Hyaluronic acid fermentation process online monitoring method based on near infrared spectrum and multi-model fusion
By constructing an online near-infrared spectroscopy acquisition system and a multi-model fusion method, the problem of insufficient accuracy in determining the molecular weight of hyaluronic acid in traditional methods was solved, enabling real-time monitoring of the hyaluronic acid fermentation process and improving the yield and quality of hyaluronic acid.
Patent Information
- Application Number
- CN202511621966.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-07
- Publication Date
- 2026-02-27
AI Technical Summary
Traditional methods for determining the molecular weight of hyaluronic acid are easily affected by sample impurities and require high sample purity, resulting in insufficient accuracy and making it difficult to achieve real-time monitoring of the hyaluronic acid fermentation process.
An online near-infrared spectroscopy acquisition system was constructed. Combining a multi-model fusion method, the hyperparameters were optimized by improving the frilled lizard optimization algorithm. Multiple basic models were constructed and the predicted values were merged. A multiple linear regression model was trained to monitor the molecular weight of hyaluronic acid in real time.
This technology enables real-time and accurate monitoring of molecular weight during hyaluronic acid fermentation, improving monitoring timeliness and accuracy, providing timely feedback for fermentation process optimization, and enhancing the yield and quality of hyaluronic acid.
Smart Images
Figure CN121583331A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of biomanufacturing technology, and in particular to an online monitoring method for hyaluronic acid fermentation process based on near-infrared spectroscopy and multi-model fusion. Background Technology
[0002] Hyaluronic acid (HA) is a polysaccharide composed of glucuronic acid and N-acetylglucosamine, naturally occurring in various parts of the human body, such as the skin, synovial fluid, and eyes. This polysaccharide plays a crucial role in maintaining tissue health due to its strong water-retention capacity and its regulatory effect on the extracellular environment. The molecular chains of hyaluronic acid can absorb and lock in moisture, providing lubrication and elasticity to the skin, while also reducing joint friction and protecting cartilage.
[0003] Hyaluronic acid has a wide range of applications in the medical and cosmetic fields. High molecular weight hyaluronic acid is often used for injections to fill skin depressions or as a joint lubricant; while low molecular weight hyaluronic acid is more commonly used in skin care products due to its better permeability, which helps promote skin absorption. The molecular weight of hyaluronic acid has a significant impact on its efficacy; therefore, accurate determination of its molecular weight is crucial for ensuring product quality and therapeutic effects.
[0004] Traditional methods for determining the molecular weight of hyaluronic acid may include gel permeation chromatography (GPC) and light scattering (LST). GPC determines molecular weight by utilizing the different retention times of molecules with different molecular weights in a gel column. However, this method can be affected by impurities in the sample, leading to poor separation and affecting the accuracy of molecular weight determination. LST calculates molecular weight by measuring the scattering of light by hyaluronic acid molecules. However, this method requires high sample purity and sophisticated equipment, as well as specialized operating skills. Furthermore, LST is susceptible to interference from other large molecules in the sample. Summary of the Invention
[0005] The purpose of this invention is to provide an online monitoring method for the fermentation process of hyaluronic acid based on near-infrared spectroscopy and multi-model fusion, so as to realize the real-time and accurate monitoring of molecular weight, a key parameter in the fermentation process of hyaluronic acid, and provide strong technical support for optimizing the fermentation process and improving the yield and quality of hyaluronic acid.
[0006] To achieve the above objectives, the present invention provides the following solution: A near-infrared spectroscopy online acquisition system for the hyaluronic acid fermentation process was constructed, and appropriate instrument parameters were set to ensure the quality of the spectra. Near-infrared spectral data at different fermentation time points were collected online, and the collection operation was repeated for multiple batches of fermentation processes; Divide the dataset into training set and test set according to certain proportion; Based on the preprocessed near-infrared spectrum and the known HA molecular weight, a plurality of basic models are constructed by using a plurality of machine learning algorithms; The hyperparameters of the basic model are optimized by using the improved umbrella lizard optimization algorithm and the cross-validation method to determine the optimal hyperparameters; The basic model after hyperparameter optimization is used for prediction; Try different combinations of basic models, combine the corresponding multiple prediction values as new input features, and train a multiple linear regression model to generate the final HA molecular weight prediction value.
[0007] Optionally, the online near-infrared spectrum instrument parameters include wavelength range, wavelength interval, integration time, and scanning times.
[0008] Optionally, the SPXY algorithm is used for data set division.
[0009] Optionally, the combination of Savitzky-Golay (SG) smoothing method and standard normal variate transformation (SNV) method is used for near-infrared spectrum preprocessing, wherein the window width of the SG smoothing method is 7 and the order is 2.
[0010] Optionally, the plurality of machine learning algorithms include partial least squares (PLS) algorithm, support vector machine (SVM) algorithm, artificial neural network (ANN) algorithm, gradient boosting decision tree (GBDT) algorithm, and random forest (RF) algorithm.
[0011] Optionally, the improved umbrella lizard optimization algorithm introduces two optimization strategies, namely adaptive inertia weight strategy and elite reverse learning strategy.
[0012] Optionally, the formula of the adaptive inertia weight strategy is as follows:
[0013] Optionally, after introducing the adaptive inertia weight strategy, the position update formula is as follows:
[0014] Optionally, the formula of the elite reverse learning strategy is as follows:
[0015] Optionally, the optimization object of the improved chameleon optimization algorithm is an SVM, an ANN, a GBDT or an RF model.
[0016] Optionally, the hyperparameters of the SVM include a penalty coefficient and a kernel function coefficient; the hyperparameters of the ANN include an iteration number and a learning rate; the hyperparameters of the GBDT include a number of learners and a learning rate; and the hyperparameters of the RF include a number of decision trees and a maximum depth of the decision trees.
[0017] Optionally, the optimization object of the cross-validation method is a PLS model, and the optimization parameter is a number of latent variables (LVs).
[0018] Optionally, the number of the basic models is set to 3 based on a task complexity, data diversity and a requirement of a computing resource.
[0019] Optionally, the RMSE and the RMSEP are used as the indexes for evaluating the models.
[0020] Compared with the prior art, the present application has the following advantages: 1. The present application successfully constructs an online near-infrared spectrum acquisition system, realizes real-time monitoring of the molecular weight in the fermentation process of hyaluronic acid, greatly improves the timeliness of the monitoring compared with the traditional offline detection method, and can timely feedback the changes in the fermentation process, thereby providing a more timely basis for process adjustment.
[0021] 2. The present application adopts a multi-model fusion method. Different algorithms can capture different data features and patterns, and the prediction results of these models can be fused to more comprehensively and accurately reflect the changes in the molecular weight in the fermentation process of hyaluronic acid, thereby providing more accurate guidance for optimizing the fermentation process. BRIEF DESCRIPTION OF DRAWINGS
[0022] Figure 1 A schematic diagram of online monitoring of the fermentation process of hyaluronic acid based on near-infrared spectrum and multi-model fusion is provided for the embodiments of the present application; Figure 2 A raw near-infrared spectrum diagram is provided for the embodiments of the present application; Figure 3 A preprocessed near-infrared spectrum diagram is provided for the embodiments of the present application; Figure 4 A test set prediction scatter plot of the best single model is provided for the embodiments of the present application; Figure 5 A test set prediction scatter plot of the best fusion model is provided for the embodiments of the present application. DETAILED DESCRIPTION
[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0024] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0025] The online monitoring method for hyaluronic acid fermentation based on near-infrared spectroscopy and multi-model fusion provided in this invention is illustrated in the flowchart below. Figure 1 As shown, it includes: Construct an online near-infrared spectral acquisition system for the hyaluronic acid fermentation process; Specifically, when constructing a near-infrared spectroscopy online acquisition system, selecting a suitable spectral acquisition instrument is crucial. Parameters such as wavelength range, wavelength spacing, integration time, and number of scans must be comprehensively considered to accurately capture effective information during the fermentation process. Given the high temperature, high pressure, and complex chemical composition of the fermentation environment, the probe must possess excellent corrosion resistance and resistance to high temperature and pressure.
[0026] Near-infrared spectra and molecular weights were collected from different batches at different fermentation time points. Specifically, different batches of fermentation samples were prepared to ensure that the starting conditions for each batch remained consistent during fermentation. Based on the expected fermentation progress and past experience, multiple different fermentation time points were set for data collection. For each batch, data was collected at the same preset time point. When acquiring near-infrared spectra online, the spectral data were associated and saved with the corresponding batch number and time point information. The molecular weight of the samples was determined using gel permeation chromatography as a reference value for modeling.
[0027] The dataset is divided into training and testing sets according to a certain ratio; Specifically, a suitable partitioning ratio is determined by comprehensively considering factors such as data size, data distribution characteristics, and model complexity. To maintain the consistency of data distribution and avoid significant deviations in data features between the training and test sets, the SPXY algorithm is chosen for partitioning.
[0028] Preprocess the raw near-infrared spectrum; Specifically, a combination of SG smoothing and SNV methods was used to preprocess the near-infrared spectra. The SG smoothing method had a window width of 7 and an order of 2. The SG smoothing method mainly focuses on removing noise and making the spectral curve smoother, while the SNV method is mainly used to eliminate baseline drift.
[0029] A plurality of basic models are established by using the pretreated spectrum and known molecular weight; Specifically, the plurality of basic models include a PLS model, a SVM model, an ANN model, a GBDT model, and an RF model.
[0030] The hyperparameters of the basic models are optimized by improving the umbrella lizard optimization algorithm and the cross-validation method. Specifically, the improved umbrella lizard optimization algorithm introduces two optimization strategies, namely, an adaptive inertia weight strategy and an elite reverse learning strategy. The optimization objects of the algorithm are the SVM, ANN, GBDT, and RF models. Among them, the hyperparameters of the SVM include a penalty coefficient and a kernel function coefficient; the hyperparameters of the ANN include an iteration number and a learning rate; the hyperparameters of the GBDT include a learner number and a learning rate; and the hyperparameters of the RF include a decision tree number and a maximum depth of the decision tree. In addition, the optimization object of the cross-validation method is the PLS model, and the optimization parameter is the number of latent variables.
[0031] Further, the formula of the adaptive inertia weight strategy is as follows:
[0032] Further, after introducing the adaptive inertia weight strategy, the position update formula is as follows:
[0033] Further, the formula of the elite reverse learning strategy is as follows:
[0034] Different combinations of the basic models are tried, the corresponding multiple prediction values are combined as new input features, and a multiple linear regression model is trained to generate the final molecular weight prediction value.
[0035] Specifically, the prediction values of each basic model are regarded as a characteristic variable, and multiple characteristic variables are combined to form a meta-feature. A multiple linear regression model is constructed by using the meta-feature and the known molecular weight. Different combinations of characteristic variables are tried, and the best basic model combination is selected as the optimal model. EMBODIMENT
[0036] The online monitoring method for the hyaluronic acid fermentation process based on near-infrared spectroscopy and multi-model fusion includes: (1) Construction of a near-infrared spectrum online acquisition system: A custom-made high-temperature and corrosion resistant immersion diffuse reflectance probe (Beijing Zhikong Technology Co., Ltd.) was installed at the electrode port of the fermenter and connected to the PRO-1900 online near-infrared spectrometer through an optical fiber. The obtained spectrum contained 701 variables, with a wavelength range of 950-1650 nm. The integration time and scan number were 10 ms and 50 times, respectively.
[0037] (2) HA sample collection: To ensure that the model built can adapt to the determination of hyaluronic acid molecular weight under various conditions, it is necessary to collect samples from multiple fermentation batches and the whole fermentation process. In one fermentation batch, the sampling time interval is 1 h, and 24 samples are collected continuously for 24 h. In 4 independent fermentation batches, a total of 96 representative samples are finally collected. The batch number and time are recorded when sampling.
[0038] (3) Near-infrared spectrum acquisition: At the same time of collecting samples, the HxNIR System Online 2000 software is used to collect near-infrared spectra, and the batch and time information of the spectra are saved, so that the spectra can be corresponded with the molecular weight in the future.
[0039] (4) Molecular weight determination: The molecular weight of all samples is determined by gel permeation chromatography. The determination instrument used in the present application includes Waters 600E high performance liquid chromatography system and DAWN EOS multi-angle laser light scattering detector. The sample is separated on an SB-806 HQ gel filtration column, and the column temperature is maintained at 35 °C during the process. The mobile phase is selected as 0.1 mol / L sodium chloride solution, and the flow rate is set to 0.6 mL per minute. The sample injection amount is 200 μL each time. During the analysis process, the specific refractive index increment value used is 0.160.
[0040] (5) Divide the data set: The data set is divided into training set and test set by SPXY algorithm. Compared with random division, Kennard-Stone algorithm and other methods, this method considers both X variables and Y variables, which can more effectively cover the multi-dimensional vector space, increase the difference and representativeness between samples, and thus improve the stability of the model. The ratio of training set and test set is 4:1, wherein the training set contains 77 samples and the test set contains 19 samples. Both have similar statistical parameters such as mean and standard deviation, which ensures the effectiveness and reliability of model evaluation.
[0041] (6) Pretreatment of near infrared spectrum: first, the original near infrared spectrum is processed by SG smoothing method to remove the noise existing in the spectrum, so that the spectrum curve is smoother. Among them, the window width is set to 7, and the order is 2. Then, the smoothed spectrum is pretreated by SNV to eliminate baseline drift.
[0042] (7) Construction of base model: In order to overcome the limitations of single algorithm and improve the prediction accuracy and generalization ability of quantitative model, the method of multi-model fusion is introduced. By training multiple base models, the outputs of these base models are used as new inputs to train a multiple linear regression model to realize accurate and robust prediction. The present application constructs 5 base models with different characteristics and advantages. Among them, the advantage of PLS model lies in the strong linear relationship search ability, which can accurately mine linear information; SVM model can use kernel function to process nonlinear relationship; ANN model has good nonlinear fitting ability and can process highly complex nonlinear data; GBDT model can capture small changes in data after optimization; RF model has good robustness and generalization ability. Table 1 shows the modeling results of different base models.
[0043]
[0044] R2 is the determination coefficient of the training set; RMSEC (%) is the root mean square error of the training set; R2 is the determination coefficient of the test set; RMSEP (%) is the root mean square error of the test set.
[0045] (8) Optimization of base model hyperparameters: The improved chameleon optimization algorithm is used to optimize the model hyperparameters, including the penalty coefficient and kernel function coefficient of the SVM model, the iteration number and learning rate of the ANN model, the number of learners and learning rate of the GBDT model, and the number of decision trees and the maximum depth of the decision tree of the RF model. The target is to reduce the RMSECV under five-fold cross validation, and 50 iterations are carried out. After iteration, the prediction accuracy of different models is improved to a certain extent, which shows that the adjustment of hyperparameters can make the model better adapt to the characteristics and rules of data, so as to improve the accuracy and reliability of the model in practical application. At the same time, the number of latent variables of PLS model is optimized by five-fold cross validation method. Table 2 shows the best hyperparameters of different models.
[0046]
[0047] (9) Multi-model fusion: try different combinations of basic models to build the final multiple linear regression model. Table 3 summarizes the prediction results of the optimized single model and the fusion model. Compared with the single model, the prediction accuracy of the fusion model has been significantly improved. This shows that this strategy maximizes the advantages and characteristics of different algorithms, enabling the model to more comprehensively and accurately capture the complex patterns and rules in the data, overcoming the one-sidedness and limitations of single models. In addition, the combination of SVM, GBDT and PLS models shows the best prediction performance and generalization ability. The determination coefficient of the training set of the optimal model is 0.9378, the root mean square error of the training set is 10.54%, the determination coefficient of the test set is 0.9274, and the root mean square error of the test set is 11.26%.
[0048]
[0049] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, devices and units can refer to the corresponding processes in the foregoing method embodiments, which will not be repeated here.
[0050] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the scope of protection of the present application.
Claims
1. A method for online monitoring of hyaluronic acid fermentation process based on near-infrared spectroscopy and multi-model fusion, characterized in that, include: Step 1: Construct an online near-infrared spectroscopy acquisition system for the hyaluronic acid fermentation process; Step 2: Collect near-infrared spectra and molecular weight data at different fermentation time points for different batches; Step 3: Divide the dataset into training and test sets according to a certain ratio; Step 4: Preprocess the raw near-infrared spectrum; Step 5: Establish multiple basic models using the preprocessed spectra and known molecular weights; Step 6: Optimize the hyperparameters of the basic model by improving the frilled lizard optimization algorithm and using cross-validation; Step 7: Try different combinations of basic models, merge the corresponding multiple predictions as new input features, and train a multiple linear regression model to generate the final molecular weight prediction value.
2. The method according to claim 1, characterized in that, The construction of the near-infrared spectroscopy online acquisition system in step 1 includes selecting a suitable spectrometer, setting key parameters such as wavelength range, wavelength spacing, integration time, and number of scans.
3. The method according to claim 1, characterized in that, In step 2, the starting conditions for different batches of fermentation were kept consistent. A series of equally spaced sampling time points were set to ensure coverage of the entire fermentation process; for each batch, samples and data were collected at the same preset time points. Molecular weight was determined using gel permeation chromatography, and the measurement results were used as reference values.
4. The method according to claim 1, characterized in that, In step 3, the SPXY algorithm is used to divide the training set and the test set so that they have similar data distributions, thereby ensuring the reliability of the model evaluation.
5. The method according to claim 1, characterized in that, In step 4, the preprocessing of the raw near-infrared spectra uniformly employs a combination of SG smoothing and SNV processing. SG smoothing removes noise, making the spectral curves smoother, while SNV processing is mainly used to eliminate baseline drift. The combination of the two can effectively improve the quality of spectral data, laying a solid foundation for subsequent analysis and modeling.
6. The method according to claim 1, characterized in that, The basic models established in step 5 include the PLS model, SVM model, ANN model, GBDT model, and RF model.
7. The method according to claim 1, characterized in that, Step 6 improves the frilled lizard optimization algorithm by introducing two strategies: an adaptive inertia weighting strategy and an elite back-learning strategy. The optimization targets include the penalty coefficient and kernel function coefficient of the SVM model, the number of iterations and learning rate of the ANN model, the number of learners and learning rate of the GBDT model, and the number of decision trees and the maximum depth of the decision trees in the RF model. Additionally, the number of latent variables in the PLS model is optimized using five-fold cross-validation.
8. The method according to claim 7, characterized in that, The formula for the adaptive inertia weighting strategy is as follows:
9. The method according to claim 7, characterized in that, After introducing the adaptive inertia weighting strategy, the position update formula is as follows:
10. The method according to claim 7, characterized in that, The formula for the elite reverse learning strategy is as follows:
11. The method according to claim 1, characterized in that, In step 7, the prediction results from different base models are combined as new input features to train a multiple linear regression model to produce the final prediction result. By comparing the modeling performance of different combinations of base models, the optimal combination is selected to build a model with good prediction performance and strong generalization ability.