Pulse type variable-temperature fermentation process for feed yeast culture

Abstract

The invention relates to the technical field of feed fermentation, and discloses a pulsating variable-temperature fermentation process for feed yeast culture, which comprises the following steps: S1, establishing a coupling model for coupling a physical environment model of a fermentation tank with a biological response model of yeast, the physical environment model being used for predicting temperature field distribution in the fermentation tank, and the biological response model being used for predicting the temperature field distribution in the fermentation tank; the biological response model is used for predicting the physiological metabolism state of the yeast at a specific temperature; s2, based on the coupling model and a predefined multi-objective function, determining an optimal control track for regulating and controlling the fermentation temperature; and S3, in the fermentation process, executing the optimal control track, and obtaining fermentation data in real time. By constructing a physical and biological deep coupling dynamic model and calculating a dynamic optimal control track throughout the whole fermentation period on the basis, the defect that physiological needs and economic cost of different stages cannot be considered due to dependence on a fixed temperature set value in a traditional method is overcome.

Description

Technical Field

[0001] This invention relates to the field of feed fermentation technology, specifically to a pulsed variable temperature fermentation process for feed yeast cultivation. Background Technology

[0002] Bio-fermentation is the core of modern industrial biotechnology and is widely used in medicine, food, chemical industry, and agriculture. In various fermentation processes, temperature, as a fundamental and critical environmental parameter, directly affects the growth rate of microorganisms, metabolic pathways, the synthesis efficiency of target products, and even the quality of the final product through precise control. Therefore, designing and implementing efficient and precise temperature control strategies is an important technical means to improve the overall efficiency of the fermentation process and ensure production stability.

[0003] However, current temperature control practices in industrial fermentation processes still suffer from significant technical limitations. Traditional control methods typically simplify the large, physicochemically complex fermenter into an idealized, uniformly lumped-parameter system. This approach ignores the spatial non-uniformity prevalent in large industrial fermenters, such as the gradient distribution of temperature, substrate concentration, and dissolved oxygen across different regions due to fluid dynamic constraints. Control strategies based on this simplified model have decision-making criteria that deviate significantly from the actual, non-uniform process conditions, making it difficult to guide the entire microbial community to its optimal physiological state.

[0004] At the execution level of control strategies, existing technologies generally rely on setting one or several fixed temperature values ​​and maintaining them through classic PID controllers. This static or segmented static control method is essentially a passive adjustment, designed to eliminate deviations rather than actively seek optimization, making it difficult to adapt to the dynamic changes in the environmental temperature requirements of microorganisms at different growth stages during fermentation. Fermentation is a highly nonlinear and complex biochemical reaction process with long time delays. Simple constant-value control is often a compromise between multiple conflicting objectives such as yield, energy consumption, and production cycle, and cannot achieve global process optimization.

[0005] Furthermore, the inherent randomness and time-varying nature of biological processes present deeper challenges to precise control. Even for the same process, differences in raw material composition between batches, fluctuations in the activity of inoculated strains, and even the physiological degradation of strains during successive passages can cause the actual fermentation process to deviate from the preset ideal state. Existing control systems generally lack online learning and adaptive capabilities, making it impossible to identify and correct deviations between the model and reality in real time during the process, especially lacking a deep understanding of the dynamic biological characteristics of microorganisms. At the same time, production experience between batches often relies on manual summarization, lacking an effective mechanism to quantify, solidify, and automatically transfer the optimization results of a single batch to subsequent batches, resulting in long process optimization cycles and difficulty in ensuring long-term production consistency. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a pulsed variable-temperature fermentation process for feed yeast cultivation. This process solves the problem that existing fermentation temperature control methods, due to oversimplification of models, use of static settings, and lack of learning mechanisms, cannot effectively cope with the spatiotemporal nonuniformity, dynamic changes, and batch-to-batch differences in the fermentation process, resulting in poor control robustness and difficulty in achieving optimal overall efficiency.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a pulsed, variable-temperature fermentation process for feed yeast cultivation, comprising the following steps:

[0008] S1. Establish a coupled model that couples the physical environment model of the fermenter with the biological response model of yeast. The physical environment model is used to predict the temperature field distribution inside the fermenter, and the biological response model is used to predict the physiological metabolic state of yeast at a specific temperature.

[0009] S2. Based on the coupling model and the predefined multi-objective function, determine an optimal control trajectory for regulating fermentation temperature;

[0010] S3. During the fermentation process, the optimal control trajectory is executed, and fermentation data is acquired in real time; based on the deviation between the real-time fermentation data and the predicted value of the coupled model, the coupled model is corrected online and the subsequent control trajectory is re-optimized in a rolling manner.

[0011] Preferably, in step S1, the physical environment model is specifically a three-dimensional model of the fermenter established based on computational fluid dynamics, used to simulate the flow field, temperature field and control energy consumption within the fermenter.

[0012] Preferably, in step S1, the biological response model specifically refers to a temperature-dependent whole-genome metabolic model that constructs the kinetic parameters of key enzymatic reactions as a temperature function, used to predict the specific growth rate of yeast and the specific production rate of the target product.

[0013] Preferably, in step S1, the establishment of the coupling model includes: using the distributed temperature field predicted by the physical environment model as the input condition for the operation of the biological response model, so as to obtain a global biological response prediction under a non-uniform temperature field.

[0014] Preferably, in step S2, the predefined multi-objective function is specifically a weighted function that integrates at least two of the following indicators: final biomass, total target product, total energy consumption, and total fermentation time.

[0015] Preferably, in step S2, the optimal control trajectory specifically refers to a series of timing control commands for the actuator of the fermenter temperature control system.

[0016] Preferably, in step S3, the step of rolling re-optimizing the subsequent control trajectory includes: executing only the initial part of the current optimal control trajectory, and then re-optimizing the control trajectory in the future time domain based on the corrected model at the next time step.

[0017] Preferably, in step S3, the step of online calibration of the coupling model includes: using the deviation between the real-time fermentation data and the predicted value of the coupling model to correct the state variables in the model, wherein the state variables include cell concentration, substrate concentration and product concentration.

[0018] Preferably, the step of online calibration of the coupled model further includes: using the deviation between the predicted values ​​of the coupled model, synchronously correcting the intrinsic biological parameters in the biological response model through a data assimilation algorithm.

[0019] Preferably, the fermentation process further includes a cross-batch learning step, which specifically involves using the final biological parameters, after online calibration, obtained after the current fermentation batch is completed as the initial model parameters for the next fermentation batch.

[0020] This invention provides a pulsed, variable-temperature fermentation process for cultivating feed yeast. It offers the following advantages:

[0021] 1. This invention overcomes the shortcomings of traditional methods that rely on fixed temperature setpoints and cannot take into account the physiological needs and economic costs at different stages by constructing a dynamic model that deeply couples physics and biology, and using this model to calculate the dynamic optimal control trajectory throughout the entire fermentation cycle. This results in a better balance between yield, target product quality, energy consumption and time cost, and enables holistic and predictive management of the fermentation process.

[0022] 2. This invention employs a model predictive control framework, combining offline optimization with online rolling control, which can effectively address the unpredictable disturbances and uncertainties that actually exist during the fermentation process. Through a closed-loop cycle of "prediction-optimization-execution-correction," the system can adjust the control strategy in real time and proactively to offset the negative impacts of disturbances. Compared with traditional passive adjustment methods such as PID control, this significantly enhances the robustness of the process and the stability of actual operation.

[0023] 3. The online adaptive mechanism of this invention not only corrects directly observable state variables, but also delves into the model itself. It uses real-time data to synchronously identify and correct key biological parameters characterizing the intrinsic vitality of microorganisms, enabling the coupled model to learn online and approximate the true physiological characteristics of the current batch of microorganisms. This achieves a leap from controlling appearances to regulating the intrinsic nature of the microorganisms, improving the accuracy of model predictions and the effectiveness of control decisions.

[0024] 4. This invention establishes a cross-batch learning and knowledge transfer mechanism. By using the final model parameters of the previous batch after full correction as the initial model of the new batch, the digital and automated inheritance of production experience is realized. This enables the system to automatically adapt to slow and long-cycle changes caused by factors such as changes in raw material batches and strain mutations. This ensures the consistency and efficiency of the process in long-term operation and solves the technical problem that it is difficult to guarantee the consistency of traditional processes.

[0025] 5. By introducing computational fluid dynamics in the modeling stage, this invention accurately characterizes the non-uniform physical environment inside large fermenters and combines this spatial distribution characteristic with a temperature-dependent whole-genome metabolic model, enabling optimization and control to penetrate into the microscopic level of the process, fully tapping the production potential of equipment and strains, and thus achieving higher overall benefits on a macroscopic level. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the overall process of the method of the present invention;

[0027] Figure 2 This is a schematic diagram of step S1 of the present invention;

[0028] Figure 3 This is a schematic diagram of step S2 of the present invention;

[0029] Figure 4 This is a schematic diagram of step S3 of the present invention;

[0030] Figure 5 This is a schematic diagram of the cross-batch learning process of the present invention. Detailed Implementation

[0031] 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.

[0032] Please see the appendix Figure 1 -Appendix Figure 5 This invention provides a pulsed, variable-temperature fermentation process for culturing feed yeast, comprising the following steps:

[0033] S1. Establish a coupled model that couples the physical environment model of the fermenter with the biological response model of yeast. The physical environment model is used to predict the temperature field distribution inside the fermenter, and the biological response model is used to predict the physiological metabolic state of yeast at a specific temperature.

[0034] In this embodiment, the specific purpose of step S1 is to construct a coupled model that can faithfully and dynamically reproduce the actual fermentation process before the fermentation process begins. This model serves as the computational foundation and core basis for all subsequent optimization and adaptive control steps. The implementation details of this step are as follows:

[0035] This step aims to establish a unified model that closely links the macroscopic physical environment of the fermenter with the microscopic biological response of yeast. To this end, this step constructs separate physical environment and biological response models, and deeply couples them in a specific way to form a unified digital twin, specifically through the following processes:

[0036] Construction of physical environment model

[0037] To accurately describe the non-uniform and unsteady physical fields within industrial-scale fermenters caused by the complexity of fluid dynamics, thermodynamics, and geometry, in this embodiment, the physical environment model is preferably a three-dimensional model constructed using computational fluid dynamics technology.

[0038] The construction process begins with accurate geometric modeling of the fermenter. This modeling includes not only the main body of the fermenter but also all internal components that significantly affect the flow and temperature fields within the tank, such as: different types of stirring systems (including the shape, number, layers, and shaft of the impellers), baffles used to disrupt circulation, and external cooling devices, heating jackets, and internal cooling coils, which serve as the main heat exchange units.

[0039] After completing the geometric modeling, the three-dimensional space is discretized into a high-quality mesh, that is, it is divided into tens of thousands or even millions of tiny control volume units. This step is the basis for transforming the continuous physical space into a discrete numerical computation domain. The quality of the mesh is directly related to the accuracy and convergence of subsequent calculations.

[0040] Based on this discretized model, the physical phenomena inside the tank are simulated by solving a set of partial differential equations. The core governing equations include the unsteady Navier-Stokes equations describing fluid flow and the energy transport equations describing the spatiotemporal distribution of the temperature field. The general form of the energy transport equations in Cartesian coordinates can be expressed as:

[0041]

[0042] Where T is the temperature field function T(x,y,z,t), representing the local temperature at spatial location (x,y,z) and time t;

[0043] ρ and C p These are the density and specific heat capacity of the fermentation medium, respectively, which can be defined as a constant or a function of temperature;

[0044] u is the fluid velocity vector coupled to the temperature field, obtained by solving the Navier-Stokes equations.

[0045] k is the thermal conductivity of the culture medium;

[0046] Q h The comprehensive heat source term per unit volume is a composite variable that includes: the stirring heat generated by the high-speed rotation of the stirrer on the high-viscosity fluid, the bioheat released by the metabolic activities of the yeast cells themselves, and the heat exchange introduced by the heat exchange jacket and coils through heat exchange with the external temperature control medium.

[0047] The core function of numerically solving this physical environment model is that when given the control command of the external temperature control system (e.g., the opening degree or flow rate of the cooling water valve) as the boundary condition of the model, the model can predict the instantaneous temperature of each discrete grid node in the fermenter at any future time, thus forming a complete three-dimensional unsteady temperature field distribution T(x,y,z,t). At the same time, the model can also calculate the instantaneous control energy consumption required to realize the control command.

[0048] Construction of biological response models

[0049] To describe the physiological and metabolic responses of feed yeast under different temperature conditions, in this embodiment, the biological response model is a temperature-dependent genome-wide metabolic model (T-GSMM). This model represents a significant improvement over existing technologies, abandoning simple growth models that rely on empirical fitting and instead revealing the mechanisms of temperature influence at the level of the cellular metabolic network.

[0050] The model is constructed based on a genome-wide chemostometric model of the target yeast strain. At its core is a chemostometric matrix N, which describes the stoichiometric relationships between reactants and products in hundreds or thousands of biochemical reactions within the cell; and a reaction flux vector v, representing the rate of each reaction. Under the assumption of metabolic homeostasis, the system must satisfy the fundamental constraint of mass conservation, i.e.;

[0051] N·v=0;

[0052] The core technical feature of this invention lies in constructing the key flux constraints in the above-mentioned flux balance analysis, particularly the maximum reaction flux of the rate-limiting enzymatic reaction that determines the rate of cell growth, substrate consumption, and target product synthesis, as an explicit function of temperature. A functional form that can describe enzyme activation within the physiological temperature range and inactivation at superoptimal temperatures is as follows:

[0053]

[0054] Among them, v j,max (T) represents the maximum permissible flux of the j-th enzymatic reaction at absolute temperature T, which directly defines the bottleneck of the reaction pathway;

[0055] A is the pre-exponential factor, which is related to the collision frequency of the reaction;

[0056] E a The activation energy of this reaction reflects the strength of the positive promoting effect of temperature increase on the reaction rate.

[0057] E d This is the energy constant for enzyme thermal inactivation, reflecting the intensity of the negative effects of excessively high temperatures on enzyme protein structure, leading to loss of activity.

[0058] R is the universal ideal gas constant;

[0059] B is an empirical constant related to the deactivation process;

[0060] These parameters (A,B,E) a E d (etc.) together constitute the parameter vector θ that characterizes the biological characteristics of the strain in this biological model.

[0061] In this way, the biological response model can predict the instantaneous specific growth rate of cells and the specific production rate of target products under given local temperature and substrate supply conditions by solving a linear programming problem with the optimization objective of maximizing cell growth or other physiological goals.

[0062] Formation of the coupling model

[0063] Neither a physical model nor a biological model alone can solve the technical problems faced by this invention. Therefore, the key to this step is to deeply couple the two models to form a unified coupled model capable of collaborative computation.

[0064] In this embodiment, the coupling is implemented as follows: a dynamic data interaction interface is established. During the overall simulation calculation, the physical environment model first calculates the instantaneous temperature of all discrete grid nodes inside the tank based on the control input, forming a high-resolution temperature field distribution matrix. Subsequently, this temperature field matrix is ​​transmitted to the biological response model in real time and in a distributed manner.

[0065] Specifically, it can be envisioned that an independent T-GSMM instance is deployed on each CFD mesh node. Each T-GSMM instance receives the local temperature value of its node as its core computational parameter and independently calculates the local biological response of that node, which is specifically the instantaneous specific growth rate of cells at that location and the specific generation rate of the target product.

[0066] Finally, by weighted integral or summation of the biological responses of all nodes within the entire fermenter volume, the overall macroscopic performance indicators of the fermenter under the current non-uniform temperature field can be obtained, such as the total biomass growth rate and the total product generation rate.

[0067] S2. Based on the coupling model and the predefined multi-objective function, determine an optimal control trajectory for regulating fermentation temperature;

[0068] In this embodiment, the core objective of step S2 is to transform complex, sometimes even conflicting, production demands into a structured, solvable mathematical problem, and to calculate an optimal control trajectory that can guide the operation of the temperature control system throughout the entire fermentation cycle based on the high-fidelity coupling model constructed in step S1.

[0069] This step is not simply about setting a constant temperature target, but rather about planning a temperature curve that dynamically changes over time to maximize overall benefits. This is achieved through the following processes:

[0070] Definition and Quantization of Multi-Objective Functions

[0071] To ensure that the optimization process accurately reflects the true production intentions, this invention employs a multi-objective function as the ultimate benchmark for evaluating the merits of any control strategy. This function is designed to go beyond simple technical indicators, embedding economic cost factors into the core of the optimization decision.

[0072] The multi-objective function is preferably constructed as a weighted summation, and its mathematical expression is as follows:

[0073] J total =w bio J bio +w prod J prod -w en J en -w time J time ;

[0074] In this function, the physical meaning and calculation method of each term are as follows:

[0075] Total biomass target (J) bio This represents the total yeast cell dry weight obtained at the end of fermentation, and is one of the core indicators for measuring yield. Its value depends on the prediction of the coupled model in step S1, obtained by adjusting the fermentation endpoint time t. f Cell concentration distribution X(x,y,z,t) f Integrating over the entire effective volume V of the fermenter yields:

[0076] J bio =∫ V X(x,y,z,t f )dV;

[0077] Total target of target product (J) prod This value represents the cumulative total amount of the target product that is beneficial to improving feed quality throughout the entire fermentation process. Calculation of this value requires analyzing the cell specific generation rate q predicted by the biological response model in step S1 at different spatiotemporal locations. p The product of (x,y,z,t) and the corresponding cell concentration X(x,y,z,t) over the entire fermentation time domain [0,t] f Double integration is performed within the fermenter volume V:

[0078]

[0079] Total energy cost target (J en This item represents the total energy consumption for implementing the temperature control strategy and is a key component of production costs. Its value is derived from the instantaneous control energy consumption E calculated from the physical environment model in step S1. cost(t) Integrate over the entire fermentation time:

[0080]

[0081] Total Time Cost Target (J) time This represents the total time taken for the entire fermentation batch, which is directly related to equipment turnover rate and production efficiency. Its value is the fermentation completion time.

[0082] w bio w prod w en and w time A non-negative constant pre-set by the user based on the strategic priority of specific production tasks.

[0083] Determination of the optimal control trajectory

[0084] Having defined the above optimization objective, the core task of this step is to find an optimal control input timing trajectory acting on the actuator of the temperature control system, so that under the constraints of all equations of the coupled model described in step S1 and the physical actuator, the overall objective function can reach its global maximum value.

[0085] In this embodiment, the optimization problem is typically computed globally offline before the fermentation process officially begins. Due to the complexity of the problem, a single optimization algorithm may struggle to balance global search capability with local convergence accuracy.

[0086] Therefore, a preferred implementation is to employ a hybrid optimization strategy. For example, heuristic algorithms with strong global exploration capabilities, such as genetic algorithms or particle swarm optimization, can be used to search the vast solution space and locate the potential region where the optimal solution is located with a high probability, thereby avoiding the trap of local optima.

[0087] Subsequently, using the optimal region or optimal individual obtained from the global search as the initial value, a more efficient gradient-based local optimization algorithm, such as sequential quadratic programming, is employed to refine the search and fine-tune the solution in order to obtain the optimal solution that meets the accuracy requirements.

[0088] The final output of this step is a complete, theoretically optimal sequence of control commands. This sequence details the specific operations that the temperature control system should perform at every point in time from the start to the end of fermentation. This trajectory is not the end point, but rather serves as the initial blueprint and guiding baseline for real-time adaptive control in subsequent steps, laying the foundation for the intelligent and optimized operation of the entire fermentation process.

[0089] S3. During the fermentation process, the optimal control trajectory is executed, and fermentation data is acquired in real time; based on the deviation between the real-time fermentation data and the predicted value of the coupled model, the coupled model is corrected online and the subsequent control trajectory is re-optimized in a rolling manner.

[0090] For step S3, in this embodiment, the purpose of this step is to put the optimal control trajectory planned in step S2 into practice and to deal with disturbances and model uncertainties in the fermentation process, so as to ensure that the process always operates in a state close to the optimal state.

[0091] Step S3 does not simply execute the pre-calculated entire control plan, but rather employs an advanced control framework based on model predictive control. The essence of this framework is a closed-loop cycle of "prediction-optimization-execution-correction" that rolls forward continuously over time.

[0092] In this embodiment, the closed-loop control and adaptive steps specifically include the following coordinated operations:

[0093] Execution of rolling optimization and control

[0094] This step runs periodically at a preset, fixed time step until the fermentation ends. At the beginning of each control cycle, the controller first extracts only the first control action of the current time step from the optimal control trajectory updated in step S2 or the previous cycle, and sends it as an instruction to the physical actuator of the fermenter.

[0095] This rolling strategy, which executes only the initial action rather than the entire trajectory, is one of the key features of this method. It leaves room for the system to adjust subsequent control strategies in real time based on new information, which is the foundation for achieving adaptive control.

[0096] Real-time fermentation data acquisition and feedback

[0097] During or at the end of the aforementioned control actions, the system collects measurement data reflecting the true state of the fermentation process in real time through various online sensors deployed on the fermenter.

[0098] Preferably, the sensors may include: a multi-point distributed fiber optic temperature sensor or multiple platinum resistance thermometers at different locations for accurately capturing the temperature distribution inside the tank; a turbidimeter for online monitoring of cell concentration; and pH and dissolved oxygen electrodes for monitoring chemical parameters of the culture environment. The data collected by these sensors collectively constitute a real-time measurement vector, which is input as feedback information to the system's calibration stage.

[0099] Online calibration and adaptation of coupled models

[0100] This is the core technology that distinguishes this invention from traditional control methods; its purpose is to dynamically bridge the gap between the virtual coupled model and the actual physical fermentation process. Upon acquiring real-time measurement data, the system compares it with the coupled model's prediction of the current state at the previous moment, and the deviation between the two drives a dual correction mechanism.

[0101] To clearly describe this correction process, the present invention can abstract the entire fermentation system into a nonlinear state-space model:

[0102] x sys,k =f(x) sys,k-1 ,u k-1 ,θ k-1 )+v k-1 ;

[0103] Where, x sys,k Let k be the state vector of the system at discrete time point k. This is a set containing variables that describe the core state of the fermentation system, such as cell concentration, substrate concentration, target product concentration, etc.

[0104] k is the index of the discrete time, representing the current time or the kth control cycle;

[0105] x sys,k-1 Let k be the state vector of the system at the previous discrete time point k-1;

[0106] f(·) represents the nonlinear process function or state transition function that describes the dynamic evolution of the system. This function is defined by the coupled model and describes how the state of the system evolves from time k-1 to time k under given control and parameter conditions.

[0107] u k-1 This is the control input applied to the system at time k-1, which represents the external control action, such as the set value of the cooling water flow rate of the temperature control system, the stirring speed, etc.

[0108] θ k-1 This is the system model parameter vector at time k-1. It contains parameters in the model used to characterize the intrinsic properties of the system. In this invention, it specifically refers to the intrinsic biological parameters in the biological response model, such as the activation energy and inactivation energy of an enzyme.

[0109] v k-1 Let be the process noise vector at time k-1. It represents the impact of unknown or random disturbances on the true state of the system that cannot be accurately described by the model itself. It is usually assumed to be a Gaussian white noise with zero mean.

[0110] This embodiment employs data assimilation algorithms, such as extended Kalman filtering or unscented Kalman filtering, to process the aforementioned model and prediction errors, and performs dual correction:

[0111] Firstly, the correction of state variables:

[0112] Using the prediction error, firstly, the state vector x inside the model is... sys This involves making corrections. The aim is to bring variables in the model that cannot be directly measured back to their most probable values ​​inferred from actual measurements. The update logic can be conceptually represented as:

[0113]

[0114] in, This is the corrected state estimate; For prior predictions before correction; K x,k The state Kalman gain matrix, calculated based on the uncertainties of the model and measurements, determines the extent to which new measurements are trusted; y meas,k This represents the actual measurement vector obtained from the sensors in the physical system at time point k; The predicted measurement vector is calculated by the coupled model based on prior estimates at time point k.

[0115] Secondly, the correction of intrinsic biological parameters:

[0116] Furthermore, and another key innovation of this invention, is that the prediction error is not only used to correct the state, but also to simultaneously correct the parameter vectors that characterize the fundamental characteristics of the yeast strain and are deeply embedded within the biological response model (T-GSMM). For example, when the system observes that the actual growth rate of yeast is consistently lower than the model prediction at a certain temperature, the correction algorithm will adjust the enzyme activity parameters related to that temperature range accordingly, so that the model's predicted behavior is closer to the actual activity of this batch of yeast. Its update logic is similar to that of state correction:

[0117]

[0118] Among them, K θ,k Kalman gain for parameter vector; It is the posterior estimate of the parameter vector obtained at discrete time point k by combining the measurement data at that time. It represents the prior estimate of the parameter vector at discrete time point k, based solely on information up to time k-1.

[0119] Rolling re-optimization

[0120] Once the above dual corrections are completed, the system will possess the most accurate coupled model, conditioned by real-time data. At this point, the MPC framework will use this corrected model and state as a new starting point to resolve the multi-objective optimization problem defined in step S2, thereby obtaining a new, more realistic future optimal control trajectory from the current moment to the fermentation endpoint.

[0121] Subsequently, the system will enter the next control cycle and begin executing from the first action of this new trajectory, repeating the cycle continuously. This closed loop of "execution-measurement-correction-re-optimization" ensures that the process of this invention can continuously self-correct, effectively overcome the negative impacts of various uncertainties, and guide the fermentation process to always proceed along the actual optimal path.

[0122] After a complete fermentation process is completed, this invention sets up a cross-batch learning step, which aims to establish a mechanism that can quantify, store and pass on the knowledge acquired during a single batch of fermentation, thereby enabling the entire production system to have memory and learning capabilities that go beyond a single run.

[0123] In this embodiment, the specific implementation of the cross-batch learning step is as follows:

[0124] After a complete fermentation batch is completed, the system will execute a knowledge consolidation and archiving procedure. In this procedure, the core operation object is the biological parameter vector that has been repeatedly corrected and optimized online throughout the fermentation process in the aforementioned step S3 and has finally converged.

[0125] This final parameter vector is the best posterior estimate of the model parameter vector obtained at the end of fermentation. It integrates all real-time measurement data of the batch from initial inoculation to the end of fermentation and can be considered the most accurate mathematical profile of the true biological characteristics of the yeast strain used under the current production conditions.

[0126] The system will store the final parameter vector together with other key information of the batch (such as batch number, raw material batch number, fermentation date, etc.) in a structured manner, for example, by storing it in a dedicated historical database or knowledge base.

[0127] When a new fermentation task is ready to begin, the method of this invention will no longer use a general, preset, or theoretical initial biological model parameter to initiate the modeling process in step S1. Instead, the system will execute a knowledge transfer and initialization procedure.

[0128] The program accesses the aforementioned historical database and retrieves archived parameter vectors according to preset rules or operator instructions. The system then uses this vector as the initial biological parameter vector for the new batch of coupled models. This process can be represented by the following assignment operation:

[0129]

[0130] in, This is the final biological parameter vector obtained after full online calibration at the end of batch N;

[0131] This is the initial biological parameter vector used for the upcoming batch N+1;

[0132] The := operator is a definition or assignment operation that explicitly indicates the direction of knowledge transfer from one batch to the next.

[0133] By performing this step, the present invention transforms production experience from a vague, manual-dependent form into precise, automatically inheritable digital model parameters. This means that when a new fermentation batch begins, its underlying coupled model already anticipates potential changes in raw materials or strains, and its initial state is very close to the actual process. This makes the optimal control trajectory initially calculated in step S2 more realistically feasible and allows for faster convergence in subsequent step S3, reducing the magnitude of online corrections and early process fluctuations.

[0134] By repeating this process over and over again, through this cross-batch learning step, the method of the present invention can continuously optimize itself and automatically adapt to changes over long periods, thereby ensuring that the fermentation process can maintain long-term production stability and efficiency in the face of various slow disturbances.

[0135] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various interpretations can be made without departing from the principles and spirit of the invention.

[0136] These embodiments may be subject to various changes, modifications, substitutions, and variations, and the scope of the invention is defined by the appended claims.

[0137] Claims and their equivalents are limited.

Claims

1. A pulsed, variable-temperature fermentation process for culturing feed yeast, characterized in that, Includes the following steps: S1. Establish a coupled model that couples the physical environment model of the fermenter with the biological response model of yeast. The physical environment model is used to predict the temperature field distribution inside the fermenter, and the biological response model is used to predict the physiological metabolic state of yeast at a specific temperature. S2. Based on the coupling model and the predefined multi-objective function, determine an optimal control trajectory for regulating fermentation temperature; S3. During the fermentation process, the optimal control trajectory is executed, and fermentation data is acquired in real time; based on the deviation between the real-time fermentation data and the predicted value of the coupled model, the coupled model is corrected online and the subsequent control trajectory is re-optimized in a rolling manner.

2. The pulsed variable-temperature fermentation process for feed yeast cultivation according to claim 1, characterized in that, In step S1, the physical environment model is specifically a three-dimensional model of the fermenter established based on computational fluid dynamics, used to simulate the flow field, temperature field and control energy consumption within the fermenter.

3. The pulsed variable-temperature fermentation process for feed yeast cultivation according to claim 1, characterized in that, In step S1, the biological response model specifically refers to a temperature-dependent whole-genome metabolic model that constructs the kinetic parameters of key enzymatic reactions as a temperature function, used to predict the specific growth rate of yeast and the specific production rate of the target product.

4. The pulsed variable-temperature fermentation process for feed yeast cultivation according to claim 1, characterized in that, In step S1, the establishment of the coupling model includes: using the distributed temperature field predicted by the physical environment model as the input condition for the operation of the biological response model, so as to obtain the global biological response prediction under a non-uniform temperature field.

5. The pulsed variable-temperature fermentation process for feed yeast cultivation according to claim 1, characterized in that, In step S2, the predefined multi-objective function is specifically a weighted function that integrates at least two of the following indicators: final biomass, total target product, total energy consumption, and total fermentation time.

6. The pulsed variable-temperature fermentation process for feed yeast cultivation according to claim 1, characterized in that, In step S2, the optimal control trajectory specifically refers to a series of timing control commands for the actuator of the fermenter temperature control system.

7. The pulsed variable-temperature fermentation process for feed yeast cultivation according to claim 1, characterized in that, In step S3, the step of rolling re-optimization of the subsequent control trajectory includes: executing only the initial part of the current optimal control trajectory, and then re-optimizing the control trajectory in the future time domain based on the corrected model in the next time step.

8. The pulsed variable-temperature fermentation process for feed yeast cultivation according to claim 1, characterized in that, In step S3, the step of online calibration of the coupling model includes: using the deviation between the real-time fermentation data and the predicted value of the coupling model to correct the state variables in the model, wherein the state variables include cell concentration, substrate concentration and product concentration.

9. The pulsed variable-temperature fermentation process for feed yeast cultivation according to claim 8, characterized in that, The step of online calibration of the coupled model further includes: using the deviation between the predicted values ​​of the coupled model, synchronously correcting the intrinsic biological parameters in the biological response model through a data assimilation algorithm.

10. The pulsed variable-temperature fermentation process for feed yeast cultivation according to claim 1, characterized in that, The fermentation process also includes a cross-batch learning step, which specifically involves using the final biological parameters, after online calibration, obtained after the current fermentation batch is completed as the initial model parameters for the next fermentation batch.

Images