A protein feed microbial protein yield prediction method, system, terminal device and storage medium based on machine learning

Abstract

The application relates to the technical field of machine learning, in particular to a protein feed microbial protein yield prediction method and system based on machine learning, terminal equipment and a storage medium. The method comprises the following steps: constructing a multidimensional database; using multiple machine learning algorithms and corresponding hyperparameter search intervals to respectively establish corresponding initial protein feed MCP yield prediction models, and obtaining an optimal input feature subset; based on the optimal input feature subset, using a particle swarm optimization algorithm to globally optimize the hyperparameters of the multiple initial protein feed MCP yield prediction models, and obtaining a final protein feed MCP yield prediction model; inputting the static chemical indexes of a to-be-tested protein feed into the trained final protein feed MCP yield prediction model, and outputting a microbial protein yield prediction value of the to-be-tested protein feed. The method can effectively and accurately obtain MCP yield prediction data.

Description

Technical Field

[0001] This application relates to the field of machine learning technology, specifically to a machine learning-based method, system, terminal device, and storage medium for predicting the protein yield of microbial feed. Background Technology

[0002] In ruminant nutrition, rumen microbial protein (MCP) is the main source of essential amino acids for the host, accounting for 70%-85% of the amino acids absorbed in the small intestine. Accurately predicting MCP production is crucial for assessing the value of protein feeds and achieving precise nutrient supply.

[0003] Currently, methods for predicting MCP production mainly fall into two categories:

[0004] The first type is the empirical coefficient method, represented by the US National Research Council (NRC) model. This method establishes a linear relationship between MCP yield and degradable protein (RDP) or digestible nutrients (TDN) using a fixed conversion efficiency coefficient (e.g., 0.85). While this type of model has simple parameters, it masks efficiency fluctuations caused by dietary structure, microbial dynamics, and energy-nitrogen synchronicity with a static average efficiency coefficient. This is particularly problematic when applied to unconventional protein feeds such as distillers' grains, peanut meal, and DDGS, where it exhibits significant prediction bias and insufficient robustness.

[0005] The second category is the dynamic mechanism method, represented by the Cornell Net Carbohydrate-Protein System (CNCPS), which uses mathematical equations to describe the dynamic processes of substrate degradation and microbial growth in the rumen. This type of model has strong mechanistic logic, but relies on a large number of hard-to-obtain kinetic parameters (such as the degradation rate of each component and the maximum growth rate of microorganisms). More importantly, its core parameters and computational logic are built into commercial software and not publicly available, making it impossible for researchers to independently reproduce and verify them. Its applicability is questionable, especially when dealing with unconventional raw materials whose degradation kinetics deviate from the norm.

[0006] Therefore, the main drawback of the existing technology is the lack of an MCP yield prediction method that does not rely on fixed efficiency coefficients and closed parameters, can effectively handle high-dimensional nonlinear relationships, and has good predictive accuracy and generalization ability for unconventional protein feeds. Summary of the Invention

[0007] The present invention aims to solve the problems in the prior art and provide a method, system, device and storage medium for predicting the microbial protein yield of protein feed based on machine learning, which can effectively and accurately obtain MCP yield prediction data.

[0008] To achieve the above objectives, the first aspect of this application provides a machine learning-based method for predicting the microbial protein yield of protein feed, comprising: S1. Construct a multi-dimensional database, which contains static chemical indicators of at least 10 protein feed ingredients and the corresponding microbial protein yield measured by short-term static in vitro fermentation experiments; wherein, the static chemical indicators include at least one data type, each static chemical indicator is used as an input feature, and the microbial protein yield is used as a unique output variable. S2. Multiple datasets are constructed by combining multiple static chemical indicators as input features. Each dataset includes a training set and a test set. Based on the output variables, corresponding initial protein feed MCP yield prediction models are established according to multiple machine learning algorithms and corresponding hyperparameter search intervals. The performance of each initial protein feed MCP yield prediction model is evaluated and compared to obtain the optimal initial protein feed MCP yield prediction model. Based on the particle swarm optimization algorithm and the optimal initial protein feed MCP yield prediction model, redundant and collinear input features are removed to obtain the optimal subset of input features. S3. Based on the optimal input feature subset, the hyperparameters of multiple initial protein feed MCP yield prediction models are globally optimized using the particle swarm optimization algorithm. The optimization objective is to maximize the mean R² of cross-validation. The optimal hyperparameter combination is determined, and the optimal hyperparameter combination is used to train multiple initial protein feed MCP yield prediction models to obtain optimized protein feed MCP yield prediction models. By comparison, the CatBoost prediction model is determined as the optimal protein feed MCP yield prediction model, and the CatBoost prediction model is determined as the final protein feed MCP yield prediction model. S4. Input the static chemical indicators of the protein feed to be tested into the trained final protein feed MCP yield prediction model, and output the predicted value of the microbial protein yield of the protein feed to be tested.

[0009] In one embodiment, the static chemical indicators include at least three different data types: conventional nutrient components, fractionated components of the Cornell net carbohydrate-protein system, and amino acid content.

[0010] In one embodiment, the plurality of machine learning algorithms include at least the neural network algorithm ANN, the kernel-based algorithm SVM, and the tree-based ensemble algorithms RF, GBR, AdaBoost, and CatBoost.

[0011] In one embodiment, the optimal subset of input features includes total energy, starch, acid-washable insoluble protein, soluble true protein, soluble protein, moderately degrading carbohydrate, slowly degrading carbohydrate, non-degradable carbohydrate, moderately degrading protein, unusable protein, and alanine.

[0012] In one embodiment, during S2, multiple machine learning algorithms are used to establish corresponding initial protein feed MCP yield prediction models. During this modeling process, all machine learning algorithms are trained using hyperparameters, with modeling based on the most basic hyperparameter combination. The hyperparameter search range is as follows: Number of iterations: 100–1000; Learning rate: 0.01–0.3; Tree depth: 1~10; L2 regularization coefficient: 1 to 10.

[0013] In one embodiment, in step S3, the particle swarm optimization algorithm parameters are set as follows: the number of particles is 10, the individual learning factor c1=0.1, the social learning factor c2=0.7, the inertia weight w=0.9, and a linear decreasing strategy is adopted, with a maximum number of iterations of 100.

[0014] A second aspect of this application provides a machine learning-based system for predicting the microbial protein yield of protein-rich feeds, comprising: A multi-dimensional database construction module is used to obtain static chemical indicators of at least 10 protein feed ingredients and the corresponding microbial protein yield measured by short-term static in vitro fermentation experiments; wherein, the static chemical indicators include at least one data type, each static chemical indicator is used as an input feature, and the microbial protein yield is used as a unique output variable. The input feature optimization module is used to construct multiple datasets as input features by combining multiple static chemical indicators. Each dataset includes a training set and a test set. Based on the output variables, the module establishes corresponding initial protein feed MCP yield prediction models according to multiple machine learning algorithms and corresponding hyperparameter search intervals. The module evaluates and compares the performance of each initial protein feed MCP yield prediction model to obtain the optimal initial protein feed MCP yield prediction model. Based on the particle swarm optimization algorithm and the optimal initial protein feed MCP yield prediction model, redundant and collinear input features are removed to obtain the optimal subset of input features. The prediction model determination module uses the particle swarm optimization algorithm to globally optimize the hyperparameters of multiple initial protein feed MCP yield prediction models. The optimization objective is to maximize the R² mean of cross-validation to determine the optimal hyperparameter combination. The optimal hyperparameter combination is then used to train multiple initial protein feed MCP yield prediction models to obtain optimized protein feed MCP yield prediction models. By comparison, the CatBoost prediction model is determined as the optimal protein feed MCP yield prediction model, and the CatBoost prediction model is determined as the final protein feed MCP yield prediction model. The prediction output module is used to input the static chemical indicators of the protein feed to be tested into the trained final protein feed MCP yield prediction model, and output the predicted value of the microbial protein yield of the protein feed to be tested.

[0015] A third aspect of this application provides a terminal device, comprising: At least one processor; At least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements a machine learning-based method for predicting the microbial protein yield of protein feed as described above.

[0016] A fourth aspect of this application provides a computer-readable storage medium storing a processor-executable program, which, when executed by a processor, is used to implement a machine learning-based method for predicting the yield of protein-rich microorganisms in protein feed as described above.

[0017] Compared with the prior art, this application has the following beneficial effects: In the aforementioned machine learning-based method, system, equipment, and storage medium for predicting microbial protein yield in protein feeds, the prediction method directly predicts microbial protein yield using static chemical indicators (such as conventional nutrient components) after modeling, eliminating the need for time-consuming and costly static in vitro fermentation experiments. This significantly shortens the evaluation cycle and reduces testing costs. A multi-dimensional database (containing at least 10 raw materials and their corresponding measured values) is used, combined with various machine learning algorithms to establish corresponding initial protein feed MCP yield prediction models. The optimal initial protein feed MCP yield prediction model is obtained by comparing performance. Based on particle swarm optimization and the optimal initial protein feed MCP yield prediction model, redundant and collinear input features are eliminated, retaining the optimal subset of input features most relevant to microbial protein yield. Compared to using the full input feature set, the model trained with the optimal subset of input features has higher generalization ability and prediction accuracy, avoiding overfitting and thus improving the generalization ability and accuracy of the prediction model. Building upon this foundation, Particle Swarm Optimization (PSO) was employed to globally optimize the model using the mean R² of cross-validation, avoiding the blindness and local optima issues inherent in manual parameter tuning and ensuring optimal model performance on the given dataset. The ultimately selected CatBoost prediction model effectively mitigated prediction bias through its symmetric tree structure and ordered boosting, further improving the stability and accuracy of microbial protein yield prediction. Experiments showed that on the test set, the coefficient of determination R² between predicted and measured values ​​reached 0.7393, with a root mean square error (RMSE) of 8.60 and a mean absolute error (MAE) of 6.63. This invention does not rely on empirical coefficients for specific raw materials or undisclosed closed parameters, solving the problems of large prediction bias and difficult parameter adaptation in traditional models for unconventional protein feeds such as distiller's grains, peanut meal, and DDGS. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 The schematic diagram illustrates a flowchart of a machine learning-based method for predicting the microbial protein yield of protein feed according to an embodiment of this application; Figure 2 A comparison of the predictive performance of various machine learning models under different combinations of input features; Figure 3 This is the input feature selection result of an example using Particle Swarm Optimization (CatBoost). Figure 4In the combined graph, (A) is a comparison between the predicted MCP yield and the measured MCP yield of the machine learning model in the test set, (B) is a Taylor plot comparing the performance of different MCP prediction models, and (C) is a violin plot and box plot showing the distribution of prediction errors of different machine learning models. Figure 5 A schematic diagram of a computer device according to an embodiment of this application is shown. Detailed Implementation

[0020] To facilitate understanding of this application, the following description will be more comprehensive and detailed in conjunction with the accompanying drawings and preferred embodiments, but the scope of protection of this application is not limited to the following specific embodiments.

[0021] Unless otherwise defined, all technical terms used herein have the same meaning as commonly understood by those skilled in the art. The technical terms used herein are for the purpose of describing particular embodiments only and are not intended to limit the scope of this application.

[0022] Unless otherwise specified, all raw materials, reagents, instruments and equipment used in this invention can be purchased from the market or prepared by existing methods.

[0023] Please see Figure 1 , Figure 1 The illustration shows a flowchart of a machine learning-based method for predicting the microbial protein yield of protein feed according to an embodiment of this application. The method includes the following steps: Step S1: Construct a multi-dimensional database, which contains static chemical indicators of at least 10 protein feed ingredients and the corresponding microbial protein yield measured by short-term static in vitro fermentation experiments; wherein, the static chemical indicators include at least one data type, each static chemical indicator is used as an input feature, and the microbial protein yield is used as a unique output variable. Specifically, in this embodiment, the static chemical indicators include at least one of three different data types: conventional nutritional components, Cornell Net Carbohydrate-Protein System (CNCPS) fractionation components (hereinafter referred to as: CNCPS fractionation components), and amino acid content (hereinafter referred to as AA content); Preferably, the conventional nutrient components include dry matter (DM), total energy (GE), crude protein (CP), crude fat (EE), crude fiber (CF), crude ash (Ash), nitrogen-free extract (NFE), neutral detergent fiber (NDF), acid detergent fiber (ADF), acid detergent lignin (ADL), starch, neutral detergent insoluble protein (NDFIP), acid detergent insoluble protein (ADFIP), soluble protein (SP), and non-protein nitrogen (NPN); the CNCPS fractionation components include total carbohydrates (CHO), non-structural carbohydrates (NSC), rapidly degradable carbohydrates (CA), moderately degradable carbohydrates (CB1), and slowly degradable carbohydrates. The protein composition includes: CB2 (non-degradable carbohydrates), CC (non-protein nitrogen), PA (rapidly degradable protein), PB1 (medium-degradable protein), PB2 (slowly degradable protein), PB3 (non-usable protein), and PC (unusable protein). The amino acid content includes 17 amino acids: aspartic acid (Asp), threonine (Thr), serine (Ser), glutamic acid (Glu), glycine (Gly), alanine (Ala), cystine (Cys), valine (Val), methionine (Met), isoleucine (Ile), leucine (Leu), tyrosine (Tyr), phenylalanine (Phe), lysine (Lys), histidine (His), arginine (Arg), and proline (Pro).

[0024] Specifically, in one embodiment, step S1 includes the following specific steps: 1.1 Data Acquisition for Multi-Dimensional Databases This embodiment collected 14 protein feed ingredients from 11 provinces across China, including: soybean meal (SBM), rapeseed meal (RSM), distillers' grains (CBDG), brewer's grains (BDG), palm oil meal (PKM), cottonseed meal (CSM), jasmine residue (JFM), spray-dried corn husks (CBM), sesame meal (SMP), tea oil cake (COM), flaxseed cake (LSM), sunflower cake (SFM), corn DDGS, and peanut meal (GNM). Samples were collected from at least three different origins for each ingredient, resulting in a total of 14 × 3 = 42 samples.

[0025] The sample was dried at 65℃ for 48 hours, pulverized and passed through an 18-mesh sieve (1mm), and after being mixed evenly, it was divided into two portions: one portion was used for static chemical index determination, and the other portion was used for short-term static in vitro fermentation experiment.

[0026] In this embodiment, a total of 41 static chemical indicators were measured across three different data types: Common nutritional components (15 items): Dry matter (DM), Total energy (GE), Crude protein (CP), Crude fat (EE), Crude fiber (CF), Crude ash (Ash), Nitrogen-free extract (NFE), Neutral detergent fiber (NDF), Acid detergent fiber (ADF), Acid detergent lignin (ADL), Starch, Neutral detergent insoluble protein (NDFIP), Acid detergent insoluble protein (ADFIP), Soluble protein (SP), Non-protein nitrogen (NPN). The determination methods refer to the relevant national standards.

[0027] CNCPS fractionation components (9 items): calculated according to the method of Lanzas et al., including total carbohydrates (CHO), non-structural carbohydrates (NSC), rapidly degrading carbohydrates (CA), moderately degrading carbohydrates (CB1), slowly degrading carbohydrates (CB2), non-degrading carbohydrates (CC), non-protein nitrogen (PA), rapidly degrading protein (PB1), moderately degrading protein (PB2), slowly degrading protein (PB3), and unusable protein (PC).

[0028] Amino acid content (17 items): Aspartic acid (Asp), threonine (Thr), serine (Ser), glutamic acid (Glu), glycine (Gly), alanine (Ala), cystine (Cys), valine (Val), methionine (Met), isoleucine (Ile), leucine (Leu), tyrosine (Tyr), phenylalanine (Phe), lysine (Lys), histidine (His), arginine (Arg), proline (Pro). The determination method is in accordance with GB / T18246-2019.

[0029] In this embodiment, the short-term static in vitro fermentation experiment specifically adopted the short-term static in vitro fermentation batch-increment method. Specifically, rumen fluid was collected from four Xiangxi Yellow Cattle with permanent rumen fistulas and mixed with artificial rumen buffer at a ratio of 1:2 to prepare a culture medium. 1g of feed sample and 100mL of culture medium were added to each fermentation bottle, and anaerobic fermentation was carried out in a shaking water bath at 39.5℃.

[0030] Fermentation was terminated at six time points: 3h, 6h, 9h, 12h, 18h, and 24h. Six replicates were performed for each sample, with a blank control included. After fermentation, the concentration of microbial protein (MCP) in the fermentation broth was determined using the Coomassie Brilliant Blue staining method (D-250), and the concentration was converted to the total MCP mass (mg) per 105mL culture volume as the output variable.

[0031] The specific process of constructing a dataset for a multidimensional database is as follows: Using the observations of each fermentation bottle at each sampling time point as independent sample units, a total of 504 valid samples were constructed (14 raw materials × 6 replicates × 6 time points). Each sample contains 41 input features and 1 MCP yield output variable.

[0032] The input features for subsequent modeling are derived from the determination of the nutritional components of the raw materials, including three categories: (1) 15 conventional nutritional components: DM, GE, CP, EE, CF, Ash, NFE, NDF, ADF, ADL, Starch, NDFIP, ADFIP, SP, NPN; (2) 11 CNCPS fractionation components: CHO, NSC, CA, CB1, CB2, CC, PA, PB1, PB2, PB3, PC; (3) 17 amino acid contents: Asp, Thr, Ser, Glu, Gly, Ala, Cys, Val, Met, Ile, Leu, Tyr, Phe, Lys, His, Arg, Pro. The original unit is mg / gDM (GE is J / gDM). Based on the amount of bottom material added to the fermentation flask (the mass of DM is known), the units of each input indicator are converted to mg (GE is J), and the units of the input features and output variables are kept consistent.

[0033] The output variable is the microbial protein (MCP) yield of each fermentation flask at each time point (3, 6, 9, 12, 18, 24 h). The original measurement result is the fermentation broth concentration (mg / dL), which is converted to the total mass of MCP per flask (mg) based on a culture volume of 105 mL. The converted MCP yield is used as the output variable for modeling.

[0034] Step S2: Construct multiple datasets using a combination of multiple static chemical indicators as input features. Each dataset includes a training set and a test set. Combined with the output variables, establish corresponding initial protein feed MCP yield prediction models based on multiple machine learning algorithms and corresponding hyperparameter search intervals. Evaluate and compare the performance of each initial protein feed MCP yield prediction model to obtain the optimal initial protein feed MCP yield prediction model. Based on particle swarm optimization algorithm and the optimal initial protein feed MCP yield prediction model, remove redundant and collinear input features to obtain the optimal subset of input features.

[0035] Specifically, six different machine learning algorithms, including the neural network algorithm ANN, the kernel-based algorithm SVM, and the tree-based ensemble algorithms RF, GBR, AdaBoost, and CatBoost, along with their corresponding hyperparameter search intervals, were used to establish six initial protein feed MCP yield prediction models (referred to as ANN prediction model, SVM model, RF prediction model, GBR prediction model, AdaBoost prediction model, and CatBoost prediction model).

[0036] Figure 2 This paper demonstrates the generalization performance of six initial protein feed MCP yield prediction models on rumen microbial protein yield prediction across three different datasets (input features). Results include R-values ​​for both the training and test sets. 2 RMSE and MAE results. During the modeling process, all machine learning algorithms were trained using default hyperparameters. Modeling was performed using the most basic hyperparameter combinations, and the hyperparameter search range is as follows: Number of iterations: 100–1000; Learning rate: 0.01–0.3; Tree depth: 1 to 10; L2 regularization coefficient (l2_leaf_reg): 1 to 10.

[0037] The search range for each model's hyperparameters is pre-defined based on the model's characteristics, as shown in the table below: Table 1: Search range of hyperparameters for each model

[0038] The performance of each initial protein feed MCP yield prediction model was evaluated. Specifically, the coefficient of determination (R²), root mean square error (RMSE), and mean absolute error (MAE) were used to evaluate the model performance.

[0039] Overall, changes in input features (different datasets) have a significant impact on model prediction performance. Figure 2 The static chemical index data corresponding to the three input features are as follows: 1. Only conventional nutrient components; 2. Conventional nutrient components + CNCPS fractionation components; 3. Conventional nutrient components + CNCPS fractionation components + AA content. In a dataset containing only 15 conventional nutrients, the test set R of the CatBoost prediction model... 2The training set R² was 0.7125, RMSE was 9.1392, and MAE was 7.1060. The RF prediction model had a high training set R² of 0.9333, but a test set R² of only 0.7204, indicating some overfitting. In the dataset incorporating CNCPS fractionation components, the performance of all models generally improved: the CatBoost prediction model's test set R² increased to 0.7227, while RMSE and MAE decreased to 8.9769 and 7.0053 respectively, remaining the best; the RF prediction model's test set R² remained at 0.7203, with no significant improvement in overfitting. Further incorporating a dataset containing 17 amino acids, the performance of all models generally declined. The CatBoost prediction model's test set R² decreased to 0.6933, and the RF prediction model's test set R² decreased to 0.6266, further widening the gap between the training and test sets. Under different input feature conditions, the CatBoost prediction model consistently outperformed other models on the test set, demonstrating a stable advantage. Although the performance of various initial protein feed MCP yield prediction models generally declined when using datasets including nutrient components, CNCP fractionation components, and AA content as input features, the CatBoost prediction model remained the most accurate. Therefore, the CatBoost prediction model was selected as the optimal initial protein feed MCP yield prediction model to eliminate redundant and collinear input features.

[0040] The results show that adding the content of 17 amino acids as input features did not improve the model's generalization performance; in fact, it decreased. This result may be due to three factors: First, the amino acid content is highly collinear with existing input features. Crude protein is essentially the sum of the contents of all amino acids, and the fractionation components of CNCPS are also related to the composition of specific amino acid contents. Adding amino acid content did not increase independent information but instead led to redundant input features, distracting the model and increasing complexity without improving generalization ability. Second, some amino acids have extremely low contents and little variation, such as methionine <7mg / g and cysteine ​​and tyrosine <15mg / g, with limited differences between varieties. These low-variable input features are difficult to provide effective discriminative information and are easily captured by the model as spurious associations, introducing noise and exacerbating overfitting. Third, the number of data in the dataset increased from 25 to 41 while the sample size remained the same, resulting in a deterioration in the dimensionality ratio. The model no longer fits the data based on real business patterns but instead remembers random coincidences that appeared in the training data. This set of coincidences cannot be adapted to the new data; that is, it has learned non-generalizable random patterns.

[0041] Therefore, it is necessary to first filter the input features and remove useless input features.

[0042] Specifically, redundant input features are filtered out based on the Particle Swarm Optimization (PSO) algorithm and the CatBoost prediction model to obtain the optimal subset of input features, thereby reducing model complexity and improving generalization ability.

[0043] Specifically, the particle swarm optimization algorithm is combined with the CatBoost prediction model to filter input features. Feature selection rules are defined using binarization, setting each input feature to either retain or discard only one outcome. Individual particles in the particle swarm continuously adjust the selection scheme, and the CatBoost prediction model evaluates the modeling effect of each input feature combination. All particles iteratively optimize the selection scheme based on their own historical best results and the population's best results. After iteration, redundant input features are removed, selecting the optimal subset of input features. This effectively prevents random coincidences in the model's learning dataset and improves the model's ability to adapt and predict unfamiliar data.

[0044] Figure 3 The bar chart shows the PSO-CatBoost input feature selection results, with the vertical axis representing the PSO input feature selection indicator (1 represents selection and retention, -1 represents rejection). Only a small number of input features were selected, and the vast majority were deemed redundant. A total of 11 input features were selected and retained, including GE, Starch, ADFIP, SCP, SP, CB1, CB2, CC, PB2, PC, and Ala; these features constitute the optimal subset of input features for subsequent modeling. Except for Ala, the contents of the other 16 amino acids, as well as common nutrients such as CP, CF, ADF, and ADL, were not selected, indicating that these indicators contribute little to the prediction of MCP synthesis and may even introduce noise. After introducing a large number of amino acid content indicators, the model accuracy dropped significantly, and the risk of overfitting increased significantly. Therefore, subsequent modeling only used the 11 selected input features as the optimal subset of input features.

[0045] A particle swarm optimization (PSO) approach combined with an optimal initial protein feed MCP yield prediction model was used to screen input features without altering the original sample data. Redundant and collinear features were removed to obtain the optimal subset of input features. Subsequent testing of only these 11 indicators (the optimal subset of input features) is sufficient to achieve high-precision prediction of MCP synthesis, significantly reducing detection costs and time. The optimal subset of input features points to specific metabolic pathways (energy, carbon skeleton, alanine hub, specific enzyme substrates), providing strong hypotheses for subsequent biological validation. This avoids the common problem of traditional regression models failing due to too many dependent variables.

[0046] The importance of features in the model was ranked from the chaotic features. Common nutrients (such as crude protein CP, crude fiber CF, acid detergent fiber ADF / ADL) often have a strong correlation with the content of certain amino acids (such as CP being positively correlated with most amino acids). As indicators with lower importance, they were not retained.

[0047] In the optimal input feature subset, GE (total energy) and Starch (starch) provide the precursor carbon skeleton and energy (ATP) for MCP synthesis. Ala (alanine) is the only selected amino acid because, as a glucogenic amino acid, it occupies a pivotal position in nitrogen metabolism and transamination. ADFIP, SCP, SP, CB1, CB2, CC, PB2, and PC constitute an input feature set surrounding the MCP synthase system, representing the temporal heterogeneity of nitrogen release rate (fast, medium, slow, non-degradable), structural differences in energy sources (starch and non-starch polysaccharides), and enzymatic activity related to fiber degradation.

[0048] Step S3: Based on the optimal input feature subset, the particle swarm optimization algorithm is used to globally optimize the hyperparameters of multiple prediction models. The optimization objective is to maximize the R² mean of cross-validation. The optimal hyperparameter combination is determined, and the optimal hyperparameter combination is used to train multiple prediction models to obtain the optimized protein feed MCP yield prediction model. By comparison, the CatBoost prediction model is determined as the optimal protein feed MCP yield prediction model, and the CatBoost prediction model is determined as the final protein feed MCP yield prediction model.

[0049] In this application, based on the optimal input feature subset, six different machine learning algorithms—ANN (A neural network algorithm), SVM (a kernel-based algorithm), and RF, GBR, AdaBoost, and CatBoost (an ensemble algorithm based on tree models)—were used to establish a protein feed MCP yield prediction model and their performance was compared. The model performance was evaluated using metrics such as coefficient of determination (R²), root mean square error (RMSE), and mean absolute error (MAE).

[0050] PSO parameter settings: number of particles 10, individual learning factor c1=0.1, social learning factor c2=0.7, inertia weight w=0.9, and a linear decreasing strategy is adopted, with a maximum number of iterations of 100. The fitness function uses the mean of the coefficient of determination (R²) of 5-fold cross-validation.

[0051] Particle swarm optimization (PSO) was used to globally optimize the hyperparameters of the CatBoost prediction model during training, obtaining the optimal hyperparameter combination: 1000 iterations, a learning rate of 0.0256, a tree depth of 9, and an L2 regularization coefficient of 0.9771. For model training, training and testing sets were also set up, and the dataset was randomly divided into training and testing sets in an 8:2 ratio. The training set was used for model training and validation, while the testing set was used to test the model's generalization performance.

[0052] The optimal hyperparameter combinations for each machine learning model after particle swarm optimization are shown in Table 2 below: Table 2: Optimal Hyperparameter Combinations for Each Machine Learning Model

[0053] Figure 4 The paper presents a comparison of the MCP yield prediction performance of six machine learning models under the optimal hyperparameter combination after PSO hyperparameter optimization. (A) is a scatter plot of predicted and measured values, (B) is a Taylor plot of model performance comparison, and (C) is a violin plot and box plot of prediction error distribution. Figure 4 In (A), the horizontal axis represents the measured value of MCP production, the vertical axis represents the model prediction value, and the diagonal line represents the ideal prediction line. The closer the data point is to the diagonal line, the higher the accuracy of the model prediction.

[0054] After PSO optimization, the CatBoost prediction model performed best, with an R² of 0.7393, RMSE of 8.6019, and MAE of 6.6345 on the test set. The optimized ANN prediction model had an R² of 0.6103, RMSE of 10.5181, and MAE of 7.7818 on the test set. The ANN prediction model, limited by its high input feature dimensionality, complex model structure, and hyperparameter sensitivity, struggles to achieve optimal convergence with limited sample sizes. In contrast, the CatBoost prediction model, with its symmetric tree structure and ordered boosting mechanism, exhibits more stable generalization ability on small to medium sample data. Furthermore, the SVR prediction model, after PSO optimization, had an R² of 0.6706, RMSE of 9.6698, and MAE of 6.5751 on the test set, but its prediction accuracy was still lower than that of the CatBoost prediction model (0.7393) and the RF (0.7176) ensemble learning model. The results show a clear gradient in predictive performance among the six models: CatBoost > RF > GBR > SVR > AdaBoost > ANN. (Taylor plot) Figure 4(B) illustrates the performance differences of six machine learning models in predicting rumen MCP yield from protein-rich feeds. The figure evaluates model performance using three statistical indicators: the gray curve represents RMSE, the black dashed line represents the Pearson correlation coefficient, and the blue dashed line represents the standard deviation. The closer the model point is to the reference point (black pentagram), the better its predictive performance. As shown in the figure, the CatBoost prediction model (red dot) is closest to the reference point, while the ANN prediction model (yellow dot) deviates the most, indicating that CatBoost has the best predictive performance and the ANN prediction model has the worst. Figure 4 (C) shows the violin plots and box plots of the absolute errors predicted by the six models, evaluating the predictive stability of each model from three dimensions: error distribution pattern, central tendency, and dispersion. The width of the violin plot outline reflects the error probability density, while the embedded box plots display the median error, interquartile range, and extreme value range. The CatBoost prediction model has the lowest median error, the narrowest box, the shortest whiskers, and the most slender and concentrated violin outline, indicating that it has the smallest prediction error, the most concentrated distribution, and the fewest extreme error samples. Combining the results of various scores, Taylor plots, and violin and box plots, it can be clearly concluded that the CatBoost prediction model is the optimal model for predicting rumen MCP yield in protein feeds, with accurate and stable prediction results.

[0055] Step S4: Input the static chemical indicators of the protein feed to be tested into the trained final protein feed MCP yield prediction model, and output the predicted value of the microbial protein yield of the protein feed to be tested.

[0056] The final protein feed MCP yield prediction model performs forward calculations based on the input feature vector and outputs a numerical value, which is the expected microbial protein yield of the tested protein feed under given fermentation conditions. This predicted value can be recorded, displayed, or used for subsequent feed quality evaluation, formulation optimization, and other purposes.

[0057] Compared with existing technologies, the machine learning-based method for predicting microbial protein yield in protein feed provided by this invention has the following beneficial effects: In the aforementioned machine learning-based method, system, equipment, and storage medium for predicting microbial protein yield in protein feeds, the prediction method directly predicts microbial protein yield using static chemical indicators (such as conventional nutrient components) after modeling, eliminating the need for time-consuming and costly static in vitro fermentation experiments. This significantly shortens the evaluation cycle and reduces testing costs. A multi-dimensional database (containing at least 10 raw materials and their corresponding measured values) is used, combined with various machine learning algorithms to establish corresponding initial protein feed MCP yield prediction models. The optimal initial protein feed MCP yield prediction model is obtained by comparing performance. Based on particle swarm optimization and the optimal initial protein feed MCP yield prediction model, redundant and collinear input features are eliminated, retaining the optimal subset of input features most relevant to microbial protein yield. Compared to using the full input feature set, the model trained with the optimal subset of input features has higher generalization ability and prediction accuracy, avoiding overfitting and thus improving the generalization ability and accuracy of the prediction model. Building upon this foundation, Particle Swarm Optimization (PSO) was employed to globally optimize the model using the mean R² of cross-validation, avoiding the blindness and local optima issues inherent in manual parameter tuning and ensuring optimal model performance on the given dataset. The ultimately selected CatBoost prediction model effectively mitigated prediction bias through its symmetric tree structure and ordered boosting, further improving the stability and accuracy of microbial protein yield prediction. Experiments showed that on the test set, the coefficient of determination R² between predicted and measured values ​​reached 0.7393, with a root mean square error (RMSE) of 8.60 and a mean absolute error (MAE) of 6.63. This invention does not rely on empirical coefficients for specific raw materials or undisclosed closed parameters, solving the problems of large prediction bias and difficult parameter adaptation in traditional models for unconventional protein feeds such as distiller's grains, peanut meal, and DDGS.

[0058] Example The measured values ​​of 11 optimal input feature subsets (GE, Starch, ADFIP, SCP, SP, CB1, CB2, CC, PB2, PC, Ala) selected from the static chemical indicators of the protein feed to be tested (e.g., corn DDGS) are input into the trained optimal CatBoost prediction model, and the model can output the predicted value (mg) of rumen microbial protein yield of the feed during a 24-hour fermentation cycle.

[0059] For example, the input characteristics of a batch of corn DDGS were: GE=22218 J / g, Starch=7.46 mg / g, ADFIP=4.74 mg / g, SCP=5.83 mg / g, SP=12.51 mg / g, CB1=120.14 mg / g, CB2=202.15 mg / g, CC=153.94 mg / g, PB2=37.91 mg / g, PC=47.52 mg / g, and Ala=15.23 mg / g. The model calculated a predicted output MCP value of 78.5 mg. This prediction result is in high agreement with the measured value (79.2 mg) obtained through traditional in vitro fermentation, with a relative error of less than 1%, verifying the accuracy and reliability of this method.

[0060] In one embodiment, the present invention provides a machine learning-based microbial protein yield prediction system for protein feed, comprising: A multi-dimensional database construction module is used to obtain static chemical indicators of at least 10 protein feed ingredients and the corresponding microbial protein yield measured by short-term static in vitro fermentation experiments; wherein, the static chemical indicators include at least one data type, each static chemical indicator is used as an input feature, and the microbial protein yield is used as a unique output variable. The input feature optimization module is used to combine multiple static chemical indicators to construct multiple datasets as input features. Each dataset includes a training set and a test set. Multiple machine learning algorithms are used to build corresponding initial protein feed MCP yield prediction models. The performance of each initial protein feed MCP yield prediction model is evaluated and compared to obtain the optimal initial protein feed MCP yield prediction model. The optimal initial protein feed MCP yield prediction model is used to remove redundant and collinear input features to obtain the optimal input feature subset. The prediction model determination module is used to perform global optimization of the hyperparameters of multiple prediction models based on the optimal input feature subset, with the optimization objective of maximizing the R² mean of cross-validation, to determine the optimal combination of hyperparameters, and to train multiple prediction models using the optimal combination of hyperparameters to obtain an optimized protein feed MCP yield prediction model. By comparison, the CatBoost prediction model is determined as the optimal protein feed MCP yield prediction model, and the CatBoost prediction model is determined as the final protein feed MCP yield prediction model. The prediction output module is used to input the static chemical indicators of the protein feed to be tested into the trained final protein feed MCP yield prediction model, and output the predicted value of the microbial protein yield of the protein feed to be tested.

[0061] In one embodiment, the present invention provides a terminal device comprising: at least one processor; at least one memory for storing at least one program; wherein when the at least one program is executed by the at least one processor, the at least one processor implements a machine learning-based method for predicting the yield of protein-rich feed microorganisms as described above.

[0062] In one embodiment, the present invention provides a computer-readable storage medium storing a processor-executable program that, when executed by a processor, implements a machine learning-based method for predicting the yield of protein-rich feed microorganisms as described above.

[0063] Specifically, the processor contains a kernel that retrieves the corresponding program units from memory. One or more kernels can be configured, and adjusting kernel parameters allows for the implementation of machine learning-based methods for predicting the protein yield of microbial feed.

[0064] The memory may include non-permanent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM, and the memory includes at least one memory chip.

[0065] This application provides a storage medium storing a program that, when executed by a processor, implements the aforementioned machine learning-based method for predicting the yield of microbial protein in protein feed.

[0066] This application provides a processor for running a program, wherein the program executes the above-described machine learning-based method for predicting the yield of microbial protein in protein feed.

[0067] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 5As shown. The computer device includes a processor A01, a network interface A02, memory (not shown), and a database (not shown) connected via a system bus. The processor A01 provides computing and control capabilities. The memory includes internal memory A03 and a non-volatile storage medium A04. The non-volatile storage medium A04 stores an operating system B01, a computer program B02, and a database (not shown). The internal memory A03 provides an environment for the operation of the operating system B01 and the computer program B02 stored in the non-volatile storage medium A04. The database is used to store data. The network interface A02 is used for communication with external terminals via a network connection. When executed by the processor A01, the computer program B02 implements a machine learning-based method for predicting the microbial protein yield of protein-rich feed.

[0068] Those skilled in the art will understand that Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0069] This application provides a terminal device, which includes a processor, a memory, and a program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps of any of the above-mentioned machine learning-based methods for predicting the microbial protein yield of protein feed.

[0070] This application also provides a computer program product that, when executed on a data processing device, is adapted to perform a program that initializes a method for predicting the protein yield of protein feed microorganisms based on machine learning.

[0071] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0072] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0073] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0074] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0075] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0076] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0077] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0078] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0079] The above are merely preferred embodiments of this application. It should be noted that this application is not limited to the above embodiments. For those skilled in the art, several improvements and modifications can be made without departing from the principles of this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should also be considered within the scope of protection of this application.

Claims

1. A method for predicting microbial protein yield in protein feed based on machine learning, characterized in that, include: S1. Construct a multi-dimensional database, which contains static chemical indicators of at least 10 protein feed ingredients and the corresponding microbial protein yield measured by short-term static in vitro fermentation experiments; wherein, the static chemical indicators include at least one data type, each static chemical indicator is used as an input feature, and the microbial protein yield is used as a unique output variable. S2. Multiple datasets are constructed by combining multiple static chemical indicators as input features. Each dataset includes a training set and a test set. Based on the output variables, corresponding initial protein feed MCP yield prediction models are established according to multiple machine learning algorithms and corresponding hyperparameter search intervals. The performance of each initial protein feed MCP yield prediction model is evaluated and compared to obtain the optimal initial protein feed MCP yield prediction model. Based on the particle swarm optimization algorithm and the optimal initial protein feed MCP yield prediction model, redundant and collinear input features are removed to obtain the optimal subset of input features. S3. Based on the optimal input feature subset, the hyperparameters of multiple initial protein feed MCP yield prediction models are globally optimized using the particle swarm optimization algorithm. The optimization objective is to maximize the mean R² of cross-validation. The optimal hyperparameter combination is determined, and the optimal hyperparameter combination is used to train multiple initial protein feed MCP yield prediction models to obtain optimized protein feed MCP yield prediction models. By comparison, the CatBoost prediction model is determined as the optimal protein feed MCP yield prediction model, and the CatBoost prediction model is determined as the final protein feed MCP yield prediction model. S4. Input the static chemical indicators of the protein feed to be tested into the trained final protein feed MCP yield prediction model, and output the predicted value of the microbial protein yield of the protein feed to be tested.

2. The method according to claim 1, characterized in that, The static chemical indicators include at least three different data types: conventional nutritional components, fractionated components of the Cornell net carbohydrate-protein system, and amino acid content.

3. The method according to claim 1, characterized in that, The plurality of machine learning algorithms include at least the neural network algorithm ANN, the kernel-based algorithm SVM, and the tree-based ensemble algorithms RF, GBR, AdaBoost, and CatBoost.

4. The method according to claim 1, characterized in that, The optimal subset of input features includes total energy, starch, acid-wash-insoluble protein, soluble true protein, soluble protein, moderately degrading carbohydrate, slowly degrading carbohydrate, non-degradable carbohydrate, moderately degrading protein, unusable protein, and alanine.

5. The method according to claim 1, characterized in that, In S2, multiple machine learning algorithms are used to establish corresponding initial protein feed MCP yield prediction models. During the modeling process, all machine learning algorithms are trained using hyperparameters, with the most basic hyperparameter combination used for modeling. The hyperparameter search range is as follows: Number of iterations: 100–1000; Learning rate: 0.01–0.3; Tree depth: 1~10; L2 regularization coefficient: 1 to 10.

6. The method according to claim 1, characterized in that, In step S3, the particle swarm optimization algorithm parameters are set as follows: the number of particles is 10, the individual learning factor c1=0.1, the social learning factor c2=0.7, the inertia weight w=0.9, and a linear decreasing strategy is adopted, with a maximum number of iterations of 100.

7. A machine learning-based system for predicting the microbial protein yield of protein-rich feeds, characterized in that, include: A multi-dimensional database construction module is used to obtain static chemical indicators of at least 10 protein feed ingredients and the corresponding microbial protein yield measured by short-term static in vitro fermentation experiments; wherein, the static chemical indicators include at least one data type, each static chemical indicator is used as an input feature, and the microbial protein yield is used as a unique output variable. The input feature optimization module is used to construct multiple datasets as input features by combining multiple static chemical indicators. Each dataset includes a training set and a test set. Based on the output variables, the module establishes corresponding initial protein feed MCP yield prediction models according to multiple machine learning algorithms and corresponding hyperparameter search intervals. The module evaluates and compares the performance of each initial protein feed MCP yield prediction model to obtain the optimal initial protein feed MCP yield prediction model. Based on the particle swarm optimization algorithm and the optimal initial protein feed MCP yield prediction model, redundant and collinear input features are removed to obtain the optimal subset of input features. The prediction model determination module uses the particle swarm optimization algorithm to globally optimize the hyperparameters of multiple initial protein feed MCP yield prediction models. The optimization objective is to maximize the R² mean of cross-validation to determine the optimal hyperparameter combination. The optimal hyperparameter combination is then used to train multiple initial protein feed MCP yield prediction models to obtain optimized protein feed MCP yield prediction models. By comparison, the CatBoost prediction model is determined as the optimal protein feed MCP yield prediction model, and the CatBoost prediction model is determined as the final protein feed MCP yield prediction model. The prediction output module is used to input the static chemical indicators of the protein feed to be tested into the trained final protein feed MCP yield prediction model, and output the predicted value of the microbial protein yield of the protein feed to be tested.

8. A terminal device, characterized in that, include: At least one processor; At least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements a machine learning-based method for predicting the microbial protein yield of protein feed as described in any one of claims 1-6.

9. A computer-readable storage medium storing a processor-executable program, characterized in that, The processor-executable program, when executed by the processor, is used to implement a machine learning-based method for predicting the microbial protein yield of protein feed as described in any one of claims 1-6.

Images