous detection value y 15 =2.34 mg / mL. A time-series smoothing filter was applied to the instantaneous detection values ​​of multiple continuously output polyphenol concentrations to eliminate random fluctuations and obtain a more stable instantaneous concentration sequence. A moving average filtering method with a window width of 5 was used, taking the arithmetic mean of the instantaneous detection values ​​at the current time and the four times preceding it as the smoothed output value at the current time.

[0037] The process control module establishes a dynamic correlation between the instantaneous detection value of polyphenol content and the extraction process parameters, and generates process parameter control instructions accordingly. It acquires the current process parameters of the extraction equipment in real time, including at least extraction temperature, stirring speed, and solvent flow rate. In the rosmarinic acid extraction example, the real-time acquired process parameters are: extraction temperature 65 degrees Celsius, stirring speed 300 rpm, and solvent flow rate 5 L / min. A dynamic correlation model is constructed, using the rate of change of the instantaneous detection value of polyphenol content as input and the adjustment amount of the process parameters as output. The dynamic correlation model can be a preset set of rules or a trained backpropagation neural network, with the functional relationship expressed as (ΔT, ΔR, ΔF) = G(r), where r represents the rate of change of the instantaneous detection value, ΔT, ΔR, and ΔF represent the adjustment amounts of extraction temperature, stirring speed, and solvent flow rate, respectively, and G represents the mapping rule of the dynamic correlation model. The instantaneous measured value of the smoothed polyphenol content at the current moment and its changing trend are input into the dynamic correlation model. The rate of change, r, is obtained by calculating the difference between the smoothed measured values ​​at the two most recent time points. Based on built-in rules or learned mapping relationships, the dynamic correlation model calculates the direction and amount of adjustment to each process parameter to optimize or stabilize the polyphenol dissolution efficiency. In the preset rule example, one rule is: "If the rate of change r has been continuously lower than the threshold r for the past ten minutes..." th The command will then be generated to increase the extraction temperature T by ΔT degrees Celsius. The adjustment direction and amount will be formatted into a command format that the extraction equipment control system can recognize and execute, generating a process parameter control command. For example, the command format is "SET_PARAMTEMP+2.5", which means increasing the temperature setpoint by 2.5 degrees Celsius.

[0038] In some embodiments, time-series smoothing filtering can employ an exponentially weighted moving average method, assigning higher weights to recent data to smooth noise while responding more quickly to changes. It is understood that the calculation of the instantaneous rate of change r of the detected value can be based on linear fitting of data over a longer time window to obtain a slope, thus achieving a more robust trend estimate. Optionally, the built-in rules of the dynamic correlation model G can be designed based on polyphenol dissolution kinetics, for example, increasing the temperature or stirring speed to enhance mass transfer when the rate of change of the detected value is too low; and decreasing the temperature to reduce energy consumption when approaching saturation. In some embodiments, process parameter control commands can be directly sent to the programmable logic controller of the extraction equipment via industrial communication protocols such as Modbus TCP or OPCUA. It is understood that, for safety reasons, the adjustment amounts ΔT, ΔR, and ΔF calculated by the dynamic correlation model need to undergo upper and lower limit processing before generating the final command to ensure that the process parameters are adjusted within a safe operating range. Optionally, the process control module can record each control command and the changes in the instantaneous detected values ​​before and after it, for subsequent iterative optimization of the mapping rules of the dynamic correlation model G.

[0039] See Figure 5 This is a dynamic control curve of process parameters in the extraction of plant polyphenols, showing the correspondence between real-time parameter changes and control commands for extraction temperature, solvent flow rate, and stirring speed. The curve clearly shows the control rhythm of "initially accelerating dissolution → later maintaining stability," which perfectly matches the three-stage characteristics of polyphenol extraction: "initial accelerated dissolution, stable dissolution, and dissolution saturation," verifying the effectiveness of the control algorithm. The steep changes in the dashed line command reflect the system's rapid response capability to process parameters; for example, the stirring speed can be adjusted by more than 10 rpm in a short time, ensuring dynamic optimization of dissolution efficiency. The stability of temperature and the dynamic adjustment of stirring speed not only ensure polyphenol dissolution efficiency but also avoid unnecessary energy consumption, providing data basis for subsequent fine-tuning of process parameters. For the continuous decrease in stirring speed in the later stage, an adaptive control strategy based on dissolution saturation can be explored to further improve energy utilization.

[0040] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A system for simultaneous online detection of plant polyphenol content during extraction, characterized in that, include: The spectral acquisition module is used to acquire multi-channel online spectral data of plant raw materials in the extraction equipment, and to perform preprocessing and feature extraction to obtain spectral feature vectors; The dynamic modeling module is used to parse multiple characteristic absorption peaks related to the polyphenol structure and their dynamic changes from the spectral feature vector, and to construct a dynamic content change model describing the polyphenol dissolution process based on the characteristic absorption peaks and their dynamic changes. The model matching module is used to call up a set of historical calibration models that match the current plant raw material type and extraction process conditions from the historical extraction database. The dynamic content change model is compared and matched with the set of historical calibration models in multiple dimensions to obtain the matching similarity. Based on the matching similarity, multiple historical calibration models are weighted and fused to generate an online real-time calibration model suitable for the current extraction process. The content calculation module is used to input the spectral feature vector into the online real-time calibration model in real time to calculate the instantaneous detection value of polyphenol content; The process control module is used to establish a dynamic correlation between the instantaneous detection value of the polyphenol content and the extraction process parameters, and generate process parameter control instructions accordingly. The construction of a dynamic content change model describing the polyphenol dissolution process based on the characteristic absorption peaks and their dynamic changes includes: With extraction time as the horizontal axis and the absorbance value or the rate of change of its derivative of each characteristic absorption peak as the vertical axis, multiple dynamic change curves of the characteristic absorption peaks are plotted. Calculate the rate of change of the dynamic change curve of each characteristic absorption peak at different extraction stages, including the initial dissolution acceleration stage, the stable dissolution stage, and the dissolution saturation stage; Based on the rate of change, a kinetic curve model of each characteristic absorption peak changing with extraction time is constructed using a piecewise linear or nonlinear function fitting method. The kinetic curve models of multiple characteristic absorption peaks changing with extraction time are weighted and superimposed, and the weighting coefficients are set based on the contribution of each characteristic absorption peak to the types and contents of polyphenols, thereby synthesizing a comprehensive dynamic content change model. The step of comparing and matching the dynamic content change model with the historical calibration model set using multi-dimensional features to obtain the matching similarity includes: Extract the feature parameters of the dynamic content change model, including the overall trend of the curve, the timing of the appearance of the characteristic peak, the peak intensity change pattern, and the dynamic model parameters; Extract the feature parameters of each historical calibration model in the historical calibration model set, using the same extraction method as the dynamic content change model; Within a unified feature space, calculate the Euclidean distance between the feature parameters of the dynamic content change model and the feature parameters of each of the historical calibration models. The Euclidean distance is mapped to a similarity value between zero and one using a preset similarity conversion function. This similarity value is the matching similarity with the corresponding historical calibration model.

2. The system for simultaneous online detection of plant polyphenol content during extraction according to claim 1, characterized in that, The process involves collecting multi-channel online spectral data of plant raw materials within the extraction equipment, preprocessing and extracting features to obtain spectral feature vectors, including: Real-time acquisition of multi-channel online spectral data of plant raw materials within the extraction equipment; The multi-channel online spectral data is subjected to spectral noise filtering and baseline correction to obtain preprocessed standard spectral data; Multi-band feature extraction is performed on the preprocessed standard spectral data to obtain a spectral feature vector containing the characteristic wavelength and the corresponding absorbance; The real-time acquisition of multi-channel online spectral data of plant raw materials in the extraction equipment includes: deploying online spectral probes at multiple key locations in the extraction equipment, including the area near the feed inlet, the main body area of ​​the extraction tank, the action area of ​​the stirring impeller, and the area near the liquid outlet. The online spectral probe is used to synchronously acquire the raw transmission or reflection spectra of plant raw material slurry at multiple characteristic wavelengths at a fixed frequency. The raw signals from each of the online spectral probes are amplified and converted from analog to digital to obtain a multi-channel discrete spectral data stream. The discrete spectral data streams of the multi-channel array are time-stamped and aligned, and integrated to form the multi-channel online spectral data that strictly corresponds to the extraction time axis.

3. The system for simultaneous online detection of plant polyphenol content during extraction according to claim 2, characterized in that, The process of filtering spectral noise and performing baseline correction on the multi-channel online spectral data to obtain preprocessed standard spectral data includes: The high-frequency random noise in the multi-channel online spectral data is separated and filtered out using the wavelet transform method to obtain the denoised spectral data; For the denoised spectral data, its spectral baseline is estimated by an iterative fitting algorithm. The spectral baseline reflects the non-characteristic background signal caused by light scattering and background absorption. Baseline correction is completed by subtracting the spectral baseline from the denoised spectral data; The baseline-corrected spectral data is vector-normalized to eliminate the overall intensity variation caused by differences in optical path and sample concentration, and finally the preprocessed standard spectral data is output.

4. The system for simultaneous online detection of plant polyphenol content during extraction according to claim 2, characterized in that, The step of performing multi-band feature extraction on the preprocessed standard spectral data to obtain a spectral feature vector containing characteristic wavelengths and corresponding absorbance includes: Within a preset polyphenol characteristic spectral range, the preprocessed standard spectral data is subjected to continuous first-order or second-order derivative transformations to enhance the fine structural features of the spectrum. In the derivative spectrum, all local extrema are identified and located, and the wavelength position corresponding to each extrema is marked as a potential characteristic wavelength. Based on preset peak width threshold and peak height intensity threshold, the potential characteristic wavelengths are screened to remove false peaks caused by noise; Calculate the absorbance value of each selected characteristic wavelength on the preprocessed standard spectral data; The selected characteristic wavelengths and their corresponding absorbance values ​​are arranged in wavelength order to form the spectral feature vector.

5. The system for simultaneous online detection of plant polyphenol content during extraction according to claim 1, characterized in that, The step of retrieving a set of historical calibration models from the historical extraction database that match the current plant material type and extraction process conditions includes: Receive input current detection task parameters, which include at least the variety, part, particle size, type of extraction solvent, and temperature parameters of the plant raw material; Using the current detection task parameters as query conditions, perform multi-dimensional matching and retrieval in the historical extraction database; The historical extraction database stores multiple sets of historical extraction records. Each set of records includes historical task parameters, historical spectral feature vector sequences, and sequences of true polyphenol content values ​​obtained through offline chemical analysis. From the successfully matched historical extraction records, the trained historical calibration model is extracted. The historical calibration model describes the mapping relationship between the spectral feature vector and the true value of polyphenol content under specific conditions. All successfully matched and extracted historical calibration models are collected to form the historical calibration model set.

6. The system for simultaneous online detection of plant polyphenol content during extraction according to claim 1, characterized in that, The step of weightedly fusing multiple historical calibration models based on the matching similarity to generate an online real-time calibration model suitable for the current extraction process includes: The matching similarity is normalized so that its sum is one, and this is used as the weight coefficient of each of the historical calibration models in the fusion process; Obtain the core mapping function for each of the historical calibration models. The input of the core mapping function is a spectral feature vector, and the output is the predicted value of polyphenol content. Using the weighting coefficients, the core mapping functions of multiple historical calibration models are linearly combined or more complexly fused across function spaces to generate an integrated mapping function; Using the spectral feature vectors of the small amount of initial extraction stage data collected at present and the corresponding offline analysis reference values, the integrated mapping function is fine-tuned and optimized to make the parameters more suitable for the current process. The finely tuned and optimized integrated mapping function is established as the online real-time calibration model applicable to the current extraction process.

7. The system for simultaneous online detection of plant polyphenol content during extraction according to claim 1, characterized in that, The step of inputting the spectral feature vector into the online real-time calibration model in real time to calculate the instantaneous detection value of polyphenol content includes: At the end of each new spectral acquisition and processing cycle, the newly generated spectral feature vector is obtained; The spectral feature vector is used as input and provided to the mapping function in the online real-time calibration model; The mapping function processes the spectral feature vector based on its internal algorithm and directly outputs a value representing the polyphenol concentration, which is the instantaneous detection value of the polyphenol content. The instantaneous detection values ​​of the polyphenol content in the continuous output are subjected to time-series smoothing filtering to eliminate random fluctuations and obtain a more stable instantaneous content sequence.

8. The system for simultaneous online detection of plant polyphenol content during extraction according to claim 1, characterized in that, The process of establishing a dynamic correlation between the instantaneous detection value of the polyphenol content and the extraction process parameters, and generating process parameter control instructions accordingly, includes: The current process parameters of the extraction equipment are acquired in real time, including at least the extraction temperature, stirring speed, and solvent flow rate. A dynamic correlation model is constructed, with the instantaneous change rate of the polyphenol content as input and the adjustment amount of the process parameters as output. The instantaneous detection value of the polyphenol content at the current moment and its changing trend are input into the dynamic correlation model; The dynamic correlation model calculates the direction and amount of adjustment to each process parameter to achieve the optimal or most stable polyphenol dissolution efficiency based on built-in rules or learned mapping relationships. The adjustment direction and adjustment amount are formatted into an instruction format that the extraction equipment control system can recognize and execute, thereby generating the process parameter control instruction.

Images