Fermentation process modeling method based on generative fuzzy system and expert knowledge evaluation

Abstract

The invention provides a fermentation process modeling method based on a generative fuzzy system and expert knowledge evaluation. The generation type fuzzy rule can mine own specific information for the fermentation process. The modeling method supports evaluation and optimization of fermentation control parameters by utilizing expert knowledge. Therefore, the fermentation process optimization model can be effectively suitable for a fermentation process optimization scene, and the experiment frequency and the optimization time can be remarkably reduced, so that the cost is reduced and the efficiency is improved. According to the method, the rolling learning method is innovatively designed for the fermentation process optimization model, so that the fermentation process optimization model has excellent adaptive capacity. The adaptive characteristic enables the model to flexibly respond to the dynamic change and uncertainty in the fermentation process, and the robustness of the system is remarkably improved. In addition, automatic optimization and control are achieved through the intelligent characteristic, dependence on artificial experience is reduced, and operation convenience and accuracy are improved.

Description

Technical Field

[0001] The present invention belongs to the field of bio-intelligent computing and relates to a fermentation process modeling method based on a generative fuzzy system and expert knowledge evaluation. Technical Background

[0002] The fermentation process refers to a biochemical process in which microorganisms (such as bacteria, yeast, or fungi) convert organic substrates into target products (such as alcohol, organic acids, enzymes, or other metabolites) through metabolic activities under suitable environmental conditions. This process is usually carried out under anaerobic or microaerobic conditions, but some fermentation processes can also be completed under aerobic conditions. Industrial fermentation processes are widely used in fields such as biopharmaceuticals, food processing, bioenergy, and chemical production. The core goal is to efficiently produce target products through microbial or cell culture. The core of the fermentation process is that microorganisms decompose and recombine substrates through their metabolic pathways to generate energy, cell substances, and target products. The fermentation process faces multiple challenges: First, it is difficult to precisely control fermentation conditions. Parameters such as temperature, pH, and dissolved oxygen have a significant impact on the growth and metabolism of the strain. Appropriate stirring and aeration can increase the dissolved oxygen and promote the growth of the strain, but excessive stirring may damage the strain. Second, metabolic regulation also faces huge challenges. The metabolic network of the strain is complex and difficult to regulate, and the by-products generated during the metabolic process may inhibit the growth of the strain and product synthesis. Traditional fermentation optimization methods mainly rely on empirical rules and trial-and-error experiments, which are not only time-consuming and laborious but also difficult to cope with the complex dynamic changes in the fermentation process. Therefore, more efficient and intelligent optimization strategies are needed to improve fermentation efficiency.

[0003] With the development of artificial intelligence and machine learning technologies, artificial intelligence-based optimization methods provide new solutions for the fine control of fermentation conditions and the optimization of metabolic regulation in the fermentation process.

[0004] As a type of machine learning model that can learn data distributions and generate new data, generative models play an important role in fermentation optimization. First, generative models (such as variational autoencoders VAE and generative adversarial networks GAN) can learn the complex relationships in the fermentation process from historical data, construct high-precision prediction models, and achieve data-driven modeling. Second, by simulating the fermentation results under different operating conditions using generative models, the number of actual experiments can be significantly reduced, thereby reducing the optimization cost. In addition, generative models can simultaneously predict multiple targets such as product yield, quality, and energy consumption, supporting multi-objective optimization. Finally, generative models can also evaluate the uncertainty of prediction results, provide more reliable support for decision-making, and further improve the scientificity and efficiency of the optimization process. These characteristics make generative models an important tool in fermentation process optimization.

[0005] As an effective tool for dealing with uncertainty and fuzzy knowledge, fuzzy rules can achieve smooth control and fine control in fermentation optimization. By fuzzifying input variables, establishing a rule base, applying an inference mechanism, and defuzzifying the output, it can flexibly cope with the fluctuations of environmental parameters such as temperature, pH value, and dissolved oxygen content during the fermentation process, as well as the complexity of microbial metabolism. Fuzzy rules support continuous adjustment of control parameters, avoid mutations, enhance system stability, and coordinate multi-variable control to achieve adaptive optimization. For example, in temperature control, pH adjustment, and dissolved oxygen management, fuzzy rules can dynamically adjust parameters to meet the requirements of different fermentation stages, thereby improving the stability and efficiency of the fermentation process.

[0006] The intelligent fermentation process optimization method based on generative models and fuzzy rules provides an efficient and intelligent solution for the optimization of complex fermentation processes by combining data-driven modeling and expert knowledge. This approach can not only significantly improve fermentation efficiency and product quality but also reduce costs and resource consumption, having important industrial application value and broad development prospects. Summary of the Invention

[0007] In an actual production environment, fuzzy rules can achieve fine control and optimization of the segmented control process parameters. However, the consequent of traditional fuzzy rules is a linear function, and its learning ability is limited, making it difficult to fully learn the complex biochemical characteristics of the fermentation process. By using a generative model as the consequent of fuzzy rules, the learning ability of traditional fuzzy rules is enhanced and applied to complex process control scenarios. Such fuzzy rules with a generative model as the consequent are called generative fuzzy rules. The fuzzy system based on generative fuzzy rules is called a generative fuzzy system. The fermentation process optimization model based on the generative fuzzy system realizes multi-scale optimization, taking into account both overall efficiency and detail control. It not only improves the intelligent level of fermentation optimization but also enhances the robustness and practicality of the system, providing strong support for the stability and efficiency of the fermentation process.

[0008] According to the deficiencies of traditional fermentation processes, the present invention provides a fermentation process modeling method based on generative fuzzy systems and expert knowledge evaluation. The present invention mainly includes two parts: First, the present invention designs a fermentation process modeling method based on generative fuzzy systems and expert knowledge evaluation; then, the present invention designs a rolling learning method for the fermentation process optimization model in an actual production environment.

[0009] The technical solution of the present invention is as follows:

[0010] The fermentation process modeling method based on generative fuzzy systems and expert knowledge evaluation includes the following steps:

[0011] Step 1: Determine the training data set for the fermentation process field.

[0012] Using a fermentation process dataset, which usually includes data such as biomass, substrate concentration, pH value, temperature, yeast activity, etc. during the fermentation process. The key point is that to ensure adaptation to diverse fermentation processes, all features, including input features and output optimization data, must be integrated as a whole for calculation. These data will be concatenated into sequences of variable length.

[0013] Step 2: Determine the number of rules of the fuzzy system and the generative model of the consequent according to the number of samples in the dataset and the fermentation indicators in the samples, and construct and initialize the fermentation process optimization model.

[0014] 2.1 Generative Fuzzification

[0015] The process of generative fuzzification first involves the selection of representatives of fuzzy sets, which aims to map the input data to a more representative representation. Subsequently, for each representative of the fuzzy set, the system will calculate the similarity between it and the input data, which is used as the basis for measuring the membership degree of the input. Through a series of calculations of similarity values, the system can effectively characterize the fuzzification characteristics of the input data, thus providing a basis for subsequent reasoning and decision-making.

[0016] 2.2 Generative Fuzzy Rule Consequent

[0017] The generative task is to predict the full probability distribution of all variables, which requires a relatively high learning ability of the model. The consequent of the classical fuzzy system often uses a linear function, and there are fewer parameters available for learning. For generative tasks, its learning ability is weak. In order to be able to adapt to complex and diverse downstream tasks, the present invention uses an intelligent model with strong learning ability as the consequent of the fuzzy rule, and this consequent is called the generative fuzzy rule consequent (GFRC). The generative rule consequent of the fuzzy rule can be expressed as follows:

[0018] g k = GFRC k (x) (1)

[0019] where k = 1, 2, 3,..., K, K is the number of rules of the fuzzy system, GFRC k (.) is the function of the generative rule consequent processing unit of the k-th rule, and g k is the generated data of the k-th rule.

[0020] 2.3 Generative Fuzzy Rules

[0021] For generative tasks, similar to the classical fuzzy system, a generative fuzzy system is constructed based on fuzzy sets and fuzzy logic. Its core is still the fuzzy rule base, and the k-th generative fuzzy rule of the rule base can be expressed as follows:

[0022]

[0023] where \(x\) is the input data, and \(GF(.)\) is the fuzzy set corresponding to the antecedent of the \(k\)th rule represented by .

[0024] 2.4: Generative Fuzzy Rule Combination Mechanism (GFRCM)

[0025] The generative fuzzy rule combination mechanism is an intelligent decision-making mechanism for generative tasks. Its purpose is to select applicable rule combination strategies in the face of different scenarios to produce clear and definite results. In the case of processing structured data with feature alignment, the weighted average method can be used to combine the outputs of each rule. This combination method is called weighted combination. When the output data is unstructured, the method of the maximum defuzzifier can be used to select the result of the rule with the highest membership degree as the final output. This combination method is called maximum combination. By flexibly applying these combination mechanisms, the generative fuzzy system can effectively produce clear output results in different situations, providing strong support for the reasoning of solving complex tasks. Equation (3) gives the generalized form of the generative fuzzy rule combination mechanism:

[0026]

[0027] where \(x'\) is the generated value (predicted value) of the fermentation index at the next moment, \(G = \{g 1 , g 2 , g 3 , \cdots, g K \}, is the firing strength of the \(k\)th rule, When \(G\) is structured and feature-aligned, use the firing strength as the weight to perform weighted summation on the generated result \(g k of each rule. Otherwise, select the rule output result with the maximum of all firing strengths as the final output, where the function is to obtain the index value of the maximum firing strength.

[0028] 2.5 Evaluate the fermentation index based on expert knowledge and optimize the output control parameters.

[0029] In the time dimension, the changing trend of the past fermentation index can be used to predict the future index value. The fermentation process optimization model can learn historical data and predict the index value at the next moment. Evaluate the predicted data through expert experience to output the control parameters at the next moment.

[0030] Through actual fermentation production and fermentation experts, the normal ranges of fermentation indicators are obtained. For example, the range of metabolite concentration is 0.1 - 100 g / L, the range of substrate consumption rate is 0.01 - 10 g / (L·h), the range of product quality is 90% - 99.9%, the range of moisture content is 50% - 95%, the range of protein hydrolysis index is 0.1 - 0.8, the range of pH value is 4.0 - 8.0, the dissolved oxygen concentration is 0.1 - 10 mg / L, and the range of redox potential is -400 - +200 mV. These indicator ranges are expert knowledge and are used to evaluate the quality of the fermentation indicators generated by the generative rules, as shown in Equation (4).

[0031] z = Expert(x′) (4)

[0032] According to the evaluation results, the control parameters y for the next moment are output, such as including parameters like the mixer rotation speed N, ventilation volume Q, temperature T, etc., as shown in Equation (5).

[0033] y = α × z (5)

[0034] Where Expert(.) evaluates the generated value according to the empirical range of fermentation indicators, z is the evaluation result of each indicator, α is a learnable parameter matrix, p is the number of output control parameters, q is the number of fermentation indicators, and y is the control parameter for the next moment.

[0035] Step 3: Construct the objective function of the fermentation process optimization model and optimize the model.

[0036] In addition to the construction of the rule antecedent, another important task of the generative fuzzy system is the learning of the generative fuzzy rule consequent (GFRC), which can be regarded as a typical machine learning task. In machine learning, the closer the predicted data distribution learned by the model on the training data is to the real data distribution, the better. In generative tasks, calculating the difference between the target sequence and the generated sequence is the key to learning the model parameters. From the perspective of probability distribution, the target sequence is regarded as the conditional probability distribution of the input sequence. The difference between the generated sequence and the target sequence is actually the difference between two probability distributions. The cross-entropy loss function is widely used to measure the difference between two probability distributions. Therefore, the cross-entropy loss function can be used as one of the functions for calculating the sequence difference of GFRC; from the perspective of sequence similarity, the difference between sequences can be measured by the similarity between sequences. Therefore, the similarity loss function can also be used as one of the functions for calculating the sequence difference. The fermentation process optimization model uses two losses to construct the objective function, namely the loss of fermentation indicators and the loss of control parameters. To prevent the model from overfitting, the loss function adds an L2 regularization term, and its general form is as follows:

[0037]

[0038] where y is the sequence of real control parameters at the next moment, is the generated sequence, x′ is the sequence of real fermentation indexes at the next moment, is the generated sequence of fermentation indexes at the next moment, diff(.) is the function for calculating the difference between the target sequence and the generated sequence, ‖Θ‖2 is the L2 regularization term of the consequent parameters of GFRC in GenFS in the model, and γ (γ>1) is the balance parameter.

[0039] For the generation task, the sequence data is characterized by high-dimensionality and sparsity, and the distribution is uneven. During the training process, the model can effectively learn the relevant information of high-frequency and low-frequency features by reducing the learning rate of high-frequency features and increasing the learning rate of low-frequency features. Therefore, GFRC needs to adopt an optimizer that can adaptively adjust the learning rate, such as AdaGrad, RMSprop, Adadelta, and Adam, etc., to better approximate the optimal value of the model parameters, thereby minimizing the training loss. The general form of the adaptive optimizer is shown in Equation (5):

[0040]

[0041] where AdaOptim(.) is the adaptive optimizer function, is the initial value of the learning rate, and Θ′ is the model parameters updated by the optimizer calculation.

[0042] Step 4: Train the fermentation process optimization model and output the trained model.

[0043] In the fermentation scenario, the consequent GFRC of the generative rule of GenFS and the α parameter matrix in the evaluation and optimization module are learnable. The training process is to learn the complex mapping relationship between the fermentation indexes and control parameters in the dataset. When the loss of the target loss function converges and stabilizes, end the model training and output the trained model.

[0044] Step 5: Adopt the rolling learning method of the fermentation process optimization model, including the following steps:

[0045] 5.1 Obtain real production data.

[0046] The data collection mechanism includes using IoT sensors to collect the actual parameters of the fermentation process in real time, and cleaning, denoising, and normalizing the data. At the same time, combining historical production data and process parameters to form a comprehensive dataset for model training.

[0047] 5.2 Model rolling training.

[0048] Considering the differences in the fermentation environment, equipment, and processes of actual factories, it is considered to regularly use actual production data to roll and train the model on a weekly, monthly, quarterly, or annual basis. The entire training process is consistent with the third and fourth steps of the modeling method.

[0049] 5.3 New model iteration and deployment to the production environment.

[0050] To ensure production safety and stability, after the old and new models run in a mixture of experts system (MoE) mode for a period of time, the old model stops running and only the new model is retained, completing one iteration.

[0051] Advantages of the present invention:

[0052] 1) Different from existing methods, the present invention proposes an optimization modeling method for fermentation processes based on generative models and expert knowledge evaluation. The generative fuzzy rules can mine the unique information of the fermentation process. This modeling method supports the use of expert knowledge to evaluate and optimize fermentation control parameters. Therefore, the fermentation process optimization model can be effectively applied to the scenario of fermentation process optimization, significantly reducing the number of experiments and optimization time, thereby reducing costs and improving efficiency.

[0053] 2) The present invention innovatively designs a rolling learning method for the fermentation process optimization model, enabling the model for fermentation process optimization to have excellent adaptive capabilities. This adaptive characteristic enables the model to flexibly respond to dynamic changes and uncertainties in the fermentation process, significantly enhancing the robustness of the system.

[0054] 3) In addition, the intelligent feature realizes automated optimization and control, reduces the dependence on manual experience, and improves the convenience and accuracy of operation;

[0055] 4) Finally, the scalability enables this method to be widely applied to different types of fermentation processes, demonstrating broad applicability and promotion potential. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 is the overall framework diagram of the fermentation process optimization model of the present invention.

[0057] Figure 2 is the diagram of the rolling learning method applied to this model. DETAILED DESCRIPTION OF THE INVENTION

[0058] The present invention will be described in detail below with reference to the drawings and embodiments:

[0059] As Figure 1-2 shown, the present invention realizes a generative fuzzy system modeling method for fermentation process optimization.

[0060] The method includes two main parts: an optimization modeling method for fermentation process based on generative fuzzy system and expert evaluation knowledge, and a rolling learning method for the optimization model of fermentation process. The optimization model of fermentation process integrates generative fuzzy rules. The membership degree of fermentation process characteristics is calculated through the antecedent of the rules, and the consequent is a generative model which outputs the optimization data y according to the input fermentation process characteristics x. Through the rolling learning of the actual data in the later production environment, the proposed fermentation process optimization model of the present invention can automatically adapt to the specific production environment, more efficiently mine the characteristic data of the fermentation process, and output high-precision process optimization data.

[0061] Example 1

[0062] A generative fuzzy system modeling method for fermentation process optimization includes the following steps:

[0063] The first step: Determine the number of generative rules K for the training dataset, the sample input feature x, and the output optimization data y.

[0064] The second step: Construct generative fuzzy rules, select the antecedent representative dlg for each rule and set the generative model GFRC.

[0065] The third step: Train the fermentation process optimization model on the dataset.

[0066] The fourth step: Evaluate the performance of the model with biochemical indexes and output the model with qualified performance.

[0067] In Example 1, the present invention uses two fermentation process datasets of food and beverage, namely Lactic Acid Fermentation (LAF) and Beer Fermentation (BF) datasets. LAF is a lactic acid fermentation dataset, which usually contains data such as biomass, lactic acid production, substrate concentration, pH value, temperature, etc. during the fermentation process of lactic acid bacteria, and is used for optimizing lactic acid production, studying microbial metabolism and process control. BF is a beer fermentation dataset, which contains data such as sugar concentration, alcohol content, temperature, pH value, yeast activity, etc. during the wort fermentation process, and is used for optimizing beer brewing process and quality control.

[0068] Table 1 summarizes the evaluation results of the GenFS algorithm of the invention and three other comparison algorithms (FS, RNN, Transformer). The metabolite concentration, substrate consumption rate, product quality, protein hydrolysis index, pH value, and redox potential are selected as the GenFS evaluation indicators. In the experiment, the evaluation performance of each task and the comprehensive performance of the entire dataset in the LAF and BF datasets are compared. For easy observation, the optimal values in each comparison data are marked in bold. From the experimental results in Table 1, it can be observed that: (1) On the LAF and BF datasets, the overall performance of the GenFS method is superior to the comparison algorithms, indicating that the generative method proposed in GenFS can well learn the characteristics of the complex fermentation process, finely optimize the fermentation process, and improve the quality of fermentation output. (2) When observing the pH value, it is found that the FS method directly exceeds the normal value range [4.0, 6.0], and although the RNN and Transformer methods are within the range, they are still acidic and inhibit the growth and metabolism of yeast, indicating that GenFS has strong robustness. (3) GenFS performs excellently in both fermentation datasets, indicating its good generalization ability.

[0069] Example 2

[0070] The present invention is a method for rolling learning of a fermentation process optimization model. Based on the real fermentation process data of a factory beer workshop for one year, the pre-trained generative fuzzy model is subjected to rolling training.

[0071] The results in Table 2 show that the performance of the GenFS method gradually approaches the optimal value over time, achieving adaptability to the real production environment. It is worth noting that the precise control of the pH value, with an error less than 0.1, further proves the good robustness and metabolic regulation ability of the GenFS method.

[0072] Performance comparison of different algorithms on LAF and BF datasets in Table 1

[0073]

[0074] *Common ranges of each index: metabolite concentration C m (0.1–100 g / L), substrate consumption rate R s (0.01–10 g / (L·h)), product quality Q p (90%–99.9%), moisture content W c (50%–95%), protein hydrolysis index P h (0.1–0.8), pH value (4.0–8.0), dissolved oxygen concentration O d (0.1–10 mg / L), redox potential E r(-400~+200 mV), the bolded values are the optimal values, and the values in red are outside the normal range.

[0075] Table 2 Performance of the GenFS model after weekly, monthly, quarterly, and annual rolling training in a certain beer fermentation workshop

[0076]

Claims

1. A fermentation process modeling method based on generative fuzzy system and expert knowledge evaluation, characterized in that, It includes the following steps: The first step: Determine the training dataset for the fermentation process field; The second step: According to the number of samples in the dataset and the fermentation indicators in the samples, determine the number of rules of the fuzzy system and the generative model of the consequent, and construct and initialize the fermentation process optimization model; 2.1 Generative fuzzification The process of generative fuzzification first involves the selection of representatives of fuzzy sets. Subsequently, for each fuzzy set representative, the system will calculate its similarity with the input data. Through the calculation of similarity values, the system can characterize the fuzzification characteristics of the input data, providing a basis for subsequent reasoning and decision-making; 2.2 Generative fuzzy rule consequent The generative task is to predict the full probability distribution of all variables. An intelligent model is used as the consequent of the fuzzy rule. The generative rule consequent of the fuzzy rule is expressed as follows: g k = GFRC k (x) (1) where k = 1, 2, 3, …, K, K being the number of rules of the fuzzy system, GFRC k (.) is the function of the consequential rule processing unit of the k-th rule, g k is the generated data of the k-th rule; 2.3 Generative fuzzy rules For the generative task, a generative fuzzy system is constructed based on fuzzy sets and fuzzy logic. The k-th generative fuzzy rule in the rule base is expressed by formula (2): where x is the input data, and GF(.) is the fuzzy set corresponding to the antecedent of the k-th rule represented by ; 2.4: Generative fuzzy rule combination mechanism Formula (3) gives the general form of the generative fuzzy rule combination mechanism: where x′ is the generated value (predicted value) of the fermentation index at the next moment, G = {g 1 , g 2 , g 3 , …, g K}, is the firing strength of the k-th rule, when G is structured and feature-aligned, use the firing strength as the weight to generate the result g k of each rule by weighted summation, otherwise select the rule output result with the maximum of all firing strengths as the final output, where function is to obtain the index value of the maximum firing strength; 2.5 Evaluate the fermentation indicators based on expert knowledge and optimize the output control parameters; Through actual fermentation production and fermentation experts, obtain the normal range of fermentation indicators. The indicator range is expert knowledge and is used to evaluate the quality of the fermentation indicators generated by the generative rules, as shown in formula (4); z = Expert(x′) (4) According to the evaluation results, output the control parameter y for the next moment, as shown in formula (5); y = α × z (5) Among them, Expert(.) is the generated value evaluated according to the empirical range of fermentation indicators, z is the evaluation result of each indicator, α is the learnable parameter matrix, p is the number of output control parameters, q is the number of fermentation indicators, and y is the control parameter at the next moment; The third step: Construct the objective function of the generative fuzzy system and perform model optimization; The fourth step: Train the fermentation process optimization model and output the trained model; The fifth step: Adopt the adaptive continuous learning method of the fermentation process optimization model.

2. The fermentation process modeling method based on generative fuzzy system and expert knowledge evaluation according to claim 1, characterized in that In the fifth step described above, the adaptive continuous learning method of the fermentation process optimization model includes the following steps: 5.1 Obtain real production data; The data collection mechanism includes using Internet of Things sensors to collect the actual parameters of the fermentation process in real time, and cleaning, denoising, and normalizing the data. At the same time, combining historical production data and process parameters to form a comprehensive dataset for model training; 5.2 Model rolling training; Considering the fermentation environment, equipment, and process of the actual factory, consider rolling training the model regularly by week, month, quarter, and year with actual generated data. The entire training process is consistent with the third and fourth steps of the modeling method; 5.3 New model iteration and deployment to the production environment; After the old and new models run in the way of a hybrid expert system for a period of time, the old model stops running and only the new model is retained to complete one iteration.

3. The fermentation process modeling method based on generative fuzzy system and expert knowledge evaluation according to claim 1, characterized in that For the third step described above, the specific operations are as follows: The fermentation process optimization model constructs the objective function using two losses, namely the loss of fermentation indicators and the loss of control parameters; The loss function adds an L2 regularization term, and its expression is as shown in formula (6): where y is the sequence of control parameters at the true next moment, is the generated sequence, and x ′ is the sequence of fermentation indexes at the true next moment, is the generated sequence of fermentation indexes at the next moment, diff(.) is the function for calculating the difference between the target sequence and the generated sequence, ‖Θ‖2 is the L2 regularization term of the consequent parameters of GFRC in GenFS, and γ (γ>1) is the balance parameter; GFRC needs to adopt an optimizer that can adaptively adjust the learning rate. The adaptive optimizer is expressed as shown in formula (5): Θ′ = AdaOptim(Θ, l s ) (6) where AdaOptim(.) is an adaptive optimizer function, l s is the initial value of the learning rate, and Θ′ is the model parameter updated by the optimizer calculation.

4. The fermentation process modeling method based on generative fuzzy system and expert knowledge evaluation according to claim 1, wherein For the first step described above, the specific operations are as follows: Using a fermentation process dataset that includes biomass, substrate concentration, pH value, temperature, and yeast activity data during the fermentation process; all features, covering input features and output optimization data, must be integrated as a whole for calculation, and the data will be concatenated into sequences of variable length.

5. The modeling method of the fermentation process based on the evaluation of the generative fuzzy system and expert knowledge according to claim 1, wherein The specific operations for the fourth step are as follows: In the fermentation scenario, the consequent GFRC of the generative rule of GenFS in the fermentation process optimization model and the α-parameter matrix in the evaluation and optimization module are learnable; when the loss of the target loss function converges and tends to be stable, the model training is terminated, and the trained model is output.

Images