Intelligent prediction and optimization method for microbial fermentation process based on machine learning

Abstract

The application discloses a method for intelligent prediction and optimization of a microbial fermentation process based on machine learning, which comprises the following steps: acquiring real-time multi-dimensional state data and high-dimensional spectrum data of a fermentation liquor, and performing noise elimination processing on the multi-dimensional state data and the high-dimensional spectrum data; extracting a feature vector related to a fermentation stage, and performing time series modeling on the feature vector and a state sequence to capture a dynamic evolution mode of the fermentation process; adjusting a control variable of the fermentation process to generate an optimization instruction sequence; constructing a real-time optimized fermentation state model through the optimization instruction sequence, iteratively updating subsequent data, judging whether a model error is lower than a preset threshold value, and if the model error is lower than the preset threshold value, determining a final yield prediction value and an optimization path. The application can extract and integrate multi-dimensional physiological characteristics and high-dimensional spectrum information in a fermentation process in real time, accurately identify metabolic states in different stages, and screen out key parameters that are strongly related to product synthesis.

Description

Technical Field

[0001] This invention relates to the field of microbial fermentation technology, and in particular to a method for intelligent prediction and optimization of microbial fermentation processes based on machine learning. Background Technology

[0002] Microbial fermentation is a core component of the biomanufacturing industry, directly determining the yield and quality of important products such as antibiotics, amino acids, and enzyme preparations. Accurate prediction and optimization of this process are crucial for improving production efficiency, reducing energy consumption, and ensuring product quality.

[0003] Most current fermentation control methods rely on offline detection and empirical adjustments after manual sampling, making it difficult to capture metabolic changes in real time during production. This results in delayed control and unstable product accumulation.

[0004] Fermentation broth contains a large number of metabolites, the concentration of which changes dynamically over time. At the same time, cell morphology and culture medium composition are also constantly evolving, forming a highly complex multidimensional state space.

[0005] Existing methods often struggle to effectively integrate this multidimensional information, especially lacking the ability to automatically analyze high-dimensional spectra such as chromatograms and mass spectrometry data. This results in incomplete feature extraction and makes it difficult to accurately reveal the correlation between key physiological parameters and product synthesis.

[0006] What makes things more complicated is that these spectral data are not only high in dimensionality and noise, but also have subtle differences in spectral patterns at different fermentation stages, making them susceptible to minor perturbations and resulting in poor discriminative consistency within or between batches.

[0007] This tight coupling between multidimensional features and high-dimensional spectra further amplifies the difficulty of feature selection.

[0008] Since metabolite concentration, cell morphology, and culture medium composition together constitute the physiological basis of product synthesis, if it is not possible to accurately screen out the key parameters that are truly closely related to yield from massive amounts of data, it will be difficult to establish a reliable feature-yield correlation, thus making it impossible to achieve timely identification and real-time optimization of fermentation status.

[0009] For example, in actual production, when fermentation enters the plateau phase, the intensity of some peaks on the spectrum may only shift slightly, but this may indicate a rapid decline in the product synthesis rate. However, existing methods often only detect this change in post-event detection, which means that the best intervention window has been missed.

[0010] Therefore, how to extract and integrate multidimensional physiological characteristics and high-dimensional spectral information in real time during fermentation, accurately identify the metabolic state at different stages, and screen out key parameters that are strongly correlated with product synthesis to support real-time monitoring and optimization has become a key issue in intelligent prediction and optimization of microbial fermentation processes based on machine learning. Summary of the Invention

[0011] To address the shortcomings of existing technologies, this invention provides a machine learning-based intelligent prediction and optimization method for microbial fermentation processes. This method can extract and integrate multidimensional physiological features and high-dimensional spectral information in real time during fermentation, accurately identify the metabolic state at different stages, and screen out key parameters that are strongly correlated with product synthesis.

[0012] To achieve the above objectives, the technical solution of the present invention is as follows: A machine learning-based intelligent prediction and optimization method for microbial fermentation processes, comprising: Real-time multidimensional state data and high-dimensional spectral data of fermentation broth are acquired by sensors. Noise removal processing is performed on the multidimensional state data and high-dimensional spectral data to obtain clean state sequences and spectral sequences. Based on the clean state sequence and spectral sequence, feature vectors related to the fermentation stage are extracted. The feature vectors and state sequence are then used for time-series modeling to capture the dynamic evolution pattern of the fermentation process and determine the predicted trend of product synthesis rate. Based on the predicted trends, the control variables of the fermentation process are adjusted to generate an optimized instruction sequence; A real-time optimized fermentation state model is constructed by optimizing the instruction sequence. Subsequent data is iteratively updated to determine whether the model error is lower than a preset threshold. If it is lower than the preset threshold, the final yield prediction and optimization path are determined.

[0013] Preferably, the step of acquiring real-time multidimensional state data and high-dimensional spectral data of the fermentation broth through sensors, and performing noise reduction processing on the multidimensional state data and high-dimensional spectral data, includes: Real-time data from the fermentation broth is collected using sensors to obtain an initial multidimensional state dataset and a high-dimensional spectral dataset. A filtering method was used to remove noise from the initial multidimensional state dataset and high-dimensional spectrogram dataset to obtain a clean state sequence and spectrogram sequence. For pure state sequences and spectral sequences, a preset threshold is applied to detect outliers. If a data point exceeds the threshold range, it is marked as an outlier. Based on the labeling results, abnormal data points and normal data points are separated to obtain the separated normal state sequence and normal spectrum sequence; By dividing the normal state sequence and the normal spectral sequence into time windows, local data features are obtained, and the set of local features is determined.

[0014] Preferably, the step of extracting feature vectors related to the fermentation stage based on the clean state sequence and spectral sequence includes: The clean spectral sequence is standardized to obtain the processed spectral data, and a convolutional neural network is used to extract deep features from the processed spectral data to obtain a preliminary set of spectral features. If the peak signal intensity of metabolite concentration in the preliminary spectral feature set is lower than a preset threshold, then the data in this part is subjected to signal enhancement processing to determine the enhanced peak concentration data. By extracting pattern information related to bacterial cell morphology from the enhanced concentration peak data, a subset of morphological pattern features is obtained. Based on the subset of morphological feature data and concentration peak data, a spectral feature vector associated with the fermentation stage is constructed.

[0015] Preferably, the step of performing time-series modeling of feature vectors and state sequences to capture the dynamic evolution pattern of the fermentation process includes: Physiological parameter data from feature vectors and state sequences are obtained, and a unified input dataset is obtained through standardization. A recurrent neural network is used to perform time series modeling on the input dataset, analyze dynamic changes, and determine the feature representations at each time point. Based on the feature representation, dynamic change patterns of different fermentation stages are extracted to determine the stage evolution law. If the stage evolution law does not match the preset threshold range, the data is divided into additional time windows to obtain fine-grained change patterns. A time-series prediction model is constructed through the fine-grained change patterns to determine the preliminary trend estimate of the product synthesis rate.

[0016] Preferably, the step of adjusting the control variables of the fermentation process and generating an optimized instruction sequence based on the predicted trend includes: Real-time fermentation process data, including product synthesis rate and regulatory variable data, are acquired. A time series prediction model is used to make short-term predictions of the product synthesis rate and obtain the prediction trend. Determine the direction of change in the predicted trend over a future period. If the predicted trend shows a downward trend, mark it as a downward trend. Data on regulatory variables are extracted from fermentation process data. The historical correlation between the regulatory variables and the product synthesis rate is calculated to determine the main influencing variables. Based on the main influencing variables, the corresponding optimization instruction sequence is generated.

[0017] Preferably, the step of constructing a real-time optimized fermentation state model by optimizing the instruction sequence includes: Filter instruction fragments related to yield from the optimized instruction sequence, obtain the corresponding physiological parameter data, compare the physiological parameter data with historical records, and determine the key physiological indicators related to yield. Based on key physiological indicators, the support vector machine algorithm is used to classify the real-time fermentation status data and determine whether the current state deviates from the optimization range. If the current state deviates from the optimization range, the adjustment strategy is extracted and the correction instruction is obtained. By fusing the modified instructions with the original instruction sequence, the final optimized instruction combination is determined, and a fermentation state model is constructed based on the optimized instruction combination.

[0018] Preferably, the step of iteratively updating subsequent data and determining whether the model error is lower than a preset threshold includes: The subsequent high-dimensional spectral data is processed by the fermentation state model to obtain the updated feature dataset. The error of the feature dataset is evaluated, and the deviation between the predicted value and the actual value is calculated. If the deviation is higher than the preset threshold, the data update process is triggered to obtain new high-dimensional spectral data; if the deviation is lower than the preset threshold, the yield prediction result is generated based on the feature dataset to determine the range of the prediction value. Based on the predicted value range, the optimized path of the fermentation process is calculated, and the adjusted operating parameters are obtained as guidance.

[0019] Preferably, determining the final production forecast and optimization path includes: Based on the output of the fermentation state model, obtain the state prediction value, determine whether the fermentation process has reached the expected goal, and if the state prediction value has not reached the expected goal, extract the backup plan, integrate new physiological parameter data, and update the fermentation state model. The fermentation state is monitored by the updated fermentation state model to obtain real-time optimized state data. Based on the real-time optimized state data, the final yield prediction value is determined. By using yield forecasts and status data, an optimized path guide for the fermentation process is generated. Based on the optimized path guide, the data input in the fermentation status monitoring framework is updated to complete the iterative analysis.

[0020] The technical effects and advantages of this invention are as follows: The intelligent prediction and optimization method for microbial fermentation processes based on machine learning provided in this application addresses complex business scenarios involving changes in metabolic state, prediction of product synthesis rate, and optimization of regulatory variables during fermentation. It acquires multidimensional state data and high-dimensional spectral data in real time, removes noise using filtering methods, extracts spectral feature vectors using convolutional neural networks, determines changes in metabolic state, and integrates culture medium component data to identify key physiological parameters. Then, it further utilizes recurrent neural networks for time-series modeling to predict product synthesis rate trends and generates optimization instruction sequences by adjusting regulatory variables such as temperature and oxygen supply. Finally, it constructs a real-time optimized fermentation state model and iteratively updates data to achieve yield prediction and path optimization.

[0021] This invention significantly improves the accuracy of dynamic monitoring and control of the fermentation process by integrating multidimensional data and deep learning technology, effectively resolving the contradiction between metabolic state fluctuations and yield optimization, and achieving a dual improvement in fermentation efficiency and product yield. Attached Figure Description

[0022] Figure 1 This is a flowchart of the microbial fermentation process prediction and optimization of the present invention. Detailed Implementation

[0023] This invention provides a machine learning-based intelligent prediction and optimization method for microbial fermentation processes, aiming to achieve efficient monitoring and yield improvement of the fermentation process through real-time data acquisition, feature extraction, dynamic modeling, and optimization adjustments. The technical solution of this invention is described in detail below with reference to specific embodiments to make the objectives, technical solutions, and advantages of this invention clearer.

[0024] See Figure 1 As shown, this invention provides an intelligent prediction and optimization method for microbial fermentation processes based on machine learning, applicable to industrial fermentation scenarios such as lactic acid bacteria fermentation and yeast fermentation. This method captures dynamic changes during the fermentation process through multi-dimensional data acquisition and intelligent analysis, thereby optimizing control parameters and improving product synthesis efficiency. The overall process includes steps such as data acquisition, noise processing, feature extraction, time series modeling, trend prediction, and optimization adjustment. These steps will be described in detail below to ensure the feasibility of the technical solution.

[0025] Step S1 involves acquiring real-time multidimensional state data and high-dimensional spectral data of the fermentation broth using sensors. Filtering techniques are then employed to remove noise from the multidimensional state data and high-dimensional spectral data, resulting in clean state and spectral sequences. This step collects raw data from the fermentation environment and removes interfering factors through preprocessing, laying the foundation for subsequent analysis. The state data of the fermentation broth typically includes physicochemical parameters such as temperature, pH, and dissolved oxygen concentration, while the high-dimensional spectral data involves the spectral signals of metabolites in the fermentation broth. Together, these data reflect the real-time state of the fermentation process.

[0026] Specifically, various sensor devices can be deployed in industrial fermenters, such as temperature probes, pH electrodes, dissolved oxygen sensors, and near-infrared spectrometers, to collect multidimensional state data and high-dimensional spectral data of the fermentation broth in real time. Sensors are installed at different locations within the fermenter to ensure data coverage of the entire fermentation broth's state. For example, temperature probes can be positioned at the bottom and middle of the fermenter to monitor temperature differences in different areas, while the spectrometer uses a fiber optic probe to directly contact the fermentation broth and acquire spectral signals of metabolite concentrations. The acquisition frequency can be adjusted according to the dynamic characteristics of the fermentation process, typically set to acquire data once per minute to ensure data continuity and real-time performance.

[0027] The collected multidimensional state data and high-dimensional spectral data are inevitably affected by environmental noise, equipment vibration, or impurities in the fermentation broth, thus requiring noise reduction processing. Filtering techniques are commonly used data preprocessing methods that can effectively remove high-frequency noise and baseline drift. For example, for temperature and pH data, a moving average filtering method can be used to smooth data fluctuations by calculating the average value within a certain time window; while for high-dimensional spectral data, wavelet transform filtering can be used to decompose different frequency components of the signal, remove irrelevant noise, and retain feature signals related to metabolite concentration. After filtering, clean state sequences and spectral sequences are obtained, providing a reliable data foundation for subsequent feature extraction.

[0028] In one embodiment, step S1 is implemented as follows: Step S11 involves acquiring real-time data from the fermentation broth using sensors to obtain an initial multidimensional state dataset and a high-dimensional spectral dataset. This step marks the beginning of data acquisition and aims to extract raw information from the fermentation environment. The sensor equipment must possess high precision and stability to ensure the accuracy of the acquired data. For example, in lactic acid bacteria fermentation, a temperature sensor can be accurate to 0.1 degrees Celsius, the pH electrode's measurement range covers the common acidic environment of the fermentation broth, and the spectrometer needs sufficient resolution to distinguish the spectral peaks of different metabolites. The acquired data is indexed by timestamps to form the initial multidimensional state dataset and high-dimensional spectral dataset.

[0029] In one embodiment, assuming lactic acid bacteria fermentation is carried out in a 500-liter industrial fermenter, a temperature sensor records the temperature value at the bottom of the fermentation broth every minute, obtaining a set of temperature data sequences that change over time. Simultaneously, a pH electrode monitors the acidity changes of the fermentation broth, recording pH readings every minute. Furthermore, a near-infrared spectrometer acquires the spectral signal of the fermentation broth through a probe, generating high-dimensional spectral data containing multiple wavelength channels. These data collectively constitute the initial multidimensional state dataset and high-dimensional spectral dataset, providing rich raw information for subsequent processing.

[0030] Step S12 involves using filtering methods to remove noise from the initial multidimensional state dataset and high-dimensional spectral dataset, resulting in a clean state sequence and spectral sequence. The purpose of this step is to remove interference signals from the data and improve data quality. Noise may originate from stirring vibrations within the fermenter, measurement errors of the sensors themselves, or fluctuations in ambient temperature. Different filtering methods can be selected for different types of data. For example, for temperature and dissolved oxygen data, moving average filtering can effectively smooth short-term fluctuations. The window size can be set to 5 minutes, meaning the average value within 2.5 minutes before and after the current time point is calculated as the smoothing result. For high-dimensional spectral data, the noise removal process is more complex. The spectral signal may contain baseline drift and random noise, affecting the identification of metabolite concentration peaks. Therefore, wavelet transform filtering can be used to decompose the spectral signal into components of different frequencies, removing high-frequency noise, while simultaneously eliminating low-frequency drift through baseline correction. After processing, the main peaks in the spectral signal are preserved, forming a clean spectral sequence. For example, in lactic acid bacteria fermentation, the spectral peak of lactic acid concentration usually appears in a specific wavelength range. By filtering, this signal can be clearly separated, eliminating interference from impurities or solvent background.

[0031] Step S13: For the pure state sequence and spectral sequence, apply a preset threshold to detect outliers. If a data point exceeds the threshold range, it is marked as an outlier. This step aims to identify outliers in the data to ensure the reliability of subsequent analysis. Outliers may be caused by sensor malfunction, local turbulence in the fermentation broth, or human error. The preset threshold can be determined based on historical data statistics. For example, the threshold range for temperature data can be set to the normal fluctuation range of the fermentation process, such as 20 to 40 degrees Celsius. If the temperature reading at a certain time point exceeds this range, it is marked as an outlier.

[0032] In one possible implementation, outlier detection can be further confirmed using statistical methods. For example, the standard deviation method can be used to calculate the mean and standard deviation of the state sequence within a certain time window. If the value of a data point deviates from the mean by more than three times the standard deviation, it is identified as an outlier. For spectral sequences, thresholds can be set for signal intensity at specific wavelengths. If the signal intensity at a certain time point is significantly higher or lower than the normal range, it is marked as an anomaly. For example, in lactic acid bacteria fermentation, if the lactic acid concentration signal suddenly spikes to more than 10 times the historical data at a certain moment, it may be due to a measurement error caused by the spectrometer probe being blocked by air bubbles, and should be marked as an outlier and removed.

[0033] Step S14 involves separating abnormal and normal data points based on the labeling results to obtain the separated normal state sequence and normal spectral sequence. This step further processes the outlier detection results, aiming to remove abnormal data and retain reliable data sequences. The separation process can classify data points by timestamps, removing data points marked as abnormal from the original sequence, and the remaining data points form the normal state sequence and normal spectral sequence. It should be noted that removing abnormal data points may cause temporal discontinuities in the data sequence. In this case, linear interpolation can be used to fill in missing values ​​to ensure the integrity of the sequence. For example, in the monitoring of lactic acid bacteria fermentation, if temperature data is continuously abnormal for a certain period due to sensor failure, this data segment can be removed, and linear interpolation can be performed based on the temperature values ​​of the preceding and following time points to estimate the temperature readings at the missing time points, forming a continuous normal state sequence. Similarly, for the spectral sequence, if the spectral signal at a certain moment is marked as abnormal, the data at that time point is removed, and interpolation is performed using the signal values ​​of the preceding and following time points to generate a continuous normal spectral sequence. These processes ensure that the input data for subsequent feature extraction and modeling has high reliability.

[0034] Step S15 involves dividing the normal state sequence and normal spectral sequence into time windows to obtain local data features and determine the local feature set. The purpose of this step is to divide the continuous data sequence into multiple time periods and extract local features within each time period to capture short-term change patterns in the fermentation process. The size of the time window can be adjusted according to the dynamic characteristics of the fermentation process. For example, for rapidly changing lactic acid bacteria fermentation, the window size can be set to 30 minutes, while for slower yeast fermentation, the window size can be extended to 2 hours.

[0035] In one embodiment, after dividing the time window, statistical characteristics such as mean, variance, maximum, and minimum values ​​can be calculated for the data within each window, serving as local data features. For temperature data in the state sequence, the average temperature and temperature fluctuation amplitude within each window can be calculated, reflecting the thermal stability of the fermentation broth; for the spectral sequence, the average intensity and trend of specific wavelength signals within each window can be extracted, characterizing short-term changes in metabolite concentrations. These local features collectively constitute a set of local features, providing fundamental information for subsequent feature extraction related to fermentation stages. In this way, the overall dynamics of the fermentation process can be decomposed into multiple local patterns, facilitating a more detailed analysis of the changing patterns of the fermentation state.

[0036] Step S2 involves extracting feature vectors related to the fermentation stages based on the clean state sequence and spectral sequence. This step aims to extract key features characterizing different stages of the fermentation process from the preprocessed data, providing input for subsequent dynamic modeling and trend prediction. The fermentation process typically includes cell growth, metabolite synthesis, and decay stages, each with different physiological and chemical characteristics. Therefore, it is necessary to extract feature vectors related to these stages to accurately distinguish and analyze the evolution of the fermentation process.

[0037] It is important to note that the feature extraction process requires comprehensive consideration of the physicochemical parameters in the state sequence and the metabolite signals in the spectral sequence. For example, temperature and dissolved oxygen concentration can reflect the environmental conditions during the cell growth stage, while the spectral signals of metabolites such as lactic acid or ethanol are directly related to the efficiency of the product synthesis stage. Through in-depth analysis of this data, feature vectors related to the fermentation stage can be constructed, containing key information such as cell growth rate and metabolite concentration trends. These feature vectors are not only a digital representation of the fermentation process state but also an important input for subsequent time-series modeling.

[0038] In one embodiment, step S2 is implemented as follows: Step S21 involves standardizing the cleaned spectral sequence to obtain processed spectral data. This step unifies the dimensions and distribution characteristics of the spectral data, facilitating subsequent feature extraction. Standardization typically includes normalizing the signal intensity, mapping signal values ​​from different wavelength channels to the same numerical range, such as 0 to 1, to eliminate dimensional differences between different sensors or different batches of fermentation data. Furthermore, the spectral data can be centered by subtracting the mean, making the data distribution more balanced. Standardization can also be combined with background correction methods to further improve the quality of the spectral data. For example, in lactic acid bacteria fermentation, spectral data may be affected by scattering from suspended particles in the fermentation broth, resulting in uneven signal background. Background correction can remove background noise and highlight the characteristic peaks of metabolite concentrations. The processed spectral data not only has unified dimensions but also more significant signal features, laying the foundation for subsequent deep feature extraction.

[0039] Step S22 involves using a convolutional neural network (CNN) to extract deep features from the processed spectral data, obtaining a preliminary spectral feature set. This step utilizes deep learning technology to extract hidden feature information from high-dimensional spectral data. The CNN, through multiple convolutional and pooling operations, can automatically learn local patterns and global features in the spectral data, such as the peak positions, peak widths, and correlations between metabolite concentrations. The network input is standardized spectral data, and the output is a preliminary spectral feature set containing various feature information related to the fermentation process. In monitoring lactic acid bacteria fermentation, the CNN can be trained to identify the spectral peak features corresponding to lactic acid concentration. The first convolutional layer extracts local shape features of the signal, such as the rising and falling edges of peaks; subsequent layers integrate local features through pooling operations to form more abstract global features, such as the trend of lactic acid concentration over time. After multi-layer processing, the network outputs a preliminary spectral feature set containing multiple feature dimensions related to metabolite concentration. These feature sets provide rich input information for subsequent signal enhancement and pattern extraction.

[0040] Step S23: If the peak signal intensity of metabolite concentrations in the preliminary spectral feature set is lower than a preset threshold, signal enhancement processing is performed on this part of the data to determine the enhanced peak concentration data. This step aims to solve the problem of insufficient signal intensity in spectral features and ensure the significance of key metabolite features. The preset threshold can be determined based on historical data or fermentation process requirements. For example, if the signal intensity of the lactic acid concentration peak is lower than 50% of the normal range, it is considered insufficient and enhancement processing is required. Signal enhancement processing can be achieved through smoothing and amplification operations. For example, for lactic acid concentration peaks with low signal intensity, Gaussian smoothing can be performed on this part of the data first to reduce the influence of random fluctuations; then, the peak intensity can be increased to a identifiable range by linearly amplifying the signal value. The enhanced peak concentration data not only retains the shape features of the original signal but also significantly improves the recognizability of key features. For example, in the later stage of lactic acid bacteria fermentation, the peak signal of lactic acid concentration may be weak due to decreased cell activity. Through signal enhancement processing, this feature can be clearly identified, providing a reliable basis for subsequent analysis.

[0041] Step S24 involves extracting pattern information related to cell morphology from the enhanced concentration peak data to obtain a subset of morphological pattern features. The purpose of this step is to further mine feature information related to cell morphology from metabolite concentration data. Cell morphology changes during fermentation depending on environmental conditions and metabolic state. For example, lactic acid bacteria may exhibit a rod-like morphology during the growth stage and transform into a spherical morphology during the decline stage. These morphological changes can be extracted indirectly through the metabolite concentration signal. The extraction of the subset of morphological pattern features can be based on the changing trends and periodic characteristics of the concentration peak data. For example, the periodic fluctuations in lactic acid concentration may be related to the periodic changes in cell morphology. By analyzing the frequency and amplitude of the peak signal, pattern information characterizing morphological changes can be extracted. This pattern information constitutes the subset of morphological pattern features, containing characteristic representations of cell morphology at different fermentation stages, providing an important supplement for subsequent construction of spectral feature vectors.

[0042] Step S25: Based on the morphological feature subset and concentration peak data, construct a spectral feature vector associated with the fermentation stage. This step integrates various extracted feature information into a unified feature vector to characterize different stages of the fermentation process. The spectral feature vector integrates multiple dimensions of information, such as metabolite concentration and cell morphology, and can comprehensively reflect the dynamic characteristics of the fermentation process. For example, the first few dimensions of the feature vector can characterize the changing trend of lactic acid concentration, while the latter few dimensions characterize the evolution pattern of cell morphology.

[0043] In one possible implementation, the spectral feature vector can be constructed using feature fusion methods. For example, a weighted combination of morphological feature subsets and concentration peak data can be performed, with the weights determined based on the importance of each feature in distinguishing fermentation stages. The constructed spectral feature vector not only contains information from individual features but also enhances the characterization of fermentation stages through the correlations between features. For instance, in lactic acid bacteria fermentation, the feature vector can clearly distinguish between the rapid cell growth stage and the metabolite accumulation stage, providing high-quality input data for subsequent time-series modeling.

[0044] Step S3 involves time-series modeling of the feature vectors and state sequences to capture the dynamic evolution pattern of the fermentation process and determine the predicted trend of the product synthesis rate. This step aims to utilize the temporal information in the feature vectors and state sequences to construct a dynamic model of the fermentation process and predict the future trend of the product synthesis rate. The fermentation process is a complex nonlinear dynamic system involving multiple factors such as cell growth, metabolite synthesis, and changes in environmental conditions; therefore, time-series modeling is necessary to capture its evolutionary patterns.

[0045] Time-series modeling combines spectral information from feature vectors with physicochemical parameters from state sequences to analyze the changing patterns of the fermentation process at different time scales. For example, short-term fluctuations in temperature and dissolved oxygen concentration may affect cell growth rates, while long-term trends in lactic acid concentration reflect product synthesis efficiency. By comprehensively modeling these data, predicted trends in product synthesis rates can be generated, providing a basis for subsequent optimization and adjustments. The time-series modeling process not only considers the current state but also incorporates historical data and future predictions, comprehensively characterizing the dynamic characteristics of the fermentation process.

[0046] Time-series modeling allows for customized analysis at different stages of lactic acid bacteria fermentation. For example, during cell growth, the model can focus on the effects of temperature and dissolved oxygen concentration on the growth rate; during metabolite synthesis, the model focuses more on the changing trends of lactic acid concentration. This stage-by-stage modeling allows for more accurate capture of the dynamic evolution patterns of the fermentation process and prediction of future trends in product synthesis rates. For instance, if the predicted trend indicates that the lactic acid synthesis rate may decrease in the next few hours, control measures can be taken in advance to avoid yield loss. This predictive capability significantly improves the control precision and response speed of the fermentation process.

[0047] It should be noted that time-series modeling can be further optimized through multi-timescale analysis. For example, short-term and long-term models can be built simultaneously. The short-term model predicts changes in the product synthesis rate over the next few hours, while the long-term model predicts the yield trend throughout the fermentation cycle. Combining the two models can provide more comprehensive decision support for optimizing the fermentation process. For instance, in lactic acid bacteria fermentation, the short-term model can help operators adjust the stirring rate or aeration rate in a timely manner, while the long-term model can guide the nutrient supplementation strategy throughout the fermentation cycle.

[0048] In one possible implementation, trend prediction can be generated by incorporating empirical knowledge from historical fermentation data. For example, by analyzing the changes in product synthesis rates from multiple past lactic acid bacteria fermentations, typical trend patterns under certain environmental conditions can be summarized and used as a reference to correct current predictions. This method not only improves prediction accuracy but also enhances the model's adaptability to complex fermentation environments. For instance, if historical data indicates that lactic acid synthesis rates typically accelerate within a specific temperature range, the current trend prediction can be calibrated based on this pattern, ensuring that the prediction results more closely reflect the actual fermentation process.

[0049] Step S4 involves adjusting the regulatory variables of the fermentation process based on the predicted trend to generate an optimized instruction sequence. This step utilizes the product synthesis rate prediction trend obtained from time-series modeling to make targeted adjustments to key regulatory variables in the fermentation process, forming a series of specific operational instructions to maintain or improve product synthesis efficiency. Regulatory variables in the fermentation process typically include stirring rate, aeration rate, nutrient addition rate, temperature setpoint, and pH adjustment, etc. These variables directly affect the cell growth environment and metabolic pathways, thus determining the direction of change in the product synthesis rate.

[0050] Specifically, in the process of lactic acid bacteria fermentation, the selection and adjustment of regulatory variables need to be based on a comprehensive judgment of real-time monitoring data and predicted trends. For example, when the predicted trend indicates that the lactic acid synthesis rate may decrease in the future, it is necessary to analyze the potential factors causing the decrease in a timely manner, such as excessively low dissolved oxygen concentration or excessively rapid consumption of nutrients, and generate corresponding optimized instruction sequences accordingly. These instruction sequences are arranged in chronological order, with each instruction corresponding to specific operating parameters at a specific time point, ensuring the continuity and coordination of regulatory measures.

[0051] In one embodiment, step S4 is implemented as follows: Step S41: Acquire real-time fermentation process data, including product synthesis rate and control variable data. This step extracts the latest process data from the fermenter's control system as the basis for optimization adjustments. Real-time data includes the estimated product synthesis rate at the current moment, as well as the current setpoints and actual values ​​of each control variable. For example, in industrial-scale lactic acid bacteria fermentation, the control system records the stirring rate, aeration rate, and nutrient addition pump operating status every minute, while simultaneously estimating the current lactic acid synthesis rate using spectral data. These data collectively form the input basis for optimization decisions.

[0052] Assuming the fermentation process has entered the rapid accumulation phase of metabolites, real-time data indicates that the current lactic acid synthesis rate remains high, but the dissolved oxygen concentration begins to slowly decrease. The control variables acquired at this point include a stirring rate set at 300 rpm, an aeration rate of 1.5 volumes of air per volume of fermentation broth per minute, and a glucose addition rate of 20 liters per hour. The complete acquisition of this real-time data ensures the accuracy of subsequent predictions and adjustments.

[0053] Step S42 involves using a time series prediction model to make short-term predictions of the product synthesis rate, obtaining a predicted trend. This step, based on time series modeling, further focuses on predicting changes in the product synthesis rate within a short period, typically covering a time window of 1 to 6 hours. The prediction model generates a future trend curve for the product synthesis rate by analyzing historical and current data sequences, thus providing forward-looking guidance for regulatory decisions. Specifically, the short-term prediction process first integrates key physiological parameters from the aforementioned feature vectors and state sequences, and then gradually extrapolates future states using a sliding time window. For example, in lactic acid bacteria fermentation, when the current dissolved oxygen and lactic acid concentrations show specific change patterns, the model can infer that without intervention, the lactic acid synthesis rate will decrease by 15% within the next 2 hours. This generation of short-term predictive trends allows regulatory measures to intervene before problems manifest, avoiding significant losses in fermentation efficiency.

[0054] Step S43: Determine the direction of change in the predicted trend over a future period. This step performs a directional analysis on the predicted trend curve to determine whether the product synthesis rate is increasing, stable, or decreasing. The determination of the direction of change is usually based on the slope or inflection point of the trend curve. For example, if the predicted curve shows a negative slope throughout the next 3 hours, it is judged as a downward trend; if the slope is close to zero and the fluctuation is small, it is considered a stable trend.

[0055] In one possible implementation, the determination of the direction of change can be further confirmed by combining confidence intervals. For example, in the later stages of lactic acid bacteria fermentation, if the predicted trend shows a probability of a decrease in the lactic acid synthesis rate exceeding 70%, it is explicitly marked as a downward trend. This marking result directly guides the subsequent screening and adjustment strategies for regulatory variables, ensuring the targeted nature of optimization measures.

[0056] Step S44: If the predicted trend shows a downward trend, mark it as a downward trend. This step, after confirming the downward direction, explicitly marks the current prediction result to facilitate the execution of conditional branches in subsequent processing. After marking as a downward trend, the system automatically triggers the regulatory variable influence analysis module, prioritizing variables that have a positive impact on the product synthesis rate. For example, in lactic acid bacteria fermentation, when the predicted trend shows a decrease in the synthesis rate due to impending nutrient depletion, marking the downward trend immediately switches to optimizing the nutrient addition rate and stirring rate. This marking mechanism simplifies the decision-making process and improves response speed.

[0057] Step S45 involves extracting regulatory variable data from the fermentation process data, calculating the historical correlation between the regulatory variables and the product synthesis rate, and identifying the main influencing variables. This step quantifies the influence of each regulatory variable on the product synthesis rate through statistical analysis of historical data. Historical correlation calculations can be based on data from multiple past fermentation batches, using methods such as correlation coefficients or mutual information to screen out variables with significant influence as the main influencing variables. Specifically, in the historical records of lactic acid bacteria fermentation, analysis shows that the stirring rate has the highest correlation with dissolved oxygen concentration, followed by a positive correlation between nutrient addition rate and lactic acid synthesis rate. Through this historical correlation analysis, the stirring rate and nutrient addition rate can be identified as the main influencing variables, while the temperature setpoint has a relatively low correlation and is treated as a secondary variable. The identification of the main influencing variables provides a key direction for generating the optimized instruction sequence.

[0058] In one embodiment, it is assumed that historical data analysis shows that when the stirring rate is increased by 10%, the lactic acid synthesis rate increases by an average of 8%, while the impact of aeration rate adjustment varies considerably depending on the fermentation stage. Based on this correlation, the main influencing variables are identified as the stirring rate and the nutrient addition rate, and subsequent optimization instructions will prioritize adjustments around these variables.

[0059] Step S46: Generate the corresponding optimization instruction sequence based on the main influencing variables. This step formulates a specific sequence of operational instructions based on the main influencing variables and the predicted trend. Each instruction includes the adjustment range of the control variable, the execution time point, and the expected effect. For example, when a downward trend is predicted, the optimization instruction sequence might include gradually increasing the stirring rate by 20% over the next 30 minutes, while simultaneously increasing the nutrient addition rate by 15%.

[0060] In one possible implementation, the generation of the optimization instruction sequence can also consider the coupling effect between variables. For example, in lactic acid bacteria fermentation, increasing the stirring rate may cause a slight increase in temperature; therefore, the instruction sequence will simultaneously include temperature fine-tuning instructions to maintain the stability of the overall environment. The generated optimization instruction sequence is arranged in chronological order to form a complete control scheme, which is directly input into the fermentation control system for execution.

[0061] Step S5 involves constructing a real-time optimized fermentation state model by optimizing the instruction sequence, iteratively updating the model based on subsequent data, and determining whether the model error is below a preset threshold. This step integrates the optimized instruction sequence into the dynamic description of the fermentation state, constructing a state model capable of responding to control measures in real time. It continuously iterates and updates the model based on feedback from subsequent data, ensuring the model's accuracy in tracking the actual process. The real-time optimized fermentation state model uses the optimized instruction sequence as its driving input, simulating the state evolution after the implementation of control measures. Simultaneously, as new data arrives, the model continuously compares the predicted state with the actual state, calculates the error, and determines whether further adjustments are needed. This iterative update mechanism allows the fermentation process to maintain an optimized state in a dynamic environment.

[0062] In one embodiment, step S5 is implemented as follows: Step S51 involves filtering instruction fragments related to yield from the optimized instruction sequence and obtaining corresponding physiological parameter data. This step filters the optimized instruction sequence, extracting key instruction fragments that directly affect the final yield, such as instructions involving nutrient addition and dissolved oxygen maintenance. The filtered instruction fragments correspond to specific physiological parameter changes, such as the expected increase curve of lactic acid concentration or the trajectory of cell density growth.

[0063] In one embodiment, during lactic acid bacteria fermentation, yield-related instruction fragments primarily include instructions for batch nutrient addition and instructions for maintaining aeration. Implementation of these instruction fragments is expected to cause a sustained increase in lactic acid concentration in subsequent stages, corresponding to a concentration increase pattern in physiological parameter data.

[0064] Step S52 involves identifying key physiological indicators (KPIs) related to yield by comparing physiological parameter data with historical records. This step compares the selected physiological parameter data with records from historical fermentation batches to identify KPIs that consistently exhibited characteristics in high-yield batches. For example, historically high-yield batches typically show a steady increase in lactic acid concentration and dissolved oxygen concentration maintained within a specific range; these characteristics are identified as KPIs. Specifically, the comparison process involves comparing the similarity of physiological parameter curves at each time point, using dynamic time warping to quantify the similarity. The identified KPIs include the rate of increase in lactic acid concentration, the minimum dissolved oxygen value, and the amplitude of pH fluctuations; these indicators are directly related to the final yield level.

[0065] Step S53: Based on key physiological indicators, the real-time fermentation state data is classified using a support vector machine (SVM) algorithm to determine whether the current state deviates from the optimization range. This step utilizes the SVM algorithm to map the real-time state data into a feature space, performing a binary classification judgment between the optimization range and the deviation range. The SVM algorithm is trained using historical data to learn the boundaries of the optimization range; when a real-time data point falls into the deviation region, it is determined that the current state deviates from the optimization range.

[0066] In one possible implementation, the Support Vector Machine (SVM) algorithm first uses key physiological indicators as input features. During training, it uses historical high-yield batches as positive samples and low-yield batches as negative samples. During real-time classification, if the current state is classified as a negative sample, it is confirmed that the algorithm has deviated from the optimal range. For example, in lactic acid bacteria fermentation, when the dissolved oxygen concentration remains below the lower limit of the key physiological indicator, the algorithm will quickly determine that the state has deviated and trigger subsequent corrective measures.

[0067] Step S54: If the current state deviates from the optimization range, an adjustment strategy is extracted to obtain a correction instruction. After confirming the deviation, this step extracts a suitable strategy from a pre-set adjustment strategy library and generates a specific correction instruction. For example, when a low dissolved oxygen concentration causes the deviation, the extracted adjustment strategies include increasing aeration and increasing stirring speed, generating corresponding correction instructions. Specifically, the adjustment strategy library can be stored according to the type of deviation, such as low dissolved oxygen strategy, insufficient nutrients strategy, etc. Each strategy contains multiple alternative correction instructions, and the most suitable combination is selected based on the current degree of deviation. In lactic acid bacteria fermentation, if the deviation is minor, a mild correction instruction is selected, such as increasing aeration by 10%; if the deviation is severe, a more significant adjustment is adopted.

[0068] Step S55 involves determining the final optimized instruction combination by fusing the corrected instructions with the original instruction sequence. This step integrates the corrected instructions into the original optimized instruction sequence to form the updated final instruction combination. The fusion process considers time coordination to ensure that the corrected instructions do not conflict with the original instructions. For example, the execution time of the corrected instructions is inserted into an appropriate position in the original sequence to avoid drastic fluctuations in the control variables.

[0069] In one embodiment, after a deviation in dissolved oxygen is detected in the later stages of lactic acid bacteria fermentation, a correction instruction is issued to gradually increase the aeration rate over the next hour. This instruction, combined with the nutrient addition instruction in the original sequence, forms a coordinated final optimized instruction combination. This fusion method maintains the stability of regulation while promptly correcting deviations.

[0070] Step S56: Construct a fermentation state model based on the optimized instruction combination. This step uses the final optimized instruction combination as input to construct or update the fermentation state model. The model describes the state evolution path driven by the optimized instructions, including the expected changes in various physiological parameters. This model provides a dynamic framework for subsequent iterative updates and can reflect the regulatory effects in real time.

[0071] Step S6 involves iteratively updating subsequent data and determining whether the model error is below a preset threshold. This step iteratively optimizes the fermentation state model by continuously incorporating newly collected subsequent data and assesses the degree of deviation between the model predictions and actual data. The preset threshold is set according to process requirements; for example, the error threshold for predicting lactic acid concentration can be set to 5%. Specifically, the iterative update process is executed once after each time window, updating the model parameters and recalculating the error. If the error remains below the threshold, the model remains stable; otherwise, further data updates or model adjustments are triggered. This mechanism ensures the fermentation state model's continuous adaptability to process changes.

[0072] In one embodiment, step S6 is implemented as follows: Step S61 involves processing the subsequent high-dimensional spectral data using a fermentation state model to obtain an updated feature dataset. This step inputs the newly acquired high-dimensional spectral data into the fermentation state model and, combined with the effects of optimization instructions, generates an updated feature dataset. These features reflect the latest state changes after the implementation of regulatory measures, such as the latest intensity and location of the lactic acid concentration peak.

[0073] In one possible implementation, the subsequent processing of spectral data also includes comparative analysis with historical features to further enrich the content of the updated feature dataset. For example, in lactic acid bacteria fermentation, if the new spectrum shows an increase in lactic acid peak intensity, the corresponding dimension in the updated feature set is enhanced.

[0074] Step S62 involves calculating the deviation between predicted and actual values ​​by performing an error assessment on the feature dataset. This step compares the feature values ​​predicted by the model based on the optimization instructions with the actual collected feature values ​​to calculate the magnitude of the deviation. Error assessment can be performed separately for multiple key features, such as the predicted deviation of lactic acid concentration and dissolved oxygen. Specifically, the deviation is calculated in the form of absolute error or relative error, and the overall assessment result is formed by combining the errors of multiple features. In lactic acid bacteria fermentation, if the deviation between the predicted and actual lactic acid concentration values ​​exceeds 3%, the overall error is considered too high and further processing is required.

[0075] In step S63, if the deviation exceeds a preset threshold, a data update process is triggered to acquire new high-dimensional spectral data. This step proactively expands the data acquisition range or increases the acquisition frequency when the error exceeds the threshold, acquiring more new high-dimensional spectral data for rapid model correction. This triggering mechanism avoids error accumulation and ensures model accuracy.

[0076] Step S64: If the deviation is lower than a preset threshold, a yield prediction result is generated based on the feature dataset, and the predicted value range is determined. In this step, when the model accuracy meets the requirements, the updated feature dataset is used to generate the final yield prediction result, and a confidence range is given. For example, before the lactic acid bacteria fermentation ends, the predicted final lactic acid yield is 80 grams per liter, ranging from 75 to 85 grams.

[0077] Step S65: Based on the predicted value range, calculate the optimized path for the fermentation process to obtain adjusted operating parameter guidelines. This step, based on the yield prediction range, reverse-engineers the optimal operating parameter sequence, forming a complete optimized path guide. These guidelines cover adjustments to all control variables for the remaining fermentation time.

[0078] Step S7 determines the final yield prediction and optimization path. This step integrates all the preceding analysis results, outputting the final yield prediction and the complete optimization path for the fermentation process, ensuring that fermentation achieves the expected goals. Specifically, the final yield prediction is determined by combining the model output after multiple iterations, while the optimization path covers all control measures from the current moment to the end of fermentation. In lactic acid bacteria fermentation, the final predicted value might be determined to be 82 grams of lactic acid per liter, and the optimization path includes gradually reducing the stirring rate to reduce energy consumption while maintaining appropriate nutrient addition levels.

[0079] In one embodiment, step S7 is implemented as follows: Step S71: Based on the output of the fermentation state model, obtain the state prediction value and determine whether the fermentation process has achieved the expected goal. This step extracts the final state prediction value from the model and compares it with the preset process target. For example, if the expected goal is a lactic acid concentration of 80 grams per liter or more, and the prediction value meets this target, then the goal is considered to have been achieved.

[0080] Step S72: If the predicted state value does not reach the expected target, a backup plan is extracted, new physiological parameter data is integrated, and the fermentation state model is updated. In this step, if the target is not achieved, a backup control plan is activated, such as adding extra nutrient batches or adjusting the temperature curve, and the new physiological parameter data is incorporated into the model for updating. If the predicted concentration is insufficient near the end of lactic acid bacteria fermentation, backup plans include temporarily increasing aeration to stimulate residual metabolic activity, and the updated model reassesses the state.

[0081] Step S73 involves monitoring the fermentation state using the updated fermentation state model to obtain real-time optimized state data. This step utilizes the updated model to continuously monitor the remaining process and generate real-time optimized state data, ensuring maximum efficiency in the final stage.

[0082] Step S74: Determine the final production forecast based on the real-time optimized status data. This step integrates real-time data and model predictions to lock in the final production forecast value, which serves as a reference for the end of the batch.

[0083] Step S75 generates an optimized path guide for the fermentation process based on yield forecasts and status data. This step summarizes all optimization measures to form a complete path guide document, which is used to guide the operation of current and future batches.

[0084] Step S76: Based on the optimized path guidance, update the data input in the fermentation status monitoring framework to complete the iterative analysis. This step feeds back the optimized path guidance to the monitoring framework, updates the data acquisition and analysis parameters, completes the iterative optimization analysis for the entire fermentation cycle, and provides experience accumulation for the next batch.

[0085] When applied in industrial lactic acid bacteria fermentation production, this application can significantly improve process stability and yield consistency. Through real-time data-driven optimization, the fermentation process can maintain efficient operation in complex environments, with high accuracy in final yield prediction and timely and effective control measures. This method is applicable to fermentation processes of various scales, from laboratory pilot-scale trials to industrial production of tens of thousands of tons, ensuring the broad applicability of the technical solution.

[0086] It should be noted that the implementation of the above steps can be appropriately adjusted according to the specific fermentation type. For example, in some oxygen-sensitive fermentation processes, the weight of dissolved oxygen regulation variables will be higher, while in aerobic fermentation, more emphasis will be placed on the synergistic optimization of nutrients and temperature. The combination of multiple implementation methods further enhances the flexibility and robustness of the method.

[0087] In another implementation, when the fermentation process involves parallel production across multiple tanks, this method can be extended to batch-to-batch data sharing. For example, the optimized instruction sequence and yield prediction results of one fermenter can be used as a reference to assist in real-time adjustments to other tanks, thereby improving overall production efficiency. This extended application maintains consistency in core steps while adapting to more complex industrial scenarios. Specifically, in a multi-tank lactic acid bacteria fermentation system, the data from the tank that has completed optimization first can be transmitted in real-time to the control modules of other tanks. When a similar downward trend appears in the predicted trend, the validated optimized instruction sequence can be directly invoked. This data sharing mechanism further reduces control latency and improves the overall production line's responsiveness.

[0088] The above description is only a preferred embodiment of the present invention. It should be noted that those skilled in the art can make several modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the protection scope of the present invention.

Claims

1. A machine learning-based intelligent prediction and optimization method for microbial fermentation processes, characterized in that, The method includes: Real-time multidimensional state data and high-dimensional spectral data of fermentation broth are acquired by sensors. Noise removal processing is performed on the multidimensional state data and high-dimensional spectral data to obtain clean state sequences and spectral sequences. Based on the clean state sequence and spectral sequence, feature vectors related to the fermentation stage are extracted. The feature vectors and state sequence are then used for time-series modeling to capture the dynamic evolution pattern of the fermentation process and determine the predicted trend of product synthesis rate. Based on the predicted trends, the control variables of the fermentation process are adjusted to generate an optimized instruction sequence; A real-time optimized fermentation state model is constructed by optimizing the instruction sequence. Subsequent data is iteratively updated to determine whether the model error is lower than a preset threshold. If it is lower than the preset threshold, the final yield prediction and optimization path are determined.

2. The intelligent prediction and optimization method for microbial fermentation processes based on machine learning according to claim 1, characterized in that, The process of acquiring real-time multidimensional state data and high-dimensional spectral data of the fermentation broth through sensors, and performing noise reduction processing on the multidimensional state data and high-dimensional spectral data, includes: Real-time data from the fermentation broth is collected using sensors to obtain an initial multidimensional state dataset and a high-dimensional spectral dataset. A filtering method was used to remove noise from the initial multidimensional state dataset and high-dimensional spectrogram dataset to obtain a clean state sequence and spectrogram sequence. For pure state sequences and spectral sequences, a preset threshold is applied to detect outliers. If a data point exceeds the threshold range, it is marked as an outlier. Based on the labeling results, abnormal data points and normal data points are separated to obtain the separated normal state sequence and normal spectrum sequence; By dividing the normal state sequence and the normal spectral sequence into time windows, local data features are obtained, and the set of local features is determined.

3. The intelligent prediction and optimization method for microbial fermentation processes based on machine learning according to claim 1, characterized in that, The step of extracting feature vectors related to the fermentation stage based on the clean state sequence and spectral sequence includes: The clean spectral sequence is standardized to obtain the processed spectral data, and a convolutional neural network is used to extract deep features from the processed spectral data to obtain a preliminary set of spectral features. If the peak signal intensity of metabolite concentration in the preliminary spectral feature set is lower than a preset threshold, then the data in this part is subjected to signal enhancement processing to determine the enhanced peak concentration data. By extracting pattern information related to bacterial cell morphology from the enhanced concentration peak data, a subset of morphological pattern features is obtained. Based on the subset of morphological feature data and concentration peak data, a spectral feature vector associated with the fermentation stage is constructed.

4. The intelligent prediction and optimization method for microbial fermentation processes based on machine learning according to claim 1, characterized in that, The step of performing time-series modeling of feature vectors and state sequences to capture the dynamic evolution patterns of the fermentation process includes: Physiological parameter data from feature vectors and state sequences are obtained, and a unified input dataset is obtained through standardization. A recurrent neural network is used to perform time series modeling on the input dataset, analyze dynamic changes, and determine the feature representations at each time point. Based on the feature representation, dynamic change patterns of different fermentation stages are extracted to determine the stage evolution law. If the stage evolution law does not match the preset threshold range, the data is divided into additional time windows to obtain fine-grained change patterns. A time-series prediction model is constructed through the fine-grained change patterns to determine the preliminary trend estimate of the product synthesis rate.

5. The intelligent prediction and optimization method for microbial fermentation processes based on machine learning according to claim 1, characterized in that, The step of adjusting the control variables of the fermentation process based on the predicted trend and generating an optimized instruction sequence includes: Real-time fermentation process data, including product synthesis rate and regulatory variable data, are acquired. A time series prediction model is used to make short-term predictions of the product synthesis rate and obtain the prediction trend. Determine the direction of change in the predicted trend over a future period. If the predicted trend shows a downward trend, mark it as a downward trend. Data on regulatory variables are extracted from fermentation process data. The historical correlation between the regulatory variables and the product synthesis rate is calculated to determine the main influencing variables. Based on the main influencing variables, the corresponding optimization instruction sequence is generated.

6. The intelligent prediction and optimization method for microbial fermentation processes based on machine learning according to claim 1, characterized in that, The process of constructing a real-time optimized fermentation state model by optimizing the instruction sequence includes: Filter instruction fragments related to yield from the optimized instruction sequence, obtain the corresponding physiological parameter data, compare the physiological parameter data with historical records, and determine the key physiological indicators related to yield. Based on key physiological indicators, the support vector machine algorithm is used to classify the real-time fermentation status data and determine whether the current state deviates from the optimization range. If the current state deviates from the optimization range, the adjustment strategy is extracted and the correction instruction is obtained. By fusing the modified instructions with the original instruction sequence, the final optimized instruction combination is determined, and a fermentation state model is constructed based on the optimized instruction combination.

7. The intelligent prediction and optimization method for microbial fermentation processes based on machine learning as described in claim 1, characterized in that, The step of iteratively updating subsequent data and determining whether the model error is below a preset threshold includes: The subsequent high-dimensional spectral data is processed by the fermentation state model to obtain the updated feature dataset. The error of the feature dataset is evaluated, and the deviation between the predicted value and the actual value is calculated. If the deviation is higher than the preset threshold, the data update process is triggered to obtain new high-dimensional spectral data; if the deviation is lower than the preset threshold, the yield prediction result is generated based on the feature dataset to determine the range of the prediction value. Based on the predicted value range, the optimized path of the fermentation process is calculated, and the adjusted operating parameters are obtained as guidance.

8. The intelligent prediction and optimization method for microbial fermentation processes based on machine learning as described in claim 1, characterized in that, The determination of the final production forecast and optimization path includes: Based on the output of the fermentation state model, obtain the state prediction value, determine whether the fermentation process has reached the expected goal, and if the state prediction value has not reached the expected goal, extract the backup plan, integrate new physiological parameter data, and update the fermentation state model. The fermentation state is monitored by the updated fermentation state model to obtain real-time optimized state data. Based on the real-time optimized state data, the final yield prediction value is determined. By using yield forecasts and status data, an optimized path guide for the fermentation process is generated. Based on the optimized path guide, the data input in the fermentation status monitoring framework is updated to complete the iterative analysis.

Images