Deep learning-based dynamic optimization method for metabolic feedback immune activity of fermented feed

Abstract

This invention provides a deep learning-based method for dynamically optimizing the metabolic feedback immune activity of fermented feed, comprising the following steps: S1: Collection of basic data and immune activity indicators of fermented feed; S2: Data preprocessing and feature engineering; S3: Construction of a deep learning-based immune activity prediction model; S4: Construction of a fusion framework driven by chaotic mapping—slime mold algorithm, symbiotic organism search, and fruit fly optimization algorithm; S5: Dynamic optimization of fermentation parameters based on the fusion algorithm; S6: Construction of a metabolic feedback mechanism and verification of optimization results. This invention overcomes the technical problems of inaccurate parameter control, limited optimization algorithms, lack of metabolic feedback, and insufficient immune activity prediction in existing fermented feed production.

Description

Technical Field

[0001] This invention relates to the field of intelligent optimization technology for fermented feed production, and in particular to a method for dynamic optimization of metabolic feedback immune activity in fermented feed based on deep learning. Background Technology

[0002] Fermented feed is widely used in livestock farming because it is rich in active ingredients such as amino acids and polysaccharides, and can enhance animal immunity. However, there are currently many technical bottlenecks in the production of fermented feed:

[0003] Fermentation parameter control relies on experience: In traditional production, key parameters such as temperature, humidity, and pH are mostly set based on the experience of operators, which cannot accurately match the correlation between metabolic processes and immune activity, and can easily lead to large fluctuations in the content of immune active components.

[0004] Limitations of single optimization algorithms: Existing parameter optimization often uses a single intelligent algorithm, such as the single fruit fly optimization algorithm or slime mold algorithm. These algorithms have problems such as insufficient global search capability, easy to get trapped in local optima, and difficulty in balancing convergence speed and search accuracy, and cannot achieve accurate dynamic optimization of parameters.

[0005] Lack of metabolic feedback mechanism: The closed-loop feedback between metabolite changes and immune activity during fermentation has not been established, making it impossible to respond in a timely manner to production variables such as raw material batch differences and environmental fluctuations, resulting in poor adaptability of optimization parameters;

[0006] Inaccurate prediction of immune activity: Traditional prediction methods struggle to capture the complex temporal relationships between fermentation parameters, metabolites, and immune activity. Even with simple intelligent prediction models, the lack of sufficient exploration of local correlations and long-term temporal dependencies among features prevents reliable guidance for parameter optimization, further hindering the improvement of immune activity. Summary of the Invention

[0007] This invention provides a deep learning-based method for dynamic optimization of the metabolic feedback immune activity of fermented feed, which overcomes the technical problems of inaccurate parameter control, limited optimization algorithms, lack of metabolic feedback, and insufficient prediction of immune activity in existing fermented feed production.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] A deep learning-based method for dynamic optimization of metabolic feedback-based immune activity in fermented feed includes the following steps:

[0010] S1: Collect fermentation parameters, metabolite data, and immune activity indicators of fermented feed to construct the original dataset;

[0011] S2: Preprocess and partition the original dataset to obtain the training set, validation set, and test set;

[0012] S3: Construct and train a hybrid neural network immune activity prediction model based on the training set, validation set, and test set to obtain the optimal prediction model;

[0013] S4: First, construct the basic models of the chaotic mapping-driven slime mold algorithm, the basic model of the symbiotic organism search, and the basic model of the fruit fly optimization algorithm respectively; then design a two-way bidirectional interaction mechanism between each algorithm to realize parameter feedback and position coordination between each algorithm; finally, construct a fitness function based on the predicted value of the immune activity index output by the optimal prediction model, and connect the three algorithms with the optimal prediction model through the fitness function to form a fusion framework.

[0014] S5: Based on the fusion framework and the optimal prediction model, dynamic optimization of fermentation parameters is carried out. Through population initialization, fitness calculation, iterative optimization and fusion of global optimal positions, candidate optimal fermentation parameters are output.

[0015] In this specification, the deep learning-based dynamic optimization method for metabolic feedback immune activity of fermented feed also includes S6: applying candidate optimal fermentation parameters to actual fermentation experiments, collecting experimental metabolite data and experimental immune activity indicators, constructing a metabolic feedback mechanism, determining whether to update the optimal prediction model through metabolite deviation calculation, verifying the effectiveness of the optimization results by combining the relative error between the experimental immune activity indicators and the predicted values ​​of the immune activity indicators, and forming a dynamic iterative optimization closed loop.

[0016] In this specification, the chaotic mapping-driven slime mold algorithm and the symbiotic organism search mutually feed back core parameters to dynamically adjust their respective search strategies. Specifically, the optimal fitness value of the symbiotic organism search is fed back to the chaotic mapping-driven slime mold algorithm to dynamically adjust its position update weight; the chaotic variables of the chaotic mapping-driven slime mold algorithm are fed back to the symbiotic organism search to dynamically adjust its symbiosis coefficient.

[0017] In this specification, the symbiotic organism search and the fruit fly optimization algorithm mutually feed back core parameters and location information to optimize the search target and range. Specifically, the optimal position of the fruit fly optimization algorithm is fed back to the symbiotic organism search to optimize the target position of its parasitic operation; the symbiotic coefficient of the symbiotic organism search is fed back to the fruit fly optimization algorithm to dynamically adjust its olfactory search step size.

[0018] In this specification, the chaotic mapping-driven slime mold algorithm and the fruit fly optimization algorithm feed back core parameters to balance global exploration and local convergence. Specifically, the scaling factor of the chaotic mapping-driven slime mold algorithm is fed back to the fruit fly optimization algorithm to dynamically adjust its visual search radius; the optimal olfactory concentration of the fruit fly optimization algorithm is fed back to the chaotic mapping-driven slime mold algorithm to dynamically adjust its scaling factor.

[0019] In this specification, fermentation parameters in S1 are obtained through real-time monitoring by online sensors, metabolite data are obtained through laboratory chromatographic detection, and immune activity indicators are obtained through in vitro cell culture experiments.

[0020] In this specification, the global optimal position fusion in S5 adopts a weighted fusion strategy, and the weights are determined by the proportion of the next-generation optimal fitness values ​​of the slime mold algorithm driven by chaotic mapping, the symbiotic organism search, and the fruit fly optimization algorithm.

[0021] In this specification, the convergence criterion for iterative optimization in S5 is: the absolute difference between the global optimal fitness values ​​of two adjacent rounds is less than a preset threshold, or the number of iterations reaches the preset maximum number of iterations.

[0022] In this specification, the hybrid neural network in S3 includes a feature extraction module and a temporal prediction module. The feature extraction module uses an alternating structure of convolutional layers and pooling layers to capture local feature correlations, while the temporal prediction module uses a combination of long short-term memory network layers and fully connected layers to model temporal dependencies. The training process adopts an early stopping strategy, stopping training when the loss no longer decreases after a preset number of consecutive rounds of verification.

[0023] In this specification, the metabolite bias calculation in S6 is the average absolute difference between the normalized values ​​of experimental metabolites and the normalized predicted values ​​of metabolites output by the optimal prediction model. When the metabolite bias is greater than the preset bias threshold, the experimental data is added to the original dataset to re-perform data preprocessing and training of the hybrid neural network immune activity prediction model, and the optimal prediction model is updated.

[0024] In summary, the present invention has at least the following beneficial effects:

[0025] 1. Improve the reliability of immune activity prediction: By using a hybrid intelligent prediction model that combines feature extraction and time series modeling capabilities, the complex temporal correlations and local correlations between features in the fermentation process are accurately captured. Compared with traditional prediction methods and simple intelligent models, the prediction accuracy is significantly improved, providing a reliable basis for parameter optimization.

[0026] 2. Enhance the effectiveness of parameter optimization: The three-algorithm deep integration framework achieves complementary advantages through bidirectional interaction, overcomes the shortcomings of single algorithms such as local optima and slow convergence, and can efficiently search for the optimal fermentation parameters that maximize immune activity;

[0027] 3. Enhanced production adaptability: The metabolic feedback closed-loop mechanism can respond in real time to fluctuations in raw materials and environment during the production process, and dynamically adjust and optimize parameters by updating the intelligent prediction model to ensure the stability of immune activity under different production scenarios;

[0028] 4. Enhance the level of intelligent production: Replace the traditional experience-based parameter control mode, and realize precise and intelligent control of the fermentation process through intelligent prediction and optimization algorithms, reduce reliance on manual operation, and improve production efficiency and product quality consistency;

[0029] 5. Expanding technical adaptability: This method can be transferred to the production optimization of different types of fermented feed, such as soybean meal fermented feed and straw fermented feed. It can be adapted by adjusting the model input features and parameter constraint range, and has a wide range of application scenarios. Attached Figure Description

[0030] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. 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.

[0031] Figure 1 This is a schematic diagram of the process for the dynamic optimization method of fermented feed metabolic feedback immune activity based on deep learning involved in this invention.

[0032] Figure 2 This is a schematic diagram illustrating the construction process of the three-algorithm fusion framework involved in this invention.

[0033] Figure 3 This is a schematic diagram of the process for dynamically optimizing fermentation parameters involved in this invention.

[0034] Figure 4 This is a schematic diagram of the metabolic feedback and dynamic iteration process involved in this invention. Detailed Implementation

[0035] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the embodiments of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0036] The following disclosure provides many different implementations or examples for carrying out different structures of the embodiments of the present invention. To simplify the disclosure of the embodiments of the present invention, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the embodiments of the present invention. Furthermore, reference numerals and / or reference letters may be repeated in different examples of the embodiments of the present invention; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various implementations and / or arrangements discussed.

[0037] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0038] like Figure 1 As shown, this embodiment provides a deep learning-based method for dynamically optimizing the metabolic feedback immune activity of fermented feed, including:

[0039] S1: Collection of Basic Data and Immunological Activity Indicators for Fermented Feed

[0040] Dynamic optimization of the immune activity of fermented feed requires comprehensive and accurate basic data. This step involves constructing a complete dataset through multi-dimensional data collection, providing a data foundation for subsequent model training and parameter optimization. First, the scope of data collection is defined, focusing on three core data categories: core parameters of the fermentation process, key metabolites, and immune activity indicators. All data originates from standardized fermentation workshops of fermented feed production enterprises and is obtained through a combination of online real-time monitoring and precise laboratory testing.

[0041] In the fermentation parameter acquisition phase, real-time monitoring is conducted on key variables affecting the fermentation process: temperature is acquired using a PT100 platinum resistance temperature sensor, with a monitoring range of 25-45℃ and an accuracy of ±0.1℃; humidity is acquired using a capacitive humidity sensor, with a range of 60%–85% and an accuracy of ±2%; pH value is acquired using an online pH electrode, with a range of 3.5-7.0 and an accuracy of ±0.01; fermentation time is automatically recorded by the workshop time sequence control system, in hours; and inoculum amount is accurately weighed using an electronic balance, with a range of 5%–20% and an accuracy of 0.01g. All fermentation parameters are sampled at a frequency of once per hour and transmitted in real-time to the data acquisition terminal via an industrial IoT system.

[0042] Metabolite data collection focused on substances closely related to immune activity during fermentation: Amino acids (lysine, methionine, threonine, isoleucine, leucine, phenylalanine, valine, and tryptophan) – eight essential amino acids – were detected using HPLC (Agilent 1260, C18 column, 4.6 × 250 mm, 5 μm), with acetonitrile-water as the mobile phase and a detection wavelength of 254 nm (unit: g / 100g). Polysaccharides were detected using the phenol-sulfuric acid method combined with a UV-2600 spectrophotometer (490 nm, unit: mg / g). Organic acids (lactic acid, acetic acid, and propionic acid) were detected using GC (Shimadzu GC-2030, HP-FFAP capillary column, 30 m × 0.32 mm × 0.25 μm), with a flame ionization detector (FID) and a unit of mmol / L. Alkaloids were detected using HPLC-MS (Thermo Q). Exactive, unit μg / g. Metabolite sampling is performed every 24 hours. Samples should be sent to the laboratory for processing and analysis immediately after collection to avoid sample deterioration.

[0043] Immunological activity indicators were collected through in vitro cell culture experiments: mouse macrophages (RAW264.7) and lymphocytes were selected as test cells, and fermented feed extract was co-cultured with the cells at a concentration of 10 μg / mL. Macrophage phagocytosis rate was detected using a neutral red phagocytosis assay combined with microscopic counting (unit: %). Lymphocyte transformation rate was detected using the MTT assay, calculated by measuring absorbance at 490 nm using an ELISA reader (unit: %). Cytokine IL-2 and TNF-α secretion levels were detected using an enzyme-linked immunosorbent assay (ELISA) kit from R&D Systems, with a detection wavelength of 450 nm (unit: pg / mL). Immunological activity indicators were sampled every 48 hours, with three replicates per test, and the average value was used as the final data.

[0044] Finally, construct the original dataset. The dataset has the following dimensions: ,in The sample size is defined as follows: each sample corresponds to full-time data from one complete fermentation cycle, with a total of 50 fermentation batches collected. , To determine the number of features, 25 features were included: 5 fermentation parameters, 8 amino acids, 1 polysaccharide, 3 organic acids, 1 alkaloid, and 4 immunomodulatory indicators. Each sample was labeled with a unique fermentation batch number. (1,2,...,50), simultaneously recording auxiliary information such as sampling time and detection device number, the original dataset Stored in CSV format.

[0045] S2: Data Preprocessing and Feature Engineering

[0046] The original dataset contains issues such as missing values, outliers, and feature redundancy. Preprocessing is needed to improve data quality and at the same time, effective features should be selected to reduce model complexity.

[0047] In the data cleaning stage, K-nearest neighbor interpolation was used to fill in missing values. The specific steps were: calculating the Euclidean distance between the missing sample and other samples, and selecting the 5 closest samples. Missing values ​​are calculated using a weighted average, with the weight being the reciprocal of the distance. Outliers are then addressed using... Criterion identification, calculation of the mean of each feature. and standard deviation It will exceed Data within the specified range is identified as outliers, and the corresponding samples are directly removed to avoid interference from outliers in subsequent analysis.

[0048] Data normalization processes map all feature data to For intervals, eliminating dimensional differences, the normalization formula is:

[0049] ;

[0050] in, Indicates the first The first sample Normalized values ​​of each feature Indicates the first The first sample The original values ​​of each feature, Indicates the first The minimum value of each feature. Indicates the first The maximum value of each feature, , During the normalization process, the characteristics of each feature are recorded. and This is used for the inverse normalization restoration of subsequent optimization results.

[0051] Feature selection employs the mutual information method to quantify the correlation between each feature and the immune activity indicators. A higher mutual information value indicates a more significant impact of the feature on immune activity. The specific calculation process is as follows: using four immune activity indicators as target variables, the mutual information value between each feature and each target variable is calculated. The average value is taken as the comprehensive mutual information value for that feature. Features with a comprehensive mutual information value greater than 0.3 are selected to form the effective feature set. The mutual information value is calculated using a K-nearest neighbor-based estimation algorithm, with 10 neighbors, implemented using the sklearn library in Python. The filtered effective feature set... Dimensions ( (This involves) eliminating redundant features to reduce the difficulty of model training.

[0052] The dataset was partitioned using stratified sampling, with the effective feature set divided in a 7:2:1 ratio. Divided into training set Validation set and test set Among them, the training set Contains 35 samples for iterative parameter updates and fitting training of the deep learning model; validation set. It contains 10 samples to monitor overfitting during model training and to adjust model hyperparameters such as learning rate and dropout rate; the test set... It contains 5 samples, which are used as an independent dataset to evaluate the generalization performance of the model and are not used in the model training and hyperparameter tuning.

[0053] The core of stratified sampling is to ensure that the distribution of immune activity indicators in each dataset is consistent with the original effective feature set. Specifically, the samples are divided into five strata based on the comprehensive immune activity value, and samples are drawn from each stratum in a 7:2:1 ratio to ensure consistency. , , The sample distribution is consistent, avoiding model evaluation bias due to differences in data distribution.

[0054] S3: Constructing a deep learning model for predicting immune activity

[0055] A hybrid neural network of convolutional neural network and long short-term memory network is constructed as an immune activity prediction model. This model fully utilizes the feature extraction capability of convolutional neural network (CNN) and the temporal modeling capability of long short-term memory network (LSTM) to achieve accurate prediction of immune activity indicators of fermented feed.

[0056] The model structure is divided into two parts: a feature extraction module and a temporal prediction module. The overall input is the effective feature set. The temporal feature vector, with dimension . , The time step is set to 24, meaning every 24 sampling points form a time-series sample, and the output is a vector of predicted values ​​for four immune activity indicators. The feature extraction module uses an alternating structure of 3 one-dimensional convolutional layers (Conv1d) and 2 max-pooling layers (MaxPool1d): the number of input channels in the first convolutional layer is... The output channel count is 64, the kernel size is 3, the stride is 1, the padding method is "same", and the activation function is ReLU, whose mathematical expression is:

[0057] ;

[0058] The first pooling layer has a kernel size of 2, a stride of 2, and "valid" padding. The second convolutional layer has 64 input channels, 128 output channels, a kernel size of 3, a stride of 1, "same" padding, and ReLU activation. The parameters of the second pooling layer are the same as those of the first pooling layer. The third convolutional layer has 128 input channels, 256 output channels, a kernel size of 3, a stride of 1, "same" padding, and ReLU activation. This module captures local correlations between features through convolution operations, and the pooling layers reduce feature dimensionality, improving the model's generalization ability.

[0059] The temporal prediction module consists of two LSTM layers and one fully connected layer: the first LSTM layer has an input dimension of 256, corresponding to the number of output channels of the third convolutional layer, 128 hidden units, and a dropout rate of 0.2, supporting long-range dependency modeling of temporal features; the second LSTM layer has an input dimension of 128, 64 hidden units, and a dropout rate of 0.2; the fully connected layer has an input dimension of 64 and an output dimension of 4, corresponding to four immune activity indicators, and uses the Sigmoid activation function, the mathematical expression of which is:

[0060] ;

[0061] This function maps the output to... The interval was matched with the normalized immune activity index.

[0062] The model training process aims to minimize the mean squared error between the predicted and actual values. The training parameters are set as follows: learning rate. Number of training rounds Batch size The optimizer used is the Adam optimizer, whose parameters... , , The training loss function formula is:

[0063] ;

[0064] in, For training loss values, For training set The number of time-series samples, For the first The training time series sample of the nth time series The first timing step Predicted values ​​of various immune activity indicators To correspond to the actual value, The values ​​up to 4 correspond to macrophage phagocytosis rate, lymphocyte transformation rate, IL-2 secretion, and TNF-α secretion, respectively.

[0065] Model validation and optimization employ an early stopping strategy: the validation set... Input the model from training and calculate the validation loss. Calculation method and training loss Consistent, the early stop patience value is set to 10, meaning that if the loss no longer decreases after 10 consecutive rounds of verification, and the decrease is less than [a certain value], then [the loss is considered to be stopped]. When the time is right, stop training and save the current best model. If training reaches the maximum number of rounds... If early stopping is not triggered, save the last round of the model and evaluate the performance.

[0066] Model generalization performance testing passed the test set Implement and calculate prediction accuracy. As the core evaluation indicator, the formula is:

[0067] ;

[0068] in, For the test set The number of time-series samples. When If the model meets the generalization performance requirements, it can be used for subsequent optimization; otherwise, return to the model structure design stage to adjust the number of units or the timing stride of the convolutional / LSTM layers. Retrain until the requirements are met. Optimal model. The input is a time-series feature vector of fermentation parameters and metabolites, and the output is a normalized predicted value vector of four immune activity indicators. .

[0069] S4: Construction of a Fusion Framework for Chaotic Mapping-Driven Slime Mold Algorithm-Symbiotic Organism Search-Drosophila Optimization Algorithm

[0070] 4.1 Design Ideas for the Integration Framework

[0071] To address the problem of single optimization algorithms easily getting trapped in local optima and struggling to balance convergence speed and search accuracy in fermentation parameter optimization, a deep fusion framework of three algorithms—CMSMA (Chaotic Mapping Driven), SOS (Symbiotic Search), and FOA (Fruit Fly Optimization)—is constructed. This framework uses fermentation parameters as optimization variables and achieves complementary advantages among the algorithms through a pairwise bidirectional interaction mechanism: CMSMA provides strong global search capabilities, SOS enhances local development flexibility through symbiotic relationships, and FOA improves convergence speed through an olfactory-visual search mechanism. The three algorithms work together to form an optimization closed loop of global exploration, local refinement, and rapid convergence, providing an efficient search engine for subsequent dynamic optimization. The construction process of the three-algorithm fusion framework is as follows: Figure 2 As shown.

[0072] 4.2 Basic Algorithm Model Construction and Parameter Definition

[0073] 4.2.1 Chaotic Mapping Driven Slime Mold Algorithm (CMSMA) Model

[0074] The slime mold algorithm simulates the foraging and scaling behavior of slime molds and has strong global search capabilities, but insufficient diversity in the initial population can lead to low search efficiency. To address this, a Logistic chaotic mapping is introduced to generate the initial population. The randomness and ergodicity of chaotic variables are used to improve the uniformity of population distribution. Simultaneously, core parameters are defined to adapt to the fermentation parameter optimization scenario.

[0075] The initial population generation is based on the Logistic chaotic mapping, and the formula is:

[0076] ;

[0077] in, This represents the chaotic variable value of the i-th individual in CMSMA during the t-th iteration, where i is the individual index ranging from 1 to N, and t is the iteration number ranging from 1 to N. ; The initial chaotic value ranges from 0 to 1; The chaos factor ranges from 3.57 to 4.0 and is used to adjust the fluctuation intensity of chaotic variables. The larger the value, the richer the population diversity.

[0078] Mapping the chaotic variables to the fermentation parameter constraints yields the initial population position in CMSMA:

[0079] ;

[0080] in, This represents the position vector of the i-th individual in CMSMA at the t-th iteration, corresponding to a set of fermentation parameters; The minimum value vector for fermentation parameters is set to [25, 60, 3.5, 24, 5], which correspond to temperature, humidity, pH value, fermentation time, and inoculum amount, respectively. The maximum value vector for fermentation parameters is [45, 85, 7.0, 168, 20]; N is the population size, which is 50. The maximum number of iterations is set to 100.

[0081] The CMSMA position update model simulates the slime mold's contractile foraging behavior, and the formula is:

[0082] ;

[0083] in, Let represent the position vector of the i-th individual in CMSMA after the update in the (t+1)-th iteration; The position update weight for the t-th iteration ranges from 0.1 to 0.9, used to balance the influence of an individual's historical position and its optimal position; The value of the slime mold stretching coefficient for the t-th iteration ranges from 0 to 1. The larger the coefficient, the stronger the tendency of the individual to move toward the optimal position. represents the optimal position vector of CMSMA in the t-th iteration, corresponding to the position of the individual with the largest fitness value in that iteration; rand(0,1) represents a random number in the range of 0 to 1, used to introduce randomness in the search.

[0084] 4.2.2 Symbiotic Search (SOS) Model

[0085] SOS simulates the symbiotic relationships between organisms in nature, including mutualism, symbiosis, parasitism, and symbiosis. It achieves a balance between local development and global exploration by dynamically switching between these three relationships. However, its fixed symbiosis coefficient can lead to insufficient search flexibility, and it needs to be dynamically adjusted in conjunction with other algorithms.

[0086] The core parameters and location update formulas for SOS are as follows:

[0087] 1. Mutually beneficial symbiotic relationship: Two individuals cooperate to improve each other's fitness; the update formula is:

[0088] ;

[0089] in, This represents the position vector of the i-th individual in SOS during the t-th iteration; Let represent the updated position vector of the i-th individual in SOS after the (t+1)th iteration; The co-occurrence coefficient for the t-th iteration ranges from 0 to 2 and is used to adjust the intensity of the influence of the symbiotic relationship. The mutual benefit coefficient for the t-th iteration ranges from 0 to 1 and is used to allocate the cooperative benefits between the two individuals. represents the optimal position vector of SOS in the t-th iteration; rand(0,1) represents a random number ranging from 0 to 1.

[0090] 2. Beneficial symbiotic relationship: One individual benefits while the other is unaffected; the update formula is:

[0091] ;

[0092] The parameters in this formula are defined in the same way as in a mutually beneficial relationship, the difference being the removal of the benefit coefficient. Individuals directly learn from the optimal position to improve their own adaptability.

[0093] 3. Parasitic Relationship: Parasitic individuals replace inferior individuals by imitating the characteristics of the optimal individual. The update formula is:

[0094] ;

[0095] in, This represents the position vector of a randomly selected individual in SOS during the t-th iteration; rand(-1,1) represents a random number ranging from -1 to 1. The parasitic probability in the t-th iteration ranges from 0 to 1. When rand(0,1) is less than Parasitic operations are performed when necessary; otherwise, mutualistic or symbiotic operations are randomly selected.

[0096] 4.2.3 Fruit Fly Optimization Algorithm (FOA) Model

[0097] FOA simulates the olfactory-visual search behavior of fruit flies. In the olfactory stage, the fly explores the world to find food sources, and in the visual stage, it refines the search to get closer to the optimal food source. However, its fixed search step size can easily lead to insufficient convergence accuracy in the later stage, so it needs to be dynamically adjusted in coordination with other algorithms.

[0098] The core parameters and position update formulas of FOA are as follows:

[0099] 1. Olfactory search phase, global exploration: Fruit flies perceive food odors through their sense of smell, expanding their search range. The formula is:

[0100] ;

[0101] in, This represents the position vector of the i-th individual in the FOA during the t-th iteration; Let represent the updated position vector of the i-th individual in the FOA after the (t+1)th iteration; The olfactory search step size for the t-th iteration ranges from 0 to 1. The larger the step size, the wider the global search range; rand(-1,1) represents a random number ranging from -1 to 1. and Defined the same as CMSMA, it is used to limit the search scope.

[0102] 2. Visual search stage, local refinement: Fruit flies observe the positions of other fruit flies visually and move closer to the optimal individual. The formula is: ;

[0103] in, The visual search radius for the t-th iteration ranges from 0 to 5; the smaller the radius, the higher the local search accuracy. Let FOA represent the optimal position vector in the t-th iteration; This represents the Euclidean norm, used to calculate the distance between an individual's position and its optimal position. To take the value of the minimum value This is used to avoid the denominator being 0.

[0104] 3. Concentration Judgment Mechanism: Fruit flies judge the quality of food sources by olfactory concentration. The formula is:

[0105] ;

[0106] in, This represents the olfactory concentration value of the i-th individual in the FOA during the t-th iteration; The fitness function; The concentration threshold for the t-th iteration ranges from 0 to 1. Greater than If the individual's position is not specified, it is retained; otherwise, it is discarded and a new individual is generated.

[0107] 4.3 Algorithm Pairwise Bidirectional Interaction Mechanism Design

[0108] To achieve deep integration of the three algorithms rather than simple superposition, a two-way interaction mechanism is designed. Through parameter sharing, position feedback, and strategy collaboration, the running status of each algorithm directly affects the core parameters or search behavior of another algorithm, forming a closed loop of mutual guidance, dynamic adjustment, and collaborative optimization.

[0109] 4.3.1 Two-way interaction between CMSMA and SOS

[0110] 1. SOS parameter feedback adjustment for CMSMA: CMSMA position update weights This directly impacts the balance between global search and local exploitation, and the optimal fitness value of SOS reflects the effectiveness of the current local search. Introducing the optimal fitness value of SOS into... The update formula enables dynamic guidance of the global search strategy based on the local search results:

[0111] ;

[0112] in, This represents the optimal fitness value of SOS in the t-th iteration; This represents the optimal fitness value of CMSMA in the t-th iteration; This represents the worst fitness value of CMSMA in the t-th iteration. The closer the optimal fitness value of SOS is to the optimal fitness value of CMSMA, the better. The larger the size, the more CMSMA tends to retain the current search direction and enhance local development capabilities; conversely, the smaller the size, the more likely it is to retain the current search direction and enhance local development capabilities. The smaller the value, the more CMSMA tends to explore globally, avoiding getting trapped in local optima.

[0113] 2. CMSMA's parameter feedback adjustment of SOS: SOS co-occurrence coefficient The strength of the influence that determines the symbiotic relationship, and the chaotic variables of CMSMA. It exhibits strong randomness, which can enhance search diversity. Chaotic variables from CMSMA are introduced. The updated formula enhances the search flexibility of SOS:

[0114] ;

[0115] in, is the chaotic variable in CMSMA; t is the current iteration number; This represents the maximum number of iterations. When the chaotic variables in CMSMA fluctuate significantly, The changes are more drastic, the symbiotic relationship of SOS is adjusted more frequently, and it can effectively escape local optima; in the later stages of iteration, t approaches... hour, Approaching 0, Stable at around 1, SOS focuses on localized refinement.

[0116] 3. Post-interaction position update formula: Combining bidirectional parameter feedback, the position update formulas of CMSMA and SOS incorporate the optimal information from each other, achieving synergy in search strategies.

[0117] CMSMA interaction post-position update formula:

[0118] ;

[0119] In this formula, the position update of CMSMA not only refers to its own optimal position, but also incorporates the optimal position of SOS. By leveraging the localized development advantages of SOS, we can optimize our search direction.

[0120] SOS interaction-based mutually beneficial update formula:

[0121] ;

[0122] In this formula, the mutually beneficial symbiotic update of SOS refers to the optimal position of CMSMA. By leveraging the global search advantages of CMSMA, we can broaden our search scope.

[0123] 4.3.2 Two-way interaction between SOS and FOA

[0124] 1. FOA Position Feedback Adjustment for SOS: The parasitic operation of SOS aims to replace inferior individuals. If only random individuals within the population are considered, the replacement effect may be poor. Introducing the optimal position of FOA into the parasitic update formula of SOS improves the effectiveness of the parasitic operation.

[0125] ;

[0126] in, Let FOA represent the optimal position vector of FOA in the t-th iteration. FOA has strong local convergence ability, and its optimal position is closer to the global optimum. By learning from the optimal position of FOA, parasitic individuals of SOS can quickly improve their fitness and replace inferior individuals more efficiently.

[0127] 2. SOS's parameter feedback adjustment of FOA: olfactory search step size of FOA The symbiosis coefficient of SOS determines the overall exploration intensity. It can reflect the diversity of the current search. The co-occurrence coefficient of SOS is introduced. The update formula dynamically balances the global exploration and local development of FOA:

[0128] ;

[0129] in, t is the co-occurrence coefficient of SOS; t is the current iteration number; This represents the maximum number of iterations. Initially, t is relatively small. Approaching 1, When the value is relatively large, FOA focuses on global exploration; as t increases in the later stages of iteration, Approaching 0, To reduce [the impact], FOA focuses on localized development. Meanwhile, The fluctuation energy is Introducing randomness helps prevent the FOA from getting trapped in local optima too early.

[0130] 3. Post-interaction location update formula: SOS and FOA optimize their respective search strategies through location sharing and parameter coordination.

[0131] Parasitic update formula after SOS interaction:

[0132] ;

[0133] This formula directly uses the optimal position of FOA as the parasitic learning target, improving the accuracy of SOS parasitic operations.

[0134] FOA interaction olfactory search formula:

[0135] ;

[0136] In this formula, the olfactory search of FOA is based not only on its own positional fluctuations, but also on the optimal position of SOS. By leveraging the symbiotic relationship advantage of SOS, the search scope can be broadened.

[0137] 4.3.3 Bidirectional Interaction between CMSMA and FOA

[0138] 1. CMSMA parameter feedback adjustment for FOA: Visual search radius of FOA The scaling factor of CMSMA determines the local search accuracy. It can reflect the current progress of the global search. The scaling factor of CMSMA is introduced. The updated formula optimizes the local search accuracy of FOA:

[0139] ;

[0140] in, t is the scaling factor of CMSMA; t is the current iteration number; This represents the maximum number of iterations. When... When the size is small, CMSMA focuses on global exploration. If the area is large, FOA expands the local search range and, in conjunction with CMSMA, explores new areas; when When the size is large, CMSMA focuses on local development. With a smaller value, FOA focuses on searching around the optimal solution, improving convergence accuracy.

[0141] 2. FOA's feedback adjustment of CMSMA parameters: CMSMA's scaling factor The tendency to move towards the optimal position is determined by the FOA (Focus of Odor) concentration, which reflects the quality of the current local optimum. The optimal FOA concentration is then introduced into... The updated formula improves the convergence speed of CMSMA:

[0142] ;

[0143] in, This represents the optimal individual olfactory concentration value of FOA in the t-th iteration. A higher optimal olfactory concentration in FOA indicates a better quality local optimum. As the CMSMA value increases, its position updates tend to move closer to the optimal position, accelerating convergence; conversely, as the value decreases... By reducing the size of the CMSMA, the global exploration capability is maintained.

[0144] 3. Post-interaction position update formula: CMSMA and FOA achieve a balance between global exploration and local convergence through parameter coordination and policy complementarity.

[0145] Visual search formula after FOA interaction:

[0146] ;

[0147] In this formula, the visual search reference for FOA is the optimal position of CMSMA. By leveraging the global search advantage of CMSMA, local optima can be avoided.

[0148] CMSMA interaction post-position update formula:

[0149] ;

[0150] In this formula, the position update of CMSMA introduces the optimal position of FOA. It leverages the local convergence advantage of FOA to improve its search accuracy.

[0151] 4.4 Construction of Fitness Function for Collaborative Optimization by Three Algorithms

[0152] The fitness function is the optimization objective guide of the fusion algorithm. It is necessary to transform the immune activity indicators predicted by the deep learning model into quantifiable optimization objectives of the algorithm, while taking into account the differences in importance of each immune activity indicator, so as to ensure that the optimization results meet the needs of practical applications.

[0153] The fitness function is based on a deep learning model. The predicted output is constructed with the goal of maximizing the comprehensive value of immune activity indicators, as shown in the formula:

[0154] ;

[0155] in, The fitness function value ranges from 0 to 1. The input fermentation parameter vector corresponds to the individual position vectors of the three algorithms; This is the normalized value of macrophage phagocytosis rate predicted by the deep learning model; This is the normalized value of lymphocyte transformation rate predicted by the deep learning model; The raw values ​​of IL-2 secretion predicted by the deep learning model; The raw values ​​of TNF-α secretion predicted by the deep learning model; The weighting coefficients are set to 0.3, 0.3, 0.2, and 0.2 respectively, satisfying that the sum of the weighting coefficients is 1. The weighting is determined based on the importance attached to immune activity indicators by the local fermented feed industry. Macrophage phagocytosis rate and lymphocyte transformation rate are assigned higher weights as core indicators. The maximum value for IL-2 secretion was set at 200 pg / mL; The maximum value of TNF-α secretion was taken as 150 pg / mL, which was derived from historical data statistics of the local fermented feed industry.

[0156] The core function of this fitness function is to integrate multi-dimensional immune activity indicators into a single quantitative indicator, providing a clear optimization direction for the fusion algorithm. Among these, and The data has been preprocessed and normalized to the 0-1 range. and Normalization is achieved by dividing by the industry statistical maximum value, ensuring that the values ​​of the four indicators are within the same range and avoiding the impact of differences in units on the optimization results. The input to the fitness function is the individual position vector of the three algorithms, and the output is the comprehensive immune activity value. The algorithm continuously iterates and updates the individual positions to find the fermentation parameter vector that maximizes the fitness function value, thereby achieving the optimization goal of immune activity.

[0157] Through the detailed design of S4 described above, the fusion framework realizes the deep coupling and bidirectional interaction of the three algorithms CMSMA-SOS-FOA. The core parameters and search behavior of each algorithm are affected by the other algorithms. At the same time, the fitness function closely links the algorithm optimization with the prediction of the deep learning model, providing logically rigorous and collaboratively efficient technical support for the subsequent dynamic optimization of fermentation parameters.

[0158] S5: Dynamic optimization process of fermentation parameters based on fusion algorithm

[0159] Based on the three-algorithm fusion framework built with S4, combined with the immune activity prediction model trained with S3, dynamic optimization of fermentation parameters is achieved, outputting the optimal fermentation parameters that maximize the comprehensive value of immune activity indicators. The dynamic optimization process of fermentation parameters is as follows: Figure 3 As shown.

[0160] First, the population is initialized, with the population size for all three algorithms set to [value missing]. The optimization variable is the fermentation parameter vector. The variable constraint range is consistent with S1 ( , Among them, the initial population of the Chaotic Map-Driven Slime Mold Algorithm (CMSMA) is generated through the Logistic Chaotic Map defined by S4 to ensure population diversity; the initial populations of the Symbiotic Organism Search (SOS) and Fruit Fly Optimization Algorithm (FOA) adopt the uniform random generation method, which randomly samples and generates the initial population within the constraints to avoid the initial population concentration leading to search blind spots.

[0161] In the initial fitness calculation stage, the initial population position vectors (i.e., combinations of candidate fermentation parameters) of the three algorithms are converted into temporal feature vectors (according to the time step size of S3). (Construct), input the optimal prediction model This yields a vector of predicted immune activity values ​​for each candidate parameter. Then, the fitness function defined in S4 is substituted into it. The fitness value of each individual is calculated to obtain the initial fitness set of CMSMA. SOS and FOA The initial optimal position for each algorithm is selected based on its fitness value. , , This serves as the optimal reference for the first iteration.

[0162] The iterative optimization process is executed cyclically according to the following steps, with a certain number of iterations. From 1 to :

[0163] 1. Parameter Interaction Update: Execute the pairwise bidirectional interaction mechanism defined in S4 to update the core parameters of the three algorithms respectively: update the optimal fitness value based on SOS, update the position of CMSMA, and update the weights. Update the co-occurrence coefficient of SOS using chaotic variables based on CMSMA The parasitic operation target of SOS is updated based on the optimal position of FOA, and the olfactory search step size of FOA is updated based on the co-occurrence coefficient of SOS. ; Update the visual search radius of FOA based on the scaling factor of CMSMA The scaling factor of CMSMA is updated based on the optimal olfactory concentration of FOA. The parameter update order is CMSMA interacts with SOS → SOS interacts with FOA → CMSMA interacts with FOA, ensuring that the updates of each parameter reflect the current state of other algorithms.

[0164] 2. Population Position Update: Based on the updated core parameters, the position update formulas of the three algorithms (including the interaction terms defined in S4) are executed respectively to generate the next generation of population positions. , , Truncate the position vectors that exceed the constraint range after the update: if the value of a certain variable is less than... Then set it as If greater than Then set it as This ensures that all candidate parameters meet the actual fermentation requirements.

[0165] 3. Next-Generation Fitness Calculation: The truncated next-generation population position vector is converted into a temporal feature vector, and then input... Obtain the corresponding predicted value of immune activity Substitute the values ​​into the fitness function to calculate the fitness values ​​of the next generation of individuals for each algorithm, and then select the optimal positions of the next generation for each algorithm. , , .

[0166] 4. Global Optimal Position Fusion: A weighted fusion strategy is used to calculate the global optimal position. The weights are determined by the proportion of the best fitness values ​​of each new generation of algorithms, using the following formula:

[0167] ;

[0168] in, , , This ensures that algorithms with higher fitness occupy a higher weight in the globally optimal position, achieving complementary advantages.

[0169] 5. Convergence criterion: Calculate the global optimal fitness value. If the absolute difference between the global optimal fitness values ​​of two adjacent rounds is less than (Right now ), or the number of iterations reaches If the iteration terminates, then let the iteration terminate; otherwise, let Return to step 1 and continue iterating.

[0170] After the iteration terminates, output the global optimal position. By inverse normalization (using S2 records) and To restore the actual fermentation parameter values, the inverse normalization formula is:

[0171] ;

[0172] Finally, the optimized fermentation parameter vector is obtained. This is used for subsequent actual fermentation experiments for verification.

[0173] S6: Construction and Optimization Results of Metabolic Feedback Mechanism

[0174] A metabolic feedback loop was constructed, and the effectiveness of the optimized parameters was verified through actual fermentation experiments. Simultaneously, experimental data was used to update the predictive model, achieving continuous dynamic optimization of immune activity and ensuring that the optimization results adapt to actual production needs. The metabolic feedback and dynamic iteration process is as follows: Figure 4 As shown.

[0175] The actual fermentation experiment was conducted in a 50L stainless steel fermentation tank (model BI-50A) in the fermented feed production workshop, using three parallel batches to eliminate random errors. Each batch of fermentation was strictly carried out according to the optimized parameters. Control: The fermentation temperature is stabilized at a constant temperature using a temperature control system. The humidity control system maintains the humidity at The online pH adjustment system controls the pH value in real time. Fermentation time is precisely controlled to Inoculation volume according to Accurate weighing and addition. During fermentation, fermentation parameters and metabolite data are collected in real time according to the S1 data collection specifications (to verify parameter control accuracy). and immune activity indicators After each batch of experiments is completed, a complete experimental dataset is obtained.

[0176] In the metabolic feedback data processing stage, the metabolite data collected in the experiment are first processed. Perform normalization on S2 to obtain Calculate the metabolite bias value. The formula for quantifying the difference between experimental and predicted values ​​is:

[0177] ;

[0178] in, Characteristic quantity of metabolites, To be enter The obtained number Normalized predicted values ​​of metabolites. This reflects the model's accuracy in predicting changes in metabolites, when If the model fails to fit the metabolic process under the current fermentation conditions, it indicates that the model is not sufficiently accurate and needs to be updated.

[0179] The model update process is as follows: Newly collected experimental data... , , Add the original dataset The data was cleaned, normalized, and feature-selected again using the method in S2 to obtain the updated dataset. ;Will The training set was re-divided according to a 7:2:1 ratio. Validation set and test set To keep the training parameters consistent with the original model, the learning rate... Number of training rounds Batch size The CNN-LSTM model was retrained based on the new dataset to obtain the updated optimal model. This is used for fitness calculation in the next round of optimization, enabling dynamic iterative optimization of the model.

[0180] Optimization results validation using experimental immune activity indicators Compared with model predictions ( enter The relative error obtained is used as the judgment criterion, and the formula for relative error is:

[0181] ;

[0182] in, The numbers up to 4 correspond to four immune activity indicators. If all... This indicates that the optimized parameters are effective, and the output is... As the final fermentation parameter; otherwise, As the new true value, return to S5 to restart the fusion algorithm optimization, adjust the candidate parameter range and iterate again until all relative errors meet the requirements.

[0183] To ensure the long-term stability of immune activity indicators, a dynamic iterative optimization mechanism was established: after applying the final fermentation parameters to actual production, the experimental verification and model update process of S6 was repeated every 10 batches produced. The prediction model was continuously corrected through metabolic feedback data, and the fermentation parameters were dynamically adjusted to cope with fluctuations in actual production such as differences in raw material batches and changes in environmental temperature, so as to ensure that the immune activity of fermented feed is always at the optimal level.

[0184] In some embodiments, to further enhance the targeting of temporal feature extraction and the generalization ability of the model, a gated attention fusion unit (GAFU) can be added between the CNN feature extraction module and the LSTM temporal prediction module. The core of this unit is to dynamically allocate the weights of the local features extracted by the CNN and the temporal features captured by the LSTM through a gating mechanism. Specifically, the 256-dimensional feature map output by the CNN is first subjected to global average pooling to obtain a 1×256 global feature vector. The 128-dimensional temporal feature vector output from the first layer of the LSTM Dimension augmentation (increasing the dimension to 256 through 1×1 convolution) yields Then construct the gate function. ,in The weight matrix is ​​256×512. It is a 256-dimensional bias vector. The Sigmoid function is used; the final fused features are... ,Will Input the second layer LSTM. This design avoids information redundancy caused by simply concatenating local and temporal features. It adaptively focuses on the more critical feature components for predicting immune activity through an attention mechanism, making it particularly suitable for scenarios where feature correlations change dynamically during fermentation.

[0185] In some embodiments, to address the issue of missing short-period fluctuation features caused by a fixed time step size during model training, a dynamic time step size adjustment mechanism and a meta-learning initialization strategy can be introduced. The dynamic time step size adjustment mechanism calculates the feature change rate between adjacent sampling points. , For the first The first sample Step 1 eigenvalues, when At that time, the time step size is temporarily reduced to 12 to enhance the capture of features during periods of intense fluctuation; when To reduce data redundancy during the stationary phase, the time-series step size was expanded to 36. The meta-learning initialization strategy employed the MAML (Model-Agnostic Meta-Learning) algorithm, utilizing small sample data from different batches of fermentation feedstocks. Five time-series samples from each batch were used for meta-training to obtain initial prior values ​​for the model parameters, which were then fine-tuned based on the target fermentation feedstock data. This improvement allows the model to quickly adapt to different feedstock characteristics, reduces the amount of training data required for new feedstock scenarios, and enhances its adaptability to the characteristics of different stages of the fermentation process.

[0186] In some embodiments, to address the premature decline in population diversity caused by fixed symbiotic relationships in the SOS algorithm, a dynamic symbiotic evolution strategy enhanced by chaotic perturbation can be introduced. Specifically, the symbiotic relationships in SOS in S4—mutualism, symbiotic-parasitism, and parasitism—are changed from a fixed traditional ratio of 3:1:1 to dynamic adjustment, based on the chaotic entropy value of the population. , For the first The proportion of chaotic mapping values ​​for each individual: when When population diversity is insufficient, the proportion of mutualistic symbiosis is increased to 50%, partial symbiosis to 30%, and parasitism to 20%, and a chaotic perturbation is applied to the position vector of each individual after the symbiotic operation. ,in The disturbance coefficient, with a value ranging from 0.05 to 0.15, varies with... Decrease and increase Random numbers in the interval [0,1] generated by the Logistic chaotic mapping; when When population diversity is sufficient, the traditional ratio is restored, and chaotic perturbation is stopped. This strategy breaks local optima through chaotic perturbation, while the dynamic coexistence ratio balances exploration and exploitation capabilities, preventing the algorithm from converging prematurely.

[0187] In some embodiments, to further improve the collaborative efficiency of the three algorithms, a cross-algorithm feature migration (CAFM) module can be added on the basis of the two-way bidirectional interaction mechanism. The core of this module is to extract the "dominant feature vectors" of each algorithm population and migrate them across algorithms: First, principal component analysis (PCA) is performed on the population position vectors of CMSMA, SOS, and FOA respectively, and the first three principal components are extracted to form the dominant feature vectors of each algorithm. , , Subsequently, a feature transfer matrix is ​​constructed. ,in , , The transfer weights are determined by the proportion of the current optimal fitness values ​​of each algorithm, consistent with the calculation logic of the global optimal fusion weights in S5; finally, the transfer matrix is... Perform matrix dot products with the population position vectors of each algorithm to obtain the migrated population position vectors. This improvement enables each algorithm to share the advantageous search features of other algorithms, avoiding the search blind spots of a single algorithm. At the same time, it combines the chaotic perturbation and dynamic symbiosis strategy of the previous embodiment to form a collaborative optimization closed loop of feature sharing, dynamic cooperation, and perturbation exploration.

[0188] In some embodiments, to address the lag issue in adjusting the olfactory search step size of the FOA, a step size update mechanism optimized by fractional calculus (FC) can be introduced. The olfactory search step size of the FOA in S4 is then adjusted. The update formula has been changed from integer-order iteration to fractional-order iteration, specifically using the step-size update formula defined by Caputo's fractional derivative: ,in The learning rate is 0.01 to 0.03. The fractional order (0.3–0.7, optimized through the validation set). For the optimal fitness sequence of FOA Fractional derivatives. Fractional derivatives can capture long-range dependency information of the fitness sequence, enabling step size adjustments to respond in advance to changes in fitness trends and avoid search oscillations caused by lag. At the same time, combined with the dynamic co-occurrence and feature transfer strategies of the previous two embodiments, the search accuracy and convergence speed of the fusion algorithm are further improved.

[0189] In some embodiments, to address the issues of low proportion of high-quality candidate parameters and slow convergence speed caused by random generation of the initial population, an initial population screening mechanism based on metabolite fingerprinting can be introduced. The specific steps are as follows: First, the original dataset collected from S1... In the selection process, the top 20% of samples by overall immune activity value were chosen as a high-quality sample set. Metabolite characteristics were extracted from these samples to construct a metabolite fingerprint library. Subsequently, candidate fingerprint profiles were constructed based on the predicted metabolite values ​​corresponding to the initial populations randomly generated by SOS and FOA. ;calculate and The cosine similarity of each metabolite fingerprint is used to select candidate parameters with a similarity greater than 0.8 as valid initial population. If there are fewer than 50 such parameters, additional parameters are generated and re-selected. This mechanism uses the similarity matching of metabolite fingerprints to pre-select candidate parameters that are more likely to be close to the optimal solution, thereby improving the quality of the initial population and shortening the iteration convergence time.

[0190] In some embodiments, to achieve real-time adaptation of the optimization process to the dynamic fluctuations of the fermentation process, an online incremental learning (OIL) module can be added during the iterative optimization process. The core of this module is to utilize new data collected in real time during the fermentation process, collecting data once every five iterations to update the prediction model. First, the newly acquired time-series data is preprocessed using S2 to obtain incremental samples. Incremental learning algorithms, such as incremental CNN-LSTM, are employed. These algorithms freeze existing convolutional layer parameters and fine-tune only the LSTM and fully connected layers. The update is performed, and the loss function for the update process is... , This represents the loss of the original training set. To avoid catastrophic forgetting, the updated model will use incremental sample loss. Fitness calculations are used in subsequent iterations. This improvement, combined with the high-quality initial population selection of the previous embodiment, enables the optimization process to converge quickly based on high-quality initial points, and also improves the real-time performance and accuracy of optimization parameters by updating the model online to adapt to the dynamic changes in the fermentation process.

[0191] In some embodiments, to address the issue that weight allocation during global optimal position fusion relies too heavily on the current fitness and is susceptible to noise interference, a weighted fusion strategy optimized by Bayesian inference can be introduced. The weight calculation for the global optimal position in S5 is changed from the current fitness percentage to the Bayesian posterior probability percentage: first, the optimal fitness value sequence of each algorithm is treated as the observed value, and a Bayesian model is constructed. ,in The weights of each algorithm, Let be the likelihood function. The prior probability is initially set to a uniform distribution; the posterior probability is solved iteratively using the Markov Chain Monte Carlo (MCMC) algorithm. The mean of the posterior probabilities is used as the final weight; the global optimal position is calculated based on this weight. This improvement combines the high-quality initial population selection and online incremental learning of the previous two embodiments to form a full-process optimization mechanism of "high-quality initialization - dynamic model update - robust weight fusion", which further improves the reliability and stability of the optimization results, while enhancing the ability to resist fitness noise.

[0192] The embodiments described above are for illustrative purposes only and are not intended to limit the invention. Therefore, any changes in numerical values ​​or substitutions of equivalent elements should still fall within the scope of this invention.

[0193] The above detailed description will enable those skilled in the art to understand that the present invention can indeed achieve the aforementioned objectives and has complied with the provisions of the Patent Law.

[0194] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the invention. The above descriptions are merely preferred embodiments of the invention and are not intended to limit the invention. It should be noted that any modifications, equivalent substitutions, and improvements made within the spirit and principles of the invention should be included within the scope of protection of the invention.

[0195] It should be noted that the above description of the process is for illustrative purposes only and does not limit the scope of this specification. Those skilled in the art can make various modifications and changes to the process under the guidance of this specification. However, these modifications and changes remain within the scope of this specification.

[0196] The basic concepts have been described above. Obviously, for those skilled in the art who have read this application, the above disclosure is merely illustrative and does not constitute a limitation of this application. Although not explicitly stated herein, those skilled in the art may make various modifications, improvements, and corrections to this application. Such modifications, improvements, and corrections are suggested in this application, and therefore, such modifications, improvements, and corrections still fall within the spirit and scope of the exemplary embodiments of this application.

[0197] Furthermore, this application uses specific terms to describe its embodiments. For example, "an embodiment," "one embodiment," and / or "some embodiments" refer to a particular feature, structure, or characteristic related to at least one embodiment of this application. Therefore, it should be emphasized and noted that "an embodiment," "one embodiment," or "an alternative embodiment" mentioned twice or more in different positions in this specification do not necessarily refer to the same embodiment. In addition, certain features, structures, or characteristics in one or more embodiments of this application can be appropriately combined.

[0198] Furthermore, those skilled in the art will understand that aspects of this application can be described and illustrated through several patentable types or situations, including any new and useful combination of processes, machines, products, or substances, or any new and useful improvements thereof. Therefore, aspects of this application can be implemented entirely in hardware, entirely in software (including firmware, resident software, microcode, etc.), or a combination of hardware and software. All of the above hardware or software can be referred to as a “unit,” “module,” or “system.” Furthermore, aspects of this application can take the form of a computer program product embodied in one or more computer-readable media, wherein computer-readable program code is contained therein.

[0199] The computer program code required for the operation of each part of this application can be written in any one or more programming languages, including object-oriented programming languages ​​such as Java, Scala, Smalltalk, Eiffel, JADE, Emerald, C++, C, VB.NET, and Python; general programming languages ​​such as C; Visual Basic, Fortran2103, Perl, COBOL2102, PHP, and ABAP; dynamic programming languages ​​such as Python, Ruby, and Groovy; or other programming languages. This program code can run entirely on the user's computer, or as a standalone software package on the user's computer, or partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the latter case, the remote computer can be connected to the user's computer via any network, such as a local area network (LAN) or wide area network (WAN), or connected to an external computer (e.g., via the Internet), or in a cloud computing environment, or used as a service such as Software as a Service (SaaS).

[0200] Furthermore, unless expressly stated in the claims, the order of processing elements and sequences, the use of numbers and letters, or other names described in this application are not intended to limit the order of the processes and methods of this application. Although some currently considered useful embodiments of the invention have been discussed in the foregoing disclosure by way of various examples, it should be understood that such details are for illustrative purposes only, and the appended claims are not limited to the disclosed embodiments; rather, the claims are intended to cover all modifications and equivalent combinations that conform to the substance and scope of the embodiments of this application. For example, although the implementation of the various components described above can be embodied in a hardware device, it can also be implemented as a purely software solution, such as an installation on an existing server or mobile device.

[0201] Similarly, it should be noted that, in order to simplify the description of the present application and thus aid in the understanding of one or more embodiments of the invention, the foregoing description of the embodiments of the present application sometimes combines multiple features into a single embodiment, drawing, or description thereof. However, this approach of the present application should not be construed as reflecting an intention that the claimed subject matter requires more features than expressly recited in each claim. Rather, the subject of the invention should possess fewer features than in any single embodiment described above.

Claims

1. A deep learning-based method for dynamic optimization of metabolic feedback immune activity in fermented feed, characterized in that, Includes the following steps: S1: Collect fermentation parameters, metabolite data, and immune activity indicators of fermented feed to construct the original dataset; S2: Preprocess and partition the original dataset to obtain the training set, validation set, and test set; S3: Construct and train a hybrid neural network immune activity prediction model based on the training set, validation set, and test set to obtain the optimal prediction model; S4: First, construct the basic models of the chaotic mapping-driven slime mold algorithm, the basic model of the symbiotic organism search, and the basic model of the fruit fly optimization algorithm respectively; then design a two-way bidirectional interaction mechanism between each algorithm to realize parameter feedback and position coordination between each algorithm; finally, construct a fitness function based on the predicted value of the immune activity index output by the optimal prediction model, and connect the three algorithms with the optimal prediction model through the fitness function to form a fusion framework. S5: Based on the fusion framework and the optimal prediction model, dynamic optimization of fermentation parameters is carried out. Through population initialization, fitness calculation, iterative optimization and fusion of global optimal positions, candidate optimal fermentation parameters are output.

2. The method for dynamic optimization of metabolic feedback immune activity in fermented feed based on deep learning according to claim 1, characterized in that, It also includes S6: applying the candidate optimal fermentation parameters to the actual fermentation experiment, collecting experimental metabolite data and experimental immune activity indicators, constructing a metabolic feedback mechanism, determining whether to update the optimal prediction model through metabolite deviation calculation, verifying the effectiveness of the optimization results by combining the relative error between the experimental immune activity indicators and the predicted values ​​of the immune activity indicators, and forming a dynamic iterative optimization closed loop.

3. The method for dynamic optimization of metabolic feedback immune activity in fermented feed based on deep learning according to claim 1, characterized in that, The chaotic mapping-driven slime mold algorithm and the symbiotic organism search mutually feed back core parameters to dynamically adjust their respective search strategies. Specifically, the optimal fitness value of the symbiotic organism search is fed back to the chaotic mapping-driven slime mold algorithm to dynamically adjust its position update weight; the chaotic variables of the chaotic mapping-driven slime mold algorithm are fed back to the symbiotic organism search to dynamically adjust its symbiosis coefficient.

4. The method for dynamic optimization of metabolic feedback immune activity in fermented feed based on deep learning according to claim 3, characterized in that, The symbiotic search and the fruit fly optimization algorithm feed back core parameters and location information to each other to optimize the search target and range. Specifically, the optimal position of the fruit fly optimization algorithm is fed back to the symbiotic search to optimize the target position of its parasitic operation; the symbiotic coefficient of the symbiotic search is fed back to the fruit fly optimization algorithm to dynamically adjust its olfactory search step size.

5. The method for dynamic optimization of metabolic feedback immune activity in fermented feed based on deep learning according to claim 1, characterized in that, The chaotic mapping-driven slime mold algorithm and the fruit fly optimization algorithm feed back core parameters to balance global exploration and local convergence. Specifically, the scaling factor of the chaotic mapping-driven slime mold algorithm is fed back to the fruit fly optimization algorithm to dynamically adjust its visual search radius; the optimal olfactory concentration of the fruit fly optimization algorithm is fed back to the chaotic mapping-driven slime mold algorithm to dynamically adjust its scaling factor.

6. The method for dynamic optimization of metabolic feedback immune activity in fermented feed based on deep learning according to claim 1, characterized in that, Fermentation parameters in S1 were obtained through real-time monitoring using online sensors, metabolite data were obtained through laboratory chromatographic detection, and immunological activity indicators were obtained through in vitro cell culture experiments.

7. The method for dynamic optimization of metabolic feedback immune activity in fermented feed based on deep learning according to claim 1, characterized in that, In S5, the global optimal position fusion adopts a weighted fusion strategy, and the weights are determined by the proportion of the next-generation optimal fitness values ​​of the chaotic mapping-driven slime mold algorithm, symbiotic organism search, and fruit fly optimization algorithm.

8. The method for dynamic optimization of metabolic feedback immune activity in fermented feed based on deep learning according to claim 1, characterized in that, The convergence criteria for iterative optimization in S5 are: the absolute difference between the global optimal fitness values ​​of two adjacent rounds is less than a preset threshold, or the number of iterations reaches the preset maximum number of iterations.

9. The method for dynamic optimization of metabolic feedback immune activity in fermented feed based on deep learning according to claim 1, characterized in that, The hybrid neural network in S3 includes a feature extraction module and a temporal prediction module. The feature extraction module uses an alternating structure of convolutional and pooling layers to capture local feature correlations, while the temporal prediction module uses a combination of long short-term memory network layers and fully connected layers to model temporal dependencies. The training process employs an early stop strategy, stopping training when the loss no longer decreases after a preset number of consecutive rounds.

10. The method for dynamic optimization of metabolic feedback immune activity in fermented feed based on deep learning according to claim 2, characterized in that, In S6, the metabolite bias is calculated as the average absolute difference between the normalized values ​​of experimental metabolites and the normalized predicted values ​​of metabolites output by the optimal prediction model. When the metabolite bias is greater than the preset bias threshold, the experimental data is added to the original dataset to re-process the data and retrain the hybrid neural network immune activity prediction model to update the optimal prediction model.

Images