A method for real-time measurement of the content of components in crystallization and fermentation processes based on infrared spectroscopy instant learning technique

Abstract

The present application belongs to the field of industrial process detection, and relates to a method for real-time measurement of component content in crystallization and fermentation process based on infrared spectrum instant learning technology. The present application uses an infrared spectrum analyzer equipped with an immersion probe to build an online monitoring experimental platform for crystallization and fermentation process, so as to measure infrared spectrum data of a solution in a reaction process in real time and in situ. First, spectrum data and reference data are collected, second, a wavelet base function is determined and historical spectrum data is processed by function, then, similar data sets are selected for a current test sample, and then a functional partial least squares model is established based on the similar data sets, and finally, the current sample is brought into the established model to calculate the predicted value of the component concentration. The present application can achieve automatic and rapid detection of biomass, substrate concentration and product concentration, and is convenient for actual industrial application and popularization.

Description

A method for real-time measurement of component content during crystallization and fermentation processes based on infrared spectroscopy real-time learning technology Technical Field

[0001] This invention belongs to the field of industrial process detection, and relates to the detection of component content in crystallization and fermentation processes. It relates to a method for real-time measurement of component content in crystallization and fermentation processes based on infrared spectroscopy real-time learning technology. Background Technology

[0002] Crystallization and fermentation engineering are widely used in biopharmaceuticals, food production, and chemical industries. The crystallization and fermentation processes involve numerous parameters. In engineering applications, the production process is mainly controlled and optimized by adjusting the chemical and physical parameters. However, for component content parameters, offline measurement techniques are still used, which are time-consuming, labor-intensive, and resource-intensive, making online real-time detection difficult. Infrared spectroscopy, due to its advantages of being in-situ, rapid, non-destructive, and unaffected by particle size, has been increasingly used in recent years for detecting component content in crystallization and fermentation processes, offering better real-time performance compared to offline detection techniques. Under light irradiation, molecules undergo energy level transitions, leading to the absorption or emission of light. Taking near-infrared spectroscopy as an example, vibrational energy level transitions cause vibrational spectra, and photons of different frequencies possess different energy values. When the energy value of a photon is equal to that of an electronic energy level in a molecule or atom, light absorption occurs. The absorbing molecule or atom changes from its ground state to an excited state, forming an absorption spectrum. The near-infrared spectral band has a wavelength range of 780 nm to 2500 nm. This spectral region mainly contains the combination and overtone characteristics of the vibrations of hydrogen groups such as CH, NH, and OH. Since the substrates and products of crystallization and fermentation processes contain such hydrogen-containing groups, near-infrared spectroscopy is suitable for analyzing the content of components that are directly or indirectly related to hydrogen-containing groups during crystallization and fermentation processes.

[0003] Currently, although infrared spectroscopy has begun to be applied to detect the concentration of some substances in the crystallization and fermentation processes, it does not take into account the impact of data differences caused by process changes and the continuous variation characteristics of spectral data, resulting in low measurement accuracy of the detection model. Summary of the Invention

[0004] This invention provides a real-time detection and modeling method based on infrared spectroscopy real-time learning technology to predict the component content during crystallization and fermentation.

[0005] The technical solution adopted in this invention is as follows:

[0006] A method for real-time measurement of component content during crystallization and fermentation processes based on infrared spectroscopy real-time learning technology includes the following steps:

[0007] (1) Collect spectral data and reference data of the solution in the reaction vessel.

[0008] To eliminate noise from the external environment affecting the measured spectrum, background subtraction is employed. An optically stable medium is selected as the reference background. The near-infrared spectral data for each sampling is obtained by subtracting the apparent spectrum measured by the instrument from the reference background spectrum, thus eliminating measurement errors caused by noise. Simultaneously, the reaction solution is sampled and pretreated using centrifugation or dilution. Offline measurement methods are then used to determine the concentrations of each component in the solution.

[0009] (2) Determine wavelet basis functions and spectral data functionalization

[0010] Since each spectrum is a smooth curve, it can be regarded as a continuous smooth function. The approximate function of each spectrum is calculated by using wavelet basis function approximation. Wavelets with multi-scale, orthogonal and compact support properties are used as basis functions for approximation. Then, the approximate function corresponding to each historical spectrum sample is calculated by using the determined wavelet basis function and the least squares method.

[0011] (3) Select similar sample sets

[0012] For the current spectral sample, first calculate its corresponding approximate function, and then use a distance-based similarity index to select several spectral data with the smallest distance from the historical data and their corresponding quality data as the modeling training dataset.

[0013] (4) Functional partial least squares modeling

[0014] Based on the selected similar sample set, a predictive model between spectrum and mass is established using the functional partial least squares method, and the model parameters are determined using cross-validation.

[0015] (5) Real-time detection of component concentration

[0016] The sample to be tested is fed into the established functional partial least squares model to obtain the corresponding component concentration prediction value, thereby realizing real-time concentration detection.

[0017] The beneficial effects of this invention are as follows: This invention can detect biomass, substrate concentration, and product concentration in real time during crystallization and fermentation. Considering the differences in samples collected at different times during fermentation, a real-time learning modeling method is adopted. The most similar dataset is selected for modeling using a distance similarity index to ensure the training accuracy of the model. Furthermore, a functional data analysis method is used to effectively solve the problems of high-dimensional and nonlinear modeling of spectral data. This method is highly operable, requires less experience and technical expertise, and can achieve automatic and rapid detection of biomass, substrate concentration, and product concentration, making it convenient for practical industrial applications and promotion. Attached Figure Description

[0018] Figure 1 is a flowchart of the modeling process.

[0019] Figure 2 is a schematic diagram of an online monitoring experimental platform for the ethanol fermentation process based on near-infrared spectroscopy.

[0020] Figure 3 shows the near-infrared spectrum of the ethanol fermentation process. Detailed Implementation

[0021] This example demonstrates online monitoring of biomass, glucose concentration, and ethanol concentration during batch ethanol fermentation. The established online monitoring experimental platform for the ethanol fermentation process is shown in Figure 2. The fermenter has a capacity of 2.5L, and the temperature, pH value, and agitator speed within the tank are controlled by fermentation control equipment. A platinum thermometer (PT100) was used to measure the temperature in the fermenter, which was controlled by a heating device and circulating cooling water. The pH of the fermentation broth was adjusted using NaOH solution, and the pH value in the fermenter was monitored in real-time by a pH meter. An electric agitator was used to uniformly stir the solution in the fermenter. The strain used in this experiment was *Saccharomyces cerevisiae* 4126. Before the fermentation experiment, the inoculum needed to be cultured and activated before being inoculated into the fermentation medium.

[0022] The specific implementation steps of this invention are as follows:

[0023] Step 1: Acquisition of spectral and reference data

[0024] Near-infrared spectra were acquired using a Fourier transform near-infrared spectrometer and a matching immersion diffuse reflectance probe. Spectral acquisition settings: measurement wavenumber range of 4000 cm⁻¹. -1 -12000cm -1 The instrument resolution is 16cm. -1 The spectral scans were performed 64 times, the detector gain was 237.84, and air was used as the reference background for near-infrared spectral acquisition. The glucose concentration in the fermentation broth was measured using liquid chromatography, the ethanol concentration using gas chromatography, and the biomass (expressed as OD) using an enzyme-linked immunosorbent assay (ELISA) reader. All acquired spectral data were recorded as a matrix. Quality data is recorded as N represents the number of samples, and M and J represent the wavenumber and the number of mass variables in the spectrum, respectively.

[0025] Step 2: Determine the wavelet basis functions

[0026] Each spectrum can be represented as a smooth, continuous function, and expressed using wavelet basis functions.

[0027]

[0028] These are wavelet basis functions. These are the corresponding coefficients, K is the number of product functions, the superscript 'T' indicates transpose, and s represents the independent variable in the continuous function, for example, s represents the process running time. The wavelet basis functions can be determined using traditional wavelet transform methods. Then, the least squares method is used to calculate the fitting coefficient vector c. n .

[0029] Step 3: Select a similar sample set

[0030] For newly collected samples First, its functionalized expression is determined using the defined wavelet basis functions and the least squares method.

[0031]

[0032] c q These are the corresponding approximation coefficients. Calculate x. q The distances of (s) to all historical data samples are as follows:

[0033] d q,n = < x q (s),x n (s)>=∫ Ω x q (s)x n (s)ds (3)

[0034] Where d q,n It is sample x q Let (s) be the distance between the nth historical sample and the nth historical sample, where n = 1, 2, ..., N. The first N1 samples selected in ascending order of distance are denoted as the similar sample set. The corresponding quality data is denoted as

[0035] Step 4: Functional partial least squares modeling

[0036] Based on similar dataset x q,rep (s) and Y q,rep The least squares model for the function row is established as follows:

[0037]

[0038] Where <·, ·> are integration operators, for example <w(s), w(s)> = ∫ Ω Let w(s)ds, where w(s) is the load function, v is the load vector, and Ω is the domain. It is a coefficient vector, which can be obtained by solving the following model using the nonlinear iterative least squares method.

[0039]

[0040] The model parameters are then recalculated for subsequent online computation.

[0041]

[0042]

[0043] Where p j It is the j-th load vector, r i It is the first i There are t regression vectors, I is the identity matrix, and A is the number of features retained.

[0044] Step 5: Real-time detection of component concentration

[0045] For sample x q (s), the corresponding quality prediction value The calculation is as follows:

[0046]

[0047] The current sample x can be predicted using the trained parameters Θ. q Predicted quality value of (s)

Claims

1. A method for real-time measurement of component content during crystallization and fermentation processes based on infrared spectroscopy real-time learning technology, characterized by including: The following steps are taken: (1) Collect spectral data and reference data of the solution in the reaction vessel. In order to eliminate the noise caused by the external environment to the measured spectrum, the background subtraction method is used to remove noise. A medium with stable optical properties is selected as the reference background. The near-infrared spectral data of each sampling is obtained by the difference spectrum between the apparent spectrum measured by the instrument and the reference background, thereby eliminating the measurement error caused by noise. At the same time, the reaction solution is sampled and pretreated by centrifugation or dilution. The concentration data of each component in the solution is measured by offline measurement method. Near-infrared spectra are collected by Fourier transform near-infrared spectrometer and matching immersion diffuse reflection probe. The near-infrared spectra are collected with air as the reference background. The reference data are detected by offline detection method. Among them, the glucose concentration of the fermentation broth is measured by liquid chromatography, the ethanol concentration of the fermentation broth is measured by gas chromatography, and the biomass of the fermentation broth is measured by enzyme-linked immunosorbent assay (ELISA). All collected spectral data are recorded as a matrix. Quality data is recorded as N represents the number of samples, and M and J represent the wavenumber and the number of mass variables of the spectrum, respectively; (2) Determine the wavelet basis function and functionize the spectral data. Each spectrum is a smooth curve, which is regarded as a continuous smooth function. The approximate function of each spectrum is calculated by using the wavelet basis function approximation method; the wavelet with multi-scale, orthogonal and compact support characteristics is used as the basis function for approximation, and then the determined wavelet basis function and the least squares method are used to calculate the approximate function corresponding to each historical spectral sample; each spectrum is represented as a smooth continuous function and represented by the wavelet basis function. (1) Among them, These are wavelet basis functions. These are the corresponding coefficients, K is the number of product functions, and the superscript 'T' indicates transpose. The process runtime is represented; wavelet basis functions are determined using traditional wavelet transform methods. Then, the least squares method is used to calculate the fitting coefficient vector. (3) Selecting a similar sample set: For the current spectral sample, first calculate its corresponding approximate function, and then use the distance-based similarity index to select several spectral data with the smallest distance and their corresponding quality data from the historical data as the modeling training dataset; For newly collected samples First, its functionalized expression is determined using the defined wavelet basis functions and the least squares method. (2) Among them, These are the corresponding approximation coefficients; calculation The distances to all historical data samples are as follows (3) It is a sample The distance between the nth historical sample and the nth historical sample, n=1,2,...,N; select the previous samples in ascending order of distance. Let be the set of similar samples denoted as . The corresponding quality data is denoted as (4) Functional partial least squares modeling: Based on the selected similar sample set, a predictive model between spectrum and mass is established using functional partial least squares, and the model parameters are determined using cross-validation; based on similar datasets and The least squares model for the function row is established as follows: (4) It is the integration operator, and its definition is... ， It is a load function. It is a load vector. It is the domain; let ， These are coefficient vectors, obtained by solving the following model using the nonlinear iterative least squares method; (5) The coefficient matrix is ​​calculated from the selected similar historical spectral samples; then the model parameters are calculated for subsequent online computation. (6) (7) It is the j-th load vector. It is the first i A regression vector, It is the identity matrix. (5) Real-time detection of component concentration: Substitute the sample to be tested into the established functional partial least squares model to obtain the corresponding component concentration prediction value, and realize real-time concentration detection; For the sample The corresponding quality prediction value The calculation is as follows: (8) Utilize the trained parameters That is, the current sample can be predicted. Quality prediction 。 2. The method for real-time measurement of component content during crystallization and fermentation processes based on infrared spectroscopy real-time learning technology according to claim 1, is further characterized in that, In the first step, the spectral acquisition settings are as follows: the measurement wavenumber range is 4000 cm⁻¹. -1 -12000cm -1 The instrument resolution is 16 cm. -1 The number of spectral scans was 64, and the detector gain was 237.84.

Images