CN-122021456-A - Machine learning driven fermentation device hydrodynamic multi-field coupling simulation method

CN122021456ACN 122021456 ACN122021456 ACN 122021456ACN-122021456-A

Abstract

The invention provides a machine learning driven fermentation device hydrodynamic multi-field coupling simulation method which comprises the following steps of S1, biological-physical mapping modeling and multi-dimensional data preprocessing, S2, adaptive multi-scale grid generation and field intensity association optimization, S3, multi-field coupling control equation correction and interaction modeling, S4, machine learning prediction and real-time updating of dynamic boundary conditions, S5, multi-field coupling simulation parallel calculation and machine learning iteration optimization, and S6, simulation result multi-dimensional verification and machine learning model calibration. According to the invention, through constructing a bidirectional coordination mechanism among algorithms, field intensity quantification and grid generation adaptation, data driving and physical priori fusion, dynamic boundary condition accurate prediction and full-flow model iterative optimization are realized, and finally, the accurate and efficient simulation of multi-field coupling characteristics in the fermentation process is realized, so that a core technical support is provided for fermentation process optimization and intelligent control.

Inventors

LIN JUN
LIU ZHENXING
LI MINGHAI
CHANG MAI

Assignees

南京师范大学常州创新发展研究院

Dates

Publication Date: 20260512
Application Date: 20260320

Claims (10)

1. The machine learning driven fermentation device hydrodynamic multi-field coupling simulation method is characterized by comprising the following steps of: s1, collecting experimental measurement data and real-time monitoring data of microbial activity in a fermentation process, preprocessing, and constructing a biological-physical mapping model to obtain a biological-physical association weight matrix; s2, based on a biological-physical association weight matrix, combining three-dimensional geometric characteristic data of a fermentation device, performing bidirectional interaction with a graph annotating force enhanced generation countermeasure network through an anisotropic kernel density estimation algorithm to generate a self-adaptive multi-scale grid and field intensity association coefficient; S3, constructing a control equation correction model based on the preprocessed experimental measurement data, the self-adaptive multi-scale grid and the field intensity association coefficient, performing bidirectional interaction with a sparse self-encoder guided by physical information through a hierarchical Bayesian variation inference algorithm, finishing correction of a multi-field coupling control equation, and outputting the corrected multi-field coupling control equation; S4, acquiring a calculation result of the corrected multi-field coupling control equation, inputting real-time monitoring data of microbial activity into a long-period memory network model to obtain a dynamic boundary condition predicted value, comparing calculation errors with real-time boundary condition data acquired by an online sensor, completing real-time updating of the dynamic boundary condition or calibration of the long-period memory network model according to the error result, and outputting the updated dynamic boundary condition; S5, constructing a parallel computing framework based on the self-adaptive multi-scale grid, solving a corrected multi-field coupling control equation by combining with updated dynamic boundary conditions, realizing data communication and field quantity fusion between subareas through a GPU shared memory to obtain a global multi-field distribution result, feeding the global multi-field distribution result back to a biological-physical mapping model and a control equation correction model, completing iterative optimization, and outputting a simulation result; And S6, performing multidimensional verification on simulation results by adopting experimental data of independent fermentation batches, constructing an error source analysis model to determine error types, performing targeted calibration on each machine learning model according to the error types, repeating the steps S2 to S5 until the verification reaches the standard, and outputting a final simulation result.
2. The machine learning driven fermentation device hydrodynamic multi-field coupling simulation method of claim 1 is characterized in that in S2, the bidirectional interaction process of an anisotropic kernel density estimation algorithm and a graph meaning force enhancement generation countermeasure network is specifically that the anisotropic kernel density estimation algorithm utilizes a biological-physical association weight matrix to initialize field intensity association coefficients, the graph meaning force enhancement generation countermeasure network takes the field intensity association coefficients as weight items of a graph attention layer, the density distribution of a self-adaptive multi-scale grid is restrained and the grid is generated, the graph meaning force enhancement generation countermeasure network calculates error average values of the generation grid and theoretical grids derived based on the field intensity association coefficients, the error average values are fed back to the anisotropic kernel density estimation algorithm for updating an anisotropic bandwidth matrix of the anisotropic kernel density estimation algorithm, and the interaction process is repeated until a field intensity density estimation loss function and a generation loss function meet preset convergence conditions.
3. The machine learning driven fermentation device hydrodynamic multi-field coupling simulation method of claim 1, wherein the bidirectional interaction process of the hierarchical Bayesian variation inference algorithm and the physical information guided sparse self-encoder in S3 is specifically that the likelihood function correction process of the low-dimensional physical feature extracted by the physical information guided sparse self-encoder is integrated into the hierarchical Bayesian variation inference algorithm, the inference direction of the correction coefficient of the control equation is constrained, the optimal correction coefficient output by the hierarchical Bayesian variation inference algorithm is integrated into the encoder input of the physical information guided sparse self-encoder, the physical feature extraction direction is optimized, and the interaction process is repeated until the evidence lower bound of the hierarchical Bayesian variation inference algorithm is converged and the total loss function of the physical information guided sparse self-encoder meets the preset convergence condition.
4. The machine-learning-driven fermentation device hydrodynamic multi-field coupling simulation method of claim 1, wherein the closed loop process and the end condition of the iterative optimization in S5 are specifically that the field quantity errors are calculated by comparing the global multi-field distribution result with experimental measurement data, the weight parameters of the biological-physical mapping model are updated based on the field quantity errors, the global multi-field distribution result is supplemented as a training sample of a control equation correction model and the correction coefficient is updated, the iterative process is repeated, and when the field quantity error drop amplitude of 3 continuous iterations is smaller than a preset threshold value, the iterative optimization is stopped and the current global multi-field distribution result is output.
5. The machine-learning-driven fermentation device hydrodynamic multi-field coupling simulation method of claim 1 is characterized in that in the step S6, the multi-dimensional verification and targeted calibration process comprises the steps of calculating mean square errors and correlation coefficients of simulation results and independent fermentation batch experimental data from four dimensions of a fluid field, a temperature field, a concentration field and biological-physical coupling, taking verification indexes of each dimension, field intensity correlation coefficients and boundary condition prediction errors of an error source analysis model as input, outputting error source types, and respectively adjusting structural parameters or training data of a corresponding machine learning model according to the error source types to finish targeted calibration.
6. The machine learning driven fermentation device hydrodynamic multi-field coupling simulation method of claim 1 is characterized in that the real-time updating of the dynamic boundary condition or the calibration process of the long-short-period memory network model in S4 is specifically that a preset error threshold value is adopted, when the prediction error is larger than the preset threshold value, the weight parameter of the long-short-period memory network model is updated by adopting a gradient descent method to complete the model calibration, and when the prediction error is smaller than or equal to the preset threshold value, the predicted value of the dynamic boundary condition is directly used as the simulation boundary condition of the next moment, so that the real-time updating of the dynamic boundary condition is realized.
7. The machine learning driven fermentation device hydrodynamic multi-field coupling simulation method of claim 1, wherein the generation process of the self-adaptive multi-scale grid in S2 further comprises dividing strong, medium and weak coupling areas based on field intensity correlation coefficient sum outputted by an anisotropic nuclear density estimation algorithm, respectively corresponding to different grid node distances, and ensuring that the grid density of the strong coupling area is greater than that of the medium and weak coupling areas.
8. The machine learning driven fermentation device hydrodynamic multi-field coupling simulation method of claim 1, wherein the multi-dimensional data preprocessing process in S1 is specifically that abnormal values are removed by adopting a3 sigma criterion, missing values are filled by adopting a linear interpolation method or an average value in the same fermentation stage, and all parameter data are mapped to a unified interval by adopting a standardized method to obtain a standardized data set for training of a biological-physical mapping model.
9. The machine learning driven fermentation device hydrodynamic multi-field coupling simulation method of claim 1 is characterized in that data communication and field quantity fusion among subregions of a parallel computing framework in S5 are specifically realized by sharing memory through a GPU, boundary field quantity data interaction of subregions corresponding to each computing core is realized, for public boundary nodes of adjacent subregions, a weighted average method is adopted to fuse the field quantity data, weight coefficients are determined by field strength association coefficients, and the larger the field strength association coefficients are, the larger the weight coefficients are, so that the continuity of global field quantity distribution is ensured.
10. The method for simulating the hydrodynamic multi-field coupling of the machine learning driven fermentation device according to claim 1, wherein the specific judgment standard for verifying the standard in S6 is that the mean square error of the four dimensions of the fluid field, the temperature field, the concentration field and the bio-physical coupling is smaller than a corresponding preset error threshold value, the correlation coefficient is larger than a corresponding preset correlation coefficient threshold value, and the iterative optimization simulation process is stopped at the moment to output a final simulation result.

Description

Machine learning driven fermentation device hydrodynamic multi-field coupling simulation method Technical Field The invention relates to the technical field of hydrodynamic simulation, in particular to a machine learning driven fermentation device hydrodynamic multi-field coupling simulation method. Background The fermentation process is a complex process of microorganism metabolism and fluid flow, heat transfer and mass transfer multi-field interaction, and the multi-field coupling characteristic directly determines the fermentation efficiency and the product quality, so that the accurate multi-field coupling simulation is a core support for fermentation process optimization. The existing fluid mechanics multi-field coupling simulation technology of the fermentation device has the following key defects: The grid division level is that the traditional simulation adopts uniform grids or empirical non-uniform grids, so that the spatial heterogeneity of the multi-field coupling strength cannot be adapted, the accuracy of a strong coupling region is insufficient or the calculation redundancy of a weak coupling region is caused, and the simulation accuracy and the calculation efficiency are difficult to balance; The control equation correction level is that the existing correction method is used for carrying out parameter fitting driven by pure data by relying on experimental data, ignoring basic physical law constraints such as hydrodynamics, thermodynamics and the like, and easily generating the problem of 'physical inconsistency', so that a simulation result deviates from an actual fermentation process; The boundary condition processing layer is mainly set by adopting a static boundary condition, and the dynamic change of working conditions such as feeding, cooling, stirring and the like in the fermentation process cannot be adapted in real time, so that the simulated working condition is disjointed with the actual fermentation working condition; The model cooperation level is that modules such as mapping of biological parameters and physical parameters, grid generation, equation correction, boundary prediction and the like are mutually independent, a cooperation optimization mechanism is lacked, and the overall simulation precision and the robustness are limited. These defects lead to the difficulty in accurately describing the multi-field coupling dynamic characteristics of the fermentation process in the prior art, and the failure to provide reliable theoretical support for fermentation process optimization restricts the intelligent upgrading of the fermentation industry. Disclosure of Invention The invention provides a machine learning driven fermentation device hydrodynamic multi-field coupling simulation method, which overcomes the defects of poor grid adaptability, lack of physical constraint in correction of control equations, static boundary conditions and insufficient model synergy in the existing fermentation device hydrodynamic multi-field coupling simulation technology, and realizes field intensity quantization and grid generation adaptation, fusion of data driving and physical priori, accurate prediction of dynamic boundary conditions and full-flow model iterative optimization by constructing a bidirectional coordination mechanism between algorithms, thereby finally realizing accurate and efficient simulation of multi-field coupling characteristics in the fermentation process and providing core technical support for fermentation process optimization and intelligent control. In order to achieve the above purpose, the invention adopts the following technical scheme: a machine learning driven fermentation device hydrodynamic multi-field coupling simulation method comprises the following steps: s1, collecting experimental measurement data and real-time monitoring data of microbial activity in a fermentation process, preprocessing, and constructing a biological-physical mapping model to obtain a biological-physical association weight matrix; s2, based on a biological-physical association weight matrix, combining three-dimensional geometric characteristic data of a fermentation device, performing bidirectional interaction with a graph annotating force enhanced generation countermeasure network through an anisotropic kernel density estimation algorithm to generate a self-adaptive multi-scale grid and field intensity association coefficient; S3, constructing a control equation correction model based on the preprocessed experimental measurement data, the self-adaptive multi-scale grid and the field intensity association coefficient, performing bidirectional interaction with a sparse self-encoder guided by physical information through a hierarchical Bayesian variation inference algorithm, finishing correction of a multi-field coupling control equation, and outputting the corrected multi-field coupling control equation; S4, acquiring a calculation result of the corrected multi-field coupling control equati