A machine learning-based fermentation fiber feed ratio optimization method and system

Abstract

The present application relates to the technical field of machine learning, and discloses a fermentation fiber feed proportioning optimization method and system based on machine learning, which monitors temperature time series data, microbial activity and raw material synergistic effect change data. The raw material synergistic effect score is calculated based on the similarity of the molecular structure of the raw material; the correlation degree of the raw material synergistic effect is quantified by the absolute value of the ratio of the score change rate to the microbial activity change rate; the model evaluation unit judges the stability of the prediction confidence and the consistency of the feature importance; the regulation unit executes the strategy of stable state temperature sensitive state synergistic effect fluctuation state and synergistic effect failure state according to the evaluation result; and the closed-loop feedback unit updates the judgment condition to optimize the proportioning to realize dynamic self-adaptive regulation and control of the fermentation process, thereby significantly improving the proportioning accuracy and fermentation stability. The present application can improve the accuracy and stability of fermentation proportioning optimization, reduce batch quality fluctuations and ensure the consistency of feed quality.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and more specifically, to a method and system for optimizing the formulation of fermented fiber feed based on machine learning. Background Technology

[0002] Fermented fiber feed technology utilizes microbial fermentation to biotransform high-fiber raw materials, representing a key technological pathway to improve feed nutritional value and animal digestibility. In traditional feed production, high-fiber raw materials such as wheat bran, soybean hulls, and straw commonly suffer from high lignin content and poorly digestible crude fiber, resulting in low overall feed utilization. Fermentation technology leverages the synergistic effect of complex enzyme preparations and complex microbial communities to effectively degrade macromolecules such as cellulose and hemicellulose, significantly improving feed palatability and promoting intestinal health in animals. In recent years, machine learning and artificial intelligence technologies have demonstrated application potential in fermentation process optimization. By constructing data-driven models, real-time monitoring and dynamic control of fermentation parameters can be achieved, and optimization algorithms can determine the optimal ingredient ratios and process conditions, thus overcoming the limitations of traditional methods relying on experience and trial-and-error. International research has applied advanced learning technologies to fermentation processes, improving production efficiency. In the feed industry, machine learning technology is beginning to be used for component classification and formulation optimization, but its systematic application for optimizing the formulation of fermented fiber feeds remains in the exploratory stage.

[0003] While current fermentation technology for fiber feed has made progress, with synergistic fermentation of bacteria and enzymes effectively degrading crude fiber and increasing crude protein content, applying machine learning to this field presents a significant challenge: the models cannot adapt to the dynamic changes in the fermentation process in real time. Specifically, the models lack adaptability during fermentation, struggling to dynamically adjust formulations based on temperature fluctuations, changes in microbial activity, and synergistic effects of raw materials. This leads to inaccurate formulations and large batch-to-batch quality fluctuations, impacting the stability of fermentation results and the consistency of feed quality. Summary of the Invention

[0004] The purpose of this invention is to provide a machine learning-based method and system for optimizing the formulation of fermented fiber feed, which has the advantages of improving the accuracy and stability of fermentation formulation optimization, reducing batch quality fluctuations, and ensuring the consistency of feed quality.

[0005] On one hand, the present invention provides a machine learning-based fermented fiber feed formulation optimization system, comprising: The fermentation process dynamic monitoring unit is used to collect real-time data on temperature time series, microbial activity, and changes in raw material synergistic effects during the fermentation process. The raw material synergistic effect change data includes raw material synergistic effect score and raw material synergistic effect correlation degree. The raw material synergistic effect score is calculated based on the similarity of raw material molecular structure. The raw material synergistic effect correlation degree is calculated based on the correlation between the rate of change of raw material synergistic effect score and the rate of change of microbial activity during the continuous fermentation period. The correlation calculation is specifically: the absolute value of the ratio of the rate of change of raw material synergistic effect score to the rate of change of microbial activity is used as the correlation quantification value. The model operation status evaluation unit is used to determine the stability of the prediction confidence and the consistency of feature importance of the machine learning model in the fermentation scenario based on the raw material synergistic effect change data. The fermentation scenario control unit is used to determine the current process state of the fermentation scenario based on the judgment result of the model operation state evaluation unit, and to execute a model control strategy that matches the process state. The closed-loop optimization feedback unit is used to receive and output the adjusted fermentation ratio and the fermentation ratio implementation effect data, and update the judgment conditions of the model operation status evaluation unit.

[0006] Furthermore, the model operation status evaluation unit includes: The confidence stability assessment module is used to calculate temperature fluctuation characteristics based on the temperature time series data, and to calculate the rate of change of the prediction confidence distribution of the machine learning model based on the raw material synergistic effect score and the temperature fluctuation characteristics. The temperature fluctuation characteristics are the temperature standard deviation over 10 consecutive sampling periods, and the rate of change of the prediction confidence distribution is the absolute value of the difference between the current prediction interval width and the historical benchmark prediction interval width / the historical benchmark prediction interval width. The feature consistency assessment module is used to calculate the volatility of the feature importance ranking of the machine learning model for changes in the raw material synergy effect based on the correlation degree of the raw material synergy effect. The volatility of the feature importance ranking is the absolute value of the difference between the current feature importance ranking position and the historical average ranking position / the total number of features.

[0007] Furthermore, the fermentation scene control unit includes: The process status determination module is used to determine the process status of the current fermentation scenario based on the predicted confidence distribution change rate and the feature importance ranking volatility, wherein: When both the rate of change of the predicted confidence distribution and the volatility of the feature importance ranking are not greater than the preset benchmark value, the process is determined to be in a stable state. When the rate of change of the predicted confidence distribution is greater than the preset benchmark value, and the volatility of the feature importance ranking is not greater than the preset benchmark value, the process state is determined to be a temperature-sensitive state. When the rate of change of the predicted confidence distribution is not greater than the preset benchmark value, and the volatility of the feature importance ranking is greater than the preset benchmark value, the process state is determined to be a synergistic effect fluctuation state. When both the rate of change of the predicted confidence distribution and the volatility of the feature importance ranking are greater than the preset benchmark value, the process state is determined to be a state of synergistic effect failure. The control strategy execution module is used to trigger the corresponding model control strategy based on the process state determination result.

[0008] Furthermore, the triggering logic of the control strategy execution module is as follows: When the process is determined to be in a stable state, the model parameter preservation strategy is executed. When the process state is determined to be temperature sensitive, the temperature-feature weight adjustment strategy is executed. When the process status is determined to be a state of synergistic effect fluctuation, a synergistic effect local calibration strategy is executed. When the process status is determined to be a synergy failure state, the synergy recalibration strategy is executed.

[0009] Furthermore, the temperature-feature weight adjustment strategy is implemented as follows: Calculate the ratio of the fermentation temperature fluctuation range to the historical average fluctuation range; The adjusted feature weights are obtained by multiplying the feature weights associated with temperature-sensitive raw materials by the ratio.

[0010] Furthermore, the synergistic effect recalibration strategy is implemented as follows: Based on the raw material synergy score in the raw material synergy change data, the correlation degree of raw material synergy is recalculated. Based on the aforementioned raw material synergy correlation, the raw material synergy matrix is ​​reconstructed and replaced with the original matrix as the model input feature; The synergistic effect local calibration strategy is implemented as follows: Based on the volatility of the aforementioned feature importance ranking, the feature weights related to synergistic effects are dynamically adjusted.

[0011] Furthermore, the closed-loop optimization feedback unit includes: The dynamic benchmark update module is used to dynamically adjust the preset benchmark value based on the implementation effect data of the regulated fermentation ratio; the implementation effect data of the fermentation ratio includes the change rate of crude fiber digestibility and the change rate of microbial activity stability. The dynamic correlation strength generation module is used to dynamically generate the correlation strength threshold between the feature importance ranking volatility and the microbial activity change rate based on historical fermentation scenario data.

[0012] Furthermore, the fermentation process dynamic monitoring unit includes: A multi-point temperature monitoring array is deployed at different positions along the axial and radial axes of the fermenter to collect spatial distribution data of fermentation temperature. The online microbial activity analysis module includes an optical density sensor and a metabolite gas chromatography unit for real-time calculation of microbial activity; The real-time analysis module for raw material molecular structure includes a Raman spectrometer and a molecular docking calculation unit, which is used to calculate the similarity of raw material molecular structure in real time. Specifically, the similarity of raw material molecular structure is the matching degree of cellulase hydrolysis sites, which is calculated based on the matching between the raw material molecular structure and the cellulase active site.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: By collecting fermentation data in real time, evaluating model status, adjusting process status, and providing closed-loop feedback, the fermentation ratio is dynamically optimized. This has the advantages of improving the accuracy and stability of fermentation ratio optimization, reducing batch quality fluctuations, and ensuring the consistency of feed quality.

[0014] On the other hand, the present invention also provides a method for optimizing the formulation of fermented fiber feed based on machine learning, comprising: S1: Real-time acquisition of temperature time-series data, microbial activity, and raw material synergistic effect changes during the fermentation process; The raw material synergistic effect change data includes raw material synergistic effect score and raw material synergistic effect correlation degree. The raw material synergistic effect score is calculated based on the similarity of raw material molecular structure. The raw material synergistic effect correlation degree is calculated based on the correlation between the rate of change of raw material synergistic effect score and the rate of change of microbial activity during the continuous fermentation period. The correlation calculation is specifically: the absolute value of the ratio of the rate of change of raw material synergistic effect score to the rate of change of microbial activity is used as the correlation quantification value. S2: Based on the data on the changes in the synergistic effect of the raw materials, determine the stability of the prediction confidence of the machine learning model in the fermentation scenario and the consistency of the feature importance. S3: Based on the judgment result of step S2, determine the process state of the current fermentation scenario and trigger the model control strategy that matches the process state. S4: Receive and output the adjusted fermentation ratio and the fermentation ratio implementation effect data, and update the judgment conditions of the model operation status evaluation unit.

[0015] It should be noted that the beneficial effects of the machine learning-based fermented fiber feed formulation optimization system and its method provided by this invention are the same in specific terms, and will not be repeated here. Attached Figure Description

[0016] 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.

[0017] Figure 1 A flowchart illustrating a machine learning-based method for optimizing the formulation of fermented fiber feed, as provided in an embodiment of the present invention. Figure 2 This is a functional framework diagram of a machine learning-based fermented fiber feed formulation optimization system provided in an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0020] See Figure 1 As shown, this embodiment of the invention provides a machine learning-based fermented fiber feed formulation optimization system, comprising: The fermentation process dynamic monitoring unit is used to collect real-time data on temperature time series, microbial activity, and changes in raw material synergistic effects during the fermentation process. The data on changes in raw material synergy effects include raw material synergy effect scores and raw material synergy effect correlation. The raw material synergy effect score is calculated based on the similarity of raw material molecular structures. The raw material synergy effect correlation is calculated based on the correlation between the rate of change of raw material synergy effect scores and the rate of change of microbial activity during the continuous fermentation period. The correlation calculation is specifically as follows: the absolute value of the ratio of the rate of change of raw material synergy effect scores to the rate of change of microbial activity is used as the quantified value of the correlation. The model operation status evaluation unit is used to determine the stability of the prediction confidence and the consistency of feature importance of the machine learning model in the fermentation scenario based on the data on changes in raw material synergistic effects. The fermentation scenario control unit is used to determine the current process state of the fermentation scenario based on the judgment result of the model running status evaluation unit, and to execute the model control strategy that matches the process state. The closed-loop optimization feedback unit is used to receive and output the adjusted fermentation ratio and the implementation effect data of the fermentation ratio, and update the judgment conditions of the model operation status evaluation unit.

[0021] Specifically, the fermentation process dynamic monitoring unit is used to collect key data during the fermentation process in real time. For example, manual inspections can be conducted, periodically using a handheld thermometer to measure the temperature at different locations within the fermenter and recording it as time-series temperature data. Microbial activity can be obtained by periodically sampling and sending samples to the laboratory for culture counting or biochemical analysis. Data on changes in feedstock synergy effects can be obtained by manually analyzing the chemical composition table of the feedstocks and, based on experience, judging their molecular structural similarity, thereby providing a feedstock synergy effect score. The correlation degree of feedstock synergy effects can be roughly estimated by manually calculating a simple ratio between the feedstock synergy effect score and the change in microbial activity in continuous fermentation batches. However, this manual or offline data collection method may have limitations such as untimely data updates, limited accuracy, and an inability to fully reflect the complex dynamics of the fermentation process.

[0022] The model performance evaluation unit, based on the collected data on the changes in raw material synergistic effects, assesses the stability of the machine learning model's prediction confidence and the consistency of feature importance in the fermentation scenario. Specifically, a fixed time window can be set, and the variance of the model's predictions of the fermentation ratio within that window can be calculated to evaluate the stability of the prediction confidence. A larger variance indicates poorer stability. For the consistency of feature importance, the model can be retrained periodically, and the importance ranking of each raw material feature in the model can be manually compared across different training periods to determine its fluctuations. For example, if the importance ranking of a certain raw material feature changes frequently and significantly across different training periods, its consistency is considered poor. This evaluation method may not capture subtle changes in model performance and is highly dependent on human intervention.

[0023] The fermentation scenario control unit determines the current process state of the fermentation scenario based on the judgment results of the model operation status evaluation unit, and executes a model control strategy matching the process state. For example, several simple thresholds can be preset. When the fluctuations in prediction confidence variance and feature importance ranking are both below a certain threshold, it is determined to be a stable state; otherwise, it is determined to be an unstable state. For a stable state, a model parameter preservation strategy can be implemented, i.e., no adjustments are made. For an unstable state, a general model retraining strategy is implemented. This simple state determination and strategy execution method may not be able to fine-tune the handling of specific problems that occur during fermentation (such as abnormal temperature or specific raw material synergy issues), resulting in poor control effects.

[0024] The closed-loop optimization feedback unit receives and outputs the adjusted fermentation ratio and its implementation effect data, and updates the judgment conditions of the model operation status evaluation unit. For example, after implementing a new fermentation ratio, the crude protein and crude fiber content of the final feed product can be manually recorded as implementation effect data. Then, based on this effect data, expert experience determines whether it is necessary to manually adjust the thresholds used in the model operation status evaluation unit to judge confidence stability and feature importance consistency. Although this feedback mechanism achieves a closed loop, its process of updating judgment conditions relies on human experience and lagging effect data, which may lead to a slow response speed of the system to changes in the fermentation environment, making it difficult to achieve true dynamic adaptive optimization.

[0025] The following example will provide a more detailed explanation of the above technical solution: Suppose a fermented fiber feed production base aims to continuously optimize feed formulation to improve production efficiency and product quality. This base has deployed the machine learning-based fermented fiber feed formulation optimization system proposed in this application.

[0026] First, the dynamic monitoring unit for the fermentation process begins operation. This unit automatically collects temperature data every hour using multiple temperature sensors installed inside the fermenter, generating time-series temperature data. Simultaneously, staff at the base regularly extract samples from the fermenter daily and send them to the laboratory for microbial activity testing to obtain microbial activity data. Regarding the data on changes in the synergistic effect of raw materials, the base periodically performs molecular structure analysis on new batches of raw materials and calculates the molecular structural similarity between raw materials based on the analysis results, thus obtaining a synergistic effect score. Subsequently, the system calculates the correlation degree of the synergistic effect based on the rate of change of the synergistic effect score and the rate of change of microbial activity over several consecutive days. For example, when the absolute value of the ratio of the rate of change of the synergistic effect score to the rate of change of microbial activity for a certain raw material is found to be large, it indicates a strong correlation between the synergistic effect and microbial activity of that raw material.

[0027] Next, the model performance evaluation unit receives and processes this dynamic monitoring data. This unit continuously monitors the confidence distribution of the machine learning model's predictions of the fermentation ratio. For example, if the prediction interval width suddenly increases when predicting the content of a key nutrient, it indicates a decrease in prediction confidence stability. Simultaneously, this unit also analyzes the importance ranking of various raw material features in the model. For example, if the importance ranking of a certain raw material feature frequently fluctuates between first and fifth place in the model over the past week, it indicates poor consistency in feature importance.

[0028] Subsequently, the fermentation scenario control unit determines the current process state of the fermentation scenario based on the judgment results of the model operation status evaluation unit. For example, if the prediction confidence stability decreases but the consistency of feature importance is still acceptable, the system may determine that the current state is temperature-sensitive, meaning the model is highly sensitive to temperature changes, leading to increased prediction uncertainty. In response to this state, the fermentation scenario control unit will trigger corresponding model control strategies. For example, a temperature-feature weight adjustment strategy can be implemented, which dynamically adjusts the weights of features related to temperature-sensitive raw materials in the model based on the current temperature fluctuation range, so that the model can better adapt to temperature changes.

[0029] Finally, the closed-loop optimization feedback unit receives and outputs the adjusted fermentation ratio. The base staff then proceeds with production according to the new ratio. After production is complete, the closed-loop optimization feedback unit receives data on the implementation effect of this fermentation batch, such as the crude fiber digestibility and microbial activity stability of the final feed. This implementation effect data is used to update the judgment conditions of the model's operational status evaluation unit. For example, if, after adjustment, the crude fiber digestibility significantly improves and the microbial activity stability is good, the system will fine-tune the thresholds for evaluation confidence stability and feature importance consistency based on this positive feedback, making them more consistent with the actual situation of the current production environment, thereby achieving continuous learning and optimization of the system. Through this closed-loop feedback mechanism, the system can continuously improve itself, ensuring that it always provides optimal feed ratio recommendations in a dynamically changing fermentation environment.

[0030] Based on the above examples, the machine learning-based fermented fiber feed formulation optimization system proposed in this application demonstrates significant technological contributions. In traditional fermented fiber feed production, formulation optimization often relies on manual adjustments by experienced technicians or time-consuming trial-and-error experiments. This approach is not only inefficient but also struggles to respond in real-time to complex dynamic changes during fermentation, such as temperature fluctuations, changes in microbial communities, and batch-to-batch variations in raw materials, leading to inaccurate formulations and inconsistent batch quality.

[0031] In contrast, the system in this application, by introducing a dynamic monitoring unit for the fermentation process, achieves real-time, multi-dimensional acquisition of key fermentation parameters, particularly the quantification of data on changes in the synergistic effects of raw materials. This includes synergistic effect scores calculated based on the similarity of raw material molecular structures and the correlation degree of synergistic effects with changes in microbial activity. This enables the system to capture subtle changes that are difficult to detect using traditional methods, providing a precise data foundation for subsequent intelligent decision-making. For example, in the above example, the system can monitor temperature, microbial activity, and raw material synergistic effects in real time, while traditional methods may only be able to perform periodic offline monitoring, failing to detect and respond to anomalies in the fermentation process in a timely manner.

[0032] Furthermore, the introduction of a model runtime evaluation unit enables the system to proactively assess the reliability of the machine learning model in dynamic fermentation scenarios. By evaluating the stability of prediction confidence and the consistency of feature importance, the system can identify when the model may become fatigued or inaccurate, thus avoiding blindly using potentially ineffective models for decision-making. This contrasts sharply with existing machine learning applications in fermentation, which suffer from poor model adaptability and difficulty in adjusting proportions in real time according to dynamic changes. In the example, the system can detect an increase in the model's prediction interval width or fluctuations in feature importance rankings, which cannot be quantified and evaluated in traditional solutions.

[0033] The fermentation scenario control unit intelligently determines the process state and executes matching model control strategies based on the evaluation results, achieving refined management of the fermentation process. This ability to trigger specific control strategies (such as temperature-feature weight adjustment strategies) based on specific process states (such as temperature-sensitive states) far surpasses the single or coarse adjustment methods in traditional solutions. Traditional methods may only be able to perform simple parameter adjustments, but cannot delve into the model level for adaptive optimization.

[0034] Ultimately, the closed-loop optimization feedback unit establishes a mechanism for continuous learning and self-improvement. By receiving data on the effects of adjusted fermentation ratios and dynamically updating the judgment criteria for model operation status evaluation, the system can continuously learn from actual production, improving its adaptability and optimization capabilities. This data-driven closed-loop feedback effectively addresses the pain points of traditional methods, such as the lack of quantitative feedback and difficulty in continuous improvement during the optimization process, ensuring the stability and efficiency of the system in long-term operation. Overall, the system in this application, through the organic combination of data acquisition, model evaluation, intelligent control, and closed-loop feedback, provides an innovative solution that can adapt to dynamic changes in the fermentation process in real time, significantly improving the accuracy of ratios and batch quality.

[0035] In some of the solutions mentioned above in this application, a model operation status evaluation unit is proposed to judge the stability of the prediction confidence and the consistency of feature importance of the machine learning model in the fermentation scenario. However, in this process, due to the lack of a specific evaluation module, it is impossible to accurately quantify the changes in the distribution of prediction confidence and the fluctuations in the ranking of feature importance. As a result, the evaluation results rely on subjective experience or simple thresholds, making it difficult to dynamically capture the temperature fluctuations and changes in the synergistic effect of raw materials during the fermentation process, thus failing to provide a reliable basis for subsequent regulation.

[0036] To address this, this application further proposes a model operation status evaluation unit comprising: a confidence stability evaluation module, used to calculate temperature fluctuation characteristics based on temperature time-series data, and to calculate the rate of change of the prediction confidence distribution of the machine learning model based on the raw material synergy effect score and temperature fluctuation characteristics, wherein the temperature fluctuation characteristic is the temperature standard deviation over 10 consecutive sampling periods, and the rate of change of the prediction confidence distribution is the absolute value of the difference between the current prediction interval width and the historical baseline prediction interval width / the historical baseline prediction interval width; and a feature consistency evaluation module, used to calculate the volatility of the feature importance ranking of the machine learning model for changes in raw material synergy effect based on the correlation degree of raw material synergy effect, wherein the volatility of the feature importance ranking is the absolute value of the difference between the current feature importance ranking position and the historical average ranking position / the total number of features.

[0037] The confidence stability assessment module aims to quantify the reliability fluctuations of machine learning model predictions. It can be a standalone software module, implemented using programming languages ​​such as Python or Java, integrated into the model runtime evaluation unit. This module receives data from the fermentation process dynamic monitoring unit and executes pre-defined calculation logic. Alternatively, it can be designed as a standalone service based on a microservice architecture, interacting with other system components and making function calls through clearly defined API interfaces, thus achieving modular deployment and elastic scaling. Temperature time-series data refers to the continuous record of temperature changes over time during fermentation, serving as the fundamental input for assessing the stability of the fermentation environment. This data can be collected by a multi-point temperature monitoring array deployed at different axial and radial positions in the fermenter, continuously sampled using sensors such as thermocouples and PT100 resistance thermometers, and stored in a time-series database with timestamps. Alternatively, non-contact temperature measurement can be performed using an infrared thermometer, combined with a data acquisition card and host computer software for data recording and transmission to obtain more comprehensive temperature distribution information. Temperature fluctuation characteristics are used to characterize the severity or stability of temperature changes during fermentation. Besides using the temperature standard deviation over 10 consecutive sampling periods for quantification, temperature fluctuations can also be reflected by calculating the mean absolute deviation (MAD) or coefficient of variation (CV) over N consecutive sampling periods. Furthermore, signal processing methods such as wavelet analysis or Fourier transform can be used to extract specific frequency components or energy distribution features from temperature time-series data to more precisely characterize temperature fluctuation patterns. The raw material synergy score quantifies the degree to which different raw materials mutually promote or inhibit each other during fermentation, and is a key indicator for evaluating the effectiveness of raw material ratios. In addition to calculations based on the similarity of raw material molecular structures, a series of in vitro fermentation experiments can be conducted to measure the impact of different raw material combinations on microbial growth rate, product yield, or substrate degradation efficiency, and then score the results using statistical methods. Additionally, chemometric methods can be used to analyze the relationship between the chemical composition of raw materials (such as cellulose, hemicellulose, and lignin content) and fermentation effects, constructing more complex synergy score models. The rate of change in prediction confidence distribution quantifies the dynamic changes in the uncertainty range of machine learning model prediction results, reflecting the model's adaptability to the current fermentation scenario. Besides calculating the prediction interval width by dividing the absolute value of the difference between the current prediction interval width and the historical baseline prediction interval width by the historical baseline prediction interval width, it can also be measured by calculating the ratio of the current prediction interval width to the mean of the historical prediction interval widths. Furthermore, methods such as Bayesian neural networks can be used to directly output the prediction distribution and calculate the rate of change of its entropy value, or techniques such as Monte Carlo dropout can be used to obtain the prediction distribution through multiple forward propagations and calculate the rate of change of its variance or standard deviation.

[0038] The feature consistency assessment module aims to evaluate the stability of the machine learning model's ranking of the importance of key features (especially those related to feedstock synergy) under different fermentation scenarios. It can be a standalone software component responsible for receiving feature importance data after model training (e.g., obtained through SHAP values, Permutation Importance, etc.), comparing it with historical records, and calculating the ranking volatility. Alternatively, this module can be deployed as a containerized service (e.g., Docker), receiving data via a message queue and asynchronously executing assessment tasks to improve system response speed and processing capacity. The feedstock synergy correlation metric quantifies the correlation between changes in feedstock synergy and changes in microbial activity, serving as a crucial basis for determining the effectiveness of the synergy. Besides using the absolute value of the ratio of the feedstock synergy score change rate to the microbial activity change rate as the correlation metric, statistical methods such as Pearson correlation coefficient and Spearman rank correlation coefficient can be used to calculate the linear or nonlinear correlation between the two. Furthermore, time-series analysis methods such as Granger causality tests can be used to determine whether changes in feedstock synergy are the cause of changes in microbial activity, thereby quantifying the correlation more deeply. Feature importance ranking volatility quantifies the stability of the importance ranking of features (especially those related to feedstock synergistic effects) in a machine learning model, reflecting the model's robustness in identifying key factors in the fermentation process. Besides calculating it using the absolute value of the difference between the current feature importance ranking position and the historical average ranking position divided by the total number of features, Kendall's Tau or Spearman's Rho rank correlation coefficients can be used to compare the similarity between the current feature importance ranking and the historical average ranking. Additionally, the L1 or L2 norm change of the feature importance ranking can be calculated to quantify the degree of ranking volatility.

[0039] This application's solution introduces a confidence stability assessment module and a feature consistency assessment module to conduct a refined and quantitative evaluation of the machine learning model's operational status in a fermentation scenario. Specifically, the confidence stability assessment module first calculates temperature fluctuation characteristics based on temperature time-series data provided by the fermentation process dynamic monitoring unit. This characteristic is characterized by the temperature standard deviation over 10 consecutive sampling periods, enabling real-time capture of the temperature dynamics of the fermentation environment. Building upon this, the module further combines the raw material synergy effect score and the calculated temperature fluctuation characteristics to calculate the rate of change in the machine learning model's prediction confidence distribution. This rate of change is quantified by the ratio of the absolute value of the difference between the current prediction interval width and the historical baseline prediction interval width to the historical baseline prediction interval width, thus dynamically reflecting the model's prediction uncertainty in temperature-sensitive or raw material synergy effect change scenarios. Simultaneously, the feature consistency assessment module calculates the volatility of the machine learning model's feature importance ranking in response to changes in raw material synergy effects, based on the raw material synergy effect correlation degree provided by the fermentation process dynamic monitoring unit. This volatility is quantified by the ratio of the absolute value of the difference between the current feature importance ranking position and the historical average ranking position to the total number of features. This accurately identifies the model's stability in recognizing key influencing factors in the fermentation process (especially feedstock synergistic effects). These two evaluation modules work together to transform key dynamic factors in the fermentation process (temperature fluctuations and changes in feedstock synergistic effects) into quantifiable indicators of model performance. In this way, the model performance evaluation unit can comprehensively and objectively assess the adaptability of the machine learning model in complex fermentation scenarios from two dimensions: the reliability of prediction results and the stability of feature recognition. This refined quantitative evaluation overcomes the limitations of traditional evaluation methods that rely on subjective experience or simple thresholds. It provides a solid data foundation and decision-making basis for the subsequent fermentation scenario control unit to execute matching model control strategies based on the model's performance status, thereby ensuring the robustness and effectiveness of the fermentation ratio optimization system.

[0040] The following is a concrete example. In a machine learning-based fermented fiber feed formulation optimization system, the model operation status evaluation unit can integrate a confidence stability evaluation module and a feature consistency evaluation module. One specific implementation example is as follows: The confidence stability evaluation module can be a software component written in Python. It continuously receives time-series temperature data from a multi-point temperature monitoring array, for example, collecting temperature data every 5 minutes. This module maintains a sliding window, for example, containing the most recent 50 temperature sampling points (corresponding to 10 consecutive sampling periods, each period being 5 minutes), and uses the NumPy library to calculate the temperature standard deviation of these sampling points as a temperature fluctuation feature. Simultaneously, this module also receives the raw material synergistic effect score calculated by the raw material molecular structure real-time analysis module. When the machine learning model (e.g., a fermentation effect prediction model based on XGBoost) completes a prediction, this module obtains the model's prediction result and its confidence interval (e.g., calculated from the model's output prediction mean and standard deviation). It compares the width of the current prediction interval with the pre-set historical baseline prediction interval width to calculate the rate of change of the prediction confidence distribution. The feature consistency evaluation module can be a separate Java service. It periodically acquires data on the correlation between raw material synergistic effects from the fermentation process dynamic monitoring unit. When the machine learning model is retrained or periodically evaluated, this module obtains the model's importance ranking for all input features (e.g., calculated using SHAP values). This module compares the current feature importance ranking with the stored historical average ranking. Through the collaborative work of these two modules, the system can obtain real-time, quantitative indicators of the machine learning model's prediction confidence stability and feature importance consistency in the current fermentation scenario, providing a precise basis for subsequent intelligent regulation.

[0041] This application further proposes a fermentation scenario control unit comprising: a process state determination module, used to determine the current fermentation scenario's process state based on the predicted confidence distribution change rate and the feature importance ranking volatility, wherein: when both the predicted confidence distribution change rate and the feature importance ranking volatility are not greater than a preset benchmark value, the process state is determined to be a stable state; when the predicted confidence distribution change rate is greater than the preset benchmark value, and the feature importance ranking volatility is not greater than the preset benchmark value, the process state is determined to be a temperature-sensitive state; when the predicted confidence distribution change rate is not greater than the preset benchmark value, and the feature importance ranking volatility is greater than the preset benchmark value, the process state is determined to be a synergistic effect fluctuation state; when both the predicted confidence distribution change rate and the feature importance ranking volatility are greater than the preset benchmark value, the process state is determined to be a synergistic effect failure state; and a control strategy execution module, used to trigger the corresponding model control strategy based on the process state determination result.

[0042] The core function of the process state determination module is to identify the specific process state of the current fermentation process based on the stability of the prediction confidence and the consistency of feature importance in the evaluation results of the machine learning model in the fermentation scenario, namely the rate of change of prediction confidence distribution and the volatility of feature importance ranking. Its role is to provide a basis for subsequent precise control decisions. As one implementation method, this module can be a rule-based expert system that predefines the mapping relationship between different indicator combinations and process states; or it can be a lightweight classification model, such as a decision tree or support vector machine, which learns and establishes the correlation pattern between the rate of change of prediction confidence distribution, the volatility of feature importance ranking, and various process states by training on historical fermentation data. The control strategy execution module is responsible for triggering predefined model control strategies based on the output of the process state determination module, i.e., the process state of the current fermentation scenario. Its role is to ensure that the system can specifically address different fermentation scenario problems and achieve dynamic adaptive adjustment of the model. As one implementation approach, this module can be a strategy scheduler, internally maintaining a mapping table from process states to specific control strategies. When a specific process state instruction is received, the corresponding strategy function is invoked for processing. Alternatively, it can be an event-driven system where, when the process state determination module issues a specific state event, the control strategy executor that has pre-subscribed to the event is activated, thereby executing the corresponding control operation. This application maps the numerical relationship between the rate of change of predicted confidence distribution and the volatility of feature importance ranking to four specific process states by setting a series of condition combinations: stable state, temperature-sensitive state, synergistic effect volatility state, and synergistic effect failure state. These logical rules are key to the system's accurate identification of different abnormal modes in the fermentation process. As one implementation approach, these determination logics can be implemented using conditional statements in the software (such as if-else if-else structures) to compare the magnitude of two indicators with a preset benchmark value; or, they can be implemented using lookup tables or decision matrices to directly map the discretized intervals of two indicators to the corresponding process states. The preset benchmark value is a key threshold used to distinguish between normal and abnormal fluctuations in the fermentation process, and its setting directly affects the sensitivity and accuracy of process state determination. As one approach, the preset benchmark value can be determined by statistical analysis of a large amount of historical fermentation data and by combining the experience of domain experts. For example, the mean and standard deviation of the rate of change of prediction confidence distribution and the volatility of feature importance ranking in normal fermentation batches can be calculated, and the benchmark value can be set as the mean plus or minus a certain multiple of the standard deviation. Alternatively, it can adopt an adaptive adjustment mechanism, which can dynamically learn and adjust through machine learning methods (such as anomaly detection algorithms) to adapt to the long-term changing trend of the fermentation process and environmental disturbances.

[0043] This application's solution constructs an intelligent control mechanism capable of dynamically adapting to changes in the fermentation process by introducing a process state determination module and a control strategy execution module. Specifically, the system first receives two key indicators output by the model operation status evaluation unit: the rate of change of prediction confidence distribution and the volatility of feature importance ranking. Subsequently, the process state determination module compares these two indicators with preset benchmark values ​​according to preset logical rules. For example, when both the rate of change of prediction confidence distribution and the volatility of feature importance ranking are not greater than the preset benchmark values, the system determines that the current fermentation scenario is in a stable state, indicating that the model prediction is stable and the consistency of feature importance is good; if the rate of change of prediction confidence distribution is greater than the preset benchmark value but the volatility of feature importance ranking is not greater than the preset benchmark value, it is determined to be a temperature-sensitive state, indicating that the stability of model prediction is affected by temperature fluctuations; if the rate of change of prediction confidence distribution is not greater than the preset benchmark value but the volatility of feature importance ranking is greater than the preset benchmark value, it is determined to be a synergy effect fluctuation state, indicating that the feature importance of the raw material synergy effect has changed significantly; if both are greater than the preset benchmark values, it is determined to be a synergy effect failure state, indicating that there are serious problems with both model prediction and feature consistency. Once the process state is accurately determined, the control strategy execution module will trigger a specific model control strategy matching the determination result. This mechanism enables the system to specifically identify different abnormal patterns in the fermentation process and quickly initiate corresponding countermeasures, thereby avoiding the problems of inaccurate ratio optimization and large batch quality fluctuations caused by the lack of accurate state identification and targeted control in traditional methods. Through this close collaboration, the solution in this application achieves refined management and dynamic optimization of the fermentation process, significantly improving the adaptability and robustness of machine learning models in complex fermentation scenarios.

[0044] The following is a concrete example to illustrate this. Assume that during a fermentation process, the model operation status evaluation unit continuously monitors and outputs the rate of change of predicted confidence distribution and the volatility of feature importance ranking. The preset baseline value is set to 0.1. At one point in time, the process status determination module receives a predicted confidence distribution change rate of 0.05 and a feature importance ranking volatility of 0.03. Since both values ​​are not greater than the preset baseline value of 0.1, the process status determination module classifies the current fermentation scenario as a stable state. At this time, the control strategy execution module will trigger the model parameter maintenance strategy, that is, maintain the parameters of the current machine learning model unchanged and continue to optimize the fermentation ratio. At another point in time, the temperature during the fermentation process fluctuates significantly, causing the predicted confidence distribution change rate calculated by the model operation status evaluation unit to be 0.15, while the feature importance ranking volatility remains at 0.04. At this time, the process status determination module finds that the predicted confidence distribution change rate (0.15) is greater than the preset baseline value (0.1), while the feature importance ranking volatility (0.04) is not greater than the preset baseline value (0.1). Therefore, the process state determination module classifies the current fermentation scenario as a temperature-sensitive state. Subsequently, the control strategy execution module immediately triggers a temperature-feature weight adjustment strategy to address the impact of temperature fluctuations on the model's predictive stability. For example, in the later stages of fermentation, due to batch differences in raw materials or changes in the microbial community structure, the synergistic effect of raw materials changes significantly. The prediction confidence distribution change rate calculated by the model operation status evaluation unit is 0.08, but the feature importance ranking volatility rate rises to 0.12. Based on this, the process state determination module determines the current fermentation scenario as a synergistic effect fluctuation state. The control strategy execution module then triggers a synergistic effect local calibration strategy to adjust the model feature weights related to the synergistic effect, ensuring that the model can accurately capture new synergistic effect patterns. As can be seen from the above examples, the solution of this application can accurately identify different process states based on the real-time dynamic changes in the fermentation process and automatically trigger corresponding control strategies, thereby achieving intelligent management of the optimized formulation of fermented fiber feed.

[0045] Through the above technical solution, this application can accurately identify different process states during fermentation, including stable state, temperature-sensitive state, synergistic effect fluctuation state, and synergistic effect failure state, based on the prediction confidence distribution change rate and feature importance ranking volatility provided by the model operation state evaluation unit. This refined state determination mechanism enables the system to trigger model control strategies that match the current process state, thereby effectively solving the problems of lack of accurate state identification and targeted control in traditional methods. The solution of this application significantly improves the adaptability and robustness of machine learning models in complex fermentation scenarios, ensures the accuracy and batch stability of fermentation ratio optimization, avoids the problems of inaccurate ratios and large quality fluctuations caused by dynamic changes in the fermentation process, and thus improves the production efficiency and product quality of fermented fiber feed.

[0046] Traditional control strategy execution modules lack explicit triggering logic to execute specific strategies for different process states, resulting in imprecise and inefficient execution of control strategies and an inability to adapt to the dynamic changes in the fermentation process. To address this, this application proposes a triggering logic for the control strategy execution module, specifically: when the process state is determined to be stable, a model parameter maintenance strategy is executed; when the process state is determined to be temperature-sensitive, a temperature-feature weight adjustment strategy is executed; when the process state is determined to be in a state of fluctuating synergistic effects, a synergistic effect local calibration strategy is executed; and when the process state is determined to be in a state of synergistic effect failure, a synergistic effect recalibration strategy is executed.

[0047] The triggering logic of the control strategy execution module refers to a series of pre-defined conditional judgment rules used to determine which model control strategy should be activated under specific fermentation process conditions. Its function is to map complex fermentation process states to corresponding model intervention measures in a one-to-one or many-to-one manner, ensuring the system can respond accurately based on real-time conditions. This logic can be implemented in the software program based on conditional statements (such as if-else structures) or through a lookup table method (using process state as the key and control strategy as the value) for rapid lookup and execution. The model parameter preservation strategy aims to maintain all parameters of the machine learning model unchanged. When the fermentation process is in an ideal or acceptable stable state, no adjustments to the model are needed to avoid unnecessary computational overhead and the introduction of potential instability. One implementation is that after receiving a stable state determination, the system skips all subroutines for model updates or adjustments and continues to use the current model for prediction. Another implementation is that the system can periodically perform minor confirmatory checks on the model parameters without making substantial modifications. The temperature-feature weight adjustment strategy is used to dynamically adjust the weights of temperature-related features in the model to address the impact of temperature fluctuations on model predictions during fermentation. The aim is to compensate for model bias caused by temperature changes, ensuring the model maintains high predictive accuracy even under temperature-sensitive conditions. One approach is to use a pre-defined list of temperature-sensitive features and adjust their weights multiplicatively or additively based on the ratio of temperature fluctuation amplitude to historical average fluctuation amplitude. Another approach is to utilize adaptive learning algorithms to dynamically optimize the weights of relevant features based on the impact of temperature fluctuations on model output error, such as through gradient descent or reinforcement learning mechanisms, enabling the model to maintain optimal performance under new temperature conditions. The synergistic effect local calibration strategy addresses situations where the synergistic effect of raw materials fluctuates locally but has not completely failed during fermentation. Its goal is to fine-tune the weights of features related to the synergistic effect in the model without global model reconstruction to adapt to such local changes. One approach is to identify key features causing synergistic effect fluctuations and, based on their importance, rank the fluctuation rates and then slightly increase or decrease the weights of these specific features. Another approach is to utilize online learning or incremental learning methods to perform small-batch model updates only on input data related to synergistic effects, achieving local calibration. This involves periodically introducing small amounts of new data to fine-tune the model, rather than performing a full retraining. Synergistic effect recalibration strategies are used to address the severe situation where the synergistic effect of raw materials completely fails during fermentation. In this case, the model's understanding of the synergistic effect has deviated significantly from reality, requiring a comprehensive recalibration. The goal is to reconstruct the input features related to the synergistic effect, enabling the model to recapture the correct synergistic relationships.One approach is for the system to trigger a recalculation of the correlation degree of raw material synergy effects, reconstruct the raw material synergy effect matrix based on the new correlation degree, and then replace the model's input features with this new matrix. Another approach is to initiate a completely new model training process, but only iteratively optimize the feature engineering parts related to synergy effects, and validate them using historical data. For example, transfer learning can be used to unfreeze and retrain the layers related to synergy effects in the pre-trained model.

[0048] This application's solution achieves refined and adaptive management of the machine learning model during fermentation by introducing explicit triggering logic for the control strategy execution module within the fermentation scenario control unit. This triggering logic works closely with the process state determination module to form an intelligent decision-making closed loop. Specifically, when the process state determination module determines the current fermentation scenario's process state based on the rate of change in prediction confidence distribution and the volatility of feature importance ranking, the control strategy execution module immediately maps the determination result to the uniquely corresponding model control strategy according to the preset triggering logic. For example, if the state is determined to be stable, a model parameter maintenance strategy is executed, avoiding unnecessary model intervention and maintaining system stability and efficiency. If the state is determined to be temperature-sensitive, a temperature-feature weight adjustment strategy is triggered, indicating that the system recognizes that temperature fluctuations have a significant impact on model predictions and requires targeted adjustments to temperature-related feature weights to compensate for this impact and ensure the model's robustness under temperature changes. When the state is determined to be a synergistic effect fluctuation state, the system executes a synergistic effect local calibration strategy, indicating that the synergistic effect between raw materials has undergone local changes but has not completely failed. In this case, local fine-tuning rather than global reconstruction can efficiently restore the model's accuracy. When a synergistic effect is deemed ineffective, a synergistic effect recalibration strategy is implemented. This indicates a fundamental change in the raw material synergistic effect, requiring a complete reconstruction and calibration of the synergistic effect-related input features in the model to ensure it can recapture the true mechanism of the fermentation process. This differentiated triggering logic based on process state allows the system to intelligently select the most appropriate intervention measures according to the dynamic characteristics of the fermentation process, avoiding the one-size-fits-all control approach of traditional methods and significantly improving the adaptability, accuracy, and stability of the machine learning model in complex fermentation environments. In this way, the system can effectively cope with various uncertainties such as temperature fluctuations and changes in raw material synergistic effects during fermentation, ensuring the continuous effectiveness of fermentation ratio optimization.

[0049] The following is a concrete example to illustrate this. Assume that during the production of fermented fiber feed, the fermentation process dynamic monitoring unit continuously collects time-series data on temperature, microbial activity, and changes in raw material synergistic effects. Based on this data, the model operation status evaluation unit calculates the rate of change in the prediction confidence distribution and the volatility of the feature importance ranking of the machine learning model. For example, in a certain fermentation batch, the process status determination module, based on the calculation results, determines that the current fermentation scenario is in a temperature-sensitive state. This means that the rate of change in the prediction confidence distribution is greater than a preset benchmark value, while the volatility of the feature importance ranking is not greater than a preset benchmark value. At this time, the triggering logic of the control strategy execution module will recognize this state and immediately trigger the temperature-feature weight adjustment strategy. Specifically, the system will calculate the ratio of the current fermentation temperature fluctuation amplitude to the historical average fluctuation amplitude. For example, if the current fluctuation amplitude is 1.5 times the historical average, then this ratio is 1.5. Subsequently, the system will identify the features in the model related to temperature-sensitive raw materials and multiply their weights by 1.5, thereby dynamically adjusting the importance of these features in the model prediction. For example, in another fermentation batch, the process state determination module determines that the current fermentation scenario is in a state of synergistic effect failure, which means that the rate of change in the prediction confidence distribution and the volatility of the feature importance ranking are both greater than the preset benchmark values. In this case, the triggering logic of the control strategy execution module will initiate a synergistic effect recalibration strategy. The system will recalculate the correlation degree of raw material synergistic effects based on the raw material synergistic effect score provided by the fermentation process dynamic monitoring unit. Then, using this new correlation degree data, the system will reconstruct a matrix reflecting the current raw material synergistic relationship and replace the original input feature matrix of the machine learning model with this new matrix. Subsequently, the model will be retrained or fine-tuned based on this updated synergistic effect matrix to adapt to the fundamental changes in the raw material synergistic effect, thereby restoring its prediction accuracy. As can be seen from the above examples, the solution of this application can intelligently select and execute the most appropriate model control strategy according to the real-time dynamic changes in the fermentation process, thereby ensuring the continuous and efficient operation of the fermentation ratio optimization system.

[0050] In some of the solutions described above in this application, a temperature-feature weight adjustment strategy is proposed to optimize model weights when the process is in a temperature-sensitive state. However, in its implementation, a specific implementation method is lacking, resulting in an inability to effectively quantify the impact of temperature fluctuations on feature weights. This may lead to a highly subjective and unadaptable adjustment strategy, failing to accurately respond to dynamic temperature changes during fermentation, thereby affecting the stability of model predictions and the accuracy of ratio optimization. To address this, this application proposes an implementation method for the temperature-feature weight adjustment strategy, which includes: calculating the ratio of the fermentation temperature fluctuation amplitude to the historical average fluctuation amplitude; multiplying the feature weights related to temperature-sensitive raw materials by the ratio to obtain the adjusted feature weights.

[0051] This scheme calculates the ratio of fermentation temperature fluctuation amplitude to the historical average fluctuation amplitude. The aim is to quantify the severity of temperature changes during the current fermentation process and compare it with historical norms, thus providing an objective and quantitative basis for subsequent feature weight adjustments. One implementation method is to obtain the fermentation temperature fluctuation amplitude by calculating the standard deviation or variance of temperature time-series data over a continuous sampling period. For example, a sliding time window (such as the past 1 or 2 hours) can be set, and the standard deviation of temperature readings within that window can be used as the current fluctuation amplitude. The historical average fluctuation amplitude can be obtained by statistically averaging the temperature fluctuation amplitudes within the same time window from multiple past fermentation batches or long-term operational data. Another implementation method is to define the fermentation temperature fluctuation amplitude as the difference between the maximum and minimum temperature values ​​within a specific time period. The historical average fluctuation amplitude is then the historical average of these differences. By calculating the ratio of the current fluctuation amplitude to the historical average fluctuation amplitude, the relative intensity of the current temperature fluctuation can be intuitively reflected.

[0052] Further, the feature weights related to temperature-sensitive raw materials are multiplied by the aforementioned ratio to obtain the adjusted feature weights. The role of this feature is to dynamically adjust the importance of input features related to temperature-sensitive raw materials in the machine learning model based on the quantified temperature fluctuations, ensuring that the model fully considers the impact of temperature changes on the performance of these raw materials when predicting fermentation ratios. As one implementation, in machine learning models (e.g., tree-based models, neural networks, or linear models), each input feature has a weight representing its contribution to the model output. The system pre-identifies, through domain knowledge or data analysis, which raw material characteristics (such as the activity of specific enzymes, the degradation efficiency of nutrients, etc.) exhibit high sensitivity to fermentation temperature changes. After obtaining the aforementioned ratio, the system multiplies the current weights of these identified temperature-sensitive features by this ratio. For example, if the ratio is greater than 1, the weights of these features are increased, indicating that the temperature fluctuations are large and the model should pay more attention to these features; if the ratio is less than 1, the weights are decreased. Another implementation is to establish a dynamic weight adjustment mechanism containing a mapping table that associates specific features of temperature-sensitive raw materials with their initial weights. Once the ratio is calculated, the system updates the weights of these features using a predefined function (e.g., a linear, exponential, or piecewise function) based on that ratio and the mapping table. This adjustment ensures that the model can respond to temperature changes in the fermentation environment in real time and accurately, thereby improving the accuracy and stability of predictions.

[0053] This application's solution addresses the lack of objective basis and insufficient adaptability in model weight adjustment when the fermentation process is in a temperature-sensitive state by providing a specific and quantifiable temperature-feature weight adjustment strategy. Specifically, when the system determines that the current fermentation process is in a temperature-sensitive state through the model operation status evaluation unit, the fermentation scenario control unit triggers a corresponding model control strategy. At this point, this solution first calculates the ratio of the current fermentation temperature fluctuation amplitude to the historical average fluctuation amplitude. This ratio, as a key quantitative indicator, accurately reflects the relative intensity of the current temperature fluctuation, providing data support for subsequent weight adjustments and avoiding bias from subjective judgment. Subsequently, the system multiplies the feature weights related to temperature-sensitive raw materials in the machine learning model by this ratio. This multiplicative adjustment mechanism ensures that the change in feature weights is proportional to the amplitude of temperature fluctuations, enabling the model to dynamically and accurately adjust its focus on temperature-sensitive features. For example, when temperature fluctuations increase, the ratio increases accordingly, thereby increasing the feature weights related to temperature-sensitive raw materials. This makes the model pay more attention to these temperature-dependent factors when predicting fermentation ratios. Conversely, when temperature fluctuations decrease, the ratio decreases, and the weights also decrease accordingly. In this way, this scheme enables the machine learning model to adapt to the dynamic temperature changes during fermentation in real time and quantitatively, significantly improving the stability of the model's prediction confidence and the consistency of feature importance under temperature-sensitive conditions, thus ensuring the accuracy and reliability of fermentation ratio optimization. This weight adjustment mechanism based on quantitative ratios, combined with the system's ability to identify temperature-sensitive conditions, forms an efficient and adaptive closed-loop control system, ensuring the stability of the fermentation process and product quality.

[0054] The following is a concrete example to illustrate this. Consider a fermented fiber feed production scenario where the activity of a specific cellulase is highly sensitive to temperature fluctuations, and this enzyme activity is a crucial input feature in a machine learning model. The system continuously monitors the time-series temperature data within the fermenter. Assume that within a consecutive one-hour period, the system calculates the standard deviation of the fermentation temperature to be 0.8℃, while historical data shows the average standard deviation for this fermentation stage is 0.5℃. The system then calculates the ratio of the fermentation temperature fluctuation amplitude to the historical average fluctuation amplitude, i.e., 0.8 / 0.5 = 1.6. Subsequently, the system identifies features related to the cellulase activity (e.g., enzyme activity indices, substrate degradation rate, etc.) as temperature-sensitive features. Assume that in the current machine learning model, the weight of this cellulase activity feature is 0.6. According to this approach, the system multiplies this feature weight by the calculated ratio, i.e., 0.6 * 1.6 = 0.96. Therefore, the weight of this cellulase activity feature in the model is adjusted to 0.96. This adjusted weight will be used for subsequent machine learning model prediction and fermentation ratio optimization. In this way, when fermentation temperature fluctuates significantly, the importance of features related to temperature-sensitive enzyme activity in the model is dynamically increased, enabling the model to more fully consider the impact of temperature fluctuations on enzyme activity when generating optimized formulations, thereby improving the accuracy of formulations and the stability of fermentation results.

[0055] Through the above technical solution, this application provides a specific and quantifiable temperature-feature weight adjustment strategy, effectively solving the problems of lack of specific implementation methods, susceptibility to subjective influence, and insufficient adaptability in model weight adjustment when the fermentation process is in a temperature-sensitive state. By calculating the ratio of the fermentation temperature fluctuation amplitude to the historical average fluctuation amplitude, this solution provides objective data basis for model weight adjustment, avoiding the uncertainty brought about by traditional experience-based judgment. Multiplying the feature weights related to temperature-sensitive raw materials by this ratio ensures that the model's response to temperature fluctuations is accurate and proportional, enabling the model to more sensitively and accurately capture the impact of temperature changes on key features of the fermentation process. This quantifiable adjustment mechanism, combined with the system's ability to identify temperature-sensitive states, significantly enhances the adaptability and robustness of machine learning models in dynamic fermentation environments, thereby improving the predictive confidence stability and accuracy of fermentation ratio optimization, ultimately contributing to more efficient and stable fermented fiber feed production.

[0056] In some of the embodiments described above in this application, a control strategy execution module is proposed to execute the model control strategy according to the process state. However, when the synergy effect fails, the existing strategy may not be able to effectively recalibrate the synergy effect correlation, resulting in inaccurate model input features, which in turn affects the real-time performance and accuracy of fermentation ratio optimization and causes batch quality fluctuations.

[0057] In response, this application further proposes the following triggering logic for the control strategy execution module: when the process state is determined to be a synergy failure state, a synergy recalibration strategy is executed; the synergy recalibration strategy is implemented as follows: based on the raw material synergy score in the raw material synergy change data, the correlation degree of the raw material synergy effect is recalculated; the raw material synergy effect matrix is ​​reconstructed based on the correlation degree of the raw material synergy effect, and the original matrix is ​​replaced as the model input feature; the synergy local calibration strategy is implemented as follows: based on the feature importance ranking volatility, the feature weights related to the synergy effect are dynamically adjusted.

[0058] The regulation strategy execution module is the core component of the system, taking corresponding actions based on the judgment results of the model operation status evaluation unit. Its triggering logic defines which regulation strategy should be executed under specific process conditions. When the system determines that the current fermentation scenario's process state is a synergy failure state, it means that the synergy between raw materials during fermentation has seriously deviated from expectations, severely impacting the machine learning model's ability to optimize the fermentation ratio. At this critical moment, executing the synergy recalibration strategy is necessary to thoroughly correct the model's understanding of the synergy effect and ensure that the model can perform subsequent optimizations based on accurate synergy information. This triggering logic can be implemented in several ways. For example, a rule-based expert system can be used, with a series of pre-set conditional judgment statements. When both the rate of change in the prediction confidence distribution and the volatility of the feature importance ranking are greater than the pre-set benchmark value, it is determined to be a synergy failure state, thus triggering the synergy recalibration strategy. Alternatively, it can be implemented through a state machine mechanism, defining different process states as different state nodes. When the system detects that the conditions for a synergy failure state are met, the state machine automatically transitions to the state of executing the synergy recalibration strategy.

[0059] The synergy recalibration strategy aims to thoroughly reconstruct the machine learning model's understanding of feedstock synergy. Feedstock synergy scores reflect the similarity of feedstock molecular structures and are fundamental data for assessing synergy among feedstocks. When the fermentation process enters a state of synergy failure, the model's understanding of the synergy may be severely distorted. Therefore, it is necessary to reassess the degree of correlation between feedstocks using the latest and most reliable feedstock synergy score data. Recalculation ensures that the obtained correlation accurately reflects the actual synergistic effect in the current fermentation scenario, thus providing an accurate basis for subsequent model correction. Various methods can be used to recalculate feedstock synergy correlation. For example, a sliding window averaging method can be used, collecting the latest feedstock synergy scores from consecutive fermentation batches and combining them with the rate of change in microbial activity to recalculate the absolute value of the ratio as a quantifiable correlation measure. Another approach is to use an exponentially weighted moving average method, weighting historical and current feedstock synergy scores to more sensitively reflect recent trends and recalculating the correlation accordingly, thus ensuring the timeliness and accuracy of the correlation calculation.

[0060] The raw material synergy matrix is ​​a key input feature for machine learning models to understand and utilize raw material synergies. In the synergy recalibration strategy, after the synergy correlation is recalculated, these new correlation values ​​need to be organized into a matrix and provided as updated input features to the machine learning model. The purpose of replacing the original matrix is ​​to ensure that the model can learn and predict based on the latest and most accurate synergy information, avoiding the use of outdated or incorrect features, thereby improving the model's adaptability and prediction accuracy in dynamic fermentation environments. Specifically, the raw material synergy matrix can be reconstructed by filling the recalculated synergy correlations into an N*N symmetric matrix (where N is the number of raw material types), where the elements (i, j) represent the correlation between raw material i and raw material j. Alternatively, a sparse matrix can be constructed to store only raw material pairs with significant correlations, reducing computation and storage space, and then used as model input, thereby improving model efficiency.

[0061] The synergy effect local calibration strategy is used to fine-tune the machine learning model when the synergy effect fluctuates but has not completely failed. Feature importance ranking volatility reflects the stability of the model's focus on synergy effect features. When this volatility is high, it indicates that the model's understanding of synergy effect features may be biased, but not yet to the point of requiring full recalibration. In this case, dynamically adjusting the weights of these synergy effect-related features in the model can enhance the model's adaptability to changes in synergy effect, allowing it to maintain good predictive performance without large-scale retraining. Dynamically adjusting feature weights can employ various algorithms. For example, gradient descent can be used to iteratively adjust the weights of synergy effect-related features based on the magnitude of the feature importance ranking volatility, making them more consistent with the actual situation of the current fermentation scenario. Another approach is to use reinforcement learning-based methods, using feature importance ranking volatility as a reward signal to dynamically optimize the weights of synergy effect-related features through learning, thereby achieving more intelligent and adaptive weight adjustments.

[0062] This application's solution employs a refined control strategy to ensure that the machine learning model can adapt in real-time to the dynamic changes in raw material synergy effects during the optimization of fermented fiber feed formulations. When the model operation status evaluation unit determines, based on raw material synergy effect change data, that the machine learning model's prediction confidence stability and feature importance consistency in the fermentation scenario are problematic, the fermentation scenario control unit will determine the current fermentation scenario's process status according to the judgment result. Specifically, when the process status determination module determines that the current fermentation scenario's process status is a synergy effect failure state, the control strategy execution module will immediately trigger a synergy effect recalibration strategy. This strategy first recalculates the raw material synergy effect correlation degree based on the raw material synergy effect score in the raw material synergy effect change data collected in real-time by the fermentation process dynamic monitoring unit. This step ensures the timeliness and accuracy of the correlation degree data, avoiding bias caused by data lag. Subsequently, the system will reconstruct the raw material synergy effect matrix based on these recalculated raw material synergy effect correlation degrees and replace the original model input feature matrix with it. Through this series of operations, the machine learning model can obtain the latest and most accurate raw material synergy effect information as input, thereby thoroughly correcting the model's understanding of the synergy effect and enabling it to accurately optimize the fermentation formulation again. Furthermore, when the process state determination module determines that the current fermentation scenario is in a state of fluctuating synergistic effects, the control strategy execution module will trigger a local calibration strategy for synergistic effects. This strategy dynamically adjusts the feature weights related to synergistic effects based on the feature importance ranking volatility provided by the model running state evaluation unit. This local adjustment mechanism allows the model to perform refined optimization when there are slight fluctuations in synergistic effects, enhancing the model's robustness and adaptability to subtle changes without the need for comprehensive recalibration. Through the synergistic application of the above two strategies, the solution of this application can take appropriate and efficient control measures for synergistic effect problems of different degrees. Whether it is a thorough recalibration or a refined local calibration, the aim is to ensure that the machine learning model can always accurately capture the raw material synergistic effects in the fermentation process, thereby maintaining the real-time performance and accuracy of fermentation ratio optimization and effectively avoiding batch quality fluctuations caused by synergistic effect problems.

[0063] The following is a concrete example. Assume a fermentation process for fiber feed is underway, with corn stalks, soybean meal, and wheat bran undergoing mixed fermentation in a fermenter. The fermentation process dynamic monitoring unit continuously collects time-series data on temperature, microbial activity, and changes in raw material synergistic effects. In a particular fermentation batch, the model operation status evaluation unit detects a significant decrease in both the stability of the machine learning model's prediction confidence and the consistency of feature importance; for example, the rate of change in prediction confidence distribution and the volatility of feature importance ranking both exceed preset benchmark values. Based on this, the process status determination module determines that the current fermentation scenario is in a synergistic effect failure state. At this point, the control strategy execution module immediately triggers a synergistic effect recalibration strategy. The system obtains the latest raw material synergistic effect score from the fermentation process dynamic monitoring unit. This score may be calculated based on the matching degree between the raw material molecular structure and the cellulase active site, as analyzed in real-time by a Raman spectrometer and a molecular docking calculation unit. For example, the system finds a significant change in the matching degree of cellulase hydrolysis sites between corn stalks and soybean meal. Based on these updated scores, the system recalculates the synergistic effects of corn stalks with soybean meal, corn stalks with wheat bran, and soybean meal with wheat bran. For example, it calculates the absolute value of the ratio of the change rate of their respective synergistic effect scores to the change rate of microbial activity. Subsequently, the system uses these new correlation values ​​to construct a new 3x3 synergistic effect matrix and replaces the original synergistic effect input matrix in the machine learning model. Through this process, the model can optimize the fermentation ratio based on the latest and most accurate synergistic effect information. Alternatively, if the model operation status evaluation unit only detects that the feature importance ranking volatility exceeds a preset benchmark value, while the prediction confidence distribution change rate is still within an acceptable range, the process status determination module will determine that the current fermentation scenario is in a synergistic effect fluctuation state. At this time, the control strategy execution module will trigger a synergistic effect local calibration strategy. The system will dynamically adjust the feature weights related to the synergistic effects of corn stalks, soybean meal, and wheat bran based on the feature importance ranking volatility provided by the feature consistency evaluation module. For example, if the feature importance ranking of the synergistic effect between corn stalks and soybean meal fluctuates significantly, the system may slightly increase the weight of this feature in the model to enhance its influence on the model's prediction results, thereby responding more sensitively to such fluctuations without requiring a complete model reconstruction.

[0064] Through the above technical solution, this application effectively solves the problem that when the synergy effect fails, the machine learning model cannot effectively recalibrate the correlation of the synergy effect, resulting in inaccurate model input features, which in turn affects the real-time performance and accuracy of fermentation ratio optimization, leading to batch quality fluctuations. Specifically, when the fermentation scenario is determined to be in a state of synergy effect failure, the control strategy execution module can promptly trigger the synergy effect recalibration strategy, recalculate the correlation based on the latest raw material synergy effect score, and update the model input features, ensuring that the model always uses accurate and timely synergy effect information for prediction. This avoids model bias caused by data lag or errors, significantly improving the accuracy of fermentation ratio optimization. Simultaneously, for synergy effect fluctuations, the synergy effect local calibration strategy proposed in this application can dynamically adjust the weights of relevant features based on the volatility rate ranked by feature importance, achieving more refined fine-tuning of the model. This hierarchical control mechanism enables the system to take the most appropriate response according to the severity of the problem, avoiding unnecessary full recalibration while ensuring the robustness of the model in the face of minor fluctuations. Overall, the proposed solution significantly enhances the adaptability of the machine learning model to the dynamic changes in the synergistic effect of raw materials during fermentation by flexibly applying synergistic effect recalibration and local calibration strategies. This ensures the real-time, accuracy, and stability of the fermentation fiber feed formulation optimization and effectively reduces the risk of batch quality fluctuations.

[0065] In some of the embodiments described above in this application, a closed-loop optimization feedback unit is proposed to update the judgment conditions of the model running status evaluation unit. However, in its implementation, the preset benchmark value and the correlation strength threshold are fixed and cannot adapt to the dynamic changes in the fermentation process, resulting in unstable model optimization effect and increased batch-to-batch quality fluctuations.

[0066] To address this, this application further proposes an improvement to the closed-loop optimization feedback unit. This unit receives and outputs data on the adjusted fermentation ratio and its implementation effect, and updates the judgment conditions of the model operation status evaluation unit. To solve the aforementioned problems, the closed-loop optimization feedback unit specifically includes a dynamic benchmark update module and a dynamic correlation strength generation module.

[0067] The dynamic benchmark update module is a key component, responsible for adaptively adjusting the preset benchmark values ​​used within the system based on actual fermentation performance data. This module can be a software algorithm unit, for example, by implementing an adaptive algorithm based on feedback control, such as a proportional-integral-derivative (PID) controller or a fuzzy logic controller. It calculates and outputs the adjusted preset benchmark values ​​in real time based on the deviation between the received fermentation ratio implementation performance data (including the rate of change in crude fiber digestibility and the rate of change in microbial activity stability) and the preset targets. Alternatively, this module can employ statistical process control (SPC) methods, dynamically updating the upper and lower limits of the control chart through cumulative analysis of historical fermentation batch performance data to obtain new preset benchmark values. Fermentation ratio implementation performance data is a key indicator for measuring the effectiveness of fermentation process optimization. The rate of change in crude fiber digestibility reflects the degree of improvement in feed nutritional value, while the rate of change in microbial activity stability indicates the health of the fermentation environment and the vitality of the microbial community.

[0068] The dynamic association strength generation module is responsible for dynamically generating association strength thresholds between feature importance ranking volatility and microbial activity change rate based on historical fermentation scenario data. This module can be a data analysis and modeling unit; for example, by deploying a time series prediction model (such as an ARIMA model or a Long Short-Term Memory (LSTM) network) to learn the long-term trends and interrelationships between feature importance ranking volatility and microbial activity change rate in historical fermentation scenarios, and dynamically predict or adjust the association strength threshold accordingly. Alternatively, this module can employ clustering algorithms from machine learning to perform pattern recognition on historical data, classify data under different association strength patterns, and select the most suitable association strength threshold based on the real-time characteristics of the current fermentation scenario.

[0069] The following is a specific example to illustrate this. As a concrete implementation, the dynamic baseline update module can employ an adaptive adjustment strategy based on performance indicators. For example, after implementing a single fermentation ratio adjustment, the system monitors that the change rate of crude fiber digestibility is lower than the target value by 2%, and the change rate of microbial activity stability also shows a downward trend. At this time, the dynamic baseline update module will, according to preset adjustment rules, for example, lower the preset baseline value used to judge the change rate of prediction confidence distribution by 10%, to improve the model's sensitivity to fluctuations in prediction uncertainty, prompting the model to identify potential risks earlier in subsequent evaluations. Simultaneously, the dynamic correlation strength generation module can periodically (e.g., every 10 batches of fermentation or every week) re-analyze historical fermentation scenario data. For example, this module can run a sliding window-based multiple regression model to analyze the correlation between the volatility of feature importance ranking and the change rate of microbial activity in the past 50 batches of fermentation data. If the correlation between the two is found to be significantly weakened within a certain period of time, it indicates that the model's assessment of feature importance may be out of touch with actual changes in microbial activity. In this case, the module will appropriately increase the correlation strength threshold to prompt the model operation status assessment unit to more rigorously examine the fluctuations in feature importance ranking, thereby avoiding misjudgments caused by weakened correlation.

[0070] Through the above technical solutions, this application effectively solves the problem that fixed preset benchmark values ​​and correlation strength thresholds cannot adapt to dynamic changes in the fermentation process, leading to unstable model optimization results and large batch-to-batch quality fluctuations. The dynamic benchmark update module can adaptively adjust the preset benchmark value based on the implementation effect data of the adjusted fermentation ratio, such as the change rate of crude fiber digestibility and the change rate of microbial activity stability. This allows the system to self-correct based on actual fermentation performance, avoiding evaluation bias caused by fixed thresholds. The dynamic correlation strength generation module dynamically generates the correlation strength threshold between the feature importance ranking fluctuation rate and the microbial activity change rate based on historical fermentation scenario data. This allows the system to fully utilize historical experience and flexibly respond to correlation changes under different fermentation stages and environmental conditions, improving the model's accuracy in identifying fluctuations in the importance of key features. Overall, these dynamic adjustment mechanisms significantly enhance the judgment accuracy and robustness of the model's operational status evaluation unit, enabling the entire fermented fiber feed ratio optimization system to adapt more accurately and stably to the complex dynamic changes in the fermentation process, thereby effectively improving the optimization effect of the fermentation ratio and the batch quality stability of the final product.

[0071] In some of the solutions mentioned above in this application, a dynamic monitoring unit for the fermentation process is proposed to collect real-time data on temperature, microbial activity, and changes in the synergistic effect of raw materials during the fermentation process. However, in this process, if the spatial coverage of the monitoring unit is insufficient, resulting in incomplete temperature data, untimely calculation of microbial activity affecting real-time performance, or inaccurate analysis of raw material structure reducing the accuracy of the synergistic effect score, it will lead to incomplete and unreliable data collection, which in turn affects the accuracy of subsequent model evaluation and control, and cannot effectively adapt to the dynamic changes in the fermentation process.

[0072] In this regard, this application further proposes that the dynamic monitoring unit for the fermentation process includes a multi-point temperature monitoring array, an online microbial activity analysis module, and a real-time raw material molecular structure analysis module. The multi-point temperature monitoring array is deployed at different positions along the axial and radial directions of the fermenter to collect spatial distribution data of fermentation temperature; the online microbial activity analysis module includes an optical density sensor and a metabolite gas chromatography unit for real-time calculation of microbial activity; the real-time raw material molecular structure analysis module includes a Raman spectrometer and a molecular docking calculation unit for real-time calculation of the similarity of raw material molecular structures, specifically the matching degree of cellulase hydrolysis sites, calculated based on the matching between the raw material molecular structure and the cellulase active sites.

[0073] The multi-point temperature monitoring array is a network of multiple temperature sensors strategically distributed at different spatial locations within the fermenter. Its function is to acquire detailed spatial distribution information of the temperature field inside the fermenter, rather than temperature values ​​at a single or a few points. This array can employ a thermocouple array, deploying thermocouples at multiple preset points along the axial and radial directions of the fermenter using fixed supports. Each thermocouple is connected to a data acquisition module, enabling synchronous acquisition of multi-channel temperature data. Alternatively, a distributed fiber optic temperature sensor can be used, laying optical fibers along the interior of the fermenter and using principles such as fiber Bragg gratings or Raman scattering to achieve continuous temperature monitoring along the line, thereby obtaining more refined spatial temperature distribution data. The online microbial activity analysis module is an integrated system specifically designed for real-time monitoring and evaluation of the physiological state and metabolic activity of microorganisms during fermentation. Its function is to provide immediate feedback on microbial activity, ensuring accurate control of the fermentation process. Specifically, the optical density sensor can reflect the microbial cell concentration by measuring the absorbance of the fermentation broth at a specific wavelength, thereby indirectly assessing the growth status and activity of the microorganisms. The gas chromatography unit for metabolites can be used to analyze the types and concentrations of volatile organic acids, alcohols, and gases produced by microorganisms during fermentation, quantifying the metabolic activity of microorganisms based on the production rate of specific metabolites. The real-time molecular structure analysis module for raw materials is a system capable of instantly acquiring molecular structure information of fermentation raw materials and assessing their interaction potential with enzymes. Its role is to ensure an accurate understanding of raw material characteristics and provide precise molecular-level data for subsequent calculations of synergistic effects. The Raman spectrometer can provide molecular structural characteristics of major components in the raw materials, such as cellulose, hemicellulose, and lignin, by analyzing Raman scattering spectra generated by molecular vibrations and rotations. The molecular docking computation unit is a computational chemistry-based method used to simulate the interaction patterns and binding strength between raw material molecules and cellulase active sites. By evaluating parameters such as the spatial matching degree and binding energy of different raw material molecules with enzyme active sites, the matching degree of cellulase hydrolysis sites is quantified.

[0074] In one specific implementation, multiple layers of temperature monitoring points can be uniformly arranged along the axial direction of a fermenter, with multiple monitoring points in the radial direction for each layer, forming a multi-point temperature monitoring array. These sensors can be connected to a data acquisition unit via an industrial bus, and the data acquisition unit periodically uploads the temperature data from all points to the central control system. Simultaneously, an optical density sensor can be integrated into the fermenter's circulation pipeline for continuous measurement of the optical density value of the fermentation broth. A gas sampling port is located at the top of the fermenter, and fermentation gases are periodically introduced into a gas chromatograph via an autosampler for analysis of specific metabolites. The optical density values ​​and gas analysis results are transmitted in real-time to a data processing unit via a data interface for calculating the microbial growth rate and metabolic intensity. Furthermore, an online Raman spectroscopy probe can be installed near the feed inlet to perform real-time spectral scanning of the feed entering the fermenter, acquiring its molecular fingerprint spectrum. A pre-trained molecular docking model runs on a backend server. This model identifies key cellulose components in the feed based on Raman spectroscopy data and calculates the matching degree between these components and preset cellulase active sites. The calculation results are output in the form of cellulose enzymatic hydrolysis site matching degree, which serves as the input for the raw material synergistic effect score.

[0075] Traditional fermented fiber feed technologies struggle to optimize formulations due to the dynamic changes in the fermentation process, making it difficult for machine learning models to adapt in real time. This results in unstable prediction confidence and inconsistent feature importance, ultimately affecting formulation accuracy and batch quality stability.

[0076] For this, please refer to Figure 2 As shown, this application proposes a machine learning-based method for optimizing the formulation of fermented fiber feed, which includes the following steps: S1: Real-time acquisition of temperature time-series data, microbial activity, and raw material synergistic effect changes during the fermentation process; The raw material synergistic effect change data includes raw material synergistic effect score and raw material synergistic effect correlation degree. The raw material synergistic effect score is calculated based on the similarity of raw material molecular structure. The raw material synergistic effect correlation degree is calculated based on the correlation between the rate of change of raw material synergistic effect score and the rate of change of microbial activity during the continuous fermentation period. The correlation calculation is specifically: the absolute value of the ratio of the rate of change of raw material synergistic effect score to the rate of change of microbial activity is used as the correlation quantification value. S2: Based on the data on the changes in the synergistic effect of the raw materials, determine the stability of the prediction confidence of the machine learning model in the fermentation scenario and the consistency of the feature importance. S3: Based on the judgment result of step S2, determine the process state of the current fermentation scenario and trigger the model control strategy that matches the process state. S4: Receive and output the adjusted fermentation ratio and the fermentation ratio implementation effect data, and update the judgment conditions of the model operation status evaluation unit.

[0077] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A machine learning-based system for optimizing the formulation of fermented fiber feed, characterized in that, include: The fermentation process dynamic monitoring unit is used to collect real-time data on temperature time series, microbial activity, and changes in raw material synergistic effects during the fermentation process. The raw material synergistic effect change data includes raw material synergistic effect score and raw material synergistic effect correlation degree. The raw material synergistic effect score is calculated based on the similarity of raw material molecular structure. The raw material synergistic effect correlation degree is calculated based on the correlation between the rate of change of raw material synergistic effect score and the rate of change of microbial activity during the continuous fermentation period. The correlation calculation is specifically: the absolute value of the ratio of the rate of change of raw material synergistic effect score to the rate of change of microbial activity is used as the correlation quantification value. The model operation status evaluation unit is used to determine the stability of the prediction confidence and the consistency of feature importance of the machine learning model in the fermentation scenario based on the raw material synergistic effect change data. The fermentation scenario control unit is used to determine the current process state of the fermentation scenario based on the judgment result of the model operation state evaluation unit, and to execute a model control strategy that matches the process state. The closed-loop optimization feedback unit is used to receive and output the adjusted fermentation ratio and the fermentation ratio implementation effect data, and update the judgment conditions of the model operation status evaluation unit.

2. The machine learning-based fermented fiber feed formulation optimization system according to claim 1, characterized in that, The model operation status evaluation unit includes: The confidence stability assessment module is used to calculate temperature fluctuation characteristics based on the temperature time series data, and to calculate the rate of change of the prediction confidence distribution of the machine learning model based on the raw material synergistic effect score and the temperature fluctuation characteristics. The temperature fluctuation characteristics are the temperature standard deviation over 10 consecutive sampling periods, and the rate of change of the prediction confidence distribution is the absolute value of the difference between the current prediction interval width and the historical benchmark prediction interval width / the historical benchmark prediction interval width. The feature consistency assessment module is used to calculate the volatility of the feature importance ranking of the machine learning model for changes in the raw material synergy effect based on the correlation degree of the raw material synergy effect. The volatility of the feature importance ranking is the absolute value of the difference between the current feature importance ranking position and the historical average ranking position / the total number of features.

3. The machine learning-based fermented fiber feed formulation optimization system according to claim 2, characterized in that, The fermentation scenario control unit includes: The process status determination module is used to determine the process status of the current fermentation scenario based on the predicted confidence distribution change rate and the feature importance ranking volatility, wherein: When both the rate of change of the predicted confidence distribution and the volatility of the feature importance ranking are not greater than the preset benchmark value, the process is determined to be in a stable state. When the rate of change of the predicted confidence distribution is greater than the preset benchmark value, and the volatility of the feature importance ranking is not greater than the preset benchmark value, the process state is determined to be a temperature-sensitive state. When the rate of change of the predicted confidence distribution is not greater than the preset benchmark value, and the volatility of the feature importance ranking is greater than the preset benchmark value, the process state is determined to be a synergistic effect volatility state. When both the rate of change of the predicted confidence distribution and the volatility of the feature importance ranking are greater than the preset benchmark value, the process state is determined to be a state of synergistic effect failure. The control strategy execution module is used to trigger the corresponding model control strategy based on the determination result of the process state.

4. The machine learning-based fermented fiber feed formulation optimization system according to claim 3, characterized in that, The triggering logic of the control strategy execution module is as follows: When the process is determined to be in a stable state, the model parameter preservation strategy is executed. When the process state is determined to be temperature sensitive, the temperature-feature weight adjustment strategy is executed. When the process status is determined to be a state of synergistic effect fluctuation, a synergistic effect local calibration strategy is executed. When the process status is determined to be a synergy failure state, the synergy recalibration strategy is executed.

5. The machine learning-based fermented fiber feed formulation optimization system according to claim 4, characterized in that, The temperature-feature weight adjustment strategy is implemented as follows: Calculate the ratio of the fermentation temperature fluctuation range to the historical average fluctuation range; The adjusted feature weights are obtained by multiplying the feature weights associated with temperature-sensitive raw materials by the ratio.

6. The machine learning-based fermented fiber feed formulation optimization system according to claim 4, characterized in that, The synergistic effect recalibration strategy is implemented as follows: Based on the raw material synergy score in the raw material synergy change data, the correlation degree of raw material synergy is recalculated. Based on the aforementioned raw material synergy correlation, the raw material synergy matrix is ​​reconstructed and replaced with the original matrix as the model input feature; The synergistic effect local calibration strategy is implemented as follows: Based on the volatility of the aforementioned feature importance ranking, the feature weights related to synergistic effects are dynamically adjusted.

7. The machine learning-based fermented fiber feed formulation optimization system according to claim 6, characterized in that, The closed-loop optimization feedback unit includes: The dynamic benchmark update module is used to dynamically adjust the preset benchmark value based on the implementation effect data of the regulated fermentation ratio; the implementation effect data of the fermentation ratio includes the change rate of crude fiber digestibility and the change rate of microbial activity stability. The dynamic correlation strength generation module is used to dynamically generate the correlation strength threshold between the feature importance ranking volatility and the microbial activity change rate based on historical fermentation scenario data.

8. The machine learning-based fermented fiber feed formulation optimization system according to claim 7, characterized in that, The fermentation process dynamic monitoring unit includes: A multi-point temperature monitoring array is deployed at different positions along the axial and radial axes of the fermenter to collect spatial distribution data of fermentation temperature. The online microbial activity analysis module includes an optical density sensor and a metabolite gas chromatography unit for real-time calculation of microbial activity; The real-time analysis module for raw material molecular structure includes a Raman spectrometer and a molecular docking calculation unit, which is used to calculate the similarity of raw material molecular structure in real time. Specifically, the similarity of raw material molecular structure is the matching degree of cellulase hydrolysis sites, which is calculated based on the matching between the raw material molecular structure and the cellulase active site.

9. A machine learning-based method for optimizing the formulation of fermented fiber feed, applied to the machine learning-based fermented fiber feed formulation optimization system described in any one of claims 1-8, characterized in that, include: S1: Real-time acquisition of temperature time-series data, microbial activity, and raw material synergistic effect changes during the fermentation process; The raw material synergistic effect change data includes raw material synergistic effect score and raw material synergistic effect correlation degree. The raw material synergistic effect score is calculated based on the similarity of raw material molecular structure. The raw material synergistic effect correlation degree is calculated based on the correlation between the rate of change of raw material synergistic effect score and the rate of change of microbial activity during the continuous fermentation period. The correlation calculation is specifically: the absolute value of the ratio of the rate of change of raw material synergistic effect score to the rate of change of microbial activity is used as the correlation quantification value. S2: Based on the data on the changes in the synergistic effect of the raw materials, determine the stability of the prediction confidence of the machine learning model in the fermentation scenario and the consistency of the feature importance. S3: Based on the judgment result of step S2, determine the process state of the current fermentation scenario and trigger the model control strategy that matches the process state. S4: Receive and output the adjusted fermentation ratio and the fermentation ratio implementation effect data, and update the judgment conditions of the model operation status evaluation unit.

Images