A method for monitoring the dynamic of a microbial community of indigenous saccharomyces cerevisiae fermentation

Abstract

The application discloses a kind of local saccharomyces cerevisiae fermentation microbial community dynamic monitoring method, it is related to microbial community dynamic monitoring technical field, the method includes the following steps: sterile sampling valve and sensor are respectively installed to current fermenter key site, sampling and data acquisition are carried out to fermenter key site in different fermentation stages, respectively carry out rapid screening and offline analysis to the sample and data obtained, and construct the mapping relationship database between sensor signal and microbial community state in the process of brewing fermentation;Biological significance features are extracted from sensor data collection, sensor feature time series are constructed, and the output microbial state is defined, and complete prediction model is trained using machine learning algorithm.The application constructs the local saccharomyces cerevisiae fermentation microbial community dynamic monitoring model adapted to industrial scene by combining offline modeling and online prediction and simplifying actual model based on conventional sensor.

Description

Technical Field

[0001] This invention relates to the field of microbial community dynamic monitoring technology, and more specifically, to a method for monitoring the dynamics of a native brewing yeast fermentation microbial community. Background Technology

[0002] As the core functional microorganism for alcoholic beverage fermentation, the community structure, activity state and dynamic changes of native brewing yeast directly determine the flavor, quality and alcohol yield of the fermentation products. Accurate monitoring of the dynamics of the microbial community is the key to optimizing the fermentation process and ensuring product stability. Current methods for monitoring microbial communities largely rely on laboratory sampling and testing to identify the state of microorganisms. While these methods can accurately capture the community status, they are time-consuming and cannot reflect the dynamic changes of microorganisms within the fermenter in real time. This leads to delays in process adjustments and makes the fermentation process and product quality susceptible to imbalances in the microbial community. Furthermore, laboratory analysis requires specialized equipment and technicians, resulting in high testing costs that are unsuitable for the continuous monitoring needs of large-scale industrial production. In addition, existing online monitoring technologies mostly focus on conventional physicochemical parameters such as temperature, pH, and dissolved oxygen, failing to establish accurate correlation models between these parameters and the state of the microbial community. This makes it impossible to indirectly deduce the dynamics of the microbial community from conventional sensor data. Although some research-based monitoring models attempt to correlate multiple parameters, their complex structures, reliance on numerous sensors, and insufficient analysis of parameter contribution mean that simplified models have not been constructed. This results in high deployment costs in practical applications and makes it difficult to adapt to the actual needs of industrial production scenarios. Summary of the Invention

[0003] In response to the problems in related technologies, this invention proposes a method for dynamic monitoring of the fermentation microbial community of native brewing yeast, in order to overcome the aforementioned technical problems existing in the existing related technologies.

[0004] Therefore, the specific technical solution adopted by the present invention is as follows: A method for monitoring the dynamics of a native brewing yeast fermentation microbial community, comprising the following steps: S1. Install sterile sampling valves and sensors at key sites of the fermenter, sample and collect data at key sites of the fermenter at different fermentation stages, and conduct rapid screening and offline analysis on the acquired samples and data to build a basic database of the mapping relationship between sensor signals and microbial community status during the brewing fermentation process. S2. Extract biologically significant features from the sensor-collected data, construct a sensor feature time series, define the output microbial state, train a complete prediction model using machine learning algorithms, and validate the model using fermentation batch data from the validation set. S3. Evaluate the contribution of each sensor feature to the prediction of biological state, construct a simplified prediction model using the selected core sensor features, and compare the prediction accuracy of the complete prediction model and the simplified prediction model on an independent validation dataset to determine the final combination of sensors to be deployed online for dynamic monitoring of the microbial community in the fermenter.

[0005] In a preferred embodiment, S1 includes the following steps: S11. Aseptic sampling valves and sensors are installed at key sites in the current fermenter. These key sites include the gas phase zone at the top of the tank, the upper liquid phase, the middle liquid phase, the lower liquid phase, and the biofilm on the tank wall. Sampling and data collection are performed at different fermentation stages in the fermenter using aseptic sampling valves and sensors. These fermentation stages include the lag phase, logarithmic phase, stationary phase, and decay phase. The fermenter is spatially discretized into five key ecological sites: the gas phase zone, the upper liquid phase, the middle liquid phase, the lower liquid phase, and the attached phase. Sensor groups are deployed at each liquid phase site to record signal vectors in real time and synchronize sensor timestamps with a hardware clock. A unique code is assigned to each sample. S12. For samples and data obtained at different key sites and fermentation stages, rapid screening and offline analysis are performed after sampling to obtain the types, quantities, and community structure of microorganisms in the collected samples. At the same time, the glucose, ethanol, lactic acid, and acetic acid contents of the samples are measured. Data from N batches of normally completed brewing fermentation and N / 3 batches of fermentation with abnormalities are collected to construct a basic database of the mapping relationship between sensor signals and microbial community status during brewing fermentation.

[0006] In a preferred embodiment, S12 includes the following steps: S121. During rapid screening, based on Raman spectroscopy, principal component analysis is used to reduce the dimensionality and identify patterns of the single-cell spectra collected by sampling. Based on the principal component scores, K-means cluster analysis is used to count the cell proportions of various microorganisms. The total cell concentration and the proportion of live cells are obtained through fluorescence staining. S122. During offline analysis, absolute quantification by qPCR is used to output the gene copy number concentration of the target microorganism. At the same time, high-throughput sequencing analysis is combined to obtain the Shannon diversity index and Bray-Curtis dissimilarity of the sample. High-performance liquid chromatography is used to determine the concentration vector of key metabolites, including glucose, ethanol, lactic acid, and acetic acid. All heterogeneous data are integrated into a unified feature-tag pair. Each row of the main table corresponds to a time-site sample, including key site number, sensor characteristics, and microbial state. The association table records the Shannon diversity index, Bray-Curtis dissimilarity, and metabolite concentration data. Multiple batches of data are merged and integrated into a basic database of mapping relationships.

[0007] In a preferred embodiment, S2 includes the following steps: S21. For the sensor data in the mapping relation database, extract biological significance features from the sensor data to obtain primary derived features and secondary statistical features. At the same time, define the output microbial state, including yeast activity ratio, bacterial contamination risk index, total active biomass and community state classification. S22. Based on primary derived features, secondary statistical features, and microbial state, the complete fermentation batch is divided into data points, and a complete prediction model is trained using machine learning algorithms. At the same time, the model is validated using fermentation batch data from the validation set.

[0008] In a preferred embodiment, S21 includes the following steps: S211. Extract the original feature vector from the timestamp synchronized in the mapping relationship database, and construct the first-level derived features, including oxygen uptake rate, carbon dioxide release rate, respiratory entropy, dissolved oxygen gradient, pH change rate, and temperature change rate. Among them oxygen uptake rate , This represents the derivative of dissolved oxygen concentration with respect to time. Represents dissolved oxygen concentration. Represents the volumetric oxygen transfer coefficient. Represents the saturated dissolved oxygen concentration; Among them, carbon dioxide release rate ,in These are the volumetric flow rates of exhaust gas and intake gas, respectively. These represent the carbon dioxide concentration in the exhaust gas and the intake gas, respectively, and the respiratory entropy. ; Dissolved oxygen gradient , These represent the dissolved oxygen concentrations measured at the top and bottom of the fermenter at time t, respectively. S212. Calculate the window mean, window variance, and window extreme values ​​within the sliding window time W, where the window time... ,in The width of the time window is represented by the total input feature vector X(t) of the model, which is the concatenation of all the above features. S213. Define the output microbial state, including the yeast activity ratio. , These represent the concentration of target yeast live cells in the fermentation broth at time t and the total concentration of live cells of all microorganisms in the fermentation broth at time t, respectively. Bacterial contamination risk index ,in , This represents the concentration of lactic acid bacteria and acetic acid bacteria in the fermentation broth at time t. b represents the weighting coefficient, and b is the bias term; Total active biomass and Maintaining consistency, community state classification , representing the community state label of the fermentation system at time t. K represents the set of all states, where K is the total number of state categories.

[0009] In a preferred embodiment, S22 includes the following steps: S221. Divide the data according to the complete fermentation batch, where the dataset is: The training set is The validation set is Where U represents the number of reserved batches; S222. A regression prediction model is constructed using LSTM to predict the yeast activity ratio, bacterial contamination risk index, and total active biomass. A classification prediction model is constructed using the random forest method to predict community state classification. In the LSTM model construction process, the input features include primary derived features and original online sensor parameters. The labels are the microbial states obtained from offline analysis. The mean squared error is used as the loss function, and the Adam optimizer is used for optimization. The root mean square error and the coefficient of determination are used as evaluation indicators. In the random forest classification prediction model construction process, secondary statistical features need to be added to the model input and the cross-entropy loss function is used. The feature matrix and labels of the training set are directly input into the model for fitting.

[0010] In a preferred embodiment, step S3 includes the following steps: S31. Perform feature importance analysis on the complete prediction model, evaluate the contribution of each sensor feature to the predicted biological state, and use the selected core sensor feature subset as input to train the simplified model. S32. On an independent validation dataset, compare the prediction accuracy of the complete prediction model and the simplified prediction model to determine the final sensor combination for online deployment and to dynamically monitor the microbial community in the fermenter.

[0011] In a preferred embodiment, S31 includes the following steps: S311. For the classification prediction model built for random forests, the total information gain of each feature in all decision trees is evaluated by calculating the Gini importance of the tree model. This includes the following steps: ; in The importance of the Gini symbol representing feature j. The total number of trees is represented by T, where T represents a single tree and g represents a split node in the tree. The reduction in Gini impurity of feature j when it splits at node g is represented by the amount of Gini impurity reduction. At the same time, the baseline performance of the model is calculated on the validation set. All values ​​of feature j are randomly shuffled to form a perturbation dataset. The perturbation performance is calculated, and the difference between the baseline performance and the perturbation performance is taken as the importance of feature j. The importance is obtained by repeating R times and taking the average. S312. For the regression prediction model built using LSTM, the SHAP value of feature j is calculated to determine the contribution of different features to the predicted biological state. This includes the following steps: ; in, The SHAP value represents feature j, F represents the entire feature set, and S represents a subset of features. This represents the model's prediction of sample x when the input only contains features of a subset S. The model prediction value after adding target feature j to the representative feature subset S is given for all samples. Take the absolute value and then average it to obtain the global importance of feature j. Select the core features from the permutation importance, Gini importance and global importance according to the preset threshold, and select the top P features to obtain the sensor type under the corresponding top P core features. S313. Select a subset of core features from the original feature set, and repeat step S2 to train a simplified prediction model using the same algorithm structure.

[0012] In a preferred embodiment, S32 includes the following steps: S321. On z samples of the independent verification batch, prediction is performed using the complete prediction model and the simplified prediction model respectively. The prediction accuracy parameter is compared. If the prediction accuracy is within the preset threshold range, it means that the current simplified model meets the requirements. The sensor type of the feature corresponding to the simplified model is used as the final sensor combination for online deployment. If the model accuracy does not meet the requirements, repeat steps S311-S313 to expand the range of P until a sensor combination for final online deployment that meets the conditions is obtained. S322. Real-time parameters collected by a combination of parameters from online-deployed sensors are combined with a simplified prediction model to output the microbial state.

[0013] The beneficial effects of this invention are as follows: This invention combines offline modeling with online prediction using conventional sensor data. It indirectly and in real-time measures the state of microorganisms through models. By rapidly screening and offline analyzing samples from different sites in multiple batches of fermenters and sensor parameters, a complete prediction model for different fermentation stages is constructed. This allows for dynamic prediction of the microbial community in the fermenter based on the sensor's collected parameters. Furthermore, by analyzing the contribution of different parameters in the complete prediction model, important features are selected and a simplified prediction model is constructed. This model is then linked to an online sensor combination to create a simplified model suitable for industrial online deployment. This invention establishes a dynamic mapping relationship between conventional sensor signals and the internal state of microorganisms, avoiding the lag in microbial community state results obtained by traditional reliance on offline sampling and laboratory analysis. This enables real-time and continuous prediction of microbial state. At the same time, by simplifying the sensor combination, it simplifies the cost of practical applications and improves practicality. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a flowchart of a method for dynamic monitoring of a native brewing yeast fermentation microbial community according to an embodiment of the present invention. Detailed Implementation

[0016] To further illustrate the various embodiments, the present invention provides accompanying drawings, which are part of the disclosure of the present invention. These drawings are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these drawings, those skilled in the art should be able to understand other possible implementation methods and the advantages of the present invention. The components in the drawings are not drawn to scale, and similar component symbols are generally used to represent similar components.

[0017] According to an embodiment of the present invention, a method for dynamic monitoring of the fermentation microbial community of native brewing yeast is provided.

[0018] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments: Example 1:

[0019] like Figure 1 As shown, a method for monitoring the dynamics of a native brewing yeast fermentation microbial community according to an embodiment of the present invention includes the following steps: S1. Install sterile sampling valves and sensors at key sites of the fermenter, sample and collect data at key sites of the fermenter at different fermentation stages, and conduct rapid screening and offline analysis on the acquired samples and data to build a basic database of the mapping relationship between sensor signals and microbial community status during the brewing fermentation process. S11. Aseptic sampling valves and sensors are installed at key sites in the current fermenter. These key sites include the gas phase zone at the top of the tank, the upper liquid phase, the middle liquid phase, the lower liquid phase, and the biofilm on the tank wall. Sampling and data collection are performed at different fermentation stages in the fermenter using aseptic sampling valves and sensors. These fermentation stages include the lag phase, logarithmic phase, stationary phase, and decay phase. The fermenter is spatially discretized into five key ecological sites: the gas phase zone, the upper liquid phase, the middle liquid phase, the lower liquid phase, and the attached phase. Sensor groups are deployed at each liquid phase site to record signal vectors in real time and synchronize sensor timestamps with a hardware clock. A unique code is assigned to each sample. It should be noted that the sampling frequency varies at different fermentation stages. Typically, samples are taken every 1-4 hours during the lag and logarithmic phases, every 6-8 hours during the stationary phase, and every 12 hours during the decay phase. The specific frequency needs to be adjusted based on the actual fermenter conditions and by consulting experts in the field using empirical methods. In cases of drastic parameter changes, such as a sudden drop in dissolved oxygen or a pH inflection point, additional sampling is required immediately. Sensors used include pH and dissolved oxygen electrodes, temperature probes, gas sensors, and pressure sensors. The specific settings need to be pre-set based on the fermentation process of different fermenters using empirical methods. At each preset time point, samples are aseptically collected simultaneously from the sampling valves at all the above-mentioned sites and immediately placed on ice or at 4°C. Analysis begins within 30 minutes. By assigning a unique code to each sample, data traceability can be ensured.

[0020] S12. For samples and data obtained at different key sites and fermentation stages, rapid screening and offline analysis are performed after sampling to obtain the types, quantities, and community structure of microorganisms in the collected samples. At the same time, the glucose, ethanol, lactic acid, and acetic acid contents of the samples are measured. Data from N batches of normally completed brewing fermentation and N / 3 batches of fermentation with abnormalities are collected to construct a basic database of the mapping relationship between sensor signals and microbial community status during brewing fermentation.

[0021] It should be noted that by collecting data from key locations and samples and sensors at different stages across multiple completed fermentation batches of the same brewing product, sufficient raw data can be accumulated to facilitate the construction of a complete prediction model. N is usually set to 6, but can be adjusted according to the actual situation. Abnormal fermentation batches are rounded up.

[0022] S121. During rapid screening, based on Raman spectroscopy, principal component analysis is used to reduce the dimensionality and identify patterns of the single-cell spectra collected by sampling. Based on the principal component scores, K-means cluster analysis is used to count the cell proportions of various microorganisms. The total cell concentration and the proportion of live cells are obtained through fluorescence staining. It should be noted that during principal component analysis, the high-dimensional spectrum of each cell is projected into a new space composed of several principal components. Based on the different biochemical compositions of cells from different species, such as protein, lipid, and nucleic acid characteristic peaks, they will cluster into different clusters on the principal component analysis score map to achieve rapid classification of different cells. By combining flow cytometry with fluorescent dyes, cell population gating analysis is performed on the collected samples based on a two-parameter scatter plot, and the absolute values ​​of total cell concentration and viable cell proportion are directly output.

[0023] S122. During offline analysis, absolute quantification by qPCR is used to output the gene copy number concentration of the target microorganism. At the same time, high-throughput sequencing analysis is combined to obtain the Shannon diversity index and Bray-Curtis dissimilarity of the sample. High-performance liquid chromatography is used to determine the concentration vector of key metabolites, including glucose, ethanol, lactic acid, and acetic acid. All heterogeneous data are integrated into a unified feature-tag pair. Each row of the main table corresponds to a time-site sample, including key site number, sensor characteristics, and microbial state. The association table records the Shannon diversity index, Bray-Curtis dissimilarity, and metabolite concentration data. Multiple batches of data are merged and integrated into a basic database of mapping relationships.

[0024] It should be noted that when using qPCR to absolutely quantify the gene copy number concentration of the target microorganism, which includes strains such as Saccharomyces cerevisiae, Lactobacillus, and Acetic Acid Bacillus, a standard curve is used to determine the cycling threshold for target species i. Converted to absolute gene copy concentration , Where k is the slope and b is the intercept, which needs to be obtained by fitting with standards of known concentration. In high-throughput sequencing analysis, the relative abundance of operational taxa or amplicon sequence variants is calculated, and the Shannon diversity index and Bray-Curtis dissimilarity are calculated to verify the targeting results of qPCR and discover potential new related species. By combining rapid screening with offline analysis, the situation in the fermenter is quickly determined based on the rapid screening results during sample collection. The results of offline analysis combined with the rapid screening results serve as the data basis, which facilitates subsequent model construction. At the same time, the results of rapid screening and offline analysis can be mutually verified to ensure the accuracy of the collected information.

[0025] S2. Extract biologically significant features from the sensor-collected data, construct a sensor feature time series, define the output microbial state, train a complete prediction model using machine learning algorithms, and validate the model using fermentation batch data from the validation set. S21. For the sensor data in the mapping relation database, extract biological significance features from the sensor data to obtain primary derived features and secondary statistical features. At the same time, define the output microbial state, including yeast activity ratio, bacterial contamination risk index, total active biomass and community state classification. S211. Extract the original feature vector from the timestamp synchronized in the mapping relationship database, and construct the first-level derived features, including oxygen uptake rate, carbon dioxide release rate, respiratory entropy, dissolved oxygen gradient, pH change rate, and temperature change rate. Among them oxygen uptake rate , This represents the derivative of dissolved oxygen concentration with respect to time. Represents dissolved oxygen concentration. Represents the volumetric oxygen transfer coefficient. Represents the saturated dissolved oxygen concentration; It should be noted that the derivative of dissolved oxygen concentration with respect to time is the instantaneous rate of change of DO concentration. The unit of dissolved oxygen concentration is usually mg / L, which refers to the actual oxygen concentration measured in the solution. The volumetric oxygen transfer coefficient is a key parameter reflecting the aeration and stirring efficiency of the reactor, which is used to represent the rate at which oxygen is transferred from the gas phase to the liquid phase. The unit of saturated dissolved oxygen concentration is the same as [DO], which refers to the theoretical maximum concentration of dissolved oxygen that can be achieved in the solution under the current temperature, pressure and culture medium composition. The oxygen uptake rate represents the amount of oxygen consumed by cells per unit time and per unit volume of culture medium.

[0026] Among them, carbon dioxide release rate ,in These are the volumetric flow rates of exhaust gas and intake gas, respectively. These represent the carbon dioxide concentration in the exhaust gas and the intake gas, respectively, and the respiratory entropy. ; Dissolved oxygen gradient , These represent the dissolved oxygen concentrations measured at the upper and lower parts of the fermenter at time t, respectively. It should be noted that the pH change rate and temperature change rate are the derivative of pH with respect to time, i.e., the instantaneous rate of change of pH, and the derivative of temperature with respect to time, i.e., the instantaneous rate of change of temperature, respectively. The unit of dissolved oxygen gradient is the same as that of DO. It is used to reflect whether the mixing in the fermenter is uniform. A large gradient indicates that there is a concentration dead zone due to poor mixing. The RQ value is different under different metabolic pathways. For example, RQ>1 for yeast pure alcohol fermentation, RQ≈1 for complete oxidation, and RQ<1 for bacterial oxidation.

[0027] S212. Calculate the window mean, window variance, and window extreme values ​​within the sliding window time W, where the window time... ,in The width of the time window is represented by the total input feature vector X(t) of the model, which is the concatenation of all the above features. It should be noted that, The width of the time window needs to be set in conjunction with the specific brewing process.

[0028] S213. Define the output microbial state, including the yeast activity ratio. , These represent the concentration of target yeast live cells in the fermentation broth at time t and the total concentration of live cells of all microorganisms in the fermentation broth at time t, respectively. It should be noted that the concentration of target live yeast cells in the fermentation broth was analyzed and sorted on a flow cytometer using flow cytometry combined with fluorescent probes in rapid screening and offline analysis. Fluorescently labeled antibodies or dyes specific to the yeast cell wall or internal structure were used, along with live / dead cell dyes.

[0029] Bacterial contamination risk index ,in , This represents the concentration of lactic acid bacteria and acetic acid bacteria in the fermentation broth at time t. b represents the weighting coefficient, and b is the bias term; It should be noted that, The weighting coefficient represents the relative importance of different contaminating bacteria in contributing to the final risk. It needs to be set based on the needs of different brewing processes. For example, in the current brewing process, the impact of lactic acid bacteria producing acid and causing flavor degradation is greater than that of acetic acid bacteria. Set greater than It is usually obtained by fitting a dataset of historical pollution events together with a bias term using machine learning methods such as logistic regression.

[0030] Total active biomass and Maintaining consistency, community state classification , representing the community state label of the fermentation system at time t. K represents the set of all states, where K is the total number of state categories.

[0031] It should be noted that the classification of community status needs to be defined by consulting experts in the field. For example, L=1 represents "the healthy fermentation period dominated by yeast", L=2 represents "the potential acidification period with active lactic acid bacteria", etc. The labels are manually labeled on each sample in the database by experts based on a comprehensive judgment of multi-source offline data such as qPCR absolute quantification, sequencing relative abundance, and organic acid concentration.

[0032] S22. Based on primary derived features, secondary statistical features, and microbial state, the complete fermentation batch is divided into data points, and a complete prediction model is trained using machine learning algorithms. At the same time, the model is validated using fermentation batch data from the validation set. S221. Divide the data according to the complete fermentation batch, where the dataset is: The training set is The validation set is Where U represents the number of reserved batches; It should be noted that the time series length of each fermentation batch may be different, so uniform processing is required. By setting a fixed sequence length, excessively long sequences are truncated, and excessively short sequences are filled with specific values. Interpolation is used to ensure that all batches have data points at the same time interval. At the same time, the sensor data is standardized and preprocessed, and statistics from the training set, such as the mean and standard deviation, are used to normalize all data to avoid data leakage.

[0033] S222. A regression prediction model is constructed using LSTM to predict the yeast activity ratio, bacterial contamination risk index, and total active biomass. A classification prediction model is constructed using the random forest method to predict community state classification. In the LSTM model construction process, the input features include primary derived features and original online sensor parameters. The labels are the microbial states obtained from offline analysis. The mean squared error is used as the loss function, and the Adam optimizer is used for optimization. The root mean square error and the coefficient of determination are used as evaluation indicators. In the random forest classification prediction model construction process, secondary statistical features need to be added to the model input and the cross-entropy loss function is used. The feature matrix and labels of the training set are directly input into the model for fitting.

[0034] It should be noted that during the LSTM model training process, the training set data is input into the model for multiple iterations. After each training cycle, the validation loss is calculated on the validation set. When the validation loss no longer decreases for several consecutive cycles, training is stopped, and the model weights with the minimum validation loss are rolled back to prevent overfitting. The validation set is used to evaluate the effect of different hyperparameter combinations. In the process of building the classification model using random forest, the hyperparameters are pre-set, including the number of trees, the maximum tree depth, and the maximum number of features considered for optimal splitting. The feature matrix and labels of the training set are directly input into the model for fitting. k-fold cross-validation is used on the training set, and hyperparameters are tuned using the validation set. Since the data is divided into batches, it is necessary to ensure that the batches are not shuffled during cross-validation. The tuned model is used to predict on the validation set, and preliminary evaluation metrics are calculated, including accuracy, precision, recall, and F1 score.

[0035] Example 2:

[0036] S3. Evaluate the contribution of each sensor feature to the prediction of biological state, build a simplified prediction model using the selected core sensor features, and compare the prediction accuracy of the complete prediction model and the simplified prediction model on an independent validation dataset to determine the final combination of sensors to be deployed online for dynamic monitoring of the microbial community in the fermenter. S31. Perform feature importance analysis on the complete prediction model, evaluate the contribution of each sensor feature to the predicted biological state, and use the selected core sensor feature subset as input to train the simplified model. S311. For the classification prediction model built on random forest, the total information gain of each feature in all decision trees is evaluated by calculating the Gini importance of the tree model. This includes the following steps: ; in The importance of the Gini symbol representing feature j. The total number of trees is represented by T, a single tree is represented by g, and a split node in the tree is represented by g. The reduction in Gini impurity of feature j when it splits at node g is represented by the amount of Gini impurity reduction. At the same time, the baseline performance of the model is calculated on the validation set. All values ​​of feature j are randomly shuffled to form a perturbation dataset. The perturbation performance is calculated, and the difference between the baseline performance and the perturbation performance is taken as the importance of feature j. The importance is obtained by repeating R times and taking the average. It should be noted that the baseline performance includes accuracy, precision, recall, etc. The greater the loss, the more important the feature is. The greater the reduction in Gini importance, the more important the feature is. The value of R is usually set to 15, but it can also be adjusted according to the actual situation.

[0037] S312. For the regression prediction model built using LSTM, the SHAP value of feature j is calculated to determine the contribution of different features to the predicted biological state. This includes the following steps: ; in, The SHAP value represents feature j, F represents the entire feature set, and S represents a subset of features. This represents the model's prediction of sample x when the input only contains features of a subset S. The model prediction value after adding target feature j to the representative feature subset S is given for all samples. Take the absolute value and then average it to obtain the global importance of feature j. Select the core features from the permutation importance, Gini importance and global importance according to the preset threshold, and select the top P features to obtain the sensor type under the corresponding top P core features. It should be noted that the preset thresholds for permutation importance, Gini importance, and global importance need to be set based on the actual needs of the brewing process, after consulting with experts in the relevant field and using an empirical method. The value of P needs to be set based on the specific situation, or all features that meet the threshold can be used as core features.

[0038] S313. Select a subset of core features from the original feature set, and repeat step S2 to train a simplified prediction model using the same algorithm structure.

[0039] S32. On an independent validation dataset, compare the prediction accuracy of the complete prediction model and the simplified prediction model to determine the final sensor combination for online deployment and to dynamically monitor the microbial community in the fermenter. S321. On z samples of the independent verification batch, prediction is performed using the complete prediction model and the simplified prediction model respectively. The prediction accuracy parameter is compared. If the prediction accuracy is within the preset threshold range, it means that the current simplified model meets the requirements. The sensor type of the feature corresponding to the simplified model is used as the final sensor combination for online deployment. If the model accuracy does not meet the requirements, repeat steps S311-S313 to expand the range of P until a final sensor combination for online deployment that meets the conditions is obtained. S322. Real-time parameters collected by a combination of parameters from online-deployed sensors are combined with a simplified prediction model to output the microbial state.

[0040] It should be noted that the prediction accuracy for regression tasks includes root mean square error, mean absolute error, and coefficient of determination. For classification models, the prediction accuracy includes precision, accuracy, recall, and F1 score. The percentage changes of different precision parameters between the complete and simplified prediction models are calculated, and a preset threshold is used to determine whether the requirements are met. The preset error threshold is usually set to 5%, but it can be adjusted according to the actual application requirements. The model prediction values ​​are compared and calibrated once every quarter or every six months using the results of offline sampling analysis.

[0041] In summary, this invention combines offline modeling with online prediction using conventional sensor data. It indirectly and in real-time measures microbial state through a model, rapidly screening and analyzing sensor parameters from multiple batches of samples collected from different sites in fermenters to construct a complete prediction model for different fermentation stages. This allows for dynamic prediction of the microbial community in the fermenter based on sensor data. Furthermore, by analyzing the contribution of different parameters in the complete prediction model, important features are selected to construct a simplified prediction model. This model is then linked to an online sensor combination to create a simplified model suitable for industrial online deployment. By establishing a dynamic mapping relationship between conventional sensor signals and the internal state of microorganisms, it avoids the lag in microbial community state results obtained through traditional offline sampling and laboratory analysis, enabling real-time and continuous prediction of microbial state. Simultaneously, by simplifying the sensor combination, it reduces practical application costs and improves usability.

[0042] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for monitoring the dynamics of a native brewing yeast fermentation microbial community, characterized in that, The method includes the following steps: S1. Install sterile sampling valves and sensors at key sites of the fermenter, sample and collect data at key sites of the fermenter at different fermentation stages, and conduct rapid screening and offline analysis on the acquired samples and data to build a basic database of the mapping relationship between sensor signals and microbial community status during the brewing fermentation process. S2. Extract biologically significant features from the sensor-collected data, construct a sensor feature time series, define the output microbial state, train a complete prediction model using machine learning algorithms, and validate the model using fermentation batch data from the validation set. S3. Evaluate the contribution of each sensor feature to the prediction of biological state, construct a simplified prediction model using the selected core sensor features, and compare the prediction accuracy of the complete prediction model and the simplified prediction model on an independent validation dataset to determine the final combination of sensors to be deployed online for dynamic monitoring of the microbial community in the fermenter.

2. The method for monitoring the dynamics of a native brewing yeast fermentation microbial community according to claim 1, characterized in that, S1 includes the following steps: S11. Aseptic sampling valves and sensors are installed at key sites in the current fermenter. These key sites include the gas phase zone at the top of the tank, the upper liquid phase, the middle liquid phase, the lower liquid phase, and the biofilm on the tank wall. Sampling and data collection are performed at different fermentation stages in the fermenter using aseptic sampling valves and sensors. These fermentation stages include the lag phase, logarithmic phase, stationary phase, and decay phase. The fermenter is spatially discretized into five key ecological sites: the gas phase zone, the upper liquid phase, the middle liquid phase, the lower liquid phase, and the attached phase. Sensor groups are deployed at each liquid phase site to record signal vectors in real time and synchronize sensor timestamps with a hardware clock. A unique code is assigned to each sample. S12. For samples and data obtained at different key sites and fermentation stages, rapid screening and offline analysis are performed after sampling to obtain the types, quantities, and community structure of microorganisms in the collected samples. At the same time, the glucose, ethanol, lactic acid, and acetic acid contents of the samples are measured. Data from N batches of normally completed brewing fermentation and N / 3 batches of fermentation with abnormalities are collected to construct a basic database of the mapping relationship between sensor signals and microbial community status during brewing fermentation.

3. The method for monitoring the dynamics of a native brewing yeast fermentation microbial community according to claim 1, characterized in that, S12 includes the following steps: S121. During rapid screening, based on Raman spectroscopy, principal component analysis is used to reduce the dimensionality and identify patterns of the single-cell spectra collected by sampling. Based on the principal component scores, K-means cluster analysis is used to count the cell proportions of various microorganisms. The total cell concentration and the proportion of live cells are obtained through fluorescence staining. S122. During offline analysis, absolute quantification by qPCR is used to output the gene copy number concentration of the target microorganism. At the same time, high-throughput sequencing analysis is combined to obtain the Shannon diversity index and Bray-Curtis dissimilarity of the sample. High-performance liquid chromatography is used to determine the concentration vector of key metabolites, including glucose, ethanol, lactic acid, and acetic acid. All heterogeneous data are integrated into a unified feature-tag pair. Each row of the main table corresponds to a time-site sample, including key site number, sensor characteristics, and microbial state. The association table records the Shannon diversity index, Bray-Curtis dissimilarity, and metabolite concentration data. Multiple batches of data are merged and integrated into a basic database of mapping relationships.

4. The method for monitoring the dynamics of a native brewing yeast fermentation microbial community according to claim 3, characterized in that, S2 includes the following steps: S21. For the sensor data in the mapping relation database, extract biological significance features from the sensor data to obtain primary derived features and secondary statistical features. At the same time, define the output microbial state, including yeast activity ratio, bacterial contamination risk index, total active biomass and community state classification. S22. Based on primary derived features, secondary statistical features, and microbial state, the complete fermentation batch is divided into data points, and a complete prediction model is trained using machine learning algorithms. At the same time, the model is validated using fermentation batch data from the validation set.

5. The method for monitoring the dynamics of a native brewing yeast fermentation microbial community according to claim 4, characterized in that, S21 includes the following steps: S211. Extract the original feature vector from the timestamp synchronized in the mapping relationship database, and construct the first-level derived features, including oxygen uptake rate, carbon dioxide release rate, respiratory entropy, dissolved oxygen gradient, pH change rate, and temperature change rate. Among them oxygen uptake rate , This represents the derivative of dissolved oxygen concentration with respect to time. Represents dissolved oxygen concentration. Represents the volumetric oxygen transfer coefficient. Represents the saturated dissolved oxygen concentration; Among them, carbon dioxide release rate ,in These are the volumetric flow rates of exhaust gas and intake gas, respectively. These represent the carbon dioxide concentration in the exhaust gas and the intake gas, respectively, and the respiratory entropy. ; Dissolved oxygen gradient , These represent the dissolved oxygen concentrations measured at the top and bottom of the fermenter at time t, respectively. S212. Calculate the window mean, window variance, and window extreme values ​​within the sliding window time W, where the window time... ,in The width of the time window is represented by the total input feature vector X(t) of the model, which is the concatenation of all the above features. S213. Define the output microbial state, including the yeast activity ratio. , These represent the concentration of target yeast live cells in the fermentation broth at time t and the total concentration of live cells of all microorganisms in the fermentation broth at time t, respectively. Bacterial contamination risk index ,in , This represents the concentration of lactic acid bacteria and acetic acid bacteria in the fermentation broth at time t. b represents the weighting coefficient, and b is the bias term; Total active biomass and Maintaining consistency, community state classification , representing the community state label of the fermentation system at time t. K represents the set of all states, where K is the total number of state categories.

6. The method for monitoring the dynamics of a native brewing yeast fermentation microbial community according to claim 5, characterized in that, S22 includes the following steps: S221. Divide the data according to the complete fermentation batch, where the dataset is: The training set is The validation set is Where U represents the number of reserved batches; S222. A regression prediction model is constructed using LSTM to predict the yeast activity ratio, bacterial contamination risk index, and total active biomass. A classification prediction model is constructed using the random forest method to predict community state classification. In the LSTM model construction process, the input features include primary derived features and original online sensor parameters. The labels are the microbial states obtained from offline analysis. The mean squared error is used as the loss function, and the Adam optimizer is used for optimization. The root mean square error and the coefficient of determination are used as evaluation indicators. In the random forest classification prediction model construction process, secondary statistical features need to be added to the model input and the cross-entropy loss function is used. The feature matrix and labels of the training set are directly input into the model for fitting.

7. The method for monitoring the dynamics of a native brewing yeast fermentation microbial community according to claim 6, characterized in that, S3 includes the following steps: S31. Perform feature importance analysis on the complete prediction model, evaluate the contribution of each sensor feature to the predicted biological state, and use the selected core sensor feature subset as input to train the simplified model. S32. On an independent validation dataset, compare the prediction accuracy of the complete prediction model and the simplified prediction model to determine the final sensor combination for online deployment and to dynamically monitor the microbial community in the fermenter.

8. The method for monitoring the dynamics of a native brewing yeast fermentation microbial community according to claim 1, characterized in that, S31 includes the following steps: S311. For the classification prediction model built for random forests, the total information gain of each feature in all decision trees is evaluated by calculating the Gini importance of the tree model. This includes the following steps: ； in The importance of the Gini symbol representing feature j. The total number of trees is represented by T, where T represents a single tree and g represents a split node in the tree. The reduction in Gini impurity of feature j when it splits at node g is represented by the amount of Gini impurity reduction. At the same time, the baseline performance of the model is calculated on the validation set. All values ​​of feature j are randomly shuffled to form a perturbation dataset. The perturbation performance is calculated, and the difference between the baseline performance and the perturbation performance is taken as the importance of feature j. The importance is obtained by repeating R times and taking the average. S312. For the regression prediction model built using LSTM, the SHAP value of feature j is calculated to determine the contribution of different features to the predicted biological state. This includes the following steps: ； in, The SHAP value represents feature j, F represents the entire feature set, and S represents a subset of features. This represents the model's prediction of sample x when the input only contains features of a subset S. The model prediction value after adding target feature j to the representative feature subset S is given for all samples. Take the absolute value and then average it to obtain the global importance of feature j. Select the core features from the permutation importance, Gini importance and global importance according to the preset threshold, and select the top P features to obtain the sensor type under the corresponding top P core features. S313. Select a subset of core features from the original feature set, and repeat step S2 to train a simplified prediction model using the same algorithm structure.

9. The method for monitoring the dynamics of a native brewing yeast fermentation microbial community according to claim 8, characterized in that, S32 includes the following steps: S321. On z samples of the independent verification batch, prediction is performed using the complete prediction model and the simplified prediction model respectively. The prediction accuracy parameter is compared. If the prediction accuracy is within the preset threshold range, it means that the current simplified model meets the requirements. The sensor type of the feature corresponding to the simplified model is used as the final sensor combination for online deployment. If the model accuracy does not meet the requirements, repeat steps S311-S313 to expand the range of P until a final sensor combination for online deployment that meets the conditions is obtained. S322. Real-time parameters collected by a combination of parameters from online-deployed sensors are combined with a simplified prediction model to output the microbial state.

Images