Fermentation parameter optimization method and system based on machine learning

Abstract

The invention provides a fermentation parameter optimization method and system based on machine learning. The method comprises the following steps: collecting fermentation characteristic parameters and preprocessing the fermentation characteristic parameters; performing feature screening on the fermentation feature parameters through a triple evaluation mechanism; constructing a first data set and a second data set; constructing a prediction model; training a prediction model; inputting a first feature or a second feature obtained in actual production into the prediction model; and outputting a predicted value of the adenosine yield. The beneficial effects of the invention are that the method combines linear correlation, non-linear correlation and marginal contribution degree evaluation, comprehensively screens key features, eliminates redundant information, and improves the modeling efficiency and prediction precision; optimal hyper-parameters are effectively screened through grid search cross validation, the over-fitting risk of the model is reduced, and the generalization ability is enhanced; the particle swarm optimization algorithm efficiently searches an optimal parameter combination, breaks through the local optimum limitation of a traditional method, and improves the efficiency and rationality of process parameter optimization.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of bio-fermentation and artificial intelligence, and in particular relates to a method and system for optimizing fermentation parameters based on machine learning. Background Technology

[0002] In industrial adenosine fermentation production, Bacillus subtilis is mainly used as the production strain. This process is a core application scenario where bio-fermentation and industrial production intersect. Its production efficiency and product quality directly depend on the precise monitoring and scientific control of the fermentation process.

[0003] However, existing technologies suffer from significant technical bottlenecks, hindering improvements in production efficiency and product quality. Key quality attributes (CQAs), such as cell density, residual sugar content, and adenosine yield, are currently primarily obtained through manual sampling or offline detection (e.g., NIR, Raman spectroscopy, manual analysis), resulting in sparsity and delays in data acquisition, making it difficult to achieve real-time monitoring and timely control of the fermentation process. Traditional process optimization methods (such as single-factor experiments and response surface methodology, RSM) are limited by their own principles and cannot effectively handle the high-dimensional nonlinear coupling relationships between multiple parameters during fermentation, resulting in limited optimization effects.

[0004] The aforementioned problems directly lead to large fluctuations in the adenosine fermentation production process, reliance on operator experience for control, low product yield, and poor batch-to-batch consistency, which seriously affect the stability and economy of industrial production. There is an urgent need for a technical solution that combines real-time performance, precision, and intelligence to break through the existing bottlenecks. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a fermentation parameter optimization method and system based on machine learning, which is particularly suitable for achieving fully automated closed-loop control through adenosine yield prediction and process parameter optimization.

[0006] The technical solution adopted in this invention is as follows: Firstly, a method for optimizing fermentation parameters based on machine learning is provided, comprising the following steps:

[0007] Collect fermentation characteristic parameters and perform preprocessing;

[0008] The fermentation characteristic parameters are screened using a triple evaluation mechanism, which includes the following steps: calculating the Pearson correlation coefficient between the fermentation characteristic parameters and adenosine yield at each sample point to obtain a linearly correlated feature set with a significant linear correlation to adenosine yield; calculating the feature importance of the fermentation characteristic parameters based on the random forest algorithm to obtain a nonlinearly correlated feature set with a significant nonlinear correlation to adenosine yield; inputting the union of the selected linearly correlated feature set and the nonlinearly correlated feature set into a temporary base model; calculating and concentrating the marginal contribution and positive / negative influence direction of each candidate feature on adenosine yield through SHAP interpretation analysis to obtain environmental features, control features, and biochemical features.

[0009] A first dataset is constructed based on a first feature, and a second dataset is constructed based on a second feature. The first feature includes the environmental feature, the control feature, and the biochemical feature, and the second feature includes the environmental feature and the control feature.

[0010] A prediction model was constructed to fit the nonlinear mapping between the high-dimensional coupled fermentation characteristic parameters and adenosine yield;

[0011] The prediction model is trained using grid search cross-validation.

[0012] Input the first feature or the second feature obtained in actual production into the prediction model;

[0013] Predicted output of adenosine production.

[0014] Furthermore, it also includes the following steps:

[0015] The prediction model is then optimized using global process parameters.

[0016] The first feature or the second feature obtained in actual production is input into the prediction model.

[0017] Furthermore, it also includes the following steps:

[0018] Output the theoretical adenosine yield and the corresponding optimal process parameters.

[0019] Furthermore, it also includes the following steps: applying the optimal process parameters to the actual fermentation process to form an adaptive closed-loop control.

[0020] Furthermore, constructing a predictive model includes the following steps:

[0021] The prediction model adopts a Grid-MLP architecture;

[0022] The activation function of the prediction model is fixed as ReLU, the optimizer as Adam, the loss function as mean squared error, and the maximum number of iterations as 1000;

[0023] Define the hyperparameter search space, which includes the hidden layer structure, L2 regularization coefficients, and initial learning rate.

[0024] Furthermore, training the prediction model includes the following steps:

[0025] The first dataset and the second dataset are divided into training set and test set;

[0026] The prediction model is subjected to grid search cross-validation, and the optimal hyperparameter combination is selected by traversing the hyperparameter search space.

[0027] The training set is input into the prediction model configured with the optimal hyperparameters, and iteratively trained according to the preset loss function and optimizer until the maximum number of iterations is reached or the model converges.

[0028] The trained prediction model is evaluated using the test set, and the prediction model with the best prediction accuracy is selected.

[0029] Furthermore, the global process parameter optimization of the prediction model includes the following steps:

[0030] The predicted adenosine production value output by the prediction model is used as the optimization objective function;

[0031] The process parameters are subject to range constraints.

[0032] The particle swarm optimization algorithm is introduced to search for the optimal process parameters.

[0033] Secondly, a fermentation parameter optimization system based on machine learning is provided, including:

[0034] The data acquisition module is used to collect fermentation characteristic parameters;

[0035] The data preprocessing module is used to preprocess the fermentation characteristic parameters;

[0036] The feature screening module is used to screen key features of the fermentation feature parameters through a triple evaluation mechanism to obtain the first feature and the second feature.

[0037] The soft measurement prediction module is used to receive the first feature or the second feature and perform adenosine production prediction, and output the adenosine production prediction value.

[0038] The global optimization module is used to search for optimal process parameters with the predicted adenosine yield as the objective function.

[0039] The control execution module is used to receive and execute the optimal process parameters, forming an adaptive closed-loop control.

[0040] Furthermore, the fermentation characteristic parameters include: environmental parameters, control variables, and biochemical characteristics. The environmental parameters include temperature, pH, and DO. The control variables include stirring speed, carbon and nitrogen source feeding rate, and ammonia water feeding rate. The biochemical characteristics include OD and residual sugar.

[0041] Furthermore, the environmental parameters and the control variables are sampled at a frequency of once per minute, and the biochemical characteristics are sampled at a frequency of once every four hours.

[0042] Furthermore, the preprocessing includes: imputing missing values ​​using the KNN algorithm and Z-score normalization.

[0043] The advantages and positive effects of this invention are as follows: By adopting the above-mentioned technical solution, combining linear correlation, nonlinear correlation, and marginal contribution evaluation, key features are comprehensively screened, redundant information is eliminated, and modeling efficiency and prediction accuracy are improved; it has strong interpretability, clearly defining the direction of influence of features on adenosine production, and providing a theoretical basis for subsequent process control; grid search cross-validation effectively screens the optimal hyperparameters, reduces the risk of model overfitting, and enhances generalization ability; the particle swarm optimization algorithm efficiently searches for the optimal parameter combination, breaking through the local optima limitation of traditional methods and improving the efficiency and rationality of process parameter optimization. Attached Figure Description

[0044] Figure 1 This is a schematic flowchart of a fermentation parameter optimization method according to an embodiment of the present invention. Detailed Implementation

[0045] The present disclosure will now be described more fully with reference to the accompanying drawings, which illustrate exemplary embodiments of the present disclosure. The technical solutions of the embodiments of the present disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present disclosure, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present disclosure without creative effort are within the scope of protection of the present disclosure.

[0046] like Figure 1 As shown, this invention provides a machine learning-based method for optimizing fermentation parameters, comprising the following steps:

[0047] S100. Collect fermentation characteristic parameters (multi-source variables in the fermentation process) and perform preprocessing;

[0048] S200. Fermentation characteristic parameters are screened using a triple evaluation mechanism, which includes the following steps: calculating the Pearson correlation coefficient between fermentation characteristic parameters and adenosine yield at each sample point to obtain a linearly correlated feature set with significant linear correlation to adenosine yield; calculating the feature importance of fermentation characteristic parameters based on the random forest algorithm to obtain a nonlinearly correlated feature set with significant nonlinear correlation to adenosine yield; inputting the union of the screened linearly correlated feature set and the nonlinearly correlated feature set into a temporary basic model (collecting pre-processed fermentation process characteristic parameters and using these original parameters to train a basic prediction model, such as RF); calculating and concentrating the marginal contribution and positive / negative influence direction of each candidate feature to adenosine yield through SHAP interpretation analysis; and obtaining environmental features, control features, and biochemical features.

[0049] S300. Construct a first dataset (full feature type) based on the first feature, and construct a second dataset (constrained type) based on the second feature for scenarios where biochemical indicator detection is limited. The first feature includes environmental features, control features, and biochemical features, and the second feature includes environmental features and control features.

[0050] S400. Construct a prediction model to fit the nonlinear mapping between the high-dimensional coupled fermentation characteristic parameters and adenosine yield;

[0051] S500: Train the prediction model through grid search cross-validation;

[0052] S600. Input the first or second feature obtained in actual production into the prediction model. First, collect the fermentation characteristic parameters in actual production, and perform feature screening through a triple evaluation mechanism. In industrial scenarios where sensing conditions are sufficient and biochemical indicators can be stably obtained, input the first feature into the prediction model; in industrial scenarios where the detection of biochemical indicators is limited, input the second feature into the prediction model.

[0053] S700, predicted output adenosine production.

[0054] By employing the above methods, standardized data acquisition and preprocessing procedures reduce issues such as missing data and inconsistent timelines, improving data quality and providing a reliable data foundation for subsequent modeling. The construction of two types of datasets adapts to different sensing conditions, enhancing the method's practicality and flexibility. A complete "data, feature, modeling, prediction" chain is formed, providing technical support for real-time adenosine yield prediction and overcoming the limitations of traditional offline detection. Combining linear correlation, nonlinear correlation, and marginal contribution assessment, key features are comprehensively screened, redundant information is eliminated, and modeling efficiency and prediction accuracy are improved. The method possesses strong interpretability, clearly defining the direction of feature influence on adenosine yield, providing a theoretical basis for subsequent process control. The screening results are adapted to the construction requirements of both types of datasets, balancing scenarios with sufficient and limited sensing capabilities, thus enhancing the method's applicability.

[0055] To address the problem of large production fluctuations and low yields caused by the lack of global optimization of process parameters, this embodiment provides an implementation method.

[0056] In one embodiment, the following steps are also included:

[0057] Optimize the global process parameters of the prediction model;

[0058] Input the first or second feature obtained from actual production into the prediction model;

[0059] Output the theoretical adenosine yield and the corresponding optimal process parameters.

[0060] By adopting the above method, global optimization of process parameters can be achieved, breaking through the limitations of local optima in traditional methods, and helping to improve adenosine yield and production consistency.

[0061] To address the problem of relying on experience-based control due to the lack of a real-time closed-loop control mechanism, this embodiment provides an implementation method.

[0062] In one embodiment, the following steps are also included:

[0063] The optimal process parameters are applied to the actual fermentation process to form an adaptive closed-loop control.

[0064] Using the above methods, an adaptive closed-loop control mechanism is constructed, which directly applies the optimized parameters to actual production, reduces human intervention and reliance on experience; it connects the entire process of "prediction, optimization and control", realizes real-time intelligent decision-making in the fermentation process, and promotes the transformation of the production mode from passive monitoring to active control.

[0065] To address the problem that prediction models struggle to fit the high-dimensional nonlinear coupling relationships in the fermentation process, and that the lack of a unified standard for model construction leads to poor prediction stability, this embodiment provides an implementation method.

[0066] In one embodiment, constructing a prediction model includes the following steps:

[0067] With adenosine production as the prediction target, the prediction model adopts the Grid-MLP architecture, which has a three-layer structure of input layer, hidden layer and output layer. The hidden layer is initially configured according to the hyperparameter search range, and the output layer is set with one neuron.

[0068] The activation function of the fixed prediction model is ReLU, the optimizer is Adam, the loss function is mean squared error, and the maximum number of iterations is 1000.

[0069] Define the hyperparameter search space, which includes the hidden layer structure, L2 regularization coefficients, and initial learning rate. Preferably, the hidden layer structure... L2 regularization coefficient Initial learning rate .

[0070] Using the above method and the Grid-MLP architecture, the nonlinear characteristics of the fermentation process are accurately adapted, laying the foundation for efficient fitting of the mapping relationship between features and adenosine production. The core configuration of the model is fixed to avoid blind design of the architecture and improve the standardization and repeatability of model construction. The search range of hyperparameters is clearly defined to provide a clear target for subsequent hyperparameter optimization and ensure the controllability of model performance.

[0071] To address the issues of weak model generalization ability and insufficient prediction reliability caused by non-standard model training processes and a lack of systematic verification, this embodiment provides an implementation method.

[0072] In one embodiment, training the prediction model includes the following steps:

[0073] The first and second datasets are divided into training and test sets, with the ratio of training to test sets being 8:2.

[0074] The prediction model employs grid search cross-validation, traversing the hyperparameter search space to select the optimal hyperparameter combination. It is then compared with other models (random forest, support vector machine, extreme gradient boosting, least squares support vector machine, etc.) using 5-fold cross-validation. Shuffle is enabled during each fold of training. The optimal combination of hyperparameters is selected based on the coefficients of determination.

[0075] The training set is input into the prediction model with the optimal hyperparameters, and the nonlinear mapping relationship between "input features and adenosine production" is fitted. The model is then iteratively trained according to the preset loss function and optimizer until the maximum number of iterations is reached or the model converges.

[0076] The trained prediction model is evaluated using a test set, and the evaluation is based on a three-fold metric, namely the coefficient of determination. The mean squared error (MSE) and mean absolute error (MAE) are used to select the prediction model with the best prediction accuracy.

[0077] By adopting the above methods, the scientific and standardized process of dataset partitioning, hyperparameter optimization, iterative training and performance verification is used to improve the scientific nature and standardization of model training; grid search cross-validation effectively selects the optimal hyperparameters, reduces the risk of model overfitting and enhances generalization ability; multi-model comparison and multi-index evaluation ensure that the selected model has the best prediction accuracy, providing reliable objective function support for subsequent process optimization.

[0078] To address the problem of blind and inefficient parameter optimization caused by the inability to handle the high-dimensional coupling relationships of fermentation parameters, this embodiment provides an implementation method.

[0079] In one embodiment, optimizing the global process parameters of the prediction model includes the following steps:

[0080] The predicted adenosine production output from the prediction model is used as the optimization objective function.

[0081] The process parameters are constrained to a range that conforms to the actual operating range of industrial fermentation. Preferably, the variable range of the process parameters is: temperature 30℃–40℃, pH 5.5–pH 7.0, DO 5%–60%, stirring speed 100rpm–300rpm, and the carbon and nitrogen source / ammonia water feeding rate is adapted to the fermenter specifications.

[0082] Particle swarm optimization (PSO) is introduced to search for optimal process parameters. The optimal PSO parameters are set as follows: number of particles 40, number of iterations 50, inertia weight 0.9, and individual learning factor [missing value]. Group learning factor .

[0083] Using the above method, the predicted adenosine yield is used as the objective function to ensure accurate optimization direction and avoid meaningless parameter searches. The parameter range is set based on actual industrial constraints to ensure the feasibility and implementability of the optimization results and adapt to industrial production needs. The particle swarm optimization algorithm efficiently searches for the optimal parameter combination, breaking through the limitations of local optima in traditional methods and improving the efficiency and rationality of process parameter optimization.

[0084] To facilitate the use of the machine learning-based fermentation parameter optimization method provided in this disclosure, this disclosure also provides a machine learning-based fermentation parameter optimization system, comprising:

[0085] The data acquisition module is used to collect fermentation characteristic parameters and connects to the fermenter control system via an industrial bus to acquire data in real time.

[0086] The data preprocessing module is used to preprocess fermentation characteristic parameters;

[0087] The feature screening module is used to screen key features of fermentation characteristic parameters through a triple evaluation mechanism to obtain the first feature and the second feature.

[0088] The soft measurement prediction module is used to receive the first feature or the second feature and perform adenosine production prediction, and output the adenosine production prediction value.

[0089] The global optimization module is used to search for optimal process parameters with the predicted adenosine yield as the objective function.

[0090] The control execution module is used to receive and execute the optimal process parameters, forming an adaptive closed-loop control.

[0091] In one embodiment, the fermentation characteristic parameters include: environmental parameters, control variables, and biochemical characteristics. The environmental parameters include temperature, pH, and dissolved oxygen (DO). The control variables include stirring speed, carbon and nitrogen source feeding rate, and ammonia water feeding rate. The biochemical characteristics include OD (cell density) and residual sugar.

[0092] In one embodiment, the environmental parameters and control variables are sampled once per minute, and the biochemical characteristics are sampled once every four hours.

[0093] In one embodiment, preprocessing includes imputing missing values ​​using the KNN algorithm and Z-score normalization. The K value for KNN imputation is 5. Hourly resampling ensures temporal consistency. High-frequency data (environmental parameters and control variables) and low-frequency data (biochemical characteristics) with different time granularities are aligned to integer multiples. The equation for Z-score normalization is... ,in These are the standardized values ​​of individual fermentation characteristic parameters. These are the original measured values ​​of fermentation characteristic parameters. The mean, The standard deviation is denoted as .

[0094] Based on embodiments of this disclosure, this disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0095] An electronic device includes at least one processor and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the machine learning-based fermentation parameter optimization method provided in this disclosure.

[0096] Electronic devices are intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. Electronic devices can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0097] A non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to cause a computer to execute the machine learning-based fermentation parameter optimization method provided in this disclosure.

[0098] The various embodiments of this disclosure can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0099] A computer program product includes a computer program / instructions that are executed by a processor, comprising the machine learning-based fermentation parameter optimization method disclosed herein.

[0100] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0101] In the context of this disclosure, machine-readable media includes machine-readable storage media and machine-readable signal media. A machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0102] The embodiments of the present invention have been described in detail above, but the content described is only a preferred embodiment of the present invention and should not be considered as limiting the scope of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the patent coverage of the present invention.

Claims

1. A machine learning based fermentation parameter optimization method, characterized by, The method comprises the following steps: Collecting fermentation characteristic parameters and preprocessing; Screening the fermentation characteristic parameters through a triple evaluation mechanism, which comprises the following steps: calculating the Pearson correlation coefficient of the fermentation characteristic parameters and the adenosine yield at each sample point to obtain a linearly correlated feature set with significant linear correlation with the adenosine yield, calculating the feature importance of the fermentation characteristic parameters based on a random forest algorithm to obtain a nonlinearly correlated feature set with significant nonlinear correlation with the adenosine yield, screening the union set of the linearly correlated feature set and the nonlinearly correlated feature set, inputting into a temporary base model, and calculating the marginal contribution and positive and negative influence direction of each candidate feature in the union set to the adenosine yield through SHAP explanation analysis, and screening environmental features, control features and biochemical features; Based on the first feature, a first data set is constructed, and based on the second feature, a second data set is constructed, wherein the first feature includes the environmental feature, the control feature and the biochemical feature, and the second feature includes the environmental feature and the control feature; A prediction model is constructed for fitting the nonlinear mapping between the high-dimensional coupled fermentation characteristic parameters and the adenosine yield; The prediction model is trained through grid search cross-validation; The first feature or the second feature obtained in actual production is input into the prediction model; The adenosine yield prediction value is output.

2. The machine learning based fermentation parameter optimization method according to claim 1, characterized in that, Further comprising the following steps: Global process parameter optimization is performed on the prediction model; The first feature or the second feature obtained in actual production is input into the prediction model; The theoretical adenosine yield and the corresponding optimal process parameters are output.

3. The machine learning based fermentation parameter optimization method of claim 2, wherein, Further comprising the following steps: The optimal process parameters are applied to the actual fermentation process to form a self-adaptive closed-loop regulation.

4. The machine learning based fermentation parameter optimization method of claim 1, wherein, The construction of the prediction model comprises the following steps: The prediction model adopts a Grid-MLP architecture; The activation function of the prediction model is fixed as ReLU, the optimizer is Adam, the loss function is mean square error, and the maximum iteration step is 1000; The hyperparameter search space is defined, which includes the hidden layer structure, the L2 regularization coefficient and the initial learning rate.

5. The machine learning based fermentation parameter optimization method of claim 4, wherein, The training of the prediction model comprises the following steps: The first data set and the second data set are divided into a training set and a test set; The prediction model is subjected to grid search cross-validation to screen the optimal hyperparameter combination by traversing the hyperparameter search space; The training set is input into the prediction model configured with the optimal hyperparameters, and the iteration training is performed according to the preset loss function and the optimizer until the maximum iteration step is reached or the model converges; The trained prediction model is evaluated using the test set to screen the prediction model with the optimal prediction accuracy.

6. The machine learning based fermentation parameter optimization method of claim 2, wherein, The global process parameter optimization of the prediction model comprises the following steps: The adenosine yield prediction value output by the prediction model is used as an optimization objective function; The process parameters are range-constrained; The particle swarm optimization algorithm is introduced to search for the optimal process parameters.

7. A machine learning based fermentation parameter optimization system, characterized in that, It comprises: A data acquisition module for collecting fermentation characteristic parameters; A data preprocessing module for preprocessing the fermentation characteristic parameters; The feature screening module is configured to screen key features of the fermentation feature parameters through a triple evaluation mechanism to obtain first features and second features. The soft measurement prediction module is configured to receive the first features or the second features and perform adenosine yield prediction to output an adenosine yield prediction value. The global optimization module is configured to search for optimal process parameters with the adenosine yield prediction value as a target function. The control execution module is configured to receive the optimal process parameters and execute them to form an adaptive closed-loop regulation.

8. The machine learning-based fermentation parameter optimization system of claim 7, wherein, The fermentation feature parameters include environmental parameters, control variables and biochemical features, the environmental parameters include temperature, pH and DO, the control variables include stirring speed, carbon and nitrogen source feeding rate and ammonia feeding rate, and the biochemical features include OD and residual sugar.

9. The machine learning-based fermentation parameter optimization system of claim 8, wherein: The sampling frequency of the environmental parameters and the control variables is one time per minute, and the sampling frequency of the biochemical features is one time per four hours.

10. The machine learning-based fermentation parameter optimization system of claim 7, wherein, The preprocessing includes using a KNN algorithm to interpolate missing values and Z-score standardization.

Images