Search

CN-122024881-A - Modeling method of fermentation dynamics model, fermentation dynamics model built by modeling method and application of fermentation dynamics model

CN122024881ACN 122024881 ACN122024881 ACN 122024881ACN-122024881-A

Abstract

The invention belongs to the technical field of brewing, and relates to a modeling method of a fermentation dynamics model, the built fermentation dynamics model and application thereof, wherein the modeling method comprises the following steps of collecting fermentation pile shape information and multisource observation data; the method comprises the steps of constructing a fermentation pile shape model, constructing a shape equation under a cylindrical coordinate system, constructing a fermentation dynamics mathematical model taking physical and chemical indexes as objects, designing a physical information neural network structure and a loss function, training a network and inverting parameters, reconstructing four-dimensional space-time distribution of the physical and chemical indexes and visualizing the four-dimensional space-time distribution. The fermentation pile is abstracted into a composite geometrical body of an upper ellipsoid and a lower cone, the model is combined with multisource observation data acquired in the pile body through a physical information neural network algorithm, four-dimensional space-time distribution of physicochemical indexes in the pile fermentation process is reconstructed, visual monitoring and early prediction of the fermentation state are finally realized, and the calculation result of the model is more in line with the actual situation.

Inventors

  • LIU YONG
  • TIAN YUAN
  • LI XIAOHONG
  • DAI BING
  • YAN SUI
  • ZHANG XIAOMING
  • Zuo Jiarong

Assignees

  • 贵州茅台酒股份有限公司

Dates

Publication Date
20260512
Application Date
20260107

Claims (10)

  1. 1. A method for modeling a fermentation kinetic model, comprising the steps of: step 1, collecting fermentation pile shape information and multi-source observation data, wherein the fermentation pile shape information comprises fermentation pile height and radius, and the multi-source observation data comprises physicochemical index data; Step 2, constructing a fermentation pile shape model, namely abstracting the fermentation pile into a composite geometrical body of an upper ellipsoid and a lower cone, building a shape equation under a cylindrical coordinate system, and simplifying three-dimensional space distribution into two-dimensional distribution of radial sections by assuming that physicochemical indexes have circumferential symmetry; step 3, constructing a fermentation dynamics mathematical model taking physical and chemical growth indexes as objects, namely, based on a fermentation principle, establishing a partial differential equation set containing diffusion terms and source terms, describing evolution rules of the physical and chemical growth indexes in radial, axial and time dimensions, and setting boundary conditions and initial conditions of a symmetrical boundary, a bottom boundary and a top boundary of a pile; Step 4, designing a physical information neural network structure and a loss function, namely constructing a multi-layer perceptron with six fully connected layers, introducing a Sigmoid-WEIGHTED LINEAR Unit activation function, and constructing a composite loss function comprising control equation residual loss, boundary condition loss, initial condition loss and observation data fitting loss; Training a network and inverting parameters, namely adopting Adaptive Moment Estimation optimizers to minimize a composite loss function, taking unknown dynamic parameters as variables to be optimized, and updating the network parameters and the dynamic parameters through back propagation iteration to realize inversion of key parameters; and 6, reconstructing four-dimensional space-time distribution of the physical and chemical indexes and visualizing, namely reconstructing dynamic distribution of the physical and chemical indexes in a three-dimensional space and one-dimensional time through a trained physical information neural network based on inverted dynamic parameters to generate a visualized result.
  2. 2. The method of modeling a fermentation pile of claim 1, wherein the physicochemical index data in step 1 includes temperature, oxygen concentration, ethanol concentration, moisture content, starch concentration, soluble sugar concentration, yeast concentration, organic acid concentration, and ester concentration.
  3. 3. The method of modeling a fermentation pile of claim 1, wherein the shape equation of the fermentation pile in step 2 is as follows: ; Wherein alpha is an ellipsoidal long half shaft, For the height of the pile of fermentation, For the radius of the bottom of the pile, Is the joint height of the ellipsoid and the cone, In the form of a radial coordinate, As the axial coordinate of the two-dimensional coordinate system, Is the circumferential coordinate.
  4. 4. The method of modeling a fermentation pile of claim 1, wherein the physicochemical index data in step 1 are from different depths and different radial positions of the fermentation pile.
  5. 5. The method of modeling a fermentation pile of claim 1, wherein the set of partial differential equations of the mathematical model of fermentation dynamics in step 3 is: ; Wherein, the Is the distribution function of the ith physicochemical index, In order for the diffusion coefficient to be the same, As a function of the source term(s), For convective mass/heat transfer coefficients, As the value of the environment outside the boundary, Is a radial cross-sectional open area, 、 、 Respectively a symmetrical boundary, a top boundary and a bottom boundary.
  6. 6. The method of modeling a fermentation pile of claim 1, wherein the composite loss function in step 4 is: ; Wherein, the In order to control the equation residual loss, For the bottom boundary loss to occur, In order for the boundary to be lost to symmetry, For the top boundary loss to be a function of, In order to achieve the loss of the initial condition, For the purpose of fitting the loss to the observed data, ~ Is the weight coefficient of the corresponding loss function.
  7. 7. The method of modeling a fermentation pile of claim 1, wherein the fermentation pile comprises at least one of a fermented grain fermentation pile and a Daqu fermentation pile.
  8. 8. The method of modeling a fermentation pile according to claim 1, wherein the visual result in step 6 includes a spatial-temporal distribution dynamic video of physicochemical indexes, which can intuitively display the spatial distribution characteristics of the physicochemical indexes in the fermentation pile at any time point.
  9. 9. A fermentation kinetics model with coupled multi-physical and chemical indexes, which is constructed by the fermentation pile modeling method according to any one of claims 1-8, wherein the fermentation kinetics model is constructed based on an ellipsoid-cone composite geometry of a fermentation pile, and the following control equation set is satisfied: ; Wherein, the Is the distribution function of the ith physicochemical index, In order for the diffusion coefficient to be the same, As a function of the source term(s), For convective mass/heat transfer coefficients, As the value of the environment outside the boundary, Is a radial cross-sectional open area, 、 、 Respectively a symmetrical boundary, a top boundary and a bottom boundary.
  10. 10. Use of a fermentation kinetics model as claimed in claim 9 in the field of white spirit brewing, wherein the fermentation kinetics model is used to demonstrate the spatial distribution characteristics of the physicochemical properties of a fermentation pile.

Description

Modeling method of fermentation dynamics model, fermentation dynamics model built by modeling method and application of fermentation dynamics model Technical Field The invention belongs to the technical field of brewing, and relates to a modeling method of a fermentation dynamics model, the constructed fermentation dynamics model and application thereof. Background The pile-up fermentation of Maotai-flavor white spirit is a core link for determining the flavor and quality of the spirit body, in the pile-up fermentation process, brewing raw materials (sorghum, wheat and the like) and Daqu are mixed and piled into approximate hemispherical piles with the height of 1 m-1.5 m and the radius of 1-2 m, and key reactions such as starch saccharification, alcohol fermentation, flavor precursor substance generation and the like are completed by utilizing the synergistic effect of natural microorganisms and Daqu microorganisms in an open environment. In order to realize accurate regulation and control of the fermentation process, the industry tries to analyze the index change rule in the pile body of the piled fermentation by a modeling method, builds a mathematical model based on classical heat transfer and mass transfer equations, but presumes the pile body as a regular cylinder or cuboid, ignores the real form of the pile body, and has larger deviation between the calculation result of the model and the actual situation. Disclosure of Invention The invention aims to provide a modeling method of a fermentation dynamics model, the built fermentation dynamics model and application thereof, wherein the modeling method of the fermentation dynamics model constructs the fermentation pile shape model by collecting fermentation pile shape information and multisource observation data, abstracts the fermentation pile into a composite geometrical body of an upper ellipsoid and a lower cone, combines the built model with multisource observation data collected in the pile body through a physical information neural network algorithm, realizes effective inversion and identification of key physicochemical parameters, reconstructs four-dimensional space-time distribution of physicochemical indexes in a pile fermentation process, finally realizes visual monitoring and advanced prediction of fermentation states, and better accords with practical conditions. In a first aspect, the present invention provides a method for modeling a fermentation kinetic model, comprising the steps of: step 1, collecting fermentation pile shape information and multi-source observation data, wherein the fermentation pile shape information comprises fermentation pile height and radius, and the multi-source observation data comprises physicochemical index data; Step 2, constructing a fermentation pile shape model, namely abstracting the fermentation pile into a composite geometrical body of an upper ellipsoid and a lower cone, building a shape equation under a cylindrical coordinate system, and simplifying three-dimensional space distribution into two-dimensional distribution of radial sections by assuming that physicochemical indexes have circumferential symmetry; step 3, constructing a fermentation dynamics mathematical model taking physical and chemical growth indexes as objects, namely, based on a fermentation principle, establishing a partial differential equation set containing diffusion terms and source terms, describing evolution rules of the physical and chemical growth indexes in radial, axial and time dimensions, and setting boundary conditions and initial conditions of a symmetrical boundary, a bottom boundary and a top boundary of a pile; Step 4, designing a physical information neural network structure and a loss function, namely constructing a multi-layer perceptron with six fully connected layers, introducing SiLU (Sigmoid-WEIGHTED LINEAR Unit) activation functions, and constructing a composite loss function containing control equation residual loss, boundary condition loss, initial condition loss and observation data fitting loss; Training a network and inverting parameters, namely adopting Adam (Adaptive Moment Estimation) optimizers to minimize a composite loss function, taking unknown dynamic parameters as variables to be optimized, and updating the network parameters and the dynamic parameters through back propagation iteration to realize inversion of key parameters; and 6, reconstructing four-dimensional space-time distribution of the physical and chemical indexes and visualizing, namely reconstructing dynamic distribution of the physical and chemical indexes in a three-dimensional space and one-dimensional time through a trained physical information neural network based on inverted dynamic parameters to generate a visualized result. In some embodiments, the physicochemical index data in step 1 includes temperature, oxygen concentration, ethanol concentration, moisture content, starch concentration, soluble sugar concentration, yeast co